# Changelog

## [Unreleased]

## [0.2.424] - 2026-11-17

### Fixed

- Recall returns actionable code locations and no longer treats empty searches as evidence. Project and global memories are ranked independently, and an empty generated memory overview is recovered from eligible entries.
- Opening a file in the Index Graph selects it after the graph loads, including files outside the initial overview.

- A late session refresh no longer replaces the older messages you are reading. Child chats continue showing new activity while earlier history loads, and reattaching a running stream preserves text already received.
- Cloud chats keep turns authored independently on different devices. Concurrent attachment selections respect the eight-file limit, a failed chat Stop no longer reports success, and Stop also cancels a global voice call that is still starting.
- The working-file diff now compares against an empty staged file rather than showing changes against the last commit instead.

## [0.2.423] - 2026-11-17

### Added

- **GPT-6.2 Astra and GPT-6.2 Sol on OpenAI and Codex.** Both models are in the OpenAI API catalog (API key) and the Codex catalog (ChatGPT plan sign-in). Each has a 1,050,000-token context window and 128,000 output tokens, with text and image input. API pricing is $8 / $40 per 1M tokens for Astra and $2 / $10 for Sol, cached input $0.40 and $0.10. Codex lists both for ChatGPT Work and Codex on supported plans. The effort picker offers Low through Max for both, without Off, because OpenAI rejects effort `none` for these models.

### Fixed

- **Streaming chat text keeps pace with long markdown replies.** A long streamed tail was parsed in the background, but each new text update cancelled the parse still running, so its result was discarded and the formatted text fell far behind and then jumped. A running parse now finishes, the newest text is parsed next, and the per-frame file-mention scan runs on the stable prefix while the reply streams.
- **Steered messages no longer stay "Delivered" without appearing in the chat.** A steer the agent consumed while the chat was not listening (for example after switching sessions) left its queue chip behind with no message in the transcript. The chat now re-reads the session from the backend, shows the message from the backend's own transcript, and clears the chip; while a delivered chip is still waiting, the session is re-read periodically.

## [0.2.422] - 2026-11-17

### Fixed

- **A stopped or timed-out terminal no longer wakes the chat over and over.** When a terminal was stopped at its time limit but its end could not be confirmed, the chat was woken again after 10, 20, 40, 80 and 160 minutes with "background shell needs inspection", and every wake started a reply that could do nothing. Such a terminal now reports once that its stop is unconfirmed and stays quiet afterwards; if it does end later, that end is still reported.

## [0.2.421] - 2026-11-17

### Added

- **Workflows: build your own way of working as a node graph.** The new Workflows dialog (composer mode menu → Manage workflows) lets you chain steps, each with its own mode, model, effort, workers, skills and instructions, and add expert teams (several models at once: competing hypotheses, debate, review panel with a vote), checks that run a command or ask an LLM judge, debug loops with a round limit and escalation, routers, user gates and sub-workflows. Activate a workflow from the mode menu and every message you send runs through it step by step; Test run plays the graph through without calling a model. Five templates are included, among them a debug loop with an expert team.

### Fixed

- **A message queued for after the run is sent even when you are in another session.** A follow-up waited in the chat window until that session's run ended, and only the open chat sent it. If you had switched to another session by then, it stayed queued until you opened the session again. It is now sent as soon as that run ends. After a window reload it still goes out when you open the session.
- **Queued messages that cannot be sent together now say so.** When queued messages from different sources could not be merged into one prompt, they stayed queued with no message. The chat now shows why they were not sent.

## [0.2.420] - 2026-11-16

### Fixed

- **Vault passwords no longer carry a stray space or line break.** A password saved with a trailing space or line break (from a paste or an import) was stored one character too long and typed that way, so set-password forms such as Hoststar's reported "the passwords are not identical". Vault passwords are now stored and filled without surrounding spaces and line breaks; existing entries are filled correctly without saving them again. API tokens keep their exact value.
- **A vault fill no longer puts the user name into the wrong field.** On a set-password form without a user-name field, the name landed in the next text field after the password, such as the mailbox size. The user name is now only filled into a field that comes before the password.

## [0.2.419] - 2026-11-16

### Fixed

- **Vault login fills now pass "passwords do not match" checks.** Some set-password forms, such as Hoststar's new mailbox dialog, only pick up a password field's content from a keystroke and compare the two fields from that. A vault fill inserted the text without any keystroke, so the form kept comparing old content and reported that the passwords are not identical, although both fields held the same value. Each filled field now also receives a harmless End keystroke, so such forms see the new content.

## [0.2.418] - 2026-11-16

### Changed

- **Several messages sent while the agent is busy arrive together.** When you send a second or third message while the agent is still working, they now reach it in the same turn as the first queued one instead of one turn each. Background notices such as finished builds or worker results are still handled separately.

## [0.2.417] - 2026-11-16

### Fixed

- **Vault login fills put exactly the stored value into every field.** On set-password forms with a "repeat password" field, a field that already held text could end up with leftovers in front of the stored value, so the two fields differed and the site reported that the passwords do not match. Unary now replaces the field's whole content directly, checks that each field holds exactly the stored value, and stops before submitting if a page alters it. The fill also signals the change, so sites that compare the two fields notice the new content.

## [0.2.416] - 2026-11-16

### Fixed

- **Switching models no longer sends the previous model's thinking as an answer.** This could make Opus reject an ordinary task as `reasoning_extraction`. Unary now keeps incompatible thinking out of the outgoing conversation while preserving normal answers and tool results. Compatible signed thinking stays intact; unsigned Anthropic thinking is no longer converted into answer text. The stored conversation is not deleted, so existing sessions benefit on their next request.

## [0.2.415] - 2026-11-16

### Added

- **Qwen4 Max on Qwen Cloud.** The generally available Qwen4 Max can now be chosen on Qwen Cloud, both pay-per-token and with the Token Plan: 1M tokens of context, up to 393K tokens of output, text, image and video input, and the effort levels low, high and xhigh.

### Fixed

- **The phone remote no longer loses a command on the direct connection.** When the phone sent two commands at almost the same moment over the direct link to the desktop, the second could overtake the first, and the desktop then discarded the first as a replay. That command was never carried out. Commands on the direct link now always leave in the order they were sent.

- **The copy button on selected chat text no longer flickers.** While you were still dragging a selection the button jumped around, its hover background let the text underneath show through, and its tooltip covered the selected text. It now appears once the selection is done, at the end of the selection, with an opaque background and without the tooltip.

- **xAI and Meta AI sign-in work again.** Since 0.2.381, the OAuth button for xAI (Grok) and Meta AI started GitHub Copilot's sign-in instead of the provider's own, so nothing useful opened and the button stayed busy. Every provider that signs in with a device code now gets its own code and page, the dialog names the provider, and the sign-in page opens in your browser. Meta AI sign-in also failed on Meta's side, because Meta now turns away sign-in requests that carry no app name; Unary now sends one. When the automatic browser return fails for other providers, Settings now says why instead of claiming that pasting the redirect URL is not supported.

## [0.2.413] - 2026-11-15

### Changed

- **API keys are entered where you expect them.** When the agent needs an API key or token, it now shows a card "API key needed" in the chat that names the variable (for example `ELEVENLABS_API_KEY`), its profile and what it is for. You paste the key, click Save to vault, and the agent is told only that the variable is set, never the value. The card can link to the page where you get the key and open the variable in the Unary Secure panel. Before, the agent showed a website-login form for this, which did not fit.
- **The ENV tab of Unary Secure is simpler.** Variables that a project needs but that have no value yet are listed first, with a field to paste the value right there. Every set variable can be changed or removed from its row menu. The form for a new variable and the `.env` import sit behind "+ Variable".

## [0.2.411] - 2026-11-15

### Fixed

- **The Unary Secure key card appears again for every model.** Some models (for example Gemini) call tools such as `secure` through a generic wrapper (`invoke_tool`). The tool ran and asked for the card, but the chat only saw the wrapper's name, so the form where you type a password or API key never showed up, and the agent kept pointing at a card that was not there. The chat now shows such calls as the tool that actually ran, so the key card, the question cards and the tool labels appear as with any other model. Older chats show the card too after the update.

## [0.2.410] - 2026-11-15

### Fixed

- **Automatic account switching no longer skips accounts that can still serve the model.** Anthropic can refuse a single model on an account (for example Fable on its own weekly window, `7d_oi`) while the account still serves Opus. Unary stored that refusal for the whole account and then skipped it for every model for days, so a turn failed with "quota exhausted" while three spare accounts would have worked. Switching by hand worked. A refusal now applies only to the model it was for. The skip message names the refused model and its reset time instead of only the weekly percentage.

## [0.2.409] - 2026-11-15

### Fixed

- **The agent no longer loops on refused contract reviews while its edits are unpublished.** With file isolation on, a review of the project rules was refused because the agent's edits still sat in its private copy, but the refusal pointed back to re-reading the rules instead of to publishing the edits. In one run that cost 24 refused calls in a row and tripped the loop guard. Every refusal now says first that the private edits must be published with `edit finish`, and names that call.

- **A message to an unknown session says which id is wrong.** An Intercom message with a mistyped or invented session id failed with "membership scan exceeded its bounded lookup", which named neither the id nor the remedy. It now names the session id and points to the session list.

## [0.2.408] - 2026-11-15

### Fixed

- **Creating a chat no longer fails at random with "inspect catalog … Zugriff verweigert".** On Windows, opening the project's workspace catalog while another part of Unary was just closing it failed, because the catalog's temporary SQLite file is briefly locked while it is deleted. Since session creation runs the project snapshot in the background, this happened often. Unary now waits that moment out the way SQLite itself does.

- **A chat knows which Blender or UnaryCad you linked to it.** Linking an app with the Link App button had no effect on the chat's agent: the button and the chat each kept their own list of links, so the agent was told no Blender was chosen and worked around it. There is now one list for the whole app. The agent also gets a short note in every turn saying which instance is linked (pid, scene, file), and Blender's status marks the linked instance. Switching Blender off in one project no longer drops the links of chats in other projects.

- **Blender no longer shows "Add-on outdated" for an up-to-date add-on.** The add-on and Unary compared their files byte for byte, so the same add-on with different line endings (for example copied from a Git checkout instead of exported from Unary) was always reported as outdated. Line endings no longer count. After updating Unary, reinstall the add-on once from Settings › Plugins › Blender.

- **New chats prepare faster, also several at once.** Before a new session could start, Unary re-checked on every creation whether the newest snapshot of the project folder was empty, which re-read the whole snapshot (about 0.6 s for 8,000 files, about 1 s for three sessions at once). The answer is now remembered per snapshot and prepared right after a snapshot is taken, so creating a session no longer waits for it.

## [0.2.407] - 2026-11-14

### Changed

- **"Set up spec files" can be switched off for good.** The banner now offers "Not for this project" and "Never ask again"; the latter stops the question in every project and can be switched back on in Settings › Rules ("Ask to set up spec files"). All top banners use the app's standard round buttons.

- **The end of a run is marked once, in the chat itself.** The answer dock above the composer is gone: it repeated the final reply that the chat already showed. Instead a thin line in the transcript marks where every run ended and what it ended as (Summary, Answer, Waiting for workers, Stopped by you, Run failed, or No closing message), with the time, how long the run took and the goals at that moment. Under the reply you find what the harness recorded (edits, checks ran and passed) plus Copy and Pin. The line always sits at the end of the run, also when a run stopped, failed or ended without a closing message; in that last case it offers "Ask for summary".

### Added

- **Blender can edit meshes live.** The agent gets `blender_mesh_edit` (as `mesh_edit_*` actions of the grouped `blender` tool): it selects vertices by index, box, radius, normal or stored vertex group, moves, scales and rotates them with falloff, adds loop cuts, subdivides, extrudes, insets, bevels and deforms (push/pull, sag, smooth, seeded noise, twist, bend), in object mode and undoable. The add-on now reports its handlers and a build fingerprint in its status, and an outdated add-on is named as such instead of failing with `unknown method`.

- **The image tool can make animated GIFs and use more GPT Image options.** With a GPT Image model (for example gpt-image-2.6), an output path ending in `.gif` produces a looping animation; frame count (2–16) and frame duration (20–2000 ms) can be set, and "extract frames" also saves every frame as its own PNG next to the GIF. WebP and JPEG output, compression and the planning effort ("complexity") are available too. A transparent background is described as real alpha, so sprites and cut-outs no longer need a green screen. Models that cannot do this (xAI, the ChatGPT account for WebP/JPEG/complexity) say so before anything is sent instead of saving a PNG under the wrong name.

- **A project folder can hold several repositories.** Open a plain folder whose subfolders are separate Git repositories — all your apps side by side, for example — as one project. A session in the folder and a session in one of its repositories now see each other's file claims, take turns on the same file and list each other as neighbours. The `git` tool reaches a repository below the folder with `git -C <path>` and says which repositories exist when you run it in the folder itself. The changes of a turn are listed per repository, and private file versions work for files in the repositories and in the folder. Checkpoints and Git worktrees stay per repository: open the repository itself to use them.

- **Projects can be nested in the sidebar.** Drag a project onto another project's header to show it as a subproject there; the header lights up and says "Move into". Dropping on the upper edge of a header still only reorders. To take a subproject out again, drag it onto "Move to top level" at the end of the list, or use "Move into project…" and "Move to top level" in the project menu. Nesting is your own arrangement and does not depend on folders: chats stay with their project, the parent's counter includes the chats of its subprojects, and removing a parent from the list moves its subprojects up one level.

### Fixed

- **A new chat is ready sooner.** Creating a chat waited up to 0.3 s for a fresh snapshot of the project folder that it then did not use, and ran a `git status` before it returned. It now starts from the last published snapshot at once while the fresh one is taken in the background, and records the Git baseline right after creation instead of during it. When the workspace catalog is busy, creation retries after 25 ms instead of 100 ms and logs each wait, so a slow start can be traced.

- **Idle chats are released from memory again.** The app keeps at most 24 chats loaded and unloads the ones not in use. Two faults stopped that: every status refresh counted as use, so no chat ever looked idle, and a chat with a finished terminal whose output was kept for recovery could never be unloaded. Only real use counts now, and a finished terminal no longer holds its chat in memory; its kept output stays on disk.

- **Ending a run with many workers no longer copies the whole worker list three times.** The check that decides whether a run is finished reads the workers once.

- **Edits with an absolute path inside the project work again on Windows.** Writing or editing a file given as `L:/Repos/…` failed with "invalid relative path: empty path, control character or non-portable separator/name" whenever the session folder was recorded in the `\\?\` form. The path is now recognised as inside the project regardless of that prefix, the slash direction or the drive-letter case.

- **Resolving a conflict works on the first try while other chats are busy.** With file isolation on, the first attempt to resolve a conflict was often refused as outdated, because any other chat or worker publishing an unrelated file in the meantime invalidated it. A conflict now only becomes outdated when one of its own files changes again. Checking an edit that has conflicts also says that resolving is the next step.

- **A worker that ended in a real conflict can be reviewed and applied.** Previewing or applying such a worker's edits failed with a workspace error, so the only way out was to let the worker fix it itself. The preview and apply now work, the worker's full report is shown with the conflict, and a worker that has since resolved and published its edits is no longer listed as a conflict.

- **File claims with isolation off: a refused writer is queued and told when the file is free.** When a file was held by another running Unary process that had written it between turns, a second writer was refused with a misleading message and never notified. It is now queued behind that holder like any other writer. A claim file that is locked or damaged no longer loses other chats' claims: the write is refused with a plain message to retry, releases are caught up, and a damaged file is kept aside.

- **Git checks no longer fail at random when many things run at once.** Under heavy load, Git status, staging and commit checks could fail with "an inherited output pipe stayed open" or because too many helper processes were running. Git output that is complete is now accepted even when another process still holds the pipe, and a short wait replaces the immediate refusal.

- **Deleting a chat no longer slows down as your history grows.** Deleting used to open every saved chat to find the files that belong to the one being removed, read that chat's file several times and rewrite its deletion record after each step. It now takes the file list from the session index, opens only files the index does not know, reads the chat once and writes the record a few times. How long each step of a delete took is written to the log.

- **The agent no longer repeats its answer when the Goals board could not be read.** When reading the board at the end of a run took too long, the agent was told that one goal was still open, although the board was empty, and wrote a second summary. It is now told that the state is unknown, checks the board, and ends with a single line when nothing is open. A slow read also logs which part of it was slow.

- **A reopened chat no longer misses its newest messages.** Switching back to a chat paints it instantly from memory and then checks it against the saved conversation. When that check could not replace the whole view (for example because the chat was still receiving updates), the view stayed as it was and the newest messages only appeared when the chat was opened a second time. The check now always finishes by adding the messages the view is missing.

## [0.2.406] - 2026-11-14

### Added

- **Blender: the agent can do much more in your open scene without writing scripts.** After comparing the current Blender MCP servers, the Unary add-on gained typed, scene-preserving actions: organise objects (delete, rename, duplicate, parent, collections, visibility), import files (GLB/glTF, FBX, OBJ, STL, PLY, USD, Alembic, or parts of another .blend), edit shader and geometry node graphs, set keyframes, and apply, reorder or change modifiers. Previews can now be framed automatically from standard views, several at once in one image, in solid, wireframe, X-ray or material shading, without needing a camera in the scene. The agent can search Blender's own API by keyword and check a file for missing textures and linked libraries. Each change is one undo step in Blender. Export the add-on again from Settings › Plugins › Blender and reinstall it to get the new actions.

- **The last answer of a run stays visible above the composer.** A new answer dock next to Changes and Goals shows the final reply of the last run with its kind (Summary, Answer, Status, or No summary), however far you have scrolled up. Click it to read the full reply with what the harness recorded for the run (edits made, checks ran and passed), jump to it, open it in a dialog or copy it. While a newer run is working or waiting for workers, the previous summary stays and the dock says what is happening now; a run that ended without a closing message says so and offers "Ask for summary".

- **The code editor completes what you type.** A suggestion list opens while you type, with entries from the language server and the words already in the file. Enter accepts a suggestion, and one that needs an import adds it for you. Brackets and quotes close themselves and the matching bracket is highlighted.

- **Parameter help, quick fixes and formatting in the editor.** Typing `(` or `,` in a call shows the function's parameters with the current one marked. `Ctrl+.` lists the fixes and refactorings the language server offers for the cursor or selection, and error tooltips carry a "Quick fix…" button. `Shift+Alt+F` formats the file or the selection.

- **Type hints, symbol colours and a symbol list in the editor.** The editor shows inferred types and parameter names inline, colours symbols by what the language server knows about them, and shows the enclosing function or type above the code. `Ctrl+Shift+O` opens a list of the file's symbols to jump to.

- **More languages get language-server support.** CSS, HTML, YAML, TOML, C/C++ and shell scripts use their language server when it is installed on the machine.

- **The Desktop editor shows find-references results.** `Shift+F12` opens the references as one stitched view inside the editor, and "Open" on an excerpt jumps to that file and line.

- **The agent can send you a short note while it keeps working.** A new `notify_user` tool lets the main agent report a finding, a warning or an interim status without stopping for an answer. Notes are capped at five per run, and an identical note sent twice in a row is shown once.

- **Notifications have an inbox and know where you are looking.** A bell in the Desktop title bar opens an inbox: requests a session is waiting on stay at the top with their answer buttons, everything else is kept as history. Approvals, questions, agent notes, results and errors share one card inside and outside the app, a session you are looking at stays quiet, several finished sessions become one notification, and an unanswered request reminds you once. Settings › Notifications chooses banner, in-app card and sound per kind; existing switches are carried over.

### Changed

- **A worker's edits no longer end as a "Merge conflict" when nothing collides.** With file isolation on, a worker that finished without publishing its edits was reported as a conflict and its work stayed out of your files. Its edits are now delivered when the worker ends; only files that really collide with newer changes are listed as a conflict. "Automatic" in the Worker issues dialog retries the delivery and shows the reason on the row when it cannot.
- **"Format on save" now formats.** The switch in Settings › Editor had no effect. With it on, saving a file first formats it through the language server; the cursor stays where it was.
- **Edit predictions work without Ollama.** With no suggestion model chosen, predictions use a local Ollama model if there is one, then the inline assistant's model, then your default chat model, then the cheapest model of a connected provider. Settings › Editor shows which one is in use.
- **New lines indent by your tab size.** The editor used two spaces whatever the "Tab size" setting said.
- **Rename symbol keeps unsaved edits.** Renaming with `F2` changes the open file in the editor instead of rewriting it on disk, and now also works with servers that answer in the newer edit format.
- **Settings always opens at full size.** The Settings dialog now fills the whole area between the app's title bar and the status bar every time it opens, instead of reopening at the last, possibly small, size. It can still be moved and resized while it is open.
- **A commit started from a chat contains only that chat's changes.** Commit, auto commit and "auto commit and push" in a chat used to stage the whole working tree, so one chat committed what other chats or you had changed. They now stage only the files this chat changed; files you staged yourself are still included. If one of those files was changed again after the chat's last edit, the chat says so after the commit. A chat without changes of its own no longer commits anything by itself.
- **A chat's change review shows what the chat changed.** The review compared a snapshot with the file as it is on disk now, so two chats working on the same file both showed the combined result. It now shows each file from before the chat's first edit to after its last one, loads a file's content only when you open it, and tells you when the file was changed again afterwards.
- **Turns that only edit files end faster in large repositories.** After such a turn the app no longer re-reads the whole working tree to work out what changed; it compares only the files the turn edited. Turns that run shell commands are checked as before.

### Fixed

- **A message sent into a running turn no longer waits behind a watched terminal.** When the agent was waiting on a background command, a message you sent could stay "Sent" for many minutes, in one case seventeen, until the wait ran out. Your message now ends that wait itself, a second check a few seconds later ends it if the first did not and records why, and a wait that cannot notice new messages is limited to one minute at a time. The command in the terminal keeps running.
- **The diff of an unstaged file shows your changes, not the whole file as new.** Opening a file from "Changes" in Source Control compared it with nothing, so every line appeared as added. It is now compared with the staged version, or with the last commit when nothing is staged.
- **An open diff follows the file.** A diff tab did not update when you staged, unstaged or committed, or when an agent changed the file. It now refreshes when that happens.
- **Deleted files open in the diff view.** A deleted file in "Changes" could not be opened. It now shows the last version against an empty file, marked as deleted. Binary files and files over 1 MB show a short note instead of a diff.
- **A chat's list of changed files is complete.** Large files, binary files and everything beyond the first 200 files or 2 MB of a turn were left out of the list without a hint. They are now listed with the reason their content is not shown, and large files can still be opened.
- **Files a worker edits show up in its changes.** A file the main agent had already touched was missing from a worker's list of changes, and after reloading a chat every change appeared as the main agent's.
- **"Open as diff" on a file in the chat opens the chat's change.** It used to open the diff of the whole working tree for that file.
- **Source Control actions in the chat work in the Desktop app.** Opening a commit's diff, the change review and the Git panel from a chat did nothing there. They now open in the Desktop's own diff tabs and Git panel.
- **The inline assistant, edit predictions and editor gutters work in the Desktop editor.** `Ctrl+Enter` opened nothing there, and the selection actions, ghost text, git markers, blame and breakpoint markers were drawn without their styles, because those lived in a stylesheet only the IDE loads.

### Removed

- **The "Optimized" badge on provider rows is gone.** It appeared on every active provider and so told nothing apart. Optimized models are still marked on the individual model rows, and unconfigured optimized providers keep their own "Unary Optimized" group.

## [0.2.405] - 2026-11-13

### Added

- **Gemini 4 Hydrogen on Google Antigravity.** Four lines appear in the Antigravity model list, Gemini 4 Hydrogen (Low), (Medium), (High) and (XHigh), each with a 1M-token context and up to 262,144 output tokens. They arrive with the official model catalog, so no app update is needed.

### Fixed

- **A checkpoint shows once in the chat, not once per hidden turn row.** A turn writes several hidden rows (request cause, tool activations, snapshots) in the same millisecond. When the chat reconciled its tail with the saved transcript, it treated all of them as one message and repeated every row after them once per hidden row, so one "before-turn-N · Undo to here" line could appear seven times. Saved rows are now matched one to one.
- **Unchanged turns no longer add duplicate checkpoints in repositories with many checkpoints.** The check that skips a checkpoint identical to the session's last one put every checkpoint's hash on one git command line. Past roughly 790 checkpoints that line was too long for Windows, the check failed silently, and every turn added a checkpoint. The check now reads all checkpoints without that limit.

## [0.2.404] - 2026-11-12

### Removed

- **Semantic and embedding search are removed; search is keyword-, symbol- and graph-based.** Code search, document search and memory search no longer use embeddings or the bundled model. Ranking uses keywords, symbols, trigrams and the code graph, and memory recall uses keyword search with recency.
- **The bundled embedding model and its sidecar are gone.** The app no longer ships or downloads the embedding model. The old model cache can be removed in Storage.
- **Retrieval-mode, hardware and model settings are removed.** The index no longer offers a retrieval profile, CPU or GPU choice, embedding model or semantic search switch. Stored values are ignored, and the index works without a model.
- **The shared embedding model setting is removed.** Memory settings no longer offer it, and first-run tools setup no longer offers a model download.
- **Cloud embedding providers are no longer used.** OpenRouter, Ollama and Google embedding providers left with the embedding code.

## [0.2.402] - 2026-11-12

### Fixed

- **A stray reasoning token no longer splits an answer in two.** Some models on the OpenAI-compatible wire (seen with Qwen on the Qwen Cloud token plan) send a lone `.` as reasoning in the middle of their answer. Unary opened a new Thinking block for it, so a reply like "Hi Silas!" appeared as "Hi Sil", an empty Thinking block and "as!". Reasoning without any letters or digits that arrives after the answer has started is now ignored; real reasoning still shows as before.

## [0.2.401] - 2026-11-12

### Fixed

- **Pausing automatic indexing now really pauses it.** The index panel showed a workspace as paused, but session start, index searches and the file watcher kept indexing it. While a workspace is paused, searches answer from the existing index and say so, and nothing reindexes it automatically; an explicit reindex still works.

- **Unary no longer fills the machine's memory when a large code index is opened.** 0.2.399 kept new scratch copies out of the code index, but an index that already held them still had to be loaded first: one status check of such an index took nine minutes and 207 GB of memory. The index status and the session start now read only the index's file list, a large index is loaded once at a time instead of once per caller, and an index whose files on disk exceed 4 GB is not loaded at all: search runs without it and the index tools report its size and how to shrink it. A refresh that finds most of the index now excluded rebuilds it instead of carrying the old entries. The 0.2.399 note that such an index drops these entries on its next refresh did not hold for large indexes.
- **Unary no longer keeps a processor core busy after every start.** The cleanup of temporary files compared every chat against every file name one by one; with 13.6 GB of chats and tens of thousands of temporary files it never finished. It now checks all names in one pass and reads chats in small pieces. The cleanup of finished worker chats now opens a chat only when one of its workers can be cleaned up, and the project cleanup skips projects it checked recently.

## [0.2.399] - 2026-11-11

### Fixed

- **Unary no longer fills the machine's memory when a folder of several projects is open.** When a session was opened on a folder that holds several projects, the code index also took in the scratch copies that agents make in Unary's own `.unary-runtime` folder, often whole repositories. On one machine that index grew to 37 GB and the app held up to 220 GB of memory after start until Windows ended it. Hidden folders and Unary's runtime folder are now left out of every index update, and an index that already contains them drops them on its next refresh.
- **Shell commands stay fast while an edit is unpublished with file isolation on.** With Smart File Isolation on, every shell command of a session with an unpublished edit first copied the whole project into a fresh private view and deleted it afterwards: about 15 seconds per command on a 7,900-file project, and the same for its workers. The view is now kept between commands, and after an edit only the changed files are updated.
- **The conflict preview of an unpublished edit shows the project as it is now.** With file isolation on, checking an edit before publishing it could report a conflict with an old version of the file that no longer existed, while publishing worked fine. The preview now compares against the current files, like publishing does.

## [0.2.397] - 2026-11-11

### Changed

- **Windows installs and updates now use the custom Unary AI setup dialog everywhere.** Fresh downloads on `dl.unary-code.app` list and serve only `UnaryAI-Setup.exe` once published for a release, and clicking **Update & restart** inside the Windows app now opens the bundled Mail-style `UnaryAI-Setup.exe` update card (verifying the signature, replacing the installation silently in the background, and relaunching Unary AI) instead of opening the classic NSIS installer window.

## [0.2.396] - 2026-11-11

### Added

- **Gemini 4 Carbon on Google Antigravity.** The model appears in the Antigravity model list with Low, Medium and High effort, a 1M-token context and up to 262,144 output tokens. Google serves it to Google AI Ultra and Enterprise accounts only.
- **The memory index at session start is budgeted by size.** It holds up to 5,000 estimated tokens, split 75 % repository and 25 % global; an unused share goes to the other side. A coding session gets the index once at start and again after each compaction.
- **A full memory index asks before it grows.** When a new entry no longer fits the index of its scope, the agent is shown similar entries and the five lowest-ranked index lines and decides: replace one of them, merge, or store the entry as search-only, where it stays findable and readable but is not listed in the index.
- **An optional memory worker briefs a new session.** It is off by default and is switched on in the memory settings. When it is on, a worker with the Editor role reads the memory index, recent sessions and the other live sessions and hands the session a short briefing shortly after its first prompt. If the worker cannot start or fails, the session gets the index instead. It adds one worker run per session.
- **A session sees what other sessions are doing.** A coding session is told which other live sessions of the same repository are working, on which goal, which files they hold and which builds they last saw red. The note appears only when that picture changes, at most every ten minutes, and it never wakes another session.
- **A new session sees what was worked on recently.** With its first prompt it receives the last main sessions of the repository with date, title, state and last written files, and up to three earlier sessions that match the request.
- **The retired session search index can be removed in Storage.** It is listed as removable, and only an explicit choice deletes it.

### Changed

- **Code search finds the right file more often, and answers faster.** On this repository's 38 test questions the owning file is now in the first ten results every time (before: 93 of 114 runs), and results from vendored code dropped from 60 to 3. Questions in German, questions about rules and specs, and Svelte components by name or by what they render are found; release notes no longer crowd out the spec that owns a topic. A warm search takes about 0.3 s. On 18 new questions that were not used for tuning the hit rate is 16 of 18.
- **"Who calls this" and "find references" answer in milliseconds.** They took about half a second each because every answer recounted all reference sites; that work now happens once per index generation. Find references also lists call sites, and asking for references of a type includes the calls of its methods, named by method.
- **An index refresh after an edit is two to three times faster.** On this repository (7,700 files) a refresh after one changed file takes about 9 s; before, its recorded steps alone took about 28 s. References are re-resolved only where the edit can change them, the lookup files and architecture relations are rewritten only for the changed file, and the index no longer copies itself before and after each refresh. The index file itself shrank from 1.1 GB to 27 MB, with the rest stored in pieces that are rewritten only when they change. Existing index stores are rebuilt once after the update.
- **Context packs for a question arrive in about half a second.** They took three to four seconds because each one rebuilt lookup tables and re-read the hierarchy; both are now kept per index generation. A question in German keeps all its matching results instead of only the first.

### Fixed

- **A repository's memory is found again after its git address changed.** A changed remote used to detach the memory; older shards are now adopted into the repository, and the original files are kept.
- **Related earlier sessions match again.** Searching sessions with several words matched nothing; each word is matched now.
- **Dream no longer skips memory entries when it could not classify them.** A run that classified nothing, for example without a model, still moved its progress mark forward, so those entries were skipped. The mark now moves only after entries were classified, and no run starts without a model. Knowledge found in a single repository can now move to global memory; before, it needed two repositories.
- **Memory no longer collects automatic error notes.** When a turn ended with an error, Unary used to write a known-bug memory entry by itself. Such entries now come only from the agent or from you.
- **The note naming the session that last changed a file appears once.** It used to repeat on every edit of the same file; now it appears on the first edit only.
- **The session overview shows who is editing.** With file isolation on, the overview listed no edited files and no last activity for any session. It now lists the files a session holds privately and when it was last active.

- **An old project setting no longer overrides the Editor model you set today.** A project file that still named one of the removed worker roles (for example `quickCoder`) silently replaced the Editor model from your settings, so workers could start on a model the provider no longer offers and fail. A configured Editor now always wins over those archived roles. The settings show where each worker model comes from (file and key), and a worker that fails on a pinned model names that file in its failure message.

- **Finished workers no longer stay "Running" in the chat.** A worker's final report could be discarded when a late progress update overtook it, and after switching back to a session the list waited up to a minute before checking with the backend. The final report is now always applied and the check runs at once.
- **A crash inside a background index refresh no longer blocks every later refresh until restart.** The failed refresh is recorded in the index error log and the next one starts normally.
- **A search that overlaps an index refresh keeps its results readable.** A reader holds the generation it loaded until it is done, and one workspace can no longer read or disturb the index of another workspace with identical files.
- **Creating a session no longer waits for a scan of the whole project.** Every new session read every tracked file of the project before it opened: 3.3 s in the middle case and up to 51 s on a project with 7,900 files, and on most creations the scan then failed and its result was thrown away. The scan now runs in the background and reads only files that changed since the last one. A new session waits at most 0.3 seconds for it and otherwise starts from the newest recorded state of the project, sessions created together share one scan, and a failed scan is logged with its full cause.

- **Publishing an edit no longer fails after a project scan that found changes made outside Unary.** In a project whose scan succeeded, files changed outside Unary could make every later publish of an edit fail with "lost the head race 16 times" or "invalid tree preimage", and the first published edit stopped all further scans of that project. Publishing now works against the project's own recorded state, and scans continue after it.

- **A new session opens without reloading the whole session list.** After creating a session the app made five calls one after another, among them a reload of every session of the project. The creation now returns the new session's entry directly, and the remaining steps run side by side.

- **The cleanup of the workspace catalog no longer runs while the app starts.** It began 32 seconds after start, ran for 20 seconds and slowed down the first session created in that time. It now waits five minutes and then for a quiet minute without session start-up, and waits again before each project it cleans.

- **A `printf` in a shell command can no longer exhaust the machine's memory.** In Unary's built-in shell, `printf -- 'format' args` took `--` as the format, and any format without a `%` placeholder that was given extra arguments repeated without end. The command never returned, could not be stopped by its timeout or by Stop, and filled memory at about 70 MB per second until Unary crashed (2026-11-10: over 200 GB within half an hour). `--` now ends the options, and such a format is printed once.

- **New models from an official catalog update reach projects with their own provider packs.** Applying a newer model catalog while Unary ran refreshed only the global provider packs; a project's untouched copy of a pack kept the old model list until restart and hid the new models, for example "unknown model anthropic/claude-fable-5-2". The project copy now refreshes with the next model lookup after the update, open sessions reload the model list before calling a model unknown, and an app release with new built-in models can no longer be outranked by an older downloaded catalog; packs you edited stay untouched.

- **A build waiting for a build slot now says so.** A build or test that waited behind the machine-wide limit showed no output and no reason in its job until it started; its job now reports `build-slot-queue` with the sessions it waits behind, and the note stays in the job's record.

## [0.2.395] - 2026-11-09

### Added

- **Claude Fable 5.2 is available through Anthropic.** It comes with a one-million-token context, a 128K output limit, the new lower pricing ($8/$40 per million input/output tokens), always-on adaptive thinking with Low through Max effort, and Fast mode, which Fable 5.1 did not have.

### Fixed

- **Project snapshots work again after an interrupted capture.** A capture that stopped halfway, because of an error, a busy writer or Unary closing, left the project's snapshot permanently locked, and every later snapshot of the project failed with "already transitioned". Projects set up with an older version could carry such a leftover for weeks. Unary now releases a leftover lock before the next snapshot and releases its own when a snapshot does not finish.

## [0.2.394] - 2026-11-09

- Chat: a final answer whose blocks differ from what streamed no longer stays hidden behind its thinking row until the next turn; the reveal snaps and logs `chat.render.stream_reveal_drift_snap`.
- Builds: at most two heavy builds/tests run at once across all sessions and workers; a third waits visibly for a slot (`UNARY_MAX_CONCURRENT_BUILDS`). An unchanged build or test re-run within 15 minutes returns the previous result and where its full output is instead of running again.
- Workers spawned with `verification: "parent"` may run one scoped package check at the end of their package; only the workspace gate (verify, workspace-wide or release builds) stays with the lead.
- Pilot: agent-desktop parity in Pilot's own Rust. Snapshot refs with background UIA delivery (no focus, pointer or keystrokes), shell surfaces, notifications, clipboard, app/display listing, close/resize/move, desktop event waits, blocked key combos and an integrity gate. Compact element results, and `launch` no longer takes over an open window.

## [0.2.393] - 2026-11-08

### Fixed

- **A busy machine no longer records your project as empty.** When Git did not answer in time under load, the workspace catalog treated the project as having no files, stored an empty snapshot, and then re-stored the whole project on the next capture. A Git timeout is now an error for that capture, and the previous snapshot stays in place.
- **The workspace catalog no longer grows by gigabytes a day.** Every change that touched several files at once also stored each half-finished intermediate version of the file tree, and only the final one was ever used. On a busy project the catalog grew from about 5 to 11 GB in a day and held its write lock for the whole write, which slowed or blocked other sessions. Only the final tree is stored now; the leftovers already on disk are removed by the next regular cleanup.
- **A lead now notices when its worker's build goes quiet for a long time.** A build that wrote its log to `$UNARY_SCRATCH` looked silent even while it compiled, because the variable only exists in the job's own environment; the log now counts as progress and its tail is shown. A job that really stays silent is reported again at twice, four times, eight times the five-minute threshold, both to the worker's own wait and to its lead, instead of once and then not for the next half hour.
- **The Workers board says why a worker was cancelled or failed.** It used to show only `status=Cancelled`, so the agent guessed ("Unary closed") when you had actually stopped the run; each such row now carries the recorded cause and the time it ended.
- **Session debug exports keep the newest workers' transcripts.** In a session with many workers, the oldest ones used up the whole export budget and the most recent workers - usually the ones the export is about - came out empty. Each worker now gets a fair share, the newest first, and a cut transcript says why it was cut.

## [0.2.392] - 2026-11-08

### Fixed

- **Opening a session no longer fails with "workspace catalog: database is locked".** When another session was writing the workspace catalog, creating a session could stop after a few seconds at the Workspace step. Every catalog write during session start now waits for the other writer for up to 30 seconds, as the first step already did.
- **Idle chats are unloaded again, and stay unloaded.** The team panel's background refresh woke up chats that had already been unloaded just to look up which chat they belong to. Each wake-up pushed another chat out, which the next refresh woke up again, so the app kept far more chats in memory than its limit and kept reloading them. The lookup now reads the saved chat instead of starting it.

## [0.2.391] - 2026-11-08

### Added

- **Pick a model while a chat is still running.** Choosing a model during a run no longer waits for the run to finish. The new "model switch during a run" setting decides when it takes over: at the next safe point inside the run (the default), or only when the run ends or is stopped.
- **Agent Desktop now matches vercel-labs/agent-desktop completely.** The agent's on-screen cursor overlay (`cursor_overlay_enable`) works inside Unary, and the app now installs the `agent-desktop` command-line tool next to itself, so `agent-desktop` and its MCP server (`agent-desktop mcp`) are available without a separate install.

### Fixed

- **The managed browser no longer takes the focus away from you.** When the agent opened its browser window, Chromium could jump to the front right after starting, so your typing went into the browser. The window now appears behind what you are doing.
- **Switching the model in a long chat is instant again.** In a chat with a large history, every model switch copied the entire conversation several times over (once per internal settings step), so a single pick took 15-30 seconds and further clicks queued behind it - the picker seemed to ignore the click or to need several. A model, thinking-level or fast-mode change no longer rebuilds the chat history, and the session state is read without copying the conversation; measured on a 412 MB chat, the switch work dropped from about 14-19 s to about 2 s in a debug build.

## [0.2.390] - 2026-11-07

### Added

- **Link App in the footer.** While the Blender or UnaryCad plugin is enabled, the Desktop footer shows a Link App button next to the focused chat. It lists every running Blender and UnaryCad instance and links one to the current Code session or Assistant chat with one click; UnaryCad links can be removed again from the same list.

- **Agent Desktop: vercel-labs/agent-desktop is built in.** A new plugin (Settings › Plugins › Agent Desktop, off by default) gives the agent the complete agent-desktop surface as one `agent_desktop` tool: accessibility-tree snapshots with refs, headless clicks, typing and value changes that move neither mouse nor focus, windows, apps, notifications, clipboard, waits, sessions, traces and batches - all 67 commands, running inside Unary. Every result says whether the action was delivered and verified and whether a retry is safe. The upstream `agent-desktop` command-line tool, its MCP server and its C library build from the same source. Unary Pilot is unchanged.

### Fixed

- **A finished worker no longer leaves its lead asleep.** When an earlier worker had already failed, a later worker that finished successfully sent its lead no notice at all, so the lead never saw the result and its run stopped as "no progress" with goals still open. Every unread worker result now reaches the lead; failures and results that finish together arrive in one notice.
- **A worker stopped by the watchdog is reported to its lead.** A hung worker that the watchdog had to terminate ended silently; the lead now gets the notice like any other failure. As a safety net, the watchdog also re-announces any worker result that finished but was never announced, so a lost notice delays the lead by at most one watchdog interval instead of parking it.

## [0.2.389] - 2026-11-06

### Added

- **Image size and quality are controllable.** The agent can now ask for a specific size (for example 2K 2048x2048 or 4K 3840x2160), quality (low up to max) and background when it generates or edits an image with a GPT Image model. Settings › AI Agent › Image generation gains Image quality and Image size defaults that apply whenever the agent names none, and every image result reports the size it actually got. The sizes take effect with an OpenAI API key, where GPT Image 2.6 Sunburst and Flare are now selectable too; the ChatGPT (Codex) image service accepts the settings but still picks its own size.
- **GPT Image 2.6 is selectable for image generation through the ChatGPT (Codex) account.** The image slot now offers GPT Image 2.6 Sunburst (detailed visual work and editing) and GPT Image 2.6 Flare (fast iteration) under the OpenAI Codex provider, next to the existing 2.5 and 2 lines; the chosen line is sent with every generate, edit and inpaint request.

## [0.2.388] - 2026-11-04

### Fixed

- Goals finished by a worker now get checked off. A worker that reported a goal as completed lost that report when it ended, so its goal went back to open even though the work had landed. When a worker ends without reporting its goals, the lead is now told which goals went back to open instead of finding them quietly reset.
- Finished workers no longer look busy while the lead waits. When the lead was inside a long wait (for example on a build), finished workers stayed "running" until that wait ended, so their results and questions reached the lead up to an hour late. Workers whose edits are already in the shared checkout now finish at once, workers with private versions end the lead's wait so they can be applied, and the lead's wait ends when all its workers have finished.
- A worker's question to its lead no longer gets lost. When a worker asked for a decision and then ended its turn, the lead's answer could be refused with "control/finalization transition in progress"; the answer now waits until the worker has finished closing and then resumes it. A worker that finished with an open question now shows that question to the lead instead of "FYI, no reply needed".

## [0.2.387] - 2026-11-04

### Added

- Remote screen view: choose which screen to watch (All screens or a single display) and switch it live, like TeamViewer/AnyDesk — in the Desktop remote panel and the mobile app, now also over the P2P (WebRTC) connection.

### Fixed

- Long sessions with workers no longer grow their transcript by gigabytes. While `wait_agent` waited for several workers, every pass rewrote the full result record of each already finished worker and woke itself again; one session reached 1.5 GB. A record is now written only when something changed, and it no longer stores the request twice.
- A session no longer gets stuck answering every message with "on-demand tool loading: activation records require one successful result for a preceding activate_tool call". After a long run with many worker records, the end of the run could write the run's earlier messages into the transcript a second time, and the doubled tool results made every later request fail. Messages are no longer written twice, a repeated identical worker finish writes no new record, and a request Unary refuses before sending it now shows as a failed run with its reason instead of as the assistant's answer.
- Compaction is faster on large sessions: it no longer serializes the whole transcript for extensions when none is installed, clones only the entries it needs, and logs how long each phase took.
- Clicking Compact while a compaction is running no longer shows "is compacting — retry when compaction completes"; the second request waits for the running one and gets its result. The Compact button stays disabled until its request has finished, even if the turn ends with an error meanwhile.
- Work done in runs started by worker notices or Intercom messages is now recorded in the session ledger. Before, only runs started by your own messages were recorded, so a compaction after hours of such runs handed over an empty ledger segment and reported it as available. An empty segment now falls back to the transcript, and turns that cannot be recorded are counted.
- The session ledger read during compaction no longer opens the turn files of already closed segments, and replaying earlier compaction boundaries no longer re-parses the whole ledger each time.

## [0.2.386] - 2026-11-03

### Fixed

- Builds started while other sessions edit no longer fail after about a minute with "current-source capture deferred: workspace busy". The end of the source capture now waits for the workspace like its start, so a build waits until its own deadline and an identical build can join the one already running.
- File claims with isolation off: a writer that was handed a file keeps its place while it is still busy in a long tool call, and a cancelled waiter leaves the queue at once so the next writer gets the file. Commands in another repository no longer stop a shell from claiming the files it writes, and a claim that cools without a turn now wakes the next writer.
- Checkout registry: registries of different repositories whose names share a prefix are no longer merged into each other. The claim-store cleanup now runs at startup and every 6 hours and always logs a summary; it removes rows of dead processes, empty registries, old spellings and expired temporary stores. A damaged registry is set aside so it can no longer block ending a session.
- Build records: a finished, failed or abandoned build removes its "prepared" record, so no build stays shown as unfinished.

## [0.2.385] - 2026-11-03

### Fixed

- macOS: Cmd+W closes the active editor tab like Ctrl+W on Windows instead of hiding the whole window, and Cmd+Q quits through the regular shutdown, so running sessions stop cleanly and the next start no longer reports an unexplained exit.
- macOS: clicking the Dock icon brings the Desktop back after its window was closed.
- macOS: confirming a Chinese, Japanese or Korean input candidate with Enter no longer sends the half-composed message.
- macOS: voice input can use the microphone in the signed app; macOS now asks for microphone access.

## [0.2.384] - 2026-11-03

### Fixed

- A folder that became a Git repository after its first session (`git init` later) opens new sessions again instead of failing with "immutable project already exists with different data"; its project and checkout records now follow the Git state of the same folder.
- New sessions on an already-opened folder no longer rebuild and discard the folder's tracked-file list while preparing the workspace, which took up to 10 s on large repositories.
- A message you sent no longer reappears in the composer when you switch back to its session.
- Stopping a background command right after starting it now stops it. A stop that arrived while the command was still launching reported "not yet complete" and the command kept running.
- File claims with isolation off: a writer queued behind another session keeps its place when its own run ends and is handed the file when the holder's turn or run ends, when the holder is closed, or when the woken writer ahead of it gives up. A waiter that is still working is told at once instead of after its run; a handover for an edit that already landed is dropped, and one release no longer wakes the same writer several times. The queue notice is about 80 words instead of 270 and no longer gives advice that lost the writer its place.
- Workers are registered under their real session id. Finished workers no longer leave checkout rows that count as running until the app exits, and workers spawned with `verification: "parent"` are refused builds and tests again.
- The checkout registry uses the claim store's canonical repository key and lock, so a different spelling of the path or a subfolder no longer splits it. Sessions that ended in this process stop counting as running, closed sessions leave both records, and empty registries are swept with the claim store. Temp repositories reached through a non-canonical path keep their claim files in the temp directory.
- Builds: identical builds started at the same time join one run instead of compiling twice in two lanes, a build that meets a busy workspace keeps waiting until its own deadline instead of failing after about a minute, and background builds in a new lane say "Cold build lane". A build that stops right after preparing records a terminal result, and unfinished records of idle lanes are closed at once.
- Smart isolation: a shared-mode edit to a file that another session holds privately now says so (that session merges it when it delivers), and changing smartFileIsolation in settings applies to running sessions at their next run.

## [0.2.383] - 2026-11-02

### Added

- UNARY PDF plugin: with UNARY PDF installed, enable it in Settings > Plugins and the agent can open, read and edit PDFs through the app's own tools (text, pages, objects, annotations, forms, export, redaction). Edits show up live when a UNARY PDF window is open.
- Spec references now connect mock variants, component owners, plans and source-bound evidence without treating selection or recorded test results as human approval. A real-component Desktop fixture makes narrow/wide layouts, themes and interaction states reproducible.

### Changed

- Overseer gets its own Desktop workspace beside Code, Assistant and Chats, visible only while Codex Voice and its Overseer option are enabled. The workspace reuses the Overseer conversation and global call, with session browsing, history reading and explicitly targeted controls.

### Fixed

- Documents refuse identity collisions, publish primary content atomically and report history/scan failures. Purges and scope moves expose deleted, retained, failed and unknown resources instead of claiming complete success after partial cleanup.
- Attachment and Research garbage collection now refuses deletion when directory enumeration, record decoding or reference coverage is incomplete.
- Memory search, FTS, recall and node reads share owner/lifecycle/withdrawal checks. Incomplete or mistyped legacy rule fragments remain available for reversible review without becoming active model guidance.
- PDF imports extract actual text; images without supplied OCR and failed or unextracted binaries retain their originals and explicit status rather than becoming knowledge from filenames or printable-byte scraps.
- Recall keeps workspace and global memory independent, filters Research by its category and deduplicates shared Docs identities only at the same owner, scope and revision.
- Worker recovery uses hook identity, epoch and canonical revisions rather than record counts; team membership, slots and pending close obligations survive restart. Plugin storage is inventoried with bounded, paginated diagnostics and conservative retention labels.
- Spec metadata now uses one YAML reader across the index, tree and contract discovery. Desktop editor responses are workspace-fenced, and governing-section reviews retain revision/read checks without certifying unread sections.

## [0.2.382] - 2026-11-01

### Fixed

- UNA Voice now knows how Unary is built: it is told that the chat runs a separate coding agent (named with its model, effort and mode) that it operates by sending messages, that quick web lookups are its own, and who "you" and "the chat" are. It no longer claims the chat's work as its own. "Search online for …" and "search it yourself, the chat is busy" are answered by the quick lookup instead of steering the chat.
- UNA Voice quick web lookups always run a real web search now: a fast classification call decides first, and only a lookup makes a second call with the search required. Before, the router sometimes answered current facts from memory or with a promise to search. Coding requests reach the chat faster (about 1–2 s routing).
- Messages UNA Voice forwards to the chat carry a short marker for the coding agent (speech transcript, "you" may mean the voice). The chat shows them with a "Voice" label, and the composer queue no longer shows the marker text.
- UNA Voice keeps its own searches apart from the chat's: its quick lookups come back labelled as its own with the query it searched, and the status it receives lists what the chat agent is searching this turn (research queries, topics and URLs) as the chat's work.
- UNA Voice knows what it is and the split of work: quick facts and lookups (weather today, "who was Steve Jobs", "what's new with X") it answers itself; research jobs ("recherchier", Deep Research, anything thorough, compared or saved) go to the chat.

## [0.2.381] - 2026-10-31

### Fixed

- Gemini 4 and other Google Antigravity / Gemini CLI models no longer abort long answers with HTTP 408: the stream watchdog now recognises Cloud Code Assist's wrapped response frames as progress, and the response-header deadline was raised from 20 s to 300 s because Cloud Code Assist sends headers only with its first chunk (the connect timeout stays at 20 s so dead hosts still fail over quickly).
- The repository build ledger no longer counts `pwsh -File scripts/verify.ps1 -List` as a build.

### Changed

- Subscription and coding-plan OAuth specifics (Claude Code, Antigravity, Gemini CLI, GitHub Copilot, Meta Muse Code, xAI, Z.ai) are now provider data instead of code: identity, headers, beta flags, request and response envelopes, endpoints, quota codes, per-model runtime ids and thinking tables, project discovery, credential exchange, subscription usage endpoints and usage-limit markers live in `providers.json`, `oauth_providers.json` and the provider packs, so a signed catalog can update them without an app release. Codex remains the only provider whose OAuth specifics stay in code.
- Anthropic OAuth tokens are recognised only by their `sk-ant-oat` prefix, not anywhere in the key.

## [0.2.380] - 2026-10-30

### Fixed

- A release build run as `cd <checkout> && …` now stays in its pinned build lane: the absolute `cd` into the checkout is redirected into the lane, so the release source guard finds its marker instead of refusing the run.
- Unary Secure pre-commit secret scanning now covers commits made through the `git` tool (`command` and `commands`), unstaged and newly added working-tree commits (`git commit -a` / `--only`, `git add … && git commit`), `git -C` target repositories, shell wrappers (`sh -c`, subshells, environment prefixes), and URA `sh` git writes; browser login autofill also checks each same-process iframe's host against the credential's allowed domains.
- The public release download page (`dl.unary-code.app/versions`) now paginates across all R2 release prefixes so current `0.2.3xx` builds appear instead of stopping at `0.2.298`, hardens byte-range and SemVer prerelease ordering, escapes card metadata via `textContent`, and serves a Content-Security-Policy header.
- MCP HTTP OAuth token storage now coordinates writes and token refreshes across processes with an exclusive file lock, atomic index replacement, and merge-on-save so concurrent sessions cannot overwrite another server's keyring entry; URA `fs.mkdir` is confined to the workspace root.
- Provider error handling now classifies HTTP 5xx responses as transient retries even when the body mentions context-window terms, bounds malformed `Retry-After` / rate-limit headers without overflow, gates mutating plugin calls on the named-pipe editor transport, caps oversized text/image file previews before reading, and skips isolated build-lane setup for `scripts/verify.ps1 -List`.

## [0.2.379] - 2026-10-29

### Fixed

- Google Antigravity models answer again instead of failing with "Resource has been exhausted": requests go to Google's daily endpoint first, Gemini 4 Argon uses the ids Google serves for its Low/Medium/High effort, Claude keeps working after a tool call, and Gemini 3 Flash no longer offers an "off" effort Google rejects. The model list now matches what an Antigravity login actually serves: Gemini 4 Argon, Gemini 3.8 Flash, Gemini 3 Flash, Claude Opus/Sonnet 5.5 (Low/Medium/High) and GPT-OSS 120B; retired models are gone.
- Google Antigravity now identifies itself like the current Antigravity desktop app (its latest release is looked up automatically), stays on Google's daily endpoint and no longer retries a real quota exhaustion: the error says when the quota resets instead of retrying in a few seconds.
- How Unary identifies itself to subscription logins (Claude Code for an Anthropic login, the Antigravity app, Gemini CLI) is now catalog data instead of code, so a new client version can ship with a catalog update instead of an app release.

## [0.2.378] - 2026-10-29

### Fixed

- An unanswered question card no longer ends the run. When the answer time runs out, the agent takes the recommended option for reversible, low-stakes choices and says so, leaves decisions too important to make for you undecided, and continues with the rest of the work. Before finishing it asks the open question once more; if that also goes unanswered, the final report lists what is still open. An unanswered offer ("Shall I fix it?") is never treated as Yes.
- The message you just sent no longer comes back into the composer. Sending now deletes the saved draft immediately, and a draft that was sent is never loaded back from the disk copy, also when you switch away and return right after sending.

## [0.2.377] - 2026-10-28

### Added

- `UnaryCode-Setup.exe`: a branded online installer in the Desktop style. It signs in with your Unary account (the app then starts signed in), downloads the current signed Windows build from the release channel, verifies its signature and installs it for your Windows account.
- Assistant and Chats lists: select several chats with Ctrl/Cmd-click or a Shift-click range, like in Code, and archive, restore, delete, move or copy them together from the selection bar or a right-click; Esc cancels the selection.

### Fixed

- A failed or cancelled release build no longer blocks the next run with "Pinned build lane … is busy": a pinned lane is freed as soon as its process tree has ended, also when the execution view is kept for review.
- A lead agent with nothing else to do no longer polls a running worker with one `wait_agent` call after another: a timed-out wait now tells it to end its turn, and the worker's completion wakes the session with the result.

## [0.2.376] - 2026-10-28

### Fixed

- Assistant and Chat sessions in the home folder no longer fail with "cannot prepare session scratch: scratch must be outside the checkout" (a regression in 0.2.375): their scratch lives in Unary's own `~/.unary/workspace-runtime`, which is never treated as session source.

## [0.2.375] - 2026-10-28

### Added

- Assistant: a chat without a project can be moved into a project from its context menu ("Move to project …"). It keeps its history and working folder; the project instructions apply from the next message.

### Fixed

- Assistant and Chat sessions in the home folder no longer fail with "cannot prepare session scratch … Zugriff verweigert": their runtime folder is no longer created in the unwritable parent of the home folder (`C:\Users`).

## [0.2.374] - 2026-10-28

### Added

- Desktop Storage offers a default-off, revocable Cargo-pruning opt-in per registered checkout, with identity checks and started/completed receipts. Interrupted cleanup is retained for review rather than silently replayed.
- Desktop Storage lists items that need your decision: Windows `nul` files in registered checkouts, Git worktrees and build directories. Each offers Delete, Keep or Always keep. Delete moves the item aside first and then purges it with a receipt; tracked files, dirty or unmerged worktrees and build directories under a live Cargo lock are refused. An interrupted deletion stays listed until you resume or restore it. "Run dry run" shows what the cache collectors would free, with the last runs kept.
- Old files in `~/.unary/tmp` and spilled tool-output logs are removed after 14 days unless a session transcript still references them.

- GPT-6.1 Terra (`gpt-6.1-terra`) for OpenAI API keys and ChatGPT/Codex sign-in: 1.05M context, 128K output, effort none..max, $1/$5 per 1M tokens. Codex Fast mode is available for it. The OpenAI API pack's six-model list now drops GPT-5.5 to make room.
- `apply_agent` integrates a finished worker's preserved package with the single-use `integrationToken` from its `wait_agent` result; `preview: true` reports what would land without writing. `wait_agent` now shows an `integration` block (state, files, token, next step) for completed and failed workers, so a conflicted or blocked package is recoverable by the lead instead of only showing `errored`.

### Changed

- Plugins no longer drop out of a session at start: one slow plugin (mostly Docs, which rewrote its manifest on every turn) used to make every plugin's session-start contribution time out. Each plugin now has its own deadline and keeps its last good contribution, Docs loads pages without writing, and session starts use a bounded share of the plugin worker pool.
- Secret placeholders fail closed: if the vault cannot resolve `{{secret:…}}` in time, the tool call is refused instead of running with the placeholder text.
- Browser: managed Chromium sessions whose browser exited are cleaned up, so the 8-session limit no longer blocks new browser sessions.
- Plugin Creator registers a new plugin crate in the right harness file and leaves nothing behind when registration fails.
- Every first-party plugin action is classified for UI receipts and unknown actions are refused; the Index and VM panels request receipts for their changes.
- Research: the Perplexity engine uses the Agent API (Sonar ended on 2026-09-27); Google Custom Search results warn that the API shuts down on 2027-01-01.
- Voice: OpenAI speech uses `gpt-realtime-2.1-mini` over the Realtime API; `gpt-4o-mini-tts`, `tts-1` and `tts-1-hd` stay selectable until their shutdown on 2027-01-06. Transcription defaults to `gpt-transcribe`.
- The Adaptive Stripe plugin id is `unary.adaptive-stripe`; existing settings move over automatically.
- Blender: read-only sessions cannot run changing Blender actions, and the tool no longer claims Python is off by default. The Coworker Blender package installs `mcp-for-blender`.
- UNA Overseer works across all repositories and has its own chat. The titlebar pill starts and stops the Overseer and opens the "UNA Overseer" chat, where everything it hears, says and does appears: your requests, its answers, actions with their results, notices from other sessions, and the answer of a session you sent a message to. You can also type there: "stopp die Session im unary-code Repo" works without a voice call, and destructive steps wait for your "ja". Its settings are global, `new` asks in which repository the session should open, and notices only name sessions it can act on.
- UNA Voice in a session stays in that session: it no longer offers Overseer actions or a "Global Overseer" start, and its conversation always appears in that session's chat, also with the Overseer switched on.
- The per-role worker "Verification" setting is removed: verifier workers have been switched off since 2026-10-10, so the control changed nothing. Saved values are ignored. Workers left waiting for a verifier by an older build now fail on the next dispatch with their package preserved for recovery.
- Dead worker and session `worktreeMode` fields are removed from Desktop and the harness; persisted settings remain readable and the old field is ignored.
- Fourteen benchmark scenarios that require the retired worker protocol are withdrawn from the selectable suite rather than advertised as runnable. Historical scoring fixtures remain; current orchestration benchmarking needs native-tool replacements.
- The hidden `subagent_dispatch`, `subagent_inspect`, `subagent_control` and `subagent_apply` tools are removed; workers are driven only by `spawn_agent`, `send_input`, `wait_agent`, `resume_agent`, `close_agent`, `apply_agent` and `spawn_team`. A URA `ura(...)` child now starts through `spawn_agent`. Old transcripts still render their calls.

### Fixed

- An agent's `fold_back` finalize no longer moves the workspace under its own running turn. For a native isolated tree and for a managed Git worktree, the request is admitted and the session folds the tree back between steps; Stop or Pause cancels it, a changed workspace supersedes it, and the result appears as a notice in the chat.
- Anthropic refusals now show the provider's reason (category and explanation) instead of "provider returned an empty error response", and are no longer resent: the same conversation used to be retried 13 times across accounts, and Anthropic refuses it again by design. `sensitive` and unknown stop reasons are reported by name the same way.
- Cargo cache pruning retains each crate's newest artifact across toolchain changes, acquires a real lock even when the lock file was absent, and preserves unknown-age or linked entries instead of treating them as stale cache. Classic and split build layouts no longer reuse one trash destination.
- A session no longer stops after acknowledging an instruction you sent while it was still running (for example "alles get live" in reply to its deploy question). The forced answer to such a message cannot use tools; it can now mark that work remains, and the run then continues with its tools instead of waiting for your next message.
- UNA Voice no longer gets stuck saying "I'm still waiting for the answer": when a spoken request does not reach the coding session within 15 seconds, the voice says so instead of waiting forever and asking you to repeat yourself. Waiting notes stop after two.
- UNA Voice reads an answer once per spoken request. It no longer reads every turn of the session aloud (typed messages, automatic continuations) or the same answer several times; those answers stay available when you ask.
- UNA Voice no longer treats its own requests as new typed messages, no longer starts a second coding turn when the same request arrives twice within 30 seconds, and on the WebSocket route no longer hears its own playback through the microphone.
- UNA Voice calls no longer end with "Voice response timed out" after a silent read-aloud, and a request the voice handed to the session is no longer shown in the chat as "answered by the voice".
- UNA Voice can report the session status ("what exactly is running?") in Chats, Cloud chats and the Assistant, not only in workspace sessions. Voice decisions are now logged under `codex_voice`, so a stuck call can be diagnosed.
- Starting or steering a worker (`spawn_agent`, `spawn_team`, `send_input`, `resume_agent`, `close_agent`) counts as progress for the stall guard again, an `apply_agent` that lands a package arms the post-apply check, and `spawn_team` now needs the subagent permission like `spawn_agent`. These rules still named the hidden tools, so the live ones fell through.
- The "Resolve in session" worker-issues prompt and the native delivery conflict text no longer send the model to tools that do not exist.
- A sent message no longer reappears in the composer when you open its session again. The composer was emptied only after the send had finished its preparation (up to several seconds while the agent was busy), so switching sessions or closing the panel in that moment kept the message as an unsent draft, including across restarts. The composer now empties the moment you send; a message that could not be sent comes back into the composer, or into its session's draft if you have switched away, ahead of anything typed since.
- Disk space no longer grows without bound (about 4.5 TB measured on one machine). Unary prunes stale Cargo units and incremental caches in the build trees it owns (build-lane outputs, `target/agent-build`, `target/verify-*`) under Cargo's own lock, removes the second intermediate tree the `cargo_check` tool used to leave in every lane, keeps each project's build lanes under 100 GiB (every idle lane goes when the volume has less than 10% free), keeps the build receipts of the newest 20 runs per lane, and reclaims the runtime folders of deleted sessions. Your own `target/debug` and `target/release` require the separate per-checkout opt-in in Desktop Storage.
- Build refusals no longer suggest building into a new output directory (`--target-dir`); an output root a session's command creates is recorded with its session and command.
- The workspace catalog is swept every six hours while the app stays open, not only at start, and a catalog above 1 GiB no longer re-runs its whole sweep on every start.
- `> nul` in the agent shell on Windows writes to the null device instead of creating an undeletable file named `nul`.
- Test runs no longer fill `~/.unary/claims` with lock files (30 000 measured).
- `diagnostics` action `storage` and the Usage dashboard show where Unary's disk space goes, per checkout and class (cache, history, work, your content), with each catalog's last sweep.

## [0.2.373] - 2026-10-26

- Builds stay warm: a build from a session with unsaved private changes now reuses the shared build lanes even when they were last used for a different build (for example tests instead of the app). Before, it started a cold lane of its own and rebuilt everything; in UnaryCad that had left 49 such lanes with 62 GB on disk. When a lane is reassigned, opened fresh or every lane is busy, the agent is told why in the build result, and a refused build names the reason and what to do.
- Old build lanes are cleaned up automatically: lanes of sessions that have ended are deleted together with their folders, and lanes of running sessions that sat unused for a day are removed. A lane a build is still using is never touched.

- Shell commands: a read-only command that contains an escaped quote (`\"`) and later a quoted search pattern naming a build tool (for example `python3 -c "…\"…" | grep -E 'cmake|ninja'`) no longer counts as a build. Before, it waited for a full build source capture (74 s in the session that prompted this) before the command even started. Real builds are recognized as before, including PowerShell paths that end in a backslash.

- Build errors in another session's files now reach that session's owner. When a build or test fails on a file another session changed, the agent is told which session to hand the error to. That now works for background builds too (the exit message and `shell_watch` results), for `svelte-check` output (`npm run check`), for edits made with absolute paths, and for edits older than a day: the erroring line is traced with `git blame` to the session whose edit the commit carries. If no session made the change, the agent gets the commit, author and date to give you. In the session that prompted this, a deploy worker's `npm run check` failed on another session's portfolio file, nobody was named, and you had to pass the error on by hand.

## [0.2.372] - 2026-10-25

- Builds: a checkout whose build was refused on every attempt with "missing durable object … pinned by `workspace-capture` (in_flight)" builds again. An earlier interrupted source capture had left one stored file behind without its catalog entry, and every later build skipped that file instead of recording it. The next build now verifies such a file against its hash and records it; a file whose bytes do not match is still refused. No restart or manual repair is needed.

## [0.2.371] - 2026-10-25

- UnaryCad, Blender and the standalone Blender MCP can now work with several open instances at once. One chat stays linked to several UnaryCad apps (`instance` makes one current, `pid` targets any per call, `disconnect` drops a link). Any `blender` action takes a `pid` that targets another Blender for that call without changing the chat's pinned one. `unary-blender-mcp` accepts repeated `--pairing` and `--discover` and routes each call by `pid`. Nothing is ever guessed between instances.

- Agents finish first and build once. When several workers serve one task, the lead now waits for all of them, builds once and sends each error to the worker that owns the file; workers writing into one shared build no longer build on their own. In the UnaryCad session that prompted this, the lead built 13 times while 7 workers were still writing (3 of those builds were green) and spent about 30 of 110 minutes waiting on builds.
- Builds: new files that git does not track yet (and that are not ignored) are now part of the build, so a freshly written source no longer fails with errors such as CMake "cannot be read" and no `git add -N` is needed. A build started while other agents are still writing now waits until their files settle instead of failing after six attempts, and names the paths and who is writing them. The "files changed during the run" note now says to build once after the writers settle instead of asking for a rerun.
- Background builds return immediately: preparing the build view (10 to 80 seconds per start in that session) now runs inside the job, and the build receipt shows how long capture, view refresh and the build itself took. If the preparation is refused, the reason arrives as the job's result.
- Commands that only mention a build tool inside a heredoc (for example a script containing "cmake") no longer run as builds.
- `shell_watch` `wait` now waits until the job ends (up to an hour) instead of returning every two minutes; a new message from you or a worker's decision request still interrupts it. The edit tool's `verify` option now says to run it once after the last edit, not after every edit.
- Desktop: when a repository's rules file is not optimal (duplicated sections, a rule that forces a build after every small change, contradictory build rules or unresolved sync conflicts), a banner says "Repository rules are not optimal" with "Improve rules", which sends a ready-made fix request to the agent, and "Ignore", which holds until the file changes. The rules dialog lists every finding with its token cost per request.
- Rules sync: UNARY.md, AGENTS.md and CLAUDE.md that differ only in their title no longer get the whole file merged in twice, and sync conflicts are now kept and shown instead of being dropped.
- Session debug bundles now include the harness notices in worker transcripts and the reason each worker stopped.
- The hint added to messages you send during a run no longer suggests running a check before the agent answers.

## [0.2.370] - 2026-10-25

- Background commands: a finished command (for example `cargo build`) no longer stays "running" when a process it started keeps its output open; the job now settles on its real exit code after a short drain, as directly spawned commands already did. This also releases the build CPU budget that such a stuck job held, which could block the next release build with "build CPU budget busy".
- Managed browser: taking over a session now interrupts a running wait immediately instead of waiting for it to time out.

- Unary Secure: stored logins now also fill a "repeat password" field on the same form, so set-password pages (for example a new Hoststar mailbox) no longer stay half empty. When the agent asks to use a login on an additional site, you get the same one-click Allow card instead of being sent to the vault panel.

## [0.2.369] - 2026-10-24

- Unary Secure: when the agent may not fill a login on the current site (for example a Hoststar login on unary-hub.app), the chat now shows a card with "Allow once", "Always allow on <site>" and "Deny" instead of a dead end; one click and the agent continues. The agent can also open a form in the chat where you store a login yourself (website, email, password); the password goes straight into the vault and the agent only learns the entry's name.
- Windows installer: refreshed installer icon and sidebar artwork.

## [0.2.368] - 2026-10-24

- Unary Secure: logins now show where website, email and password are set. A password entry opens on a new Login tab with Website, Email / username and Password (Replace…) plus the "Agent may sign in here" switch; the Logins list shows email · website instead of the import alias, and New credential asks for website and email and suggests the alias. Fixed: any edit in the panel (even a grant switch) erased the stored username, so `secure login` had nothing to type. Logins imported from Google Password Manager get their username moved out of the purpose text into the real field on first load; values in the keyring are untouched.
- Unary Secure (already in 0.2.367, not listed there): the Logins list sorts by name, newest created or recently changed; the search box no longer draws a second border; the ENV tab in the Desktop app uses the session's workspace instead of failing with "workspace root is required".

## [0.2.367] - 2026-10-24

- A chat with a large PDF attached (over the provider's limit, e.g. 59 MB on Claude) or an unsupported file type no longer fails on every following turn with "PDF exceeds 32 MB maximum". Such attachments are now left out of the request and replaced with a short note, oldest first if several together are too large; the original stays in the session. The PDF size error is also recognized as an attachment problem, so it no longer cycles through every stored account.
- Account menu: "Check for catalog updates" now says "You're up to date" when no newer model catalog exists, instead of showing nothing.

## [0.2.366] - 2026-10-24

- Managed browser: sites flag it as a bot far less often. It no longer reports `navigator.webdriver` or launches with `--enable-automation` and a fixed English locale; its custom Chromium switches had been silently ignored because of a doubled `--` prefix. Headless sessions and the research renderer now present a regular Chrome identity instead of `HeadlessChrome`. Pressing Enter now actually submits forms.
- Stored logins (`secure` `login`) now wait for late-loading forms, find fields inside iframes and shadow DOM, and complete two-step logins (username first, then password). A page without login fields now reports an error instead of a false success, and the browser guidance points the agent to stored logins.

- Session auto-rename, commit messages and other one-shot helpers no longer fail with HTTP 400 "does not support disabled thinking mode" on Kimi K3.1 Pro / K3 (Kimi Code). Their default `off` effort now goes through the same per-model normalization as chat turns, so models that cannot turn thinking off get their lowest valid effort instead.

## [0.2.365] - 2026-10-24

- Model catalog: the model catalog now updates separately from the app. At startup the app fetches the official catalog, checks its signature against the key built into the app, and offers it as an update; the account menu shows the app version and the catalog revision separately. If you have edited provider files or keep your own local packs, the update dialog asks whether your local data or the official data wins; with "official", your edited files are kept as `.bak` backups. A catalog can change models, names, prices and limits, but never where or how requests are sent; that still needs an app update. Only the Unary team publishes catalogs.

- Desktop titlebar: on narrower windows the buttons no longer overlap. The bar hides the run counts first, then the "Desktop" label, then the Overseer state word, instead of drawing "Open IDE" and "Settings" under the Overseer pill.
- Codex Voice is now called UNA Voice, and its Overseer is now UNA Overseer. The titlebar pill shows "UNA Overseer" with an on/off state, and its switch is labeled "UNA Overseer mode". Existing settings carry over unchanged.

- Providers: Kimi K3.1 Pro and K3.1 Pro HighSpeed are available on Moonshot AI (`kimi-k3.1-pro`, `kimi-k3.1-pro-highspeed`, global and CN) and the Kimi Code membership (`k3.1-pro`, `k3.1-pro-highspeed`), with 1M context and the low/high/max effort picker. They are defined as data in `crates/unc_rust_ai_catalog/model_extensions/kimi_k3_1_pro.json`; the Kimi Code effort picker now follows the catalog's reasoning-effort flag instead of the hard-coded `k3` id.

## [0.2.364] - 2026-10-24

- iPhone: Remote Control no longer drops streamed text or misses tool results. The desktop merged away text chunks while throttling the live stream, and refused events above the relay's 1 MiB frame limit, so a long tool result never reached the phone; both are fixed. After a reconnect or when you bring the app back to the front, the open session reloads instead of showing stale content. Sessions open with the newest 80 messages, with "Load older messages" for the rest, and finished steps no longer trigger a full reload each. When the desktop's connection to the relay keeps dropping right after connecting, it now backs off instead of reconnecting every second, and the log names why each connection ended; both sides also notice a dead connection within 60 s. The phone's screen view uses the relay's ICE servers and retries peer-to-peer once before falling back to the relay, and cloud-chat images ask all holder devices at once instead of one after another.

## [0.2.363] - 2026-10-23

- Providers: Google Antigravity sign-in follows the maintained `cortexkit/antigravity-auth` plugin (the previous reference was archived). Project discovery now asks `daily-cloudcode-pa.googleapis.com`, then production, with the current agy CLI identity, onboards a managed project when none exists, and model requests use the same host and identity. When Google refuses, the error now shows the HTTP status and Google's own message; before, every refusal read "returned no project id" and the actual cause was lost.

## [0.2.362] - 2026-10-23

- Sessions: text you type in a new session while it is still being created, without pressing Enter, now appears in that session's chat input once it opens. Before, it was erased on the switch to the new session; the same erase could hit an unsent draft when switching back to a session that had to be reloaded.

## [0.2.361] - 2026-10-22

- Sessions: starting a session no longer fails with "Could not prepare … missing capture: <id>" on a project whose catalog cleanup had deleted the checkout's current capture. The cleanup kept the newest 50 captures per project and could delete the one the checkout itself stands on, so every new session on that project failed the same way and Retry could not help. The cleanup now always keeps each workspace's current capture, and a catalog an earlier cleanup already damaged is repaired on the next session start (logged as a warning).

## [0.2.360] - 2026-10-22

- Sessions: asking the agent to start a session in another repository ("spawn a session in the CAD repo and hand this over") now works. It asks once with the approval card, then creates the session there and hands over the task; `session` action `create` takes `targetWorkspace`. Before, it could only create sessions in its own workspace and fell back to a hand-over banner.

- Memory: asking "what was worked on here in the last 3 days" now has a real answer source. `memory` `sessions_list` with `since` (`3d`, a date or a timestamp) lists every session of this repository in that period, live or ended, with title, goal, files written and worked-on items, from the session index plus the repository's memory ledgers, which reach back past transcript retention. With `includeGit` it also shows each session's commits, its uncommitted files, and the commits and changes no session explains. Before, the tool listed only live sessions, so agents fell back to `git log` and missed uncommitted, audit and abandoned work.

## [0.2.359] - 2026-10-22

- Steering: a foreground `bash` or `powershell` call that has run 5 s moves into the background when a human message is waiting, so the steer is read at once; its output and Stop behave as before. Peer and machine messages do not park it.

- Steering: a message you send while the agent works is delivered exactly once, also when Esc re-sends a chip the agent already read; before, it could appear twice. After answering a mid-run message the agent no longer adds a redundant "nothing left to do" reply, and a command you stopped is reported as stopped by you and is not rerun on its own after your next message. Delivered messages are recorded as delivered during the run, so a crash or restart no longer brings them back.

- Chat input: switching to another session and back while a run starts no longer shows the session as idle, so Enter steers or queues instead of sending a second prompt; after Stop, a late streaming update no longer flips the panel back to running.

- Notifications: with Delivery "Only in background", Settings › Notifications now has checkboxes for Code, Assistant, Chats and Automations. A checked surface still notifies while the app is focused on a different surface (e.g. a finished Chat or Assistant turn while you work in Code); the surface you are in stays quiet.

- Verification: a gate run inside a build lane no longer drops or fails checks for reasons that have nothing to do with the code. In a build view, git listed no files, so the JavaScript type checks, tests and bundles silently left the plan, the orphan-crate and line-budget checks covered nothing, the rules check failed, and the formatting check stopped at the Windows command-line limit. The gate now reads the checkout's file list in a view and formats in batches.

## [0.2.358] - 2026-10-22

- Setup: a first-run dialog shows what Unary's tools use on this PC and installs what is missing. Unary's own shell, file and search tools need nothing; Git is recommended; Node.js, Python, Rust and PowerShell 7 are offered only for projects that need them. Installs run through winget on Windows or Homebrew on macOS, one at a time with visible progress, and Unary picks up the new PATH without a restart. Without a package manager the dialog links the vendor download. The same list is in Settings › Tools on this PC, and Unary reminds you at startup while Git is missing (can be turned off).

- Git login: checking whether you are signed in to GitHub no longer opens a Git Credential Manager sign-in window. The check ran `git credential fill` with prompts allowed, on opening Source Control, on every window focus and in the background, so without a stored GitHub login a sign-in prompt could come back each time the window got focus again. The check now never prompts; push and pull you start yourself still can.
- Git setup: Source Control setup is now a checklist (Git, name and email, repository, GitHub, remote). It stays visible after "Initialize Repository" until a remote exists or you continue without one. GitHub sign-in uses the same dialog as the branch panel, including the personal access token fallback, and an existing remote URL can be added without signing in to GitHub.
- Without Git: Source Control says when Git is not installed, links the download and says to restart Unary; git commands report "Git was not found" instead of a raw OS error. A commit without name and email asks for them and retries; a failed remote login explains the fix, and git no longer waits on a terminal prompt nobody can answer.
- GitHub: a sign-in whose token was saved but whose account lookup failed no longer reports "Sign-in Failed". A token that could not be moved into the OS credential store is still used, and Source Control shows why.

- Desktop projects: adding or editing a project now requires the full path of an existing folder. A typed name such as `test 2` used to be joined onto the app's install folder and stored as `C:/Unary/Code/test 2`, a folder that does not exist; the app then showed "Something went wrong" (a freeze that followed in the same report is not explained yet). The dialog now flags a path that is not absolute and shows the backend's reason ("not a full folder path" or "does not exist") instead of a generic failure.

- Desktop chat: switching back to a session whose run is still streaming no longer shows the previous run's transcript as idle. The panel now takes the backend's run state after the warm restore even when live events arrive during the fetch, and the busy state is written by the run events themselves so a warm stash cannot disagree with the live panel.
- Desktop chat: the divider above each assistant turn (`turn 4 · 3m 12s · 5.1k tok`) now states the turn's own span, from its first streamed event to its completion, which the runtime stamps on every finished assistant message. It used to measure the gap to the previous visible row, so a stale window labelled a 3-minute turn "2h 52m" and a folded notice between two turns reported the earlier turn's length. A turn that is still streaming, or one stored before this version, shows no span instead of a wrong one.
- Docs: `docs/spec/known-limitations.md` no longer lists the natural-exit shell wake as a red test; `self_stop_wake_natural_exit_still_wakes_the_session_once` passes at 0.2.357 (2/2, re-measured 2026-10-21).
- Long writes: the model now knows its output budget. The per-turn session facts state the cap for each response (on Anthropic models a third of the model maximum, for example 42,666 tokens on a 128,000-token model) and say to write a large file with `edit` `write` followed by `append`. A tool call cut off at that cap used to be answered with "re-issue it", which repeated the same oversized call; the result now says to split the work into smaller calls.
- Edit tool: the new `append` action adds text to the end of an existing file, so a large file can be written in parts instead of in one response that may hit the output cap.
- Retries: a stream that drops after at least 90 seconds of output, or for the second time on the same request, is no longer replayed identically. The harness handles it like an output cap and asks the model to split the call, and the retry notice says how long the attempt streamed. In the measured case an edit of about 40k tokens was cut four times (128 s, 35 s, 99 s, 48 s) before the split request came, after about 11 minutes.
- Builds: a build or test that was cancelled, killed or kept for review no longer leaves its warm build lane busy until Unary restarts. Once every lane was held that way, each later build was refused with "all warm build lanes for this workspace are busy or quarantined" although no build was running.

- Workers: the lead now hands over what it already knows. A worker's brief lists the files the lead read that are still unchanged (paths only, never content), and the lead is asked for a "Known" block and a "Read first" list naming file and symbol instead of line numbers. Workers open the named files together in their first turn and look code up through the index instead of reading it with `sed` or `cat` in the shell. Measured on 0.2.357: a worker's first edit came at request 16 in the median, and a third of its reads before that went to files its brief already named.
- Read-pattern guard: the hint to send independent reads together now fires on the pattern it was written for. It counted read-only turns but started over at every new file, so many small reads, one per turn, never reached it (4 hints in 26 worker runs while 16 % of their turns were such reads). It now fires once per series when a read could have gone out together with the previous one, names that pair, and also applies in questions and audits.
- Shell reads: the first command in a session that only prints files (`sed -n`, `cat`, `head`, `tail`) gets one note pointing at `read` and `index`. Nothing is blocked, and the note does not repeat.

## [0.2.357] - 2026-10-21

- Long runs: a run with open goals no longer stops at the soft token budget only because two edits failed just before and were fixed right after. Failures that a later success of the same tool or file resolved no longer block the budget renewal; a 4-hour run was paused for that reason on 2026-10-20. The pause notice now says which budget applies, what renewal needs, and that a message continues the run.

## [0.2.356] - 2026-10-20

- Memory: a session ledger is marked finished only when the session ends, not when the conversation is compacted in the middle. Retention can then tell a finished digest from an open one.
- Desktop: reopening a session no longer stays blank until the whole window is reloaded. The empty snapshot from when the session was created was painted as the finished transcript; the open now fetches the latest window before painting.

## [0.2.355] - 2026-10-20

- Harness tests: dropping a session handle now stops its actor thread even while a core-services installer still holds the command sender, and a harness heartbeat no longer waits out 15 s after the harness is dropped. Serial lib runs were leaving those threads behind and growing to several GB.
- Memory: opening a workspace's code index no longer keeps the raw index file in memory while it is parsed. For this repository's 881 MB index the peak drops from about 3.2 GB to 2.3 GB, once per open workspace at startup. The host is still well above its 2 GB target.
- Model requests: no request waits for plugins that have no provider hook any more. Before, every enabled plugin got a copy of each request, a worker thread and a share of one 250 ms deadline, so wiki, sysload, snip and index logged "missed the shared deadline" (22 times in 6 minutes) without having any hook. A plugin now takes part only by returning its hooks, and the warning names why a hook was skipped.

## [0.2.354] - 2026-10-20

- Shells: when a background command runs in a private build view, its run note now names that folder, so output redirected to a file (for example a test log) can be found where it was actually written.
- Checks: rules-claim checks inside private build views no longer report tracked `.github/` and vendored files as missing.
- Workspace capture: sessions that capture the same checkout at the same time now share one capture instead of queuing behind each other. One capture of this repository takes about 30 s and held the workspace for every other session, so turn-diff baselines were skipped or ran late.
- Sessions: capturing the checkout into MAIN no longer fails once the workspace head has advanced ("writer binding record disagrees with generation/state", 45 failures in one run). Stale writer records heal on their next capture; runtime roots left behind earlier are not reclaimed yet.
- Code index: a search during a reindex no longer fails as "corrupt index snapshot". Each content store is published as its own immutable generation, and a reader uses the one its index names.
- Goals: a session with a long closure history no longer rewrites its whole goals log on every change (up to 2.7 s per goals update before).
- Build admission: PowerShell pipelines and redirects (`2>&1 | Select-String`, `> build.log`) and `cd` chains with several builds now resolve their output roots, and every unresolved warning names its command class.
- Prompt: an over-budget rules core or spec index is reported once per workspace and size instead of on every session start (12 identical warnings in 6 minutes before).

## [0.2.353] - 2026-10-20

- Sessions: a new session no longer fails with "workspace catalog: database is locked" on a busy workspace. A known workspace attaches from reads without a catalog write, a busy catalog is retried for up to 30 s, and the prepare dialog times plugins and workspace separately.
- Desktop responsiveness: the app stays responsive while many sessions run. The session list loads once per change instead of two to four times, branch information no longer waits behind long history reads, plugin calls skip unneeded settings reloads, and idle background polling slows down. The app now records in its log whenever its window thread stalls for more than 250 ms.
- Chat: one long-frame log line per frame instead of one per mounted panel (3,884 lines for 307 frames before). Stream targets no longer stay set after a failed prompt or a missed stream end, and a failed prompt keeps its text in the composer.
- Session actor: waits on turn-diff baselines are bounded (20 s), the 120 s reply timeout names the command the actor was busy with, and "snapshot lock wedged" is reported only for one hold of at least 30 s and names its real holder. Transcript reads no longer copy the whole history under the snapshot lock.
- Session index: full sweeps run at most every 30 s and log their trigger; hidden staging files no longer force sweeps (421 sweeps a day at a 7.7 s median gap before).
- Workspace catalog: lapsed runtime roots whose generation is gone are reported once instead of on every registration (5,216 warnings in one run before); orphaned capture spill files older than 24 h are removed; in-process VACUUM is bounded to 1 GiB and GC chunks checkpoint the WAL; refused view reclaims are logged once per pass and retried.
- Logs: older logs are kept for 30 days as compressed archives instead of only one previous file, so problems can still be investigated days later. Logs and sessions are never deleted automatically.
- Workers: the worker session build logs its sub-phases; a topic-overlap refusal names the holder's result; a resumed worker gets an anchor instead of its whole task again; the upstream digest no longer asks for a complete re-read.
- Prompt: the operative-core cap measures the rules text only; the spec index has its own structural check. Prompt bytes are unchanged; this repository no longer warns on every request.
- Harness: superseded goal revisions are dropped before serialization, finished-job notices the model never receives are no longer persisted, and build admission resolves `cd`/`Set-Location` chains and read-only or `tee` pipelines (183 unresolved warnings in one run before).
- Providers: new Inception provider (`INCEPTION_API_KEY`) with the Mercury 3, Mercury 2.5 and Mercury 2 diffusion models, including their instant/low/medium/high reasoning effort.
- Shells: a command started in the background (or an interactive terminal) no longer dies about a second after launch with exit 130. The runtime cancels a parallel tool batch's tokens once the batch is done, and background and PTY shells were still watching their launching call's token. They now stop only on their own Stop, their deadline or session end.
- Prompt: a configured `promptStyle: "terse"` now actually loads the terse rules core; a session-options step reset it to prose. For this repository the main prompt shrinks from 84,206 to 74,795 bytes.
- Memory: `memory search` and the sessions list no longer fail once more than 32,768 session entries exist; they scan the newest entries and say how many older ones were skipped.
- System monitor: the process breakdown no longer comes back empty when a cheap footer reading is in flight.
- Windows: four more runtime spawns set `CREATE_NO_WINDOW`; the console-window guard no longer misreads Rust lifetimes as character literals.
- Verification: `scripts/verify.ps1` builds test binaries in their own stage, no longer kills its own console host on a timeout (which broke every later stage with Win32 0xE9), and reports a launch failure as such instead of blaming a crate.
- Tests and specs: 27 of 38 load-sensitive test failures fixed (the other 11 are listed in the known limitations); memory tests no longer read the real home; the large spec units are trimmed to current contracts.

## [0.2.352] - 2026-10-20

- Research: before the research tools are loaded, the prompt now says where a research request goes (deep research for a durable cited answer, web search for quick context, fetch for a known URL) with no goal or intake call first. Before, it only pointed at a workflow that was not loaded yet.
- Goals: intake is described as optional and never a step before the work, and the question rule no longer calls request bookkeeping "necessary". Measured on the production prompt (5 runs per arm): gpt-5.6-terra no longer opens research and many single-tool tasks with a goal; its tool-hit rate rises from 78.9 % to 86.8 % (p = 0.007), and with the production tool choice it starts web research directly on both research scenarios (before: a goal first in 5/5). claude-sonnet-5 stays within noise. The earlier "research scenario regression" exists only under the eval's forced tool choice; with the production tool choice Claude went to web search before and after the cleanup.

## [0.2.351] - 2026-10-20

- Prompt: the system prompt is much smaller. The rules core and its terse register are condensed, spec-unit index lines are capped at 200 characters (guarded), and every rule now has one owner instead of being restated in the base instructions, mode suffixes, Goals and delegation playbooks. Measured for this repository, the full production main-session prompt drops from 114,563 to 83,826 bytes (about 8k fewer tokens per request) while tool, action and schema hit rates stay unchanged within run-to-run noise (claude-sonnet-5 and gpt-5.6-terra, 5 interleaved runs each). Workers get a reduced base without the lead-only sizing, Goals and ambition sections.
- Docs: each turn carries a count plus at most five doc pages ranked against your message instead of every reachable page. A page that did not help can be marked with `doc_feedback`; at most once a week, workspace pages with repeated not-useful votes are archived (restorable) and named once.
- Memory: always-load memories arrive as a compact title index (scope, title, kind, date, id) instead of full bodies; every entry carries a short title, and older untitled entries get one on migration.
- Turn context: the worker-authority notice stays silent until a worker exists, the empty Goals brief is sent once instead of every turn, and empty fields are dropped from the session-state note.
- Tools: code search is explained in a few lines with one default entry point, and in Rust projects the structured build-check tool is ready without an extra activation step; other tools stay on demand.
- Specs: obsolete units removed or merged and history sections cut; `prompt_dump` writes the production system prompt for evals.

## [0.2.350] - 2026-10-19

- Chats: the composer no longer shows the Multi Expert switch; it stays in Code and Assistant sessions. (0.2.349 was never published: its release gate caught a stale contract test.)

## [0.2.349] - 2026-10-19

- Chats: the composer no longer shows the Multi Expert switch. Multi Expert stays available in Code and Assistant sessions.

## [0.2.348] - 2026-10-19

- Browser: a visible agent Chromium no longer takes the focus when it starts. Unary now blocks foreground activation for other processes before launching it, and releases the block once the window has settled behind your work. Before, Chromium took the focus at startup and the attempt to hand it back was refused by Windows.
- Browser panel: after the agent hands the session back, the native view now opens the page the agent was on instead of staying on its example.com start page.

## [0.2.347] - 2026-10-19

- Shell: a background command that finishes on its own wakes the session again with its exit; the turn-boundary bookkeeping had withdrawn that wake, so the exit never reached the model.
- Sessions: opening a large session no longer keeps a decoded copy of every entry in memory (about 2.7x the file size before, about 1.2x now).
- Shell: a background build or test whose output goes to a file is no longer invisible to supervision. The file's growth counts as progress, its tail rides along in every stall notice, a stall after real progress wakes the session again, and a build/test job past its budget gets one evidence notice instead of waiting forever. Build and test jobs are still never killed for being quiet.
- Shell: a foreground build or test that reaches its timeout moves to the background instead of being killed, and the memory-headroom wait before a command is bounded and shown.
- Workers: a session switching into or out of its own isolated tree could freeze file edits for every session in the app until restart; it no longer can.
- Tools: a parallel tool batch is no longer cut at 15 minutes while a member still has a longer timeout, and members that miss the deadline are cancelled instead of left running.
- Sessions: a session whose only remaining process is a dev server or an open terminal no longer shows as running, removing a worker refreshes its parent's row, and Stop or a refused close names the shell that is still blocking.
- Desktop: a worker row stuck in starting or queued after its worker is gone no longer stays busy forever.

## [0.2.346] - 2026-10-19

- Workers: a writer queued behind a file held by another session can now hand its edit to the holder. The queue refusal names the exact `session` call with `scope: "project"`; an unscoped Intercom message to a session outside the sender's tree is delivered through the project room instead of being refused, and a field of another action (such as `purpose` on `intercom_send`) is named with the action it belongs to.
- Shell: a finished foreground command is no longer reported again as a finished background job at the next turn boundary.
- Edit verification: a non-build `verify` command such as `echo ok` or `git diff` no longer reserves a build lane or prints a build receipt and cold-lane note.

## [0.2.345] - 2026-10-18

- Desktop Chat: keeps formatted answer and thinking nodes stable through streaming, terminal drain and the stored-message handoff. Restored progress prose uses the same Markdown preparation keys as the live renderer.
- Desktop Chat: replaces full-history background loading with a bounded, demand-loaded window. Older tool results load only while expanded and are released on collapse; saved conversations and live command output are retained.
- Desktop Chat: uses one measured viewport anchor and one live-follow policy instead of stacked native/nonwindowed, delayed-compensation and easing paths. Column resize and history-gap changes preserve the reading position before paint; the Desktop preview uses the DOM-free Markdown worker decoder.

- Chat input: queued human steers and follow-ups survive session close and restart instead of living only in memory. Held follow-ups recover across window reload, and a pending chip is no longer shown as delivered when the backend still queues it or its state cannot be verified. Running commands still finish at the existing safe boundary before steering is delivered.

## [0.2.344] - 2026-10-18

- Desktop: long chats stop jumping while you scroll back. Answer rows were sized as a single 52 px line because the height estimate read a text field only user rows carry; it now derives the row height from the answer's text, tool cards and collapsed thinking, and a row that is measured above the visible area is compensated before the frame is drawn — also while you are scrolling — instead of a frame later or not at all. Resizing the chat column now re-wraps and re-measures at the new width instead of keeping heights wrapped for the old one.

- Git tool: `git commit -q`/`--quiet`/`-v`/`--verbose` with explicit paths is no longer refused as a blanket commit while another session holds unrelated dirty files; the parser knows these output flags and keeps the path restriction.

- Codex Voice: a session-paired call now follows what the coding session is actually doing. The silent status snapshot names the running command, file or search per tool, the live workers and what they are on, the recent changes since the last snapshot and the session's latest note, still throttled to one append per 20 s. When the user asks for details, the voice fetches a fresh read-only status of its paired session (`[VOICE SESSION CONTROL] {"action":"status"}`: recent steps with outcome, finished workers, goals, waiting decision, last error) instead of guessing or delegating the question as work. The quick router receives the newest snapshot with every routed request. No Overseer verb or cross-session access is involved.

- Desktop Chat: manually expanded tool cards stay expanded when scrolling them out of view and back. Switching sessions still clears the previous session's row state.

## [0.2.343] - 2026-10-18

- Desktop Chat: scrolling retains already-visible message rows instead of rebuilding them at every window boundary. Height measurements are batched, delayed content keeps bottom follow, and opening an existing session loads the latest messages rather than a cached middle page.
- Desktop Chat: the final response tail finishes through the existing text pacer instead of appearing in one jump. Long plain backtick code blocks update on the render clock, text-colour caching avoids repeated style lookups, and pending text deltas are delivered before a run ends or reports an error.
- Research: configured native Anthropic and Codex providers now handle both web search and page fetch. Native document content, citations and completed page-open evidence are retained, and cached continuations remain separated by provider and model.

- Desktop Chats: switching between cloud chats could crash into the root error boundary with `Cannot read properties of null (reading 'cwd')` — the Codex Voice bar's teardown re-read the dying panel's `sessionCwd` through its prop getter after the cloud binding had been reset. Teardown now stops with the owner captured at bind time, and the cloud conversation is keyed per chat so a switch mounts a fresh instance.

- Desktop Chats: a cloud chat whose conversation panel threw while the view was being replaced escalated a second, internal error (`Cannot convert a Symbol value to a string` from the boundary's `failed` snippet rendering on an already-destroyed boundary) to the root error boundary, which is the "default screen" that replaced the app. Svelte is bumped to 5.57.2 in both hosts, whose boundary ignores a destroyed boundary's late error result; both conversation boundaries now log the underlying panel error with chat and session id instead of swallowing it.

## [0.2.342] - 2026-10-18

- Workers no longer lose their shell after an app start. The start-up catalog sweep decided from a one-time snapshot of live sessions; on a registry of thousands of checkouts the walk took over a minute, and every worker spawned in that window had its workspace generation retired, failing every build and shell command with `missing workspace generation`. The sweep now re-reads the live sessions right before it retires anything, and it prunes registry checkouts whose directory no longer exists, which is what made the walk slow in the first place.
- Desktop Chats: a cloud chat no longer rebuilds its conversation panel on every account delta (including echoes of this device's own uploads), which showed as the chat reloading into an empty panel mid-conversation; the view re-binds on the chat ID alone and unchanged rows keep their object.
- Desktop Chats: automatic naming reaches the Chats rail. Local rows take the generated title from `harness-session-renamed` instead of staying "New Chat" until the next reload; a cloud chat is never renamed to the "New chat" placeholder (which locked it as manually titled and blocked auto-naming), and a generated name is carried up to the account record even while the conversation is not on screen.

## [0.2.341] - 2026-10-18

- Release: 0.2.340 was tagged on the wrong commit and never built; 0.2.341 is that release.

## [0.2.340] - 2026-10-18

- Desktop: the bottom follow treats only a content shrink as a browser-side adjustment. Dragging the scrollbar upward while a session streams (content growing) detaches the follow again as intended; 0.2.339 kept it attached in that case.
- Desktop: the Compact action above the composer uses the shared pill button like every other panel action.

## [0.2.339] - 2026-10-18

- Desktop: the chat stays at the bottom while a session runs. A row at the end that changes height (for example a tool row switching from planned to running) no longer makes the view jump back and lose the follow; the follow now reads such a browser-side adjustment as what it is, not as you scrolling up, and re-pins on every content height change. The scroller leaves scroll anchoring to the chat itself.

## [0.2.338] - 2026-10-18

- Desktop: long chats open at the bottom without flicker. Before a message is drawn, its height is now worked out from its content (text length at the current width, code blocks, tool groups, images) instead of a fixed guess; heights measured once are remembered per session, and a chat stays hidden for two frames on open until it sits at the bottom.

## [0.2.337] - 2026-10-18

- Desktop: scrolling near the end of a long session no longer jumps or bounces. Chat rows near the visible area are now drawn at their real height instead of a 108 px placeholder, so the chat no longer corrects the scroll position mid-scroll when a long answer or tool output comes into view.

## [0.2.336] - 2026-10-18

- Desktop: the Una spinner draws its outline on its own thread. The logo trace now renders through an offscreen canvas in one shared background worker instead of repainting the window on the UI thread, so a running session no longer keeps the app busy; spinners that are scrolled away or in a hidden window pause. Where that is not available, or with reduced motion, the logo gently breathes instead.

## [0.2.335] - 2026-10-18

- Desktop: long sessions no longer freeze the app. The chat now mounts only the messages near the visible area instead of every message of the session (one long session had 1,191 messages and about 38,000 page elements mounted at once, and the UI memory kept growing while it stayed open). Expanded tool output and other opened details stay open when you scroll away and back. Browser find-in-page only finds text in messages near the visible area.
- Desktop: the UI no longer redraws the whole window about 120 times a second. Always-running indicators (the Una spinner, running dots, the thinking label, loading shimmers, the microphone pulse) now animate only opacity and scale, so the GPU does the work; spinners pause when they are off screen. The Una spinner breathes instead of tracing its outline, and a session name no longer pulses (its status mark still does). Measured on a long session: drawing time fell from about 2,800 ms to about 200 ms per 4 seconds.

## [0.2.334] - 2026-10-18

- Context: every request carries less repeated text. The Goals tool answers with the changed rows and a count instead of restating the whole board (up to ~9k tokens per call before). The docs listing is sent once per session and again only after a compaction or when the listing changes (~2.5k tokens per user turn before). The session-state note is sent only when something material changes and its fixed text is one sentence; the Goals staleness reminder is shorter. Context compaction itself is unchanged.

- Workers: a lead's follow-up to its own worker now arrives as the lead's instruction, not as a peer message the worker may decline; `wait_agent` returns a resumed worker's newest final message instead of its first one; intercom and `send_input` replies are compact (per-recipient state, delivery errors only when present). A worker's spawn record is written once; later resumes add a small delta.

- Sessions on disk: pasted images and large pasted text are stored once in a shared content-addressed store (`~/.unary/blobs/`) instead of a base64 copy per session and a new file per paste; identical pastes share one file, and blobs no transcript references are removed after a day. Existing sessions keep loading their old files. The per-session evidence journal (a second copy of every tool call) is replaced by a small index into the transcript, and runtime snapshot rows are stored without escaped duplication (about 40 % smaller). Goals keep one log file per session instead of a board file plus a ledger; old files migrate on first load.

- Memory: entries are created only when the agent decides to (on your instruction or on its own), as repo or global. Unary no longer writes session.md summaries, the session search database or the session projection database, and no longer attaches known-bug evidence automatically. Dream only checks which repo entries should move to global memory. Other sessions are found through a small rebuildable index over the transcripts. `recall` searches memory and documentation through one full-text index; `dream_search` is gone. The memory data that context compaction uses stays as it is.

- Disk: `app.log` rotates at 4 MB on every write; leftover `.session-*.tmp` files older than a day, file-claim rows of dead processes and old checkpoints beyond the newest 50 (older than 30 days) are cleaned up; file claims live in one store; idle `.unary/build/shared` caches are evicted after 14 days; index evaluations and community-extension builds write under `target/`; the memory repository no longer commits its database files.

- Memory: a tool call refused by the circuit breaker is no longer counted as a failed call in the tool error audit. A refused repeat leaves no `calls.failures` or `calls.attempts` entry; it is recorded as its own `breaker_refusal` event, so failure rates now reflect only calls that ran. Refusals that name their kind (`policy_denied`, `edit_refused`, `claim_refused`, `spawn_overlap_refused`, `goals_refused`, `shell_guard_blocked`, validation rejects) are recorded with that kind, and events withheld while repository admission is missing appear as one `collector_dropped` row instead of only a counter. Destructive-git blocks by the shell guard, edit refusals (stale, never read, ambiguous, not found, unproven), patch and write claim refusals, credential-path denials and topic-overlap spawn refusals are recorded under their refusal kinds. Blocked shell commands and refused goal clears never ran, so they do not count as calls or failures.

- Chat: models that read images but not video (Claude, GPT, Grok and others) now understand attached videos too. On Windows, Unary sends them key frames instead: the first, the last and the frames where the picture changes most, each with its timestamp. Gemini still gets the real video. The number of frames is set under Settings › AI Agent (default 4).

- Chat: you can now attach videos (MP4, WebM, MOV and more, up to 20 MB) for models that understand video; today that is every Gemini model. Other models refuse the video in the composer with the reason instead of failing on "Unsupported attachment", and a video that still reaches a model without video input is replaced by a visible note. Snip recordings can go to chat from the gallery with "Send to chat".

## [0.2.333] - 2026-10-17

- Snip (macOS): snips are no longer upside down and no longer pixelated. The frozen picture now keeps your Retina screen's full resolution instead of a shrunken thumbnail.
- Cloud chats: photos reach your other devices more reliably. Any device that already has a photo can pass it on, an offline device is skipped at once, a relay takes over when a direct connection fails, and a failed send or download now says so with a Retry instead of disappearing.
- Remote: two controllers on one Desktop no longer break each other's connection, so session lists and screen sharing stop hanging.
- Remote: scrolling up in a running session's transcript now stays put; a new row arriving at that moment no longer pulls you back to the bottom.
- iPhone Remote: a message you send while an answer is running now steers that run, like the Desktop's Remote panel, and shows as queued until the Desktop has it.

## [0.2.332] - 2026-10-17

- Desktop: 0.2.331 no longer shows "Something went wrong" (effect_update_depth_exceeded) right after every start. The screenshot gallery, which waits closed in the titlebar, kept re-running its close cleanup in an endless loop since recordings were added.

## [0.2.331] - 2026-10-16

- Codex Voice: questions like "Do you see what is running?", "How far are you?", "Can you hear me?" or a thank-you no longer go to the coding session as a message; the voice answers them itself from the current status. Tasks and instructions ("do this next", "stop") still go to the session.

- Snip: new "Record region" in the Snip menu (Windows). Select a region, it records live with a red frame around it until you press Space (Escape discards); the MP4 lands in the screenshot gallery and on the clipboard as a file you can paste into Explorer or chat apps. The gallery shows recordings with their length, plays them, copies and saves them as MP4. Every snip and recording is still saved to the gallery automatically.

## [0.2.330] - 2026-10-16

- Worktrees: the session-tree menu now always offers "New Git worktree…" with your own branch name and base, whatever the worktree engine setting. When you ask for a worktree in chat, the agent can now create it for you (`session` action `create_worktree`); it still never creates one on its own.

- Codex Voice: the chat now shows the whole voice conversation. Questions the voice answers itself appear as a "Voice" card, web lookups as a "Voice · Web lookup" card, and tasks as normal messages. Overseer calls write nothing into a chat.

- Codex Voice: a question the voice looks up on the web now also appears in the chat as a "Voice · Web lookup" card with question and answer, and the coding session sees it on its next turn, like a voice turn in the Codex CLI.

## [0.2.329] - 2026-10-15

- Codex Voice: quick questions get quick answers. The voice answers general questions itself and looks up current facts (weather, news, versions, prices) with a web search in about five seconds, instead of sending every question to the coding session. The voice no longer receives an unfilled `{{ user_first_name }}` placeholder.

## [0.2.328] - 2026-10-14

- iPhone: Settings › Codex Voice now has the Desktop plugin's settings with the same defaults (mode, protocol, transport, model, voice, read aloud, microphone). Voice chat stays connected until you tap the button again, keeps the screen awake during a call and plays the spoken answer. Codex-account dictation keeps what was heard when you stop, and the composer shows when it is listening.
- iPhone Remote: a code session's history no longer shows hidden harness entries (turn guidance, receipts, assignments) as raw escaped text; they are collapsed rows, and tool rows show what the call did. Remote images say when the desktop is offline and load by themselves once it reconnects.
- iPhone: the sign-in screen keeps readable contrast in dark themes.

## [0.2.327] - 2026-10-14

- Builds: agent builds in a build lane no longer fail on Windows path length. A lane folder repeated its workspace and generation, so a lane started 177 characters deep and `pixi run build` could not install its environment there ("path not found", os error 3). New lanes and run views use a 16-character folder name; existing lanes keep working.

## [0.2.326] - 2026-10-14

- Codex Voice Overseer: with the new Overseer switch on, voice becomes your hand across all sessions. Ask which sessions are running or building, find a session by topic, hear a summary of its last answer, send it a message, change its model, effort or mode, pause, resume or stop it, answer a pending approval, or open a new session. Destructive steps need a spoken confirmation, and other sessions report approvals, errors and finished runs on request. Everything runs only through the Codex Voice plugin and only while it and the Overseer are on. Toggle it at the mic under the composer or with the new Overseer pill in the titlebar, whose hover panel holds the read-aloud level (full, summary, result only) and the proactive level.
- Codex Voice: "Preview voice" in the settings plays the sample again when dictation is selected.
- Chat: an image named only by its file name (for example after "all in `Pictures/unary/`") now opens the file the image tool actually wrote instead of failing with "File not found" in the session folder. A bare name that no tool output or workspace file matches is no longer shown as a dead link, and a missing image says which path it looked for.

## [0.2.324] - 2026-10-14

- Loop guards: browser, desktop, Blender, UnaryCad and MCP app-control calls now count as progress when the application returns something new, so testing a web app or driving an app no longer trips the tool-loop or identical-call circuit breaker. A click or observation that returns byte-identical output is still stopped.

## [0.2.323] - 2026-10-13

- Build lanes: `pixi run <task>` is now read through `pixi.toml`, so a task like `pixi run build-debug` or `pixi run run-debug` that runs a declared build command (directly or through `depends-on`) builds in its own warm lane instead of the shared checkout, and keeps the pixi environment there. A declared environment such as `.pixi` is no longer reserved for lane builds, so parallel lanes stop waiting on each other.

## [0.2.322] - 2026-10-13

- Remote control: the rail between your PCs/sessions and the remote screen can be dragged wider or narrower, and the width is remembered.
- iPhone: dictation now says why a voice session ended, voice errors no longer hide the whole message, and the composer has a paperclip button to take a photo or pick one from the library.

## [0.2.321] - 2026-10-13

- Desktop sidebar: your own worktrees (a second branch you checked out in a sibling folder, e.g. `unary-code-rust-ui` on `rust-ui`) now read as branch groups: no worktree mark, the tooltip names the branch and its folder, and they sort ahead of Unary-managed and external worktrees. What separates the two is ownership, not whether the branch was pushed, so a push no longer moves anything; the push state is a hint on the group instead ("not pushed" without an upstream, otherwise `↑n ↓m` ahead/behind), refreshed with each worktree overview.
- Desktop sidebar: every worktree group can start a session inside its folder - a "+" on the group header and "New session in this worktree" in its menu, using the same creation path as branch groups.
- Remote control: the PCs/sessions rail is resizable (240-720 px, remembered) instead of fixed at 250-320 px, so long session titles are no longer cut off.

## [0.2.320] - 2026-10-13

- Desktop sidebar: a personal or external git worktree (one you created yourself, e.g. a sibling folder on your own feature branch) no longer lands under "Needs attention" with a red "Conflict" pill. That pill came from the merge PREVIEW of folding the branch into its parent - a decision only Unary-managed session/subagent trees owe you; a personal branch keeps its own named group. A genuine pending fold-back decision still shows. Guard: `worktreeGroup.test.ts` "never turns a merge preview into a conflict on a personal or external worktree".

- Release builds: a release build stopped halfway no longer blocks every later release build until Unary restarts. The pinned build lane is freed as soon as its run is dropped, and the next release build or the periodic sweep reclaims it.

## [0.2.319] - 2026-10-13

- Desktop sidebar: the "+" on a branch group works reliably again, and a session row no longer looks like a conflict. Four causes, all fixed: a second "+" click while an earlier creation (including its list refresh) still held the single creation slot was dropped silently - it now shows "A session is still being created" and the button is greyed meanwhile; the "+" was offered for historic branch groups no folder is checked out on and started the session in some old session's folder, so the new row appeared under a different group - the "+" now appears only when a folder is on that branch, and that branch is passed along so the row is filed under the clicked group at once; the folder came from the case-folded live-branch index (`l:/repo`), which missed the settings cache and read as a foreign folder in every comparison - the real spelling is used now; and an error after the backend had created the session only reached the console while the progress view stayed on "running" - it now fails the view with the error and a retry. A mechanical "Fold-back failed" is no longer labelled "Conflict" (`worktreeAttentionFor` reports `failed`, the ⚠ "Resolve conflicts" badge is reserved for real conflicts), and the worktree glyph on a session row says "Works in its own worktree (not a conflict)". Guards: `sessionCreationSingleFlight.test.ts`, `worktreeGroup.test.ts`. Spec: `docs/spec/frontend-ui-contract.md`.

## [0.2.318] - 2026-10-13

- The plan card no longer stays up while the agent is already implementing. A plan in `ready` means the user holds the decision, but that was only enforced when the model ended its turn — a run that kept calling tools never got there, so a session marked its plan ready, asked again with its own offer card, took the Yes there and implemented for hours under an unanswered plan card (2026-10-12, plan `cbb11bd5`). Now the same plan state gates tool dispatch: file writes, non-inspect shell commands, git writes and worker spawns are refused while the plan waits, with the two exits named — end the turn for the card, or relay a go the user gave in chat with `plan` `mark_ready` `proceed: true`, which now also approves an already-ready plan, clears the card and seeds its goals. The `mark_ready` result and tool description say the same. Guards: `a_ready_plan_refuses_execution_at_dispatch_until_the_go_is_relayed`, `proceed_approves_an_already_ready_plan_and_clears_the_card`. Spec: `docs/spec/plan-system.md`.

- Meta AI provider: add Muse Wonder (`muse-wonder-1.0`, $4.00/$0.20/$20.00 per million input/cached/output tokens) and Muse Wonder Full (`muse-wonder-full-1.0`, Muse Code Power/Maximum Usage subscription only, no per-token price), both 1M context with `minimal`..`max` effort (Meta docs read 2026-10-12). Live check 2026-10-13: `muse-wonder-1.0` answers at `minimal` and `max`; `muse-wonder-full-1.0` returned `404 model_not_found` for the signed-in account, so Full is unverified until an account with Power/Maximum Usage confirms it. The meta-ai pack gets a seventh slot and the previous pack digest is recorded so untracked copies refresh.
- Z.AI (code shipped in 0.2.317, commit f7fb5a059, without a changelog line): `zai-api/glm-5.3-flashx` (GLM-5.3-FlashX, the ~200 tok/s Flash variant: text+image, 1M context, 128K output, low/high/max effort) is in the catalog; z.ai does not offer it on the Coding Plan yet, so `zai` is unchanged. Standard-API prices follow z.ai's pricing page of 2026-10-12: GLM-5.3-Flash $0.15/$0.03/$0.50 (promotion ended), GLM-5.2 and GLM-5.1 $1.40/$0.26/$4.40, GLM-4.7 $0.60/$0.11/$2.20 per million input/cached/output tokens. The `zai-api` pack gets a seventh slot so FlashX does not displace GLM-4.6, the previous pack digest is recorded so untracked copies refresh, and the GLM-5.3 family checks (effort levels, never-disable thinking) include FlashX.

## [0.2.317] - 2026-10-12

- Build lanes can no longer be escaped. A lane-placed classified build that names the shared checkout by absolute path is refused before it starts; it was only a warning before. This covers PowerShell `Set-Location`/`cd`, `cmd /c cd /d`, `-C L:/...` and MSYS `/l/...` spellings, and only the own shell's remapped top-level `cd`/`pushd` stays exempt. A failed lane open now refuses a classified build instead of silently running it in the shared checkout; before, only error texts that looked like contention refused it. A run that opens a new lane says `Cold build lane`, so an empty `CMakeCache` reads as the expected first build rather than a reason to leave the lane. Guards: `powershell_and_msys_spellings_cannot_leave_the_lane`, `a_failed_lane_open_never_runs_a_build_in_the_shared_checkout`, and the `fresh` assertions in `crates/unc_rust_workspace_engine/tests/build_view.rs`. Spec: `docs/spec/checkout-concurrency.md`.

## [0.2.316] - 2026-10-12

- The sign-in screen is readable in dark themes again: its primary button, hint box and borders now follow the active theme's colors.

## [0.2.315] - 2026-10-12

- Sessions no longer get stuck with "missing workspace generation": a session whose private workspace was cleaned up now reopens with a fresh one instead of failing on every message, and the background cleanup can no longer mistake live sessions for deleted ones (it removed them when another Unary process, such as a test run, used a different sessions folder). Technical detail: the catalog sweep refuses when the registry and sessions home differ, when no live session is found, and inside cargo test binaries; guards `liveness_from_another_home_or_an_empty_set_never_sweeps`, `a_retired_recorded_workspace_is_detected_and_its_binding_released`.

## [0.2.314] - 2026-10-12

- Worker→lead goal state is typed, not parsed from prose: `worker_milestone` has a closed kind set — `completed {goalIds}` (a claim validated against the bound subtree at call time, stored on the hook record and applied at the terminal transition; no claim releases the binding), `proposal {title, scope?}` (passive FYI to the running lead, `proposals` in the handoff and in the lead's wake) and `blocked_on_decision {question}`. The `Completed parent goal ids:` trailer parser, the `ExplicitNone`/`MissingLabel`/`UnparsedLabel` ladder, text-hint check-off (which could close foreign rows by matching text), new_parent_goal_hints and the `Decision needed:` trailer are removed; `CompletedParentGoalReport` is `Claimed | NoClaim`; the bound-hook check-off no longer refuses on handoff quality. Measured 2026-10-11 over 30 days: 2,160 handoffs, 868 structured / 1,292 fallback, content sections in ~4 % — the section parser stays a named contract. Cargo-test binaries resolve `~/.unary` to a per-process temp home (`unary_global_home`, guarded), so no test writes the live Intercom store again (75 temp-repo rows were found there). Brief-derived worker write scopes no longer take the lead's own files ("I am editing X … do not touch those") or a bare top-level directory: one such record parked three sibling editors on 2026-10-11. Spec: `docs/spec/subagents-and-worktrees.md`, `docs/spec/goals-board.md`.
- `worker_milestone` was never advertised to a worker: `model_visible()` had excluded it since 2026-09-30, so the typed goal channel above (and every `blocked_on_decision`) was unreachable — measured by a catalog probe on 2026-10-12. The tool is now model-visible, in its own always-callable `worker_report` family, exempt from role family grants, and guarded by `a_subagent_worker_catalog_always_advertises_worker_milestone_eagerly`.
- Spawn gate: every spawn stores deterministic `topicScopes` (paths, directories, backticked symbols of the brief); a mutating spawn overlapping a live editor of the same session is refused typed (`send_input` to the holder) instead of parking; foreign holders queue with `holderSessionId` and the hand-over path; an explorer whose topics overlap a live or finished worker of the session is refused once with the evidence (`overlap: "accepted"` runs it); live claims and bound board rows of other sessions become an `advisory` array in the spawn receipt, never a refusal. The lead playbook's reuse paragraphs shrink to what the gate now enforces (≈150 fewer tokens per request). Spec: `docs/spec/subagents-and-worktrees.md` §7.
- `session` action `overview` (read-only, no audit link): every live session of the workspace with state, worker role, open goals and their bound workers, live claims and red builds, paginated and `partial`-marked when a source is unavailable; `spawn_agent` gains the optional `overlap` override. Commit attribution: every commit made through the git tool or a plain `git commit` in the shell carries `Unary-Session: <id>` and, for a worker, `Unary-Worker: <hook>` trailers; wrapped, compound or reader segments are left alone with a one-line notice. The eager catalog ceiling moves 115 500 → 118 500 B and the fixed prefix 217 000 → 220 000 B (+2 498 B ≈ +625 tokens/request; ledger in `catalog_budget.rs`).
- Intercom mailbox, first half: the Desktop panel shows kind chips (request / answer / handover / fyi / wake, mapped from legacy rows), threads, a kind filter and a "show harness wakes" toggle; `intercom_send` requires `kind` and a non-empty recipient list in the schema; `intercom_ack` stays advertised as a deprecated alias. The backend half (stored `kind`, read-time legacy mapper, wake rows hidden from the model inbox, `answer` writing the recipient receipt atomically) is NOT shipped: it makes the harness `intercom` test filter hang (blocking wait in the worker→parent route fixture, measured 2026-10-12) and is parked in git stash `phase4-intercom-backend-unverified-2026-10-12`. Spec: `docs/spec/harness-session-runtime.md` "Intercom envelope kinds" (target state, marked).
- A worker whose turn never mutated the workspace never reached its terminal state (measured 2026-10-12, root cause of today's "terminal workers are not woken" incidents and of the hanging `intercom` test filter): `TurnDiffBaselineCapture::deferred` kept the turn's event sender alive when the baseline was never launched, the child's progress thread waited forever for the channel to close, the hook stayed `running` and every `wait_agent` timed out. `TurnDiffBaselineCapture::release_unlaunched` now drops that sender at turn end; guard `unlaunched_baseline_release_drops_its_event_sender`. Correction to the earlier bullet: the Intercom backend package was not the cause of the hang; it stays parked only because it was never verified on its own.

- Session pins and marks survive reload and restart: they are now backend-owned session flags (`~/.unary/session-flags.json`, `pinned`/`marked` on the session summary, `harness_set_session_flags`) instead of WebView localStorage, which had reached its 10 MiB per-origin quota (10,485,688 B measured 2026-10-11, 98 % model-catalog snapshots) so new marks were silently not saved. Existing localStorage pins and marks migrate once. The model-catalog cache is bounded by workspace count and bytes and swept at startup; every localStorage write goes through one guarded writer that evicts caches on quota errors and shows a notice when user state cannot be saved. A session the index has not seen yet is no longer reported as removed (which also dropped its pin and mark). Spec: `docs/spec/harness-session-runtime.md`, `docs/spec/frontend-ui-contract.md`.

## [0.2.313] - 2026-10-11

- Desktop host allocator: the per-allocation counting wrapper (live/cumulative bytes, owner header, per-phase allocated bytes) was removed on 2026-10-11 after measuring 6–440× overhead per allocation against plain mimalloc; the host now runs plain mimalloc on Windows, and `harness_health` reports committed and peak host heap from mimalloc process info with allocation totals `n/a`. The per-owner live-heap table is gone. Spec: `docs/spec/checkout-concurrency.md`, `docs/spec/harness-session-runtime.md`.
- Turn-diff capture reads changed blobs in one `git cat-file --batch-check` + `git cat-file --batch` instead of two `git show` per changed file (51 changed files: 2 blob-read spawns instead of ~102), and holds the workspace git gate only while walking the shadow tree, under the session's id (`turn diff final capture` was held 21.6 s, `turn diff baseline capture` 3.5 s, both `session=unbound`, measured 2026-10-11 on 0.2.312). The forced full rehash of the baseline stays on purpose (a turn diff must never miss a same-size restored-mtime edit); the baseline is shared per checkout instead of recaptured per session.
- `context_budget_status` checks its estimate cache before serialising tool schemas, memoises the stable-prefix reserve per tool-activation revision (one integer compare per call instead of `to_string()` of every schema), extends message token sums incrementally (`MessageTokenMemo`, equal to the full recomputation on a 200-message fixture) and walks the session branch by reference instead of cloning it (was 1 182 calls × ~1.5 MB in 55 min).
- Debug-only `tracing::debug!` lines removed from the lexical query path, memory maintenance and the Desktop emit forwarder; measurement `eprintln!` lines stay only inside `#[ignore]` tests.
- Desktop host speed (harness, measured 2026-10-11 on 0.2.312: `intercom_call` mean 1.4 s / max 12.3 s at ~38 MB per call, `snapshot_refresh` 6 MB per call, `debug_snapshot_build` 7.6 MB per prompt): the turn-diff baseline is captured at the first mutating tool call instead of every prompt (explicit checkpoints stay eager) and under the session's id; context-hook sidecar writes serialize outside the registry lock and cleanup no longer re-reads the sidecar; the dispatcher, idle-parent wake, mid-run result queue and dependency settle reuse one generation-guarded registry snapshot instead of cloning per candidate; intercom reads filter visibility and `LIMIT` in SQLite and parse only returned rows; intercom membership reads each session's parent id without cloning state; busy-child counting and the per-recipient worker lookup use `with_state`; `status`/`wait`/`wait_all` request the scheduler tick instead of sweeping inline; the spawn path loads settings once; a stale history is rebuilt outside the session snapshot mutex; the debug snapshot is cached per (tool-activation revision, system prompt, layer count); the goals listener copies goals before taking the mutex; peer-pending ledger lookups read a maintained in-memory index (`SessionManager::peer_pending_states`) instead of walking the transcript; 15 debug-only tracing statements removed. Still open (need core changes): the one-time transcript open per Intercom delivery to a live recipient, a per-record revision for a cached parsed worker spec in the watchdog, child-listener delta coalescing.
- Account failover now retains the switched-to account across tool-follow-up requests and bound-worker continuations. The runtime credential binding shares its non-secret failover choice with executor clones instead of retrying the originally leased account after every tool result; explicit rebinding resets that choice and independent bindings remain isolated. Verification and live cache-savings evidence are recorded separately in `docs/spec/auth-and-oauth.md`.
- Index (lexical mode) CPU peaks: the plugin freshness watcher no longer forces a full scan on a directory event or an extension-less removal (those become directory-prefix hints and removed paths), only on watcher overflow or a failed git catch-up, and `status` names the path and reason when it does; the matcher honours `.gitignore` and `.git/info/exclude`; the Desktop host's auto-index path treats content-modify events as hints and re-merges a request the busy plugin did not start. A forced full scan hashes in the parallel parse pool and re-parses only changed files (was: all 3 700 files, 93 s); architecture relations are built once per generation and reused when nothing changed (edges stage 23 s); the hierarchy skips reconcile on an unchanged delta and caches `git log` by HEAD (hierarchy stage 43 s); `chunks.dat` is rewritten only when chunk hashes changed. Over-size and binary files (an 80 MB tracked Mach-O was being chunked) are excluded at the walker and `binaries/`, `*.gz`, `*.png`, `*.bmp`, `*.lock` join the default ignore globs; `index.bin.corrupt-*` quarantine keeps one file. Each generation appends its stage timings to `generations.jsonl`. Spec: `docs/spec/index-subsystem.md` "Freshness by generation" and "Refresh, compatibility and cost".

- Workers no longer open every task with a routine `intercom_read`: the worker's first turn carries ONE Intercom line (live sessions in its room, message count, the latest three room-wide posts, project-room state) computed once at spawn, and the playbook tells it to read only when that line or an addressed message gives a reason [user decision 2026-10-11]. `diagnostics intercom_usage` reports the room's reads with/without new messages, live deliveries by urgency, idle wakes and passive rides (an estimate of saved lead turns) from the new `usage_events` table. `task_claim`/`task_update` moved out of `session` into `team_tasks`, advertised only with the Teams switch on — outside a `spawn_team` team, task ownership moves only through the lead. A non-interrupt `send_input` to a child is persisted through the Intercom room/receipt path (interrupt steers raw).

## [0.2.312] - 2026-10-11

- A message delivered to a RUNNING session (worker handoff, `send_input`, Intercom) no longer wedges that session. Since 0.2.310 ("idle steer wakes the session") `commit_pending_message` and `accept_recovered_intercom` held the session snapshot mutex while `enqueue_steering_message` → `continue_queued_if_idle` locked it again on the same thread; the non-reentrant lock never returned, and every later resume, UI read and `diagnostics health` of that session queued behind it (measured 2026-10-11 on 0.2.311: four sessions held 111–698 s, holder `unlabeled`; the Desktop reported "session … is not responding"). The four peer-ledger paths now copy the snapshot facts out under a labeled lock and release it before any ledger I/O or control call. Guard: `a_steer_delivered_to_a_running_session_returns_without_wedging_its_snapshot` (times out with the old hold).

## [0.2.311] - 2026-10-11

- The session JSONL is now the single turn ledger and every worker/debug view derives from it: run-lifecycle entries carry `runId`/`turnIndex` as paired turn boundaries (`turn_boundaries`); worker spawn and finish are recorded as `subagent_spawned`/`subagent_finished` rows with the spawn spec and a bounded result (hidden from the model); the `*.hooks.json` sidecar is rebuilt from those rows when it is missing, corrupt or stale (`rebuild_hooks_from_transcript`) and rewritten as a cache; the evidence journal is rebuilt from the transcript when missing or stale and never parses the transcript while its file is fresh. Twenty workers no longer saturate the webview: the Tauri forwarder caps each session at 120 emits/s (sliding 250 ms window) and widens the per-hook progress window instead of dropping, lifecycle events are never held; `diagnostics health` shows the per-session mode and coalesced count. Workers of one provider share a process-wide 429 backoff (exponential 2–60 s fallback, one coalesced scheduler wake per provider) and park with a `provider_backoff` reason; `diagnostics health` lists provider backoffs. 87 `eprintln!` state logs in the subagent executor, harness commands, session ownership and task store were deleted or became `tracing::warn!/debug!` (guard `harness_sections_log_through_tracing_only`).

- Ten or more workers in one session no longer wedge the session actor or make the worker strip flicker: the context-hook registry is counted and serialized under its lock instead of being cloned 3–4× per spawn (partly under the global sessions lock — the deadlock candidate), slot releases wake only parked sessions through one scheduler thread, `harness-sessions-changed` is computed per session and debounced 150 ms off the forwarder thread, the hook sidecar is written compact and coalesced (100 ms) with compaction that clones only removed records, and the Desktop worker strip keeps every live row (finished rows fill the 12-row budget), coalesces per-tool updates per frame and refreshes its snapshot on lifecycle edges instead of every 400 ms. A snapshot lock held longer than 1 s now names its holder label and thread in the resume error and in `diagnostics health`.
- Less unnecessary writing and logging: the per-request wire capture rows and per-attempt transport rows in the session JSONL are replaced by one compact cache-usage row [user decision 2026-10-11]; the self-memory session ledger is appended per turn and folded once at read or compaction instead of rewritten per turn; `session-path-hints.json` and the `sourceUsageEventId` progress field are gone; every emitted harness event was serialized twice (size probe) and is now serialized once; per-event `eprintln!` lines on the prompt listener, `list_sessions` and worktree sweep paths are debug/trace logs; the duplicated log helpers exist once; `context_budget_status` memoizes its estimate instead of walking the branch on every call.

- Desktop sidebar: sessions on another branch are grouped by the branch their folder is on NOW in every group. Editor groups and "Detached Sessions" received no live branch index and filed rows under the branch recorded at creation; a worktree's own in-place branch switch was never refetched; a session moved with `switch_workspace` carried a false "folder has since switched branch" badge; a `\\?\` cwd found no live branch; a worktree missing from the index inherited the project's branch; `.unary/worktrees/user/<slug>` was filed as a subagent; a detached HEAD showed the last recorded branch instead of "No branch" (`shared/app/desktop/components/Sidebar/worktreeGroup.ts`, `shared/desktop/components/Sidebar/Sidebar.svelte`).
- Session titles no longer fall back to "New Session": the Desktop session store never lets a placeholder or empty name overwrite a known title (hosted session switch, overseer fallbacks, bare reconcile payloads); the generated-title write is a compare-and-set that refuses when the user typed a title while the model was generating one; a rename from the sidebar, Chats or Assistant now locks the chat panel's own auto-rename request; a model answer equal to the placeholder is rejected instead of persisted; and duplicating or forking an unnamed session reads "Copy of New Session" instead of "Copy of <uuid>".

## [0.2.310] - 2026-10-11

- Opening a session whose running turn is stuck no longer ends in a bare "Session resume … timed out after 30000ms": resume waits at most 5 s for the live actor's state and then says which session is not responding, for how long, and to stop it (or restart Unary). Retries no longer queue behind the stuck open for 60 s (the join budget is 25 s, below the Desktop's 30 s guard), and a stalled actor is logged once as an `ERROR` instead of only as repeated `session state unavailable` notice failures.

- The session snapshot and turn-diff baseline no longer get skipped as "git command output exceeded the bounded capture limit": the shadow capture's listings (`ls-files`, `check-attr`, `diff --name-only`, `add -A`) have their own 256 MiB stdout budget, a stderr flood (one CRLF warning per touched file on Windows) keeps its head instead of refusing the capture, and the error names the overflowing stream and its limit (`crates/unc_rust_workspace/src/git.rs`).
- Opening a persisted session no longer re-parses its hooks sidecar (2.9 MB for the audited session) on every open; the parsed snapshot is cached by path, length and mtime. Measured with a counting allocator on the 9.6 MB transcript of session 01a127dc: a cold open peaks at +46.7 MB and a warm open at +1.4 MB, so the 12–28 GB resident peaks recorded in `last-crash.txt` before today's three unclean exits do not come from the open path; their owner is still unidentified (`diagnostics` `health` lists live heap by owner for the next occurrence).
- A new session in a branch (worktree) group of the Desktop sidebar opens again. The first adoption of a checkout whose `.unary/` holds tracked skill bundles built its capture policy before the workspace store created its creation lock (`.unary-store-create.lock` beside the catalog), so the import refused with "invalid unmanaged import: 1 unresolved path(s) remain" (measured 2026-10-11 in the linked worktree `unary-code-rust-ui`). The store's own lock and catalog directories are now named as local state before they exist (`crates/unc_rust_coding_agent_core/src/session_workspace.rs`, `NATIVE_STORE_RUNTIME_SIBLINGS`).

## [0.2.309] - 2026-10-11

- A session that moved itself into another worktree (`session` action `switch_workspace`) and was interrupted by a restart before its run settled reopened in the checkout it had left, and the Desktop sidebar filed it there (the worktree group read "No sessions in this worktree"). The switch is now recorded durably beside the transcript at the moment it happens and applied on resume after re-checking that the tree still exists. The agent-side switch also refreshes the sidebar right away, as the Desktop command already did.

## [0.2.308] - 2026-10-11

- Build joins never wait on a live owner any more: a foreground join used to hold the lead in a blocked `Bash` call for up to 600 s and then answer "not known yet" with no handle and no wake. Every join now returns at once with a joiner shell (`shellId`, `details.buildCoordination.joinerShellId`) whose exit wake carries the owner's verdict. A session is refused immediately when it re-launches its own live build (the refusal names its shell) instead of waiting on the lock it holds itself, and the source stamp is part of the build identity again so changed inputs start a new run while the earlier one is still going — two coordination tests had hung for ten minutes since the join-on-reservation change.
- `shell_watch` `timeoutSecs` above 120 clamps (with a note and `timeoutSecsClampedTo`) instead of refusing; `wait_for` property descriptions are shorter and the eager catalog ceiling moves to 114 600 B; the dialect budgets are re-frozen at +173 B (~+46 tokens per request) for the audit's schema changes.
- A stuck transcript writer no longer hangs the agent thread: `ensure_durable` times out after 60 s and names the writer step it waited on.
- Worktrees: a linked Git worktree registers as its own project (activation refused a second project sharing the Git common dir with "Git common directory already registered"), and a session that entered a sibling worktree keeps its Goals board under the main checkout, so the panel is not empty after a restart.
- Stopped-runs dialog rows are named by their session, not by the stop notice title.

## [0.2.307] - 2026-10-11

- Release 0.2.306 did not compile either (a second field rename committed without its consumer); 0.2.307 is the same release with that line fixed.

## [0.2.306] - 2026-10-10

- Release 0.2.305 did not compile (a one-line import error committed shortly before it); 0.2.306 is the same release with that line fixed.

## [0.2.305] - 2026-10-10

- Waiting and hooks (audit 2026-10-10, all mechanisms): a background `wait_for` watcher now wakes its session by default (`wakeSession` defaults to true; a goal bound through `resumeWatcherId` to a non-waking watcher is refused), a timed-out watcher no longer flips blocked goals back to `in_progress`, `sessions_idle` counts a peer with busy workers, a paused run or a pending approval as busy, and `commit`/`goal_done` are reachable from the schema. `spawn_agent` advertises `dependsOnHookIds` again (the live schema had lost it while the rules still told the lead to use it; guard `a_live_spawn_agent_advertises_dependsOnHookIds`). `wait_agent` returns its statuses at the deadline instead of running into the 180 s tool timeout, a worker already parked on a decision interrupts the wait at once, and `session watch` follows a rebuilt target instead of timing out silently. The three wait timeouts (`wait_agent timeout_ms`, `session watch timeoutMs`, `shell_watch timeoutSecs`) clamp instead of refusing (19 of 27 live `shell_watch` failures were timeout-unit refusals). Background shells: the silence lane uses the same CPU verdict as the wait lane and its wake honours `wakeOnExit`, build/test commands no longer claim a "hard deadline" they never get, deadline kills carry `timedOut`/`deadlineSecs` in the wake, an unconfirmed stop keeps being supervised, and a mid-run exit is reported once. Build joins: a foreground join past 600 s becomes a joiner shell with a wake, joiners can be killed, and an owner crash or refusal is named instead of `exit -1`. Lost wakes (enqueue failure, full command queue) are retried once and then surfaced as an error; a wake dropped while paused is re-armed by the human resume; an addressed Intercom `decision` wakes an idle recipient in every dispatch mode. `diagnostics health` reports the build commit, pid and process start; `tool_error_audit` rows carry a `build` column. Eager catalog ceiling 114 500 → 114 600 B (~+25 tokens per request).

- Diagnostics: `diagnostics health` (and `harness_health`) shows the live heap by owner — which subsystem holds how many MiB right now and how much each has allocated in total (session actor, event fold, emit, provider stream, transcript persist, session open, worker spawn, plugin by id, context budget, untagged). The 0.2.304 measurement showed the process growing from 4 to 11 GB in six minutes while the instrumented phases explained under 0.5 % of the allocation churn, so the host allocator now stamps every block with its owner and debits that owner on free, whichever thread frees it. The `emit_serialize`/`emit_dispatch` phase timings, present but unreadable in 0.2.304, now appear in the same report.

## [0.2.304] - 2026-10-10

- Remote control (Desktop): the remote host's session list renders through the same `ProjectGroup`/`SessionRow` as the local sidebar (project colour, status dot/spinner, model label, since-active, archived subgroup, Load more; row menus and pins stay off because they act on the local machine), and the remote composer is the normal chat composer — mode, model picker with the host's own model list, effort, Fast/Ultrafast and the context meter — all forwarded to the remote host through four new controller methods (`model.list`, `session.thinking.set`, `session.fast.set`, `session.ultrafast.set`). `session.list` fields the mapper used to drop (model label, open question, pending approvals, failed run, worktree data) now reach the controller, and the remote model is read from `modelLabel` instead of a field the host never sent.

- Workers: the limits are fixed numbers instead of being derived from the CPU thread budget — 20 per session and 200 across all sessions and repositories by default (it was the thread budget per session and twice it app-wide, starting in an 8/64 window that grew one worker per sample). Settings › AI / agent › Workers shows both defaults and accepts higher or lower numbers. A new *Memory brake* switch there (on by default) keeps the only resource link: sustained RAM pressure halves and sustained commit pressure pauses new worker starts; switched off, only the limits and provider rate limits decide. Running workers are never stopped either way.

- Native workspace catalogs are garbage-collected. Deleting a session retires its private workspace (catalog rows, generation directory, writer side files), and every start sweeps the catalogs of all registered checkouts in the background: workspaces of sessions that no longer exist are retired, capture history is kept to `nativeWorkspaceKeepCaptures` positions (default 50), unreferenced snapshots, tree rows and objects are deleted and the database is vacuumed, at most once a day per catalog. Nothing was ever deleted before: this repository's catalog had reached 13 GB for 2 117 session workspaces. Settings › Developer › Delete all gains the scopes Workspace catalog and Workspace registry; `scripts/verify.ps1` removes stale `unary-verify-home-*` and test `tempfile` directories older than a day.
- A renamed or moved checkout opens again instead of failing with "immutable checkout … already exists with different data": activation proves the move by native and Git identity, sets the old catalog aside and registers the folder fresh.
- Desktop host performance. The 0.2.303 baseline measured live 7.3 GiB heap live and 22.8 GiB peak with 6–8 sessions, about 2 GiB/s allocation churn and `app.emit` stalls up to 1.3 s. This round adds incremental session snapshot refresh with shared model history and an id index for branch reads, a bounded live-session byte budget (`sessions.liveByteBudgetMiB`, default 1536), non-blocking live emits, sidecar files for large transcript payloads, and lossless delivery for approval and worker-budget decisions. Dropped progress events now surface as `EventsDropped` instead of disappearing silently. The post-fix live measurement is still open.

## [0.2.303] - 2026-10-10

- `harness_health` reports the Desktop host's heap: live, peak and cumulative allocated bytes from a counting wrapper around the process allocator. It separates memory the app actually holds from what the allocator keeps reserved, after the host was measured swinging 7–17 GB per minute while idle.

- Memory: the agent now writes every memory entry itself during the session (`memory` `store`/`user_store`, repository or global scope). The model-driven post-passes are gone: session-end and checkpoint compaction, manual model compaction (`compact` now only dedupes the review queue), Phase 1 extraction, Phase 2 consolidation, the rollout → MEMORY.md distill pipeline, auto-consolidation and the scaffold reward loop; the Self Memory crate no longer calls a model at all, so the ~100k-token session-end compaction cost disappears. Deterministic capture stays only for tool errors: the failure journal and the model-free session ledger, which at session close attaches `verified_repair` / `verified_application` evidence to agent-written known bugs from an observed failure → fix → passing verifier.
- Dream is the only model-driven memory curator and no longer runs on a fixed timer only: it becomes due after 10 new entries per repository or global store, under cooldowns of 12 h per session, 1 h per repository and 30 min per app (all plugin settings), merges and generalizes existing entries, and may promote generally useful knowledge from a single repository to global. The Desktop Memory service card shows run count, tokens per run and in total, the last run and why a run was deferred.
- Memory panel: the dead Distill/Extraction/compaction cards are gone; Consolidate explains that it removes exact duplicates without a model call.

- Workers: a 2026-10-09/10 audit of 95 worker runs found four harness defects and fixed them. A worker whose run has started is no longer flipped to "failed before execution" by the 120 s credential-bind timeout (the bind waits for the child's core services before its timer starts), and a cascade retry no longer supersedes and shuts down a child that still has a live run or running shell. Registering a worker and starting it is now one transaction: a spawn that fails after registration returns the error at once instead of being settled by the watchdog 16 minutes later. The supervision ladder no longer cancels or force-terminates a worker that owns a running shell or sits in a sanctioned wait; a cancelled worker's terminal names who cancelled it (`cancelSource`), and a package parked for recovery is reported in the next spawn on that goal. Re-binding a worker to a subtask releases its parent-row binding; the never-enforced "advisory run time" note is gone from inherited role defaults; addressed intercom posts in session scope reach only their recipients and the lead.

## [0.2.302] - 2026-10-10

- Desktop no longer lags during builds. A cargo build wrote thousands of `target/` files per watcher window; every window overflowed the `workspace:changed` payload cap, collapsed to a Resync and re-bootstrapped the whole Explorer and Git panel every 2 s for the length of the build, with total CPU still below 80 %. The watcher now drops live deltas under build-artifact directories (`target`, `node_modules`, dot-prefixed framework caches; plain `build`/`dist` stay). "All files" still shows those directories after a bootstrap; only their contents stop live-updating while expanded.
- Agent-spawned processes (tool shells, background shells, PTY and sandbox launches) run at below-normal priority on Windows and `nice 10` on Unix (`unc_rust_workspace::process::apply_background_priority`), so the window keeps the scheduler during a compile.
- The Bash tool sends live output snapshots at most every 150 ms (leading + trailing) and the live snapshot carries only the last 16 KiB; the final result still holds the full output. The Desktop host coalesces `ToolCallUpdate` events per tool call to one emit per 80 ms, flushing before the call's end event, and the running-terminal row renders the last 400 lines instead of re-parsing the whole buffer. The chat's 1 s clocks and the 20 s auto-push poll pause while the window is hidden.
- A build whose warm lane or CPU budget is refused no longer runs silently in the shared checkout; the refusal names the reason and remedy. The CPU-token budget waits up to 90 s for tokens instead of refusing at once.
- Workers: one resource sample now admits four new workers instead of two, so a four-editor fan-out starts in the same window instead of parking the last two for a recheck each.
- Workers: only sustained memory pressure limits how many workers start. CPU load no longer halves the worker window, and RAM (≥ 85 %) or commit-charge peaks throttle only after they persist for 10 s across samples; a shorter build or test peak holds the window and one lower sample lifts the limit at once. Decision reasons name the RAM figure and how long the pressure has held.
- Workers: slash-joined prose such as "build/test commands" no longer becomes an editor's create target. The 0.2.300 fix only covered declared build outputs without a write cue; in the audited fan-out the noun "commit" counted as a write cue and six of eight editors queued behind each other on `build/test` for about sixteen minutes while RAM stayed below 53 %. A nonexistent path now has to name a file, an explicit directory (`dir/`) or the stem of existing files to count. `wait_agent` also shows the live queue reason when a peer message interrupts the wait.

## [0.2.301] - 2026-10-10

- GitHub: agents now see the same GitHub login as the app. The token the clone picker and Git panel already resolve (Unary sign-in, `gh`, or Git Credential Manager) reaches agent shells as `GH_TOKEN`, unless you set `GH_TOKEN`/`GITHUB_TOKEN` yourself, and is masked in shell output. Previously the app could list your repositories while agents reported "not logged in". The Git panel also re-checks sign-in when the window regains focus or after an in-app sign-in instead of keeping the state from when it opened.
- Sign-in: a refreshed OAuth credential is no longer lost when the keyring read fails or a sibling account replaces it, and a revocation keeps its (secret-free) reason so the warning after login names what actually happened.
- Chat memory: the local chat-memory commands run their SQLite work off the UI command thread, so a slow store no longer stalls the window.
- Agent tool catalog and core prompt budgets were re-frozen with the reasons recorded (core 24 846 bytes, catalog 214 717 bytes); eval git spawns no longer open console windows on Windows.

## [0.2.300] - 2026-10-10

- Index: automatic regex-shaped code searches use verified local matching with explicit native fallback when index coverage is incomplete. Boolean trigram candidate filtering handles alternation and Unicode case folding; results disclose skipped files and match limits.
- Index context: graph neighborhoods follow query intent, preserve relation direction and resolution provenance, and rank neighbors before truncation. Context packing skips oversized candidates instead of discarding later useful results.
- Retrieval evaluation: current lexical/CPU/GPU profiles, source fingerprints and pinned external corpora prevent silent embedding fallback or mixed-source scores. Agent navigation evaluation accepts explicit task sets and labels synthetic evidence rather than claiming production token savings.

- UnaryCad plugin: `unary.unarycad` (new crate `crates/unc_rust_plugin_unarycad`) drives the running UnaryCad app in-process through its agent API, replacing the external `unarycad-agent` MCP server once deployed. One grouped `unarycad` tool (status, capabilities, snapshot, command_list, command_run, call, capture, poll_events, cancel, reconnect), one actor per session, a read-only variant, at-most-once mutations that are never resent on an unknown outcome, and an image result for viewport captures. Device-free tests cover the wire, retries and schema; live pipe verification is still open. See `docs/spec/unarycad-plugin.md`.

- Shared patch journals no longer block a checkout forever. A publish error with nothing landed removes its journal. A foreign journal whose owner is not proven running is decided under the lock by its targets: untouched → closed, every target at its pre- or postimage → rolled forward, any target matching neither → archived as `pending.stale-<id8>.json` with its blobs, nothing replayed. Workers now register in the checkout registry (no git baseline), a live session whose row another session reaped is restored by its own heartbeat, and session teardown ends its row explicitly. In the audited case a worker's zero-write journal refused every multi-file patch in a C++ checkout for hours.
- A brief that merely mentions a path below a declared build output (e.g. `build/test`) no longer becomes a worker write scope, so editors no longer queue behind each other on a build directory. The debug bundle reads worker transcripts page by page (up to 1 000 entries per worker, visible cuts) and lists a cold parent's workers from persisted state without resuming it.

- Workers (audit of session 01a12223, build 0.2.298): a worker parked longer than five minutes is no longer settled as "dispatch lost" the moment it is admitted. Admission re-registered the original request, so the orphan watchdog measured its five-minute grace from submission and fired before the bootstrap added the active hook (two editors failed 2 s after their blocker released, 744 s after submission). The grace now counts from admission. A watchdog-settled stopped worker hands back its current-segment report as a failed partial report instead of a fixed text. The handover is the worker's last message; a trailer-only last message takes the report right before it, and a leading `<!-- answers: … -->` marker is dropped. The per-worker verifier is documented as disabled, matching the code.
- Mid-run answer guard: the forced answer turn now sends the normal request byte-identically and refuses (does not execute) any tool call it emits. Withholding the tools changed the tools block and invalidated the whole cached prefix; the audited run paid 3 such turns at 504–562k cache-write tokens each ($8.05, 22 % of the run). Anthropic documents that changing `tool_choice` also invalidates the messages cache, so `tool_choice: none` is not used either. Workers arm no human answer obligation (their mid-run text never reaches the parent), and turn guidance no longer ends a blocking `wait_agent`/`shell_watch` wait.
- Build lanes: a classified build or test re-observes a moving checkout up to six times (jittered 100–300 ms, outside every lock) before the lane open falls back, each rejection is traced with its reason, and a lane-placed command that names the shared checkout by absolute path carries a warning that such writes bypass the lane. Ninja, MSBuild and `pixi run <task>` are recognised by the repository ledger. Wait deadlines for `shell_watch` wait and `wait_for` are 120 s in code as in the schema, and a refused wait argument names the valid field, unit and range.
- Built-in skill `unary-build-lanes` ships with Unary: how classified builds, warm lanes and output reservations work, what their messages mean, and how to extend `.unary/build-profiles.json` per repository. It ranks below any same-name user or workspace skill; `disabledBuiltinSkills` in settings turns it off.
- Build lanes: lane builds never update the checkout's copies of their outputs, and now say so. The first successful lane build of a session names its retained roots (such as `build/debug`) and the lane path. A later command that runs a file under such a root from the checkout gets a warning when the lane's copy is newer or the checkout has none. In the audited UnaryCad checkout no lane had ever linked the app or test executable after 62 completed lane runs, so every binary the agents ran came from checkout builds. The cause: the agents prefixed their builds with `cd <checkout> &&`, and the own shell's directory remap covered only private run views, so every lane-placed build left its lane and built the shared `build/debug` while its receipt named the lane. An absolute `cd` or `pushd` to the checkout from a warm lane now stays in the lane, and the bypass warning no longer fires for it. The UnaryCad profile also matched `pixi run build` without declaring `.pixi`, which left its lanes without the pixi environment; the skill now explains that a profile replaces the inferred defaults.

- Workers: a finished worker no longer waits for its lead's whole turn. Its integration used to be serialized only after the parent's current run, and with it the handoff wake, the goal reconcile and the release of its file scope — a measured session saw two explorers reach the lead 18.6 minutes after they finished and two queued editors fail as "dispatch lost" behind a held scope. The lead now settles pending worker integrations at every mid-run safe seam (after each complete tool round), through the same claim and apply path.

- Index freshness is cheaper per edit: watcher events are filtered by the same ignore rules as the candidate walk before they queue, bursts coalesce (2 s quiet, 10 s cap) into one generation, a batch whose paths are ignored or unchanged by content hash publishes no generation, `tags.bin`/`trigrams.json` are not rewritten when nothing changed, and `save_snapshot` no longer looks up each chunk's text linearly. `index_manage status` now prints the last generation's stage timings and watcher-batch size. Measured before (2026-10-09, 3 647 files): five generations in 12 minutes for four edited files, ~2:47 min and ~1 GB rewritten each.
- Embeddings: memory and docs retrieval no longer start a second, CPU-lane sidecar beside the index's GPU sidecar under the `gpu` index mode. `EmbeddingRuntimeOptions.device` is now optional (`None` keeps the machine-wide lane published from Settings › Index); the plain `embed_query_text`/`embed_passages` helpers used to override the lane with the `Cpu` default (measured 2026-10-09: a `--device cpu` sidecar at 2 085 CPU-s next to the `--device gpu` one at 83 CPU-s).

## [0.2.299] - 2026-10-09

- Per-turn git work is no longer duplicated or on the session thread: a turn that ran no workspace-writing tool does no end-of-turn worktree capture at all and never waits for its start capture; a writing turn computes its final capture and diff on its own thread and the session applies the result on its idle tick (or before the next prompt, compaction, rewind or close, so the diff stays in its own turn), so a follow-up message is no longer queued behind a full worktree capture. A start capture a previous turn never joined is chained instead of losing its baseline to the 10 s workspace lock. The session index no longer re-sweeps all sessions every 1–3 s: its own persist temp files (extensionless, in the watched sessions root) were triggering a full sweep each time. The prompt command no longer loads settings and makes two actor round-trips only to log them.

- Memory: startup maintenance runs once per process and repository instead of after every session start. `run_startup_refresh` enqueued the full recovery ladder (seven recovery phases, shard prune, 10 s flush wait, Phase 1/2) on every successful publish, and Phase 1/2 artifacts moved the startup epoch so the follow-up refresh re-enqueued it (measured 2026-10-09: one `self-memory-maintenance` thread with 632 CPU-s after 55 min and 14 session starts). A per-`cache_key` gate re-arms after 24 h; tombstone-hinted purges bypass it. The memory plugin's `memory:lifecycle phase=…` timings now go through `tracing` into `app.log` instead of bare stderr.

- Desktop chat: background sessions' streaming deltas are coalesced per content block inside the warm tap's 100 ms window, so a session nobody is looking at costs one reducer pass per block per flush instead of one per token (measured 2026-10-09: with 7 sessions streaming, the WebView2 renderer main thread stayed at ~58 % of a core while the visible session was idle). Visible-session rendering is unchanged.

- Workers: a subagent's handoff can no longer get lost between its completion and the lead's next turn. The handoff wake used to live only in the lead's in-memory follow-up queue while the worker registry already recorded it as delivered, so an actor rebuild (live-session cap, restart) dropped it and a later `wait_agent` returned only an `outputRef`. The wake now takes the same path as shell and watcher wakes (one durable `info` row, one coalesced machine wake), the lead stamps the record when its model actually consumes the wake, and an actor built over a completed-but-unread handoff re-queues it once under the same id. The Session facts line counts a worker blocked on a parent decision as `awaiting decision` instead of `running`, and the worker-decision guidance names the revalidation actions (`wait_agent` timeout 0, `session` `observe`).

- Sending a prompt no longer waits for the git turn-diff baseline: the shadow-repo capture (`add -A` + `write-tree`) used to run on the session thread before `AgentStart` and, on a checkout where it exceeds the 30 s git timeout, stalled every prompt 30–70 s in a silent busy state while never publishing its reusable index. It now runs on its own thread, is joined only before the first workspace-mutating tool call and at turn end, gets a 10-minute capture budget so the reusable bundle is published once, and logs its skip reasons and durations in `app.log`.

- Opening a large chat session is cheaper: resume parses the transcript once instead of twice, each JSONL line is decoded once (the `Value` detour only runs for a malformed line) and large transcripts decode in parallel — a 61.8 MB / 9 800-entry transcript parses in 61 ms in release instead of 303 ms, and resume parses it once instead of twice — and only the image attachments the model can still see are inflated into memory on open instead of the whole attachment directory. Explicit `close_session` now runs the full per-session teardown (plugins, Goals store, scratchpad, tool state) like archive and eviction do, and a background session's warm transcript keeps as many rows as a cold open fetches.

## [0.2.298] - 2026-10-09

- Spec: the watcher incident in `docs/spec/tool-execution-pipeline.md` names the goals ledger directory instead of an elided runtime path, so `rules-claims` resolves every claim again; otherwise identical to 0.2.297.

## [0.2.297] - 2026-10-09

- Release repair: 0.2.296 was tagged from a tree whose harness called `clear_orphaned_watcher_bindings` before the owning `task_store.rs` change was committed, so it never built; 0.2.297 carries the complete watcher package (f01fa6d9d) and is the first build of the 0.2.294–0.2.296 fixes.

## [0.2.296] - 2026-10-09

- Gate fixes for 0.2.296: `search` action `outline` reads the session's private file versions (held paths override the checkout, tombstones disappear, private additions participate) instead of the shared disk only, with its two-mode isolation fixture; a worker's `Completed parent goal ids:` line reads a backticked id only from a leading list span, so `Goal `id` is done` prose is never applied while `` - `id` title `` entries still parse; the eager-prefix ceilings of `tool_catalog_inventory` are re-frozen at the measured 211 907 B (+212 B ≈ +53 estimated tokens per request from the research chat-memory and Live Symbol Graph surfaces); the behavior-eval production catalog is regenerated for the `index`/`search` descriptions.

- Sessions list: with ~3,200 transcripts (2.1 GB) the Desktop re-read every session file to EOF on every run start or end (measured: 45.5 s per rebuild, so the rebuild thread never stopped), because the session index was a 5 s TTL cache rebuilt from scratch and the metadata reader never stopped at its budget; archived sessions cost as much as running ones and `find_session_files` read every file's metadata for every lookup. The index is now in-process state: a `notify` watcher on `~/.unary/sessions` queues only the paths that changed, a stat-only sweep (startup, overflow, every 10 min; measured 192 ms on the same tree) reuses every transcript whose `(mtime, len)` did not move, a changed transcript is read incrementally from its scan cursor (restart on a shrunk file or rewritten header), and the index is persisted as `session-index-cache.json` so a restart lists every session without opening one. `harness-sessions-changed` carries the changed session summary so the Desktop upserts one row instead of refetching the list; the full fetch stays the reconcile path. `sessionStore` releases the transcript window of archived and no-longer-focused sessions (focused, running, pinned, approval-pending and surface-retained sessions are never released).

## [0.2.295] - 2026-10-09

- Index: the retrieval profile defaults to `lexical` (no embedding model, no sidecar); the Granite `cpu`/`gpu` lanes stay an explicit opt-in [user decision 2026-10-09]. Memory and docs degrade to keyword retrieval under that default with a "semantic ranking unavailable (lexical index mode)" note instead of starting a sidecar. The structural index is now watcher-driven (coalesced incremental generations, git content-hash catch-up, `freshness.generation`/`generationStale` in every index result, no readiness walk on the request path), Markdown spec units/rules/skills are graph nodes (`Document`/`Section`, backtick references as `references` edges with `docReference` provenance), method calls on typed receivers and calls inside macro token trees resolve, impl blocks fold under their type node, Svelte/Vue templates contribute component/handler edges, tags and graph share one extraction pass and one `symbol_id` space (`tags.bin` format 3), the legacy `repo_map.rs` ranked-files fallback is gone, and `graph_query` gains `impact_of_change` and `tests_for`. Lexical ranking: Markdown headings are a boosted BM25 field, umlaut folding, German stop words, fuzzy fallback, document-intent routing fused by reciprocal rank, resident trigram snapshots, grep hits grouped by enclosing symbol. `search` gains the index-free `outline` action and parallel `patterns` grep (+16 B catalog, ≈4 estimated tokens per request). Owner-Hit@10 on the checked-in 33-fixture corpus (`crates/unc_rust_plugin_index/eval/owner-eval.unary-code.json`, runner `retrieval_owner_eval`) moved from 66/99 at HEAD to the number recorded in `docs/spec/index-subsystem.md`.

- Watchers: a session parked on a `blocked` goal with an armed background `wait_for` watcher is no longer shed by the live-session cap, which silently cancelled the watcher and left the session waiting for a wake that could never come (`wait_14`, 2026-10-09). Binding a `resumeWatcherId` that is not a live watcher of the session is refused, a rebuilt session clears and reports the bindings its dead watchers left behind, and a cancelled watcher is logged.

## [0.2.294] - 2026-10-09

- Isolation: an isolating session without custody is given its private store before the write admission, so its first `write`/`edit`/`patch` no longer fails with “became private after preparation” (`durable_edit_latency` on-without-custody had been red since the 0.2.292 gate). Goals handoff: a `` `id` title `` list entry under `Completed parent goal ids` is parsed again (explicit `none` still wins); a brief path ending a sentence (`boolean_exact.cpp.`) is scoped without the period. Catalog budget re-frozen at the exact 2026-10-09 measurement (+146 B ≈ +39 estimated tokens per request from the `research` family's chat-memory and appendToJob growth).

## [0.2.293] - 2026-10-09

- Harness audit fixes: the investigation classifier ignores the harness pointer block and honours read-only disclaimers, so workers keep the audit exemption; a worker's `Completed parent goal ids: none` never applies an id from prose; sanctioned waits on a live child process are not a stall and a supervision park is a parent decision instead of a transient respawn; compaction renders peer coordination without an accepted ack under a no-authority heading; a PermissionRequest deny wins over a later allow and exit 2 without stderr denies; `Retry-After` is honoured as a floor; the plan bridge no longer resurrects swept steps and abandons removed ones; `dangerZone.allowBrowserFallback` is read from the user scope only; URA `sh()` consults the destructive-git guard; scheduler accounting survives config rebuilds; multicast SSRF ranges, MCP cancellation on timeout, abandoned parallel plugin slots, vendor search redirect guard, duplicate `tool_result` dedup, Responses unknown-event error text; chat memory redacts secrets and truncates oversized recall entries; BINDING_RULES needles are distinctive and complete; `verify.ps1` reports partial coverage; harness event sequence gaps trigger a session-state reconcile.

- Review changes: a file the session never touched no longer shows up as a full-file `+N/-N` diff in a `core.autocrlf=true` repository whose blobs were committed with CRLF (UnaryCad `MainWindowDraw.cpp`, `+2930/-2930`). The checkpoint shadow's reused-index capture ran `git add --renormalize`, which converted those blobs to LF while the HEAD-seeded capture kept them CRLF, so two captures of an untouched tree disagreed on every line; the reused index now rehashes under the same conversion rules and old cache bundles are rebuilt. The same dialog also listed one file twice when the edit result carried the relative path and the end-of-turn record the absolute one; both now merge into one entry.

- Chat memory: local SQLite persistence no longer requires the remote package, account sign-in or a Cloudflare service. The Desktop memory dialog can add, refresh and remove local memories; Chat-only research actions remember durable facts and recall them across chats without exposing repository memory or arbitrary filesystem access. Cloud synchronization remains separate and optional; the cloud-deployment placeholder and unimplemented Re-aggregate button are removed.

- Workers: the goal-binding refusals from `spawn_agent` now name the working remedy (one subtask per package, re-bind the live worker, spawn on a sibling) instead of only `send_input`/`resume_agent`, and the Goals prompt says so before the first spawn. Session debug exports read a running worker's transcript through a short bounded lock retry instead of a single try, and an unreadable child transcript is listed in the export's top-level `workerEvidence.omissions` with its session and hook id.

- macOS: fix new sessions failing with "Could not prepare … idempotency key was already used" after the Mac renumbered the disk a project lives on (for example after a restart). Projects are now identified by the volume's persistent UUID, and an existing project recorded with the old device number is updated in place instead of refused.

## [0.2.292] - 2026-10-08

- Anthropic OAuth: restore the byte-exact first identity block without a trailing newline. The 0.2.291 newline change caused generic HTTP 429 `rate_limit_error` responses: same-account A/B/B/A requests returned 200/429/429/200. A serialized-wire regression guard fails with the newline and passes after restoration. Post-fix production Opus 5 completed five cache turns with exact reuse and six read-only tool tasks with 6/6 successful dispatches; the capability-map rename and late prompt layers remain unchanged.

## [0.2.291] - 2026-10-08

- Research: a batch `web_fetch` with `appendToJob` reports every URL as appended or skipped with its reason (text and `details.results[]`) instead of silently dropping sources over the `maxSources` cap; compressed responses are decoded (`gzip`/`brotli`/`deflate`) and an undecodable body is a failed fetch, never evidence; the anthropic-native engine finding keeps the cited answer text (it was the preamble sentence) and `cited_text` counts as a usable excerpt that outranks snippet-less search hits; a hosted search runs under an idle bound (180 s without stream progress, 15 min ceiling) instead of a 180 s whole-call timeout, keeps partial text on a cut (`truncated`), and continues Anthropic `pause_turn` up to three times.
- Patch/claims: a fully landed multi-file patch journal left by a closed worker no longer blocks every `edit action=patch` of the checkout; a released file reserved for a queue head that never retries lapses after 90 s, and the park message names a reservation as a reservation instead of "held by session …".

- Workers: the handover is every assistant text since the last human or steer message (an earlier message carrying `Completed parent goal ids` is kept), not only the newest one; a `send_input` queued after a worker's report no longer replaces the report with the short follow-up answer. A deadline-killed shell whose exit could not be confirmed is `exit-unconfirmed` and no longer wedges the worker in `Finalizing` (one explorer sat there 25 minutes, its audit undelivered); build and test commands are not killed by the generic worker shell deadline.
- Workers: the parent wake carries the worker's last message verbatim from the durable artifact (32 KiB) instead of synthesis prose plus a 1024-byte prefix; the "Workers settled" FYI renders the real summary instead of `{summary}`; a worker may read its parent's transcript as its pointer promises; `session transcript` honours `maxCharsPerEntry`, reads a live child instead of returning zero rows, and a busy snapshot is an error, not an empty page; `wait_agent`/`shell_watch`/`wait_for` timeouts no longer count as no-progress for the repeat-call and tool-loop guards.
- Goals: a hook-bound row is checked only when the worker names it (or its subtree) under `Completed parent goal ids`; explicit none, a missing or unparsed label releases the binding to the parent instead of auto-completing. Successful checkoff clears `assignedHookId`; every non-success terminal and `close_agent` release the binding; a checkoff refused by integration state releases immediately.
- Workers: `spawn_agent` no longer refuses with "Claim-owner handoff exceeds Intercom capacity" when the brief overlaps a foreign claim; the owner FYI carries a bounded task preview and the worker starts (five editor spawns were refused on 2026-10-08).
- Prompt: date, cwd, worktree note and user name are a late section behind the cache cut (frozen `prompt_date`), `Current mode:` is gone (the mode suffix is the only statement), activated family guidance renders behind the cut, and memory/doc/MCP inventories are turn context instead of persistent layers. On Anthropic OAuth the identity block ends with a newline and the capability map uses the advertised `Read`/`Bash`/`Edit` names.
- Prompt: the Blender `SKILL.md` body is no longer inlined (the skill card is the entry point, ≈900 tokens per request); index, research, doc, decision-card, sizing, question-vs-instruction and edit-mechanics rules each have one emitter; `prompts/unary/NOTICE.txt` is gone; project rules render as plain text with a provenance header instead of a JSON string; the repo-ledger turn brief lists three errors and drops records superseded by an older HEAD.

- Builds: a `.unary/build-profiles.json` declaration now matches the literal argv prefix of any simple command in the run (redirects excluded; pipelines, `&&`/`;` lists and literal `cmd /c "…"` / `sh -c '…'` bodies walked; backslashes as `/`) instead of the whole command text verbatim, and a `cmd /c` script body discloses `cmd /c <script>.bat`. Three UnaryCad sessions had built one `build/debug` through `cmd.exe /d /c "tools\win\cmake_build.bat build/debug --target X" > log 2>&1; rg …` and corrupted each other's `.obj`/`.lib` files for nine attempts and four 60 s waits, because the wrapper classifies Unknown and an exact-text profile never matched the per-attempt redirects.

- Research: `rag_query` / `doc_rag_query` resolve credentials like every other model call (settings → stored OAuth via `AuthStorage` → env) instead of failing `openai-codex` with “requires an API key”, and an explicit `model` argument is honoured (`modelKey` in the result). `deep_research_finalize` and `promote_to_doc` persist first and embed in the background (`embedStatus` pending/embedded/failed; an embed failure no longer fails a persisted job or loses `docId`); an unchanged claim set reuses its semantic review instead of paying another model call; provider-attested native-engine sources without an excerpt are fetched by the engine planner instead of staying uncitable. Opt-in live scorecard `research_scorecard_live` records ten observed checks under `target/`.

- Chat: user-request cards redesigned after `docs/mocks/user-request-cards-bestand-vs-sota.html` [user decision 2026-10-08]. An offer's two pills are the answer (click, 1/2 or Y/N send at once; no Send/Cancel pill; custom text still needs Enter). Question cards show a radio circle for single-select and a checkbox square for multi-select with the shortcut digit at the row's end, one wrapping footer (hints left, cancel · pause · send right) instead of a clipping title-row hint, no “No answer options available” sentence, and a `1 / n` counter on batched questions. The resolved card marks answers with ✓ and collapses more than three unchosen options.

- Workers: the supervision ladder counts a worker's completed, non-repeating tool exchanges as progress for every role; an editor reading a GPL reference before its first edit was nudged at 60 s and 120 s and headed for an unconditional 10-minute park. Silence and ≥ 3 identical calls still escalate.
- Workers: the supervisor's check-in nudge now reaches a RUNNING worker at its next tool-result seam as passive context; it was queued as a follow-up that only an idle worker consumes, while the runtime snapshot told the worker to "pull with session intercom_read" (two ~8.9k-token reads for a message that tool cannot show). The snapshot line now says queued notices arrive at the next request boundary.
- Workers: a worker's `intercom_read` no longer returns its parent's harness notice rows (build tails, stall notices addressed to another session) or full sibling briefs; harness rows are visible to their recipient and the UI only, and `workers[].task` is a 240-character preview with `taskTruncated` unless the reader is that worker or its parent.
- Workers: a worker's declared file scope is now what its brief tells it to write or create (paths need not exist yet; read-governed, read-only and "next to the existing …" mentions are excluded), instead of "which live claims the brief mentions": an editor told to create four files had an empty scope while its sibling's only scope and claim was the lead-owned read-only header it was told to read.

## [0.2.290] - 2026-10-08

- Isolation: Unary's own trees are the isolation engine by default. New Settings › Git „Worktrees“ group (`worktrees.engine` unary|git, `worktrees.automaticBuildIsolation` default on); classified builds and tests run on a warm, retained Unary lane even when the session holds no private versions; Git worktrees are created only by the explicit Isolate action.
- Isolation: Smart File Isolation On records private per-file versions on every platform (custody no longer depends on the Windows AppContainer proof); the only refusal is an unavailable native store, never a silent shared write. The retired `claimReleaseWaitSecs` setting is dropped on load.
- Isolation: workers spawned from a manually isolated session work in the parent's tree and are part of its fold back; isolated-tree fixes for an unchanged executable bit on Windows, junctions in the tree, and in-memory state after a failed persist; tree status reports its kind (`fileIsolation: file|tree`) and the mode under `smartFileIsolation`.
- Workspace engine: capture writes each file once and materialization no longer fsyncs every file of a temporary view; parallel populate and hash walks. Warm capture of `crates/` 20.8 s → 1.7 s, materialize 4.9 s → 1.4 s. Lease writes are atomic, a published root is never stranded without a lease, and one corrupt build-slot record no longer disables the lane pool.
- Providers: an Anthropic tool call without input (its single `input_json_delta` carries `partial_json: ""`) no longer fails as "argument JSON is incomplete (0 bytes)"; every zero-property MCP tool (e.g. `unarycad_reconnect`) was unusable.

- Workers: an explorer's evidence-only read budget is now 8/20 turns (nudge/hard limit) instead of the shared 12/40; the hard limit recovers into the final-answer turn. Both measured explorer runs of 2026-10-08 had their answer by turn 6 and read on to 15 and 40.
- Workers: a `spawn_agent` whose files another session claims no longer refuses with "Work routed through project Intercom… await the owner"; the worker starts, queues on the claimed paths and continues on the rest, and the owner gets an FYI (never a wake). A lead had waited indefinitely behind an owner that reads Intercom only on its own next turn.
- Workers: a worker's result reaches the lead once — either as the handoff wake or as the `wait_agent` result, the other carrying only status, handoff metadata and a reference — instead of twice (up to 16 000 + 32 768 characters).
- Workers: a continued or retried worker's second completion now wakes an idle lead; its wake id is per result revision, so the stale-wake filter no longer drops it as a repeat of the first completion (the lead had sat idle for five minutes until asked what it was waiting for).
- Lead: waiting on an owned background shell or worker (`shell_watch wait`, `wait_agent`) no longer triggers the goals-staleness guard, the answer-guard tool withholding or the continuation nudge, each of which cost a full request; the Workers board drops closed or read terminal hooks instead of repeating a failed hook every turn.
- Worker transcript: a "Copy handover" button beside "Copy all" copies only the brief, the final message (or the abnormal-stop outcome) and the run line, instead of the full tool transcript.

## [0.2.289] - 2026-10-08

- Workers: the lead's `spawn_agent` message is now the worker's user turn verbatim and the harness adds only a short pointer block (goal id, final-message delivery, where to fetch repo map, ledger, Intercom and parent transcript on demand) instead of a 9.5 KB auto-generated evidence packet, role paragraphs and a competing report schema; a measured explorer brief shrinks from 14.9 KB (16 % lead-written) to the lead's text plus about 1 KB.
- Workers: a worker that trips the evidence-only circuit breaker now gets one final-answer turn ("write your answer to your parent now, no tool calls") so a handover exists; the 2026-09-30 guardrail-profile removal had silently disabled the breaker recovery for every session, which turned a 4-million-token explorer run into an `abnormal_stop` with no answer.

## [0.2.288] - 2026-10-08

- Chat: the `+N -M` badge on an edit row counts the `+`/`-` rows of the backend's unified diff (Codex-style), so collapsed and expanded rows agree and match the preview; results without a diff string and the changes strip above the composer count per hunk instead of treating every untouched line between two hunks as removed and re-added (an 8-line two-hunk insert read `+53 -45`).
- Chat: opening a pinned message no longer crashes the app with "No markdown component provided"; the pinned-message window provides the chat's markdown renderers itself.

- Verification gate: `scripts/verify.ps1` sizes cargo parallelism from free commit charge instead of core count when run outside Unary (override with `CARGO_BUILD_JOBS`), and the dev/test profile now uses `line-tables-only` debuginfo; three 0.2.286 gate runs had died of commit exhaustion (Windows event 2004: 7.8-9.5 GB per rustc at 64 jobs) with 179 GB of physical RAM free.

## [0.2.287] - 2026-10-07

- Chat: a settled `request_user_input` card no longer crashes the app with "Cannot read properties of null (reading 'description')" when its resolved state is re-derived; the resolved card is null-safe.

## [0.2.286] - 2026-10-07

- Compaction handover: a session without a Goals board (every worker) now carries its last answer and newest reasoning digest as `## Next Steps` instead of being told to continue a board it never had; workers no longer re-discover their plan after every compaction.
- `wait_for` `sessions_idle` no longer reports a quiet repository while peers run from the same checkout spelled with the Windows extended-length prefix (`\\?\L:\...` vs `L:/...`).

- Settings › AI › Agent: the Editor and Explorer worker compaction fields (auto-compaction high-water %, context budget %) are searchable top-level fields directly under the session's auto-compaction threshold instead of unlabeled inputs inside the collapsed Subagents cards; an empty field still inherits the parent session's value.

## [0.2.285] - 2026-10-07

- Tool scheduling: a read-only call scoped to a directory (`search` `path`, `ls`, `index` `path`) now waits for an earlier edit of a file beneath that directory instead of being hoisted into the parallel wave ahead of it; before, `edit <dir>/a.txt` followed by `search grep path=<dir>` in the same response reported the pre-edit content.

- Workers: a failed worker's `wait_agent` line and the lead's failure notice now name the classified cause (provider, model, HTTP 429), a provider quota or credential refusal is no longer retried on the same model, and `session` action `watch` returns immediately for idle or unknown targets and when the caller's own session has pending input. `wait_agent`/`watch` accept integral float timeouts.
- Settings › AI › Model fallbacks: every model slot (chat, Multi Expert judge, compaction, Pilot, memory, title, image, video, each worker role) can name up to three fallback models; the chat session arms the `chat` chain and a worker whose role model's provider is refused starts on the first admissible chain model instead of parking until the refusal lifts.
- Chat terminal list collapses past five running terminals and caps the expanded list at ten rows with scrolling; finished foreground shell calls no longer linger as running terminals.

- Meta AI: sign in with the Muse Code subscription through a device-code login. The minted inference key is what requests send; an inactive subscription or a mint with no key stays unauthenticated. Meta's own subscription page still says that credential is for the Muse CLI, so flat-plan billing from Unary is unverified.
- Meta AI provider: add Muse Spark 1.3, 1.3 Contributor, 1.2 and 1.2 Contributor beside 1.1 with Meta's published Standard/Contributor pricing; pass `xhigh` through instead of clamping it to `high`, and expose `max` on Standard-tier 1.3 only (other versions clamp `max` to `xhigh`).
- Meta AI provider: send the session id as `prompt_cache_key` so Muse Spark requests stay on the replica holding their prefix cache; live, follow-up turns on 1.3 now read ~5.1k cached tokens where single-turn runs without the key read none.

## [0.2.284] - 2026-10-07

- Restore refusal-aware account fallback for worker leases, with regression coverage for mixed, all-refused and expired-refusal pools; correct the finished-prerequisite Git seed fixture and record the authorized operative-core size ceiling.

- Builds started by any session now share one process-wide jobserver sized to the CPU budget (`CARGO_MAKEFLAGS`/`MAKEFLAGS`), so parallel cargo runs, rustc codegen threads, make and ninja no longer multiply the budget.

- Settings › Providers and the Usage dialog show each Codex and Claude sign-in's banked usage-limit resets (count, expiry, or the provider's reason they are unavailable) and can redeem one after confirmation.
- Anthropic OAuth requests now identify as Claude Code 2.1.292 (was 2.1.284), the current upstream release; eval and diagnostic clients use the same version.
- Tool-result hygiene no longer ages already-sent results in the production configuration; automatic compaction gains a configurable absolute context ceiling (200,000 tokens by default).
- Anthropic inline tool loading includes a stable deferred placeholder from the first request. Cache diagnostics now report long-session tool coverage and stop scheduling further fixture turns after a terminal provider failure.
- Desktop provider and model lists use Unary Optimized grouping and Optimized tags with model-specific cache notes instead of Tested badges.
- URA's dispatched host profile uses a 300-second deadline, 256 nested calls and explicit memory bounds; nested tool results remain inside the script unless returned or printed.
- Team intercom no longer refuses a sibling worker as "not a verified member" while that worker is briefly applying a stream event; refusals now name the sender room and the recipient's live lookup. Cache diagnostics fingerprint each system-prompt part so a mid-session prefix change names the part that moved.

## [0.2.283] - 2026-10-07

- Generated native shell sidecars stay out of source control, preventing platform-specific binary swaps from dirtying release builds.

- Harness audits now use explicit build, source and run provenance instead of treating older worker chats as current evidence. Captured memory turns preserve that provenance across resumed runs, with unknown legacy data kept distinct.
- Subagent supervision notices retain machine provenance, include scoped observation identity, and describe inherited budgets as advisory. Automatically supplied repository context no longer repeats long assignments in its headings and is bounded as a complete packet.
- Quiet shell jobs distinguish unknown activity from measured inactivity and no longer recommend restarting merely because output is absent. Build results disclose observed source drift instead of claiming they verify newer checkout contents. Windows worker interruption now stops observed descendants outside the shell's job as well as the job itself.
- Desktop debug exports include bounded, authorized worker evidence with explicit gaps, alongside the existing parent-session data.

- Agents now suggest Debug mode for bugs the code alone cannot pin down (intermittent, timing-dependent, hard to reproduce), and the composer offers a "Switch to Debug" button when your message describes such a bug. When the agent recommends it in a reply, a Switch to Debug button appears right under that reply. Switching stays your click.

- Windows CPU budgets now use one shared OS CPU-time ceiling for Unary and its newly launched child processes instead of relying only on per-pool defaults. Settings reports whether enforcement is active, failed, or unsupported; native terminal tabs inherit build/test parallelism defaults too.

## [0.2.282] - 2026-10-06

- Blender: Python execution is on by default; it can still be turned off in Blender's Unary sidebar.
- Desktop Chats get a Research footer button that opens the chat's research records; Assistant mode gets the Extensions panel.

## [0.2.281] - 2026-10-06

- Blender connects to Unary automatically once the add-on is installed: no pairing file, no session picker and no Start step. The chat shown inside Blender follows the Unary chat that is using Blender.
- Clicking a PDF mentioned in a chat message opens the built-in PDF reader instead of a broken text preview; mentioned videos open in the video player.

## [0.2.280] - 2026-10-06

- Agents reuse current worker findings and checks, keep informed workers for follow-ups, and verify completed work packages instead of checking after every edit. Non-blocking side findings no longer automatically extend the task.
- Inherited worker time and token defaults no longer abort long-running work; explicit limits and safety controls remain in effect. Read-only audits mentioning builds or negated changes are classified correctly.
- Superseded failures stop recurring as active build-error hints while their history remains available. Irrelevant contract candidates can be excluded without unnecessary full-document reads, and ordinary search no-match results and expected pipe closure no longer produce false masked-error notices.

## [0.2.279] - 2026-10-06

- Agent guidance reuses existing goals and tasks when they fit, without a numeric limit. Checks, retries and small corrections no longer automatically become separate tasks; new work still gets its own entry when needed.
- Quarantine recovery authenticates the current writer before replaying a stored result, and mobile open-URL commands run off the main thread.
- Completed chat tool results load when expanded and release their loaded payload when collapsed; the low-memory notice uses the 95 % threshold, while early cache relief remains unchanged.

## [0.2.278] - 2026-10-05

- Completed local chat tool results load when expanded and release their loaded payload when collapsed; reopening a session starts results collapsed. Stored results remain available for reloading.

- The "memory is running low" notice in the chat now appears only once the renderer heap is at least 95 % full, and disappears again as soon as it drops below 95 %. The early relief is unchanged: from 65 % the app still evicts caches and loads history on demand, with a neutral status saying so.

## [0.2.277] - 2026-10-05

- Fix a message you sent sometimes missing from the chat until you reloaded: a run that started while the chat was briefly not listening lost all of its live events, a message sent while the agent was working could be dropped under load, and an older copy of the conversation could overwrite the newer one on screen. All three now keep your message visible.
- Debug mode now collects probe results through its own probe tool instead of the browser console. When you ask it to fix without reproducing the bug, it keeps the probes in place and the hunt open until a reproduction confirms the fix, and its report separates confirmed causes from causes only shown in the code.

- Builds no longer freeze the PC: the Settings › System CPU budget now caps every build and test a shell starts (cargo, CMake, pixi/conda builds, test runners, OpenMP, Go), and Auto keeps a small reserve free (120 of 128 threads, 6 of 8). Machines with more than 64 threads are now counted fully on Windows instead of as 64.

## [0.2.276] - 2026-10-05

- Workers run more in parallel when Subagents or Teams are on:
  - The agent's continuation hints now name the delegation tools it actually has.
  - A finished worker whose result was already read hands its slot to the next one, and a follow-up message brings it back.
  - Closing a worker that never started works again.
  - Teams shut down cleanly and free tasks held by a teammate that crashed.
- After a restart, the Desktop lists runs that the last quit, update or crash interrupted ("Runs stopped since your last session"). Each run offers Resume, Open chat or Discard; Resume all continues every listed run. Runs you stopped yourself stay collapsed and can be hidden.

## [0.2.275] - 2026-10-05

- MCP settings are available globally without a workspace. Enabled servers allow calls by default; Custom enables visible per-tool restrictions and advisory folder context, while failed or timed-out permission prompts are no longer reported as user denials.

- Fix runs ending halfway after the agent was made to answer a message you sent while it worked: that answer is written without tools, and the run used to stop right there ("I'll write the mock next"). The agent now continues the task with its tools afterwards.
- Fix a message you sent while the agent was working staying stuck after you pressed Stop: Stop now sends it right away as the next request, like Codex. Its chip says "Sent after the next step · Stop sends now".
- Long silent thinking no longer gets cut off after two minutes: a model stream may now stay quiet for up to five minutes before it counts as stalled, and a stalled request is retried up to five times, showing "local retry n/5" (both as in Codex CLI).
- Every session now opens at the bottom: a remembered reading position used to make some sessions open at the top every time.

## [0.2.274] - 2026-10-05

- Restore already received thinking, answers and tool progress immediately when returning to a running session; retain interrupted-conversation recovery and correct the release guard for historical worker branches.

## [0.2.273] - 2026-10-05

- Fix thinking and answers disappearing when you leave a running session and return: already received text and tool progress return immediately, answered approval cards stay answered, and interrupted conversations recover their latest saved messages without losing their place.
- Large chats now unload older messages and load them back only when you scroll up; the session size where this starts is configurable in Settings (0 = automatic, based on the memory budget).

## [0.2.272] - 2026-10-04

- Teams now run under one coordinator: it starts the team's members itself, and the session that started the team hears only from the coordinator, when the team is done, blocked or needs a decision.
- Sessions now get a short, quiet heads-up when another session already fixed a failure they hit, when someone is waiting for a file they hold, or when two sessions keep hitting the same failure.
- A background command's exit the agent has already read no longer starts an extra turn, several background events that land while a session is idle now start one turn together, and agents can list the messages still waiting for them.
- Fix background commands staying "running" after they finished when a process they started kept their output open (for example MSVC's vctip.exe after cargo): the shell now waits at most two seconds for remaining output, names the leftover process, stops it and reports the command's real exit code.
- Fix the "Interrupted run" banner appearing in a live session whose workers were only waiting for a free slot, and its Discard cancelling them without warning: the banner now shows only for workers a restart left parked and says how many, and Discard first asks, naming how many waiting workers it cancels, that their goals go back to pending and that the session pauses.
- Fix web fetches of HTML pages showing only their first 1,600 characters: pages now come through whole, up to 100,000 characters per window and continued from an exact offset, together with up to 60 same-site links the text does not already show, so the agent follows real links instead of guessing URLs.
- Research records no longer use a fetch's cache or redirect note, or a page excerpt, as a source's citation note.
- Stopping a background command yourself no longer wakes the session with a needless report; a timeout, a failure or a stop from outside still does.
- Fix Codex Voice after the Desktop move: it again knows which session it belongs to, sees your typed messages, speaks the agent's final answer and gives short, honest waiting notes instead of made-up progress.
- Keep 10,000 lines of terminal history instead of 1,000, so long build and test output no longer scrolls away.
- Fix stopped terminals reappearing as running when a chat was reopened after an aborted turn or an app restart.
- Fix the Blender setup in Desktop settings failing with "workspace root is required" for sessions without an IDE window: it now uses the focused or open workspace, and asks you to open one when none is open.
- Move the Teams switch to the left of the main model in the composer's subagent controls.
- Fix a long autonomous run stalling at its second "context too large" refusal: each refusal now gets its own compact-and-retry once the previous retry went through, and the refused request's card shows it recovered instead of "did not complete".
- Fix harness notices such as a failed worker batch showing up in the composer queue as if you had written them, with "Steer now": the queue now lists only your own messages.
- Text styles can learn from files, folders, public websites, read-only SQLite mail databases and pasted exports. The extracted writing profile and short examples work independently of the original sources; add a source or reread it explicitly to refine a style. Approved edits to suggestions refine the originating style without treating intermediate keystrokes as preferences.
- Fix a worker being held up by files its brief only told it not to touch: a path right after "do not edit", "never touch" or a similar phrase no longer counts as a file the worker will change, so it no longer waits for whoever is working on that file.
- Fix the app window crashing with "out of memory" on machines with a lot of RAM: the window's caches grew with the computer's memory (up to 32 times their normal size on 256 GB), but the window itself can only use about 4 GB. They now stay within that limit and shrink on their own when the window's memory runs low, instead of crashing it.
- Fix streaming thinking looking jittery: thinking now types out at an even speed instead of racing ahead and pausing, the answer only starts once the thinking is fully shown, finished thinking is no longer greyed out under the streaming fade, and text no longer pauses and jumps at the end of each paragraph.
- Fix workers and other long-running session commands being reported as failed after two minutes while they were still working: the two-minute limit now applies only while a command waits to start, and a running turn is awaited to its real end and returns its actual result.
- Fix searches for a word in backticks being refused as a curl call; a refused curl/wget call now points to `read` for what a URL returns and to `browser` for rendered pages.
- Fix a finished background command costing an extra model request: when its exit wakes the session, the exit report now arrives in the first request instead of only after a reply without it.
- Fix commands that write a commit message or other text through a heredoc being refused as a curl or browser call when the text mentions those tools, and close a gap where escaped quotes could hide a second command from the shell guards.
- Fix web fetches of plain-text files, JSON and source code coming back as one collapsed line cut off after about 1.7 KB: textual non-HTML pages now return verbatim, line breaks intact, up to the requested size.
- Fix a message from another session that arrived during a long run waiting until the whole run had finished and then costing a model request of its own: addressed peer messages now reach the model with its next request.
- Fix a check command counting as a passed test of finished work when escaped quotes or a `$(…)` hid a file deletion or overwrite inside it.
- Fix closed workers staying parked and holding their file claims until the app restarted: when a worker's command had been moved to the background, its finish was recorded under the background id and never counted, so the worker's cleanup waited forever. A worker's shell log also no longer drops entries when several of its commands finish at the same moment. If a closed worker's cleanup still cannot be confirmed after ten minutes, the worker inspector now says so with the reason and offers Release, which frees the worker's file claims; nothing is released without your decision.
- Fix session history rewrites failing on Windows with "Access is denied" (os error 5) while another part of the app had the session file open.

## [0.2.271] - 2026-10-04

- Fix Qwen Cloud (Token Plan) rejecting valid `sk-sp-` keys with "Invalid API-key": both Qwen Cloud providers now use the current `qwencloudapi.com` endpoints, and the Token Plan model list matches the plan's supported models (adds DeepSeek V4.1 Flash, GLM-5.3 and GLM-5.2; drops ids the plan no longer serves). The 401 hint now names the expected host and key type.
- Fix "Don't show again" on the search-setup and spec-setup banners only lasting until the next launch; it is now remembered. Update and usage-limit banners say "Hide until restart" instead.
- Workers report back while the main agent is still working: a finished worker's result, a worker's question and a finished batch arrive after its next tool step instead of only when its whole run ends, and worker results, decisions and shell exits that arrive together are handled in one turn instead of one turn each.
- Progress updates no longer cost extra turns: worker milestones, partial results and FYI Intercom messages (`urgency: info`) ride along with the next request instead of waking the session, reopening a finished answer or cutting a `wait_agent` short. Addressed questions and assignments (`urgency: decision`, the default) still wake it, and `wait_agent` reports why it was interrupted.
- Settings › Subagents & parallel tools › Team: choose the model of the team coordinator (`teamCoordinatorModel`; unset uses the session model). The coordinator itself follows in a later release.

## [0.2.270] - 2026-10-03

- Streaming chat text is smoother: it types out character by character with a soft fade at the leading edge, text and its formatting advance together every frame, code turns colour line by line without a plain flash, the answer keeps its place and selection when it finishes, and the view follows new text with a soft glide that lets go the moment you scroll.

- Fix the whole app freezing after a browser-tab error: the native browser view no longer creates or moves its window inside a main-thread command, and a failed view no longer leaves an invisible window over the app that swallows clicks.
- Fix the app growing to 14–18 GB of memory over a long day and then being closed by Windows: busy workspaces with many sessions no longer pile up file-change history in memory.
- Fix session setting changes waiting up to two minutes behind a running turn and then being silently dropped; they now apply immediately.
- Slow workspace, history, system-load and writing-style commands no longer block the window while they run.
- Repeated identical tool calls are now stopped on the parallel tool path too, not only the serial one.
- Harden scripts and remote control: URA shell, HTTP and Python calls get output caps and timeouts, URA HTTP and redirected managed-browser fetches refuse private addresses, and a stalled remote client can no longer grow the server's send queue.
- Fix slow leaks: old index chunk generations are released, an oversized trigram snapshot is rebuilt instead of failing, the tool-error audit database is capped, and cancelling a one-shot shell stops its whole process tree.

## [0.2.269] - 2026-10-03

- Windows release of the 0.2.268 changes (0.2.268 shipped macOS-only).
- Fix the transcript snapping back to an old reading position after jumping to the top loaded more history.
- Codex voice: protocol v2 (the public API-key route) now defaults to `gpt-realtime-2.1`.
- Fix the release gates: the Blender panel settings menu and the iOS remote-tools panel types.

## [0.2.268] - 2026-10-03

- Fix sessions getting stuck on their workers: stopping a worker through its session is refused with a pointer to `close_agent`, and a worker already halted that way is settled by the watchdog instead of keeping its parent "Waiting for workers" forever; a worker's long silent build wakes the idle parent at most once instead of every five minutes; Stop and Esc in the Desktop now also work while the parent is idle and only its workers run; a silent background shell that never started a command now says so instead of "still running, do not relaunch".

- Remote: paired mobile controllers can read and switch a desktop session's agent mode — the controller session exposes the `session.state` field `agentMode` as `activeAgentMode`, and the new remote method `session.mode.set` accepts the five canonical modes (`setAgentMode` on the controller session).
- Remote: paired mobile controllers can load a desktop session's tools on demand exactly like the Desktop Tools dialog — new remote methods `session.tools` and `session.tools.activate` expose the catalog's callable and withheld families (`loadTools` / `activateTool` on the controller session).
- Add Desktop Source Control’s “Reset to remote…” menu action with destructive confirmation, fresh upstream fetch and a recovery snapshot before replacing local branch history and working-tree changes.

- Integrate Blender as an optional first-party Unary plugin: guided add-on ZIP export, explicit companion pairing and mirroring of an existing session, with owner-scoped connections and cleanup on disable. No automatic Blender launch, new agent or Blender-side approval path.

## [0.2.267] - 2026-10-03

- Manual Pull, Pull (Rebase) and Sync in Source Control now autostash your uncommitted changes like VS Code instead of refusing a dirty working tree. If your changes collide with incoming ones, a "Pulled with conflicts" warning names the files and your changes stay safe in the stash; Sync never pushes over that conflict. Agent Git commands are unchanged.

- Fix the Desktop composer looking dead while the agent works: a mid-run message now clears the composer as soon as its chip is queued instead of holding Enter until the backend accepted it behind every earlier mid-run send; a delivery failure still restores the text and shows the error.

## [0.2.266] - 2026-10-03

- Fix sessions with a private native workspace failing after an app restart with "Native session requires coordinator workspace activation": the `nativeWorkspaces` setting now reaches the core before any session is created or resumed, not only after the model picker was opened.

## [0.2.265] - 2026-10-03

- macOS: this build brings the Mac app up to the current release line, including the Blender companion from 0.2.264, and long-running agent shells that are parked in the background are no longer waited on through chained waits.
- macOS: the menu-bar icon shows the Unary logo again instead of a solid white square, and adapts to light and dark menu bars.

## [0.2.264] - 2026-10-02

- Add the standalone unary-blender MCP, installable Blender companion and explicit existing-session mirror; preserve scene selection and shared mesh data, validate tool arguments, and package a concise modeling skill. Live headless modeling, checkpoint and GLB round-trips are checked; native panel and send/stop UI verification remain pending.

- Diagnose accidentally tracked build outputs with grouped evidence and explicit index-only repair guidance while preserving hardlink safety; prevent broad Git-tool staging of newly added build artifacts without blocking intentionally versioned inputs.

- Tighter Desktop security: remote keyboard, mouse and clipboard control only works during a remote session you started; the developer purge tool is unavailable in release builds; saved passwords and keys are never shown inside the app window; files the app writes stay in your workspace, app data or a location you picked; web pages opened in the built-in browser cannot talk to the app.
- Agent teams: task claims, updates and the team task list read the lead's board again, and a corrupted parent link no longer hangs a claim.
- Fix the `context` contract review getting stuck forever after a folder was passed as a changed path; folders are now refused with a clear hint, and stuck entries clear themselves.
- Bound untrusted input everywhere it was open-ended: debug-adapter frames, MCP stdio lines and HTTP bodies, provider event streams; invalid UTF-8 in a stream no longer slows decoding quadratically. MCP sign-in requires its state check, and an empty refresh token from a provider no longer overwrites a good one.
- Shells on Windows stop their whole process tree, and a timed-out command can no longer desynchronize the next one; a malformed Windows path no longer panics, and web fetch blocks more private and reserved address ranges (incl. IPv6 literals).
- Memory and research storage: concurrent writers wait instead of failing, the session search index rebuilds without deleting a database in use, and the research source archive gets a bounded, reference-safe retention.
- OpenAI-family requests can no longer carry a schema `const` from the collaboration tools, and Sonnet 5.5 keeps its default effort.
- The destructive-git guard sees through wrappers (`env`, `sudo`, `nohup`, `time`, …), `bash -c`/`pwsh -Command`/`cmd /c` scripts, command substitution and git aliases; rescue-snapshot notes belong to the session that took them.
- Terminal events from concurrent producers keep their order, and a spool failure never drops queued events.

## [0.2.263] - 2026-10-02

- Desktop shows a crash screen with reload, automatic retry and expandable error details when its display crashes, instead of leaving a blank window.
- Fix updates intermittently failing with "Error opening file for writing": the installer now waits for the old executable's lock to be released before overwriting it.
- Long chats no longer shake while opening: transcript Markdown is parsed before rows mount. Pin and answer rails get their own lanes beside the transcript, the scrollbar keeps a 12px gap, and pin-card actions use the shared pill buttons.
- Sessions are auto-named after the first prompt even when you leave the chat right after sending; the composer shows "Off" for a disabled model speed.
- Global memory only takes knowledge that holds outside the current project: global writes need a stated reason, entries citing repository paths are refused, and background consolidation again requires evidence from two repositories.
- Worker coordination: no duplicate "batch finished" or late stall notices for workers already read or closed, finalizing workers are no longer flagged as silent, failures are no longer lost, and the Workers board drops workers whose results were read.
- Terminal and own shell: Ctrl+C stops only the current foreground command, PTY output is bounded, `curl -L` drops credentials on cross-origin redirects, and `cd`/archive extraction stay inside the workspace view.

## [0.2.262] - 2026-10-02

- Codex Voice speaks only the agent's answers; asking "what is running?" now gets a short status (working, running tools, goals, waiting decision) from a silent, throttled session snapshot.
- Fix Codex Voice handoff delivery: a steering request now receives the real result, a failed delivery no longer misroutes the next answer, and v2 `background_agent` calls with `input`/`query` arguments are no longer left unanswered.
- Codex Voice v1 WebRTC applies its session settings, handshake errors show the server's reason, one long message no longer drops the older call context, and a cancelled call start releases the microphone.

## [0.2.261] - 2026-10-02

- Fix Codex Voice failing with "403 Forbidden" after the WebRTC call: the realtime sideband now joins on the OpenAI realtime host with the same Codex OAuth bearer and account id that created the call, matching the Codex CLI.

- Bundle independent tool calls with known arguments in one model response while preserving result dependencies, permissions and file claims; improve batching diagnostics and edit/contract-review recovery guidance.

- Let explicit chat deletion close its actor even when historical shell cleanup cannot be confirmed; ordinary Stop, archive and eviction retain their cleanup safeguards.

## [0.2.260] - 2026-10-01

- Add clear-all controls above chat bookmarks and unread mid-run answers without deleting conversation contents; keep incoming agent messages in Intercom instead of the visible transcript.
- Add MiniMax-M3.1-Flash-Preview to the MiniMax catalog and shipped provider pack.

- Keep normal sessions under explicit human control: every agent-created session needs a one-call human approval, including in Yolo; persistent grants, permission hooks and peer-agent approvals cannot bypass it. Autonomous delegation uses subagents instead.
- Give explicitly continued peer sessions a distinct automatic title while retaining their source conversation and workspace basis; prevent repeated Sidebar creation clicks and stale optimistic editor snapshots from duplicating sessions.

## [0.2.259] - 2026-10-01

- Avoid duplicate research while an Explorer is working: the main agent sends follow-up questions to the same worker and continues independent work, then verifies the returned findings without repeating the search.

## [0.2.258] - 2026-10-01

- Replace the copied upstream base prompt with Unary's own: requests are sized (question → answer and offer, small → fix directly, medium → one goal with parallel explorers, large → teams per area reporting goals and tasks back), follow-ups go to the worker already on that topic, and every size ends with its own clean final report. Model-facing text no longer names other agent products.

- Offer a change as Yes/No buttons: after answering a question the agent asks "Shall I fix it?" with `request_user_input` `kind: "offer"`, which always waits for the click (also in Yolo) and returns the decision to the model.

- Stop waits from holding the lead: `wait_agent` defaults to 15 s (max 120 s, 0 = status check, `mode` first/all) and reports every worker's status on timeout; a worker that succeeds while the lead is mid-turn is no longer silently dropped from the wake; `shell_watch` `wait` and `wait_for` cap a single wait at 120 s.

- Restore Astra's Off → Fast → Ultrafast speed cycle in the canonical chat model picker, including capability forwarding and session-state refresh; other models retain the two-state Fast control.

## [0.2.257] - 2026-10-01

- Free warm build lanes the moment their holder is gone: a repository build lane is busy only while the process that leased it is alive (PID plus start time, its run tree contained by a kill-on-close job object), so a crashed session no longer pins both lanes for a week; a reclaimed lane keeps its warm output and is fully refreshed on the next use, and the sweep now releases a dead session's run views as soon as their own lease expires, never touching an isolated session's tree.

- Harden manually isolated session trees: deleting a session that still owns an unfolded tree is refused, a new confirmed Discard isolated tree ends the isolation without publishing, trees whose session is gone or whose record is damaged are listed in Source Control for an explicit Fold back or Discard, the tree's lease renews on activity and never lets a sweep delete it, and builds and tests of an isolated session now run on the warm build lanes from the tree's content and write their changes back into the tree.

- Make isolated run views faster: the write-back baseline now reuses the content stamps the materializer just verified instead of re-hashing the tree, a Windows change journal (`ReadDirectoryChangesW`) narrows the after-run walk to the paths a command actually touched, falling back to the full content walk on overflow or any gap, and output-evidence probes are memoized per walk.

- Faster, fewer-turn edits (Codex-style): `patch` hunks use the `replace` matcher and repair drifted files; an exact, unique whole-line match edits a file not read in this session; refusals return the current text verbatim and count it as read; `write` may overwrite an unread file; `grep` can return the enclosing declaration's current source; `read` `paths[]` entries take their own range or symbol; `replace_files` edits several files in one all-or-nothing call; `show` returns the written regions; `patch` and batch reads no longer fence the tool scheduler, and the config-write and focus gates now see a patch's files.

- Default worker roles to 12 parallel Editors and 23 Explorers (per-role limit raised to 32, explicit session-worker limit raised to 64), and give the role settings' number and text fields the same light hover/focus border as the Select.

## [0.2.256] - 2026-09-30

- Prevent unbounded wrapped Cargo/CMake builds by inheriting commit-based job defaults; retain commit-pressure evidence, correct WebView2 invalid-state diagnostics, restore missed tool results with their workspace, and allow Desktop UI recreation through a native tray double-click without aborting backend sessions.

- Keep recognized read-only PowerShell pipelines, filters and foreach/git-log queries out of build-output reservations, including null redirects and literal PowerShell command envelopes. Account for all trailing Command arguments and keep encoded/stdin bodies conservative. Preserve admission for nested writers and builds; mark truncated holder commands visibly and describe conservative holders as live runs.

- Coordinate matching live builds across sessions and processes through immediate shared observations and Intercom lifecycle notices. Keep different commands concurrent with disjoint outputs, preserve output leases, and avoid automatic background build-lock waiting.

- Add a default-off Session Map switch in Desktop Memory & Dream settings, with bounded message/tool/file links on session reads and live edit hints from the existing journal, without extra model calls.

- Prefer recent messages, reasoning and tool results across capture and compaction; preserve an original request anchor, prioritize critical instructions and errors, and report records evicted by segment caps.

- Improve the Rust terminal's Bash compatibility, command arguments, scoped timeouts and truthful file permissions; keep interactive background jobs from blocking prompts, repair awk/jq and other utility defects, share bounded HTTP access across execution paths, synchronize PTY screen/resize state, default new Desktop tabs to Unary Shell and preserve their engine on reload.

- Edit pasted and uploaded chat images through stable session attachment references without requiring another upload or workspace input file. Open sent or unsent attachments in the shared mask editor, select an available image model for that edit and enter its prompt; preserve new source bytes separately from vision fitting and previews.

- Adopt Codex CLI 0.149.1 worker handoffs and working/final-answer instructions; retire Quick/Strict presets while preserving Unary modes, file claims and private isolation records. Default file isolation to OFF and route compatible claimed work to its active owner through Intercom.
- Pin completed answers until read, show their questions together, preserve manual bookmarks and illuminate pins in blue.
- Make Intercom retries idempotent, link replies and record recipient delivery separately from explicit acceptance, blockers and completion reports. Recover addressed messages across harness processes and restarts without waking idle agents; show receipts in Desktop and route active cross-process file claims through the existing ownership path.

- Show ordinary folder paths in Assistant project headers, tooltips and project settings instead of Windows extended-length prefixes; preserve drive roots and unchanged operational paths.
- Settings: present Research and Deep Research as depths of one tool, with shared web sources and a Perplexity API key explicitly covering research and web search.
- Repair Research Studio invocation and saved defaults; continue gathering into a bound Desktop synthesis session with cancellation, visible evidence pages and resumable failures. Preserve provider comparisons and open questions in generated reports, enforce job finalization, and invalidate old evidence-review claims after document edits.
- Finish the Editor/Explorer worker refactor with preserved legacy settings, role context budgets, and durable compaction lifetime enforcement.

## [0.2.255] - 2026-09-30

- Unify Desktop chat cards with existing theme colors; add goal details and subagent navigation, keep hover previews visible outside the chat bounds, and open HTML, SVG and explicit file cards in the existing preview dialogs.

- Repair Memory consolidation for retained histories above 128 sources and old generation opt-outs; expose scoped learned-handbook discovery and section reads. Claim extraction jobs before inference with retry backoff, validate output JSON strictly, and resume partial candidate saves without losing their errors or entry identity.

- Global Memory cards show durable original files and live file/folder references, including previews, versions and unavailable links. Clear memory intent stays automatic; uncertain suggestions appear in the current chat's Memory candidates tab for review or dismissal.

- Bound complete provider requests after tool-result projection, preserve immutable conflict evidence for paging/search, and park deterministic context failures until an explicit recovery transition.

## [0.2.254] - 2026-09-30

- macOS: this build brings the Mac app up to the current release line, including GPT-6.1 Sol with Codex Fast mode and deep research that keeps its gathered findings after a timeout.

## [0.2.253] - 2026-09-29

### Added

- GPT-6.1 Sol (`gpt-6.1-sol`) for OpenAI API keys and ChatGPT/Codex sign-in: 1.05M context, 128K output, effort low..max, $2/$10 per 1M tokens. Codex Fast mode is available for it.

### Fixed

- Deep research: a host timeout no longer discards gathered findings. Start returns immediately and gathers in the background, every engine's sources are saved as soon as it finishes, and `deep_research_rerun` with `resume: true` continues an interrupted job with its recorded sources.

## [0.2.252] - 2026-09-29

### Fixed

- Queued background builds no longer starve: waiting launches take a FIFO ticket per output root and re-check every 25-250 ms, and a foreground launch gets a 750 ms grace on a busy admission gate instead of an instant false refusal. `wait_for` gains `until: output_root_free` (`outputRoot`, default `target`), a fast, non-reserving check that a build has released its output root.
- Classified builds and tests use fresh, cancellable source preparation rather than inheriting the ordinary command's ten-second capture deadline; no stale-source fallback is permitted.
- Worker coordination survives policy refreshes; missing live telemetry no longer hides recoverable status. Retained terminal handles are distinguished from running attempts, and status/send/resume share one derived control view instead of contradictory action advice.

### Changed

- Small, tightly coupled changes default to direct parent implementation. Coding workers reuse established evidence and move to a scoped change once the relevant uncertainty is resolved, without arbitrary read or time limits. Terminal worker observations retain explicit coverage limits for later effectiveness comparisons.

## [0.2.251] - 2026-09-29

- macOS: this build brings the Mac app up to the current release line — the expanded terminal utilities and the interactive Unary Shell from 0.2.250, plus the fixes for opening saved Code sessions and recovering the Assistant and Chat lists from 0.2.249.

## [0.2.250] - 2026-09-29

### Added

- The agent's own Rust shell takes over far more of the terminal. New in-process utilities — `jq`, `awk`, `curl` (on the web-fetch network policy, fail-closed without it), `tar`/`gzip`/`zip`, `patch`, checksum tools, `comm`/`join`/`split`/`column`/`expr`/`od`/`xxd`, `tree`, `nproc`, `timeout`, `nohup` — plus `pushd`/`popd`/`dirs`, the `time` keyword and `kill` for the shell's own jobs. An interactive REPL ships as `unary-sh`: the agent's interactive terminal and a new "Unary Shell" option in the Desktop terminal use it instead of Git Bash. Foreground commands reuse an unchanged source snapshot (explicit fingerprint, never a timer) instead of re-observing the tree on every call, full-screen terminal programs are readable as rendered screen snapshots, and `ls`/`find`/`stat`/`du`/`wc`/`diff` gain a `--json` output mode.

### Fixed

- Terminals no longer fail at random on Windows with "Open a chat session for this workspace before starting a terminal". The leftover ownership gate is gone; terminals have no file coordination.

## [0.2.249] - 2026-09-28

### Fixed

- Opening a Code session retries a snapshot invalidated by concurrent activity instead of silently leaving an empty conversation. Repeated refusal now shows a load error.
- Assistant and Local Chat lists recover saved conversations from persisted session kinds when browser list metadata is missing. Known cloud bindings remain excluded from the Local list.

## [0.2.248] - 2026-09-28

### Fixed

- In-app updates no longer fail after a release that was published for only one operating system: each platform is now offered its own newest build. Before, after a macOS-only release every Windows update check failed, and after a Windows-only release every Mac update check did. The fix is on the update server, so installed versions benefit without updating first.

## [0.2.246] - 2026-09-28

### Fixed

- macOS: started from Finder or the Dock, the app now picks up your login shell's `PATH` and a UTF-8 locale, so the agent's shell finds Homebrew and rustup tools like `cargo`, and interactive terminals no longer mangle non-ASCII characters.
- macOS: file isolation no longer refuses private edits when the repository path runs through a symlinked folder (such as `/var` → `/private/var`), and per-turn git diffs and session worktree status are no longer empty on macOS and Linux.
- `trap -p` and a bare `trap` in the agent shell list the installed traps again.
- The Desktop chat's worker list no longer shows stopped workers as running. Closing a worker, a watchdog stop and a verifier verdict now end its row immediately, a worker waiting for its verifier shows as Verifying instead of running, and while any row is still live the list re-checks the saved worker state every 15 seconds.
- File isolation no longer fills the system drive. Each session's and worker's private build output, temp files and caches now live next to the repository, in `<parent folder>/.unary-runtime/<repository>/`, instead of in `~/.unary/workspace-runtime` on the system drive, and they are removed about a minute after the session ends instead of after seven days. Leftovers of deleted repositories and test runs, and the build lanes of ended sessions, are removed as well.

### Changed

- Builds under file isolation are warm per repository: all sessions and workers of a checkout share a small pool of build lanes, so a new session can continue from an earlier build instead of compiling everything again. A lane nobody used for seven days is removed.

## [0.2.245] - 2026-09-27

### Added

- An offline harness coordination report and focused regression runner retain structured findings, full test logs, native exit statuses and source fingerprints. Missing evidence, failed delivery and empty test filters are not treated as successful recovery or passing tests.

### Fixed

- Blocked-only runs now get bounded prerequisite triage and a final assessment using existing coordination tools; the final notice reports declared blockers rather than asserting that no independent work is possible.
- Files explicitly released by another session no longer keep Git operations blocked. Files still owned by another session remain protected; missing ownership information is not mistaken for a release.
- Repository build-history guidance now distinguishes process presence from session reachability and explains idle-session follow-up; accepted delivery is not reported as completed work, and a past failure observer is not assumed to own the broken file.
- Archive and restore now include recovery and evidence journals in the session move, preflight destination conflicts, and attempt rollback if a later move fails. Desktop archive/restore failures are shown instead of only logged, and successful changes request a fresh shared session list; an older archive-list response can no longer overwrite a newer one.
- The Index panel's Stop no longer leaves a permanent progress bar when no refresh was running; it says so and clears the stale running state.
- The Index Scope dialog shows the workspace file tree instead of "Workspace tree unavailable", and marks gitignored paths as excluded.
- Radio buttons and checkboxes in the Desktop (for example the automation Schedule list) render correctly again instead of as padded boxes.
- Source capture no longer deadlocks on base entries that were explicitly excluded.
- New and moved session transcripts reserve their ownership before they are published, so two sessions can no longer claim the same transcript.
- macOS release builds compile the workspace-isolation engine's Unix paths and the selected-browser launcher correctly.

### Changed

- Sync Changes is now a plain merge-then-push. It never stages, stashes, rebases or force-pushes; when the merge conflicts, nothing is pushed and you can hand the conflict to a session to resolve and finish the sync.
- A foreground build no longer queues behind a busy shared build output: admission refuses immediately, naming the holder and the remedies (wait on that run or build into a separate output directory), and the cross-process admission gate is tried once instead of sleeping up to ten seconds. Background launches keep the bounded 30 s queue.
- System, worker and delegation prompts now state that a blocked build gates only its dependents: independent implementation continues while a build lock is held, and build-verdict workers are ordered behind implementing workers via `dependsOnHookIds`.

## [0.2.244] - 2026-09-27

### Fixed

- The Git graph no longer fails on Windows with "file name or extension is too long" in repositories with many checkpoints: checkpoint and selected revisions are streamed to `git log` on stdin instead of the command line.
- Saving Settings no longer reverts per-session indexing overrides made in the Assistant meanwhile; the save merges only the dialog's own edits. Saving plugin settings no longer resets the Memory thinking level, and deleted chats drop their indexing override.
- Image region editing works from file previews even when no chat composer is open: the edit is submitted to the image's own workspace, creating a session there when needed.
- Release instructions now push the lightweight version tag explicitly so the build workflow actually starts.

### Changed

- The system prompt no longer embeds a repository map; it points to on-demand index orientation instead, which keeps the prompt smaller and stable.
- Workers get up to five seconds (was 200 ms) to receive their index orientation packet at startup.

## [0.2.243] - 2026-09-27

### Fixed

- With file isolation on, workers that declare overlapping files no longer wait for each other when both write private versions; a queued worker is re-checked and admitted as soon as its blocker becomes private, and a blocker without a session still counts as a shared writer.
- A command run in a private view can no longer leave it with an absolute `cd` into the shared checkout: the target is remapped onto the view, and the command fails instead of escaping when the view lacks the directory.
- Private run views contain every tracked source again; a tracked crate that had drifted in Git was missing, so builds in the view failed to load the workspace.
- Revoked OAuth tokens are no longer reused: a leased token is resolved from the account's current credentials.
- Workers that stopped are no longer shown as running: every worker state is reconciled before the list is capped, and parked workers are distinguished from executing ones.
- Grok Responses WebSocket failures report the real parser error instead of "response consumer closed", and transport faults are no longer shown as a successful stop.
- A source-capture refusal names every unresolved path with recovery evidence, not only the first one.
- Waiting on a shell that had already exited no longer replays its exit as a fresh turn after the run finished.
- Literal process/status queries no longer become build-output writers merely because they run through PowerShell. Unresolved executable script bodies remain conservatively coordinated.
- Private build capture and write-back separate declared build outputs/environments from real source and local inputs. Legacy output entries are omitted only from the run view, without deleting catalog history or held edits.

### Added

- `wait_for` can wait for another session's work: `until: "commit"` fires when HEAD advances (optionally only when a watched path changes) and `until: "goal_done"` fires when a goal in any session closes. A goal marked blocked without a watcher now says that nothing will resume it and names the call that would.
- Cloud chats stay live on every device through the account's sync connection, and Assistant and Chats rows show the same lifecycle states as the Code sidebar.

### Changed

- Work requests, not every message, open Goals: intake and completion are bound to the active user request.
- Registered private-file builders use bounded reusable isolated lanes even after publishing their edits. Compatible lanes keep stable paths and declared build/environment directories through source refresh; occupied lanes are never concurrently overwritten.
- Project build profiles can declare exact commands, mutable roots, local inputs, lane counts, CPU-job defaults and runtime files. Runs reserve CPU-job tokens, keep foreground/background environment parity, and emit source/build receipts; explicitly declared runtime files can be staged outside mutable compiler output after a successful run. Opaque wrapper job overrides remain visible limitations rather than claimed enforcement.

## [0.2.242] - 2026-09-27

### Fixed

- Session and Repository Info always show the HTML mocks group in Sources, open by default together with Web. Mocks read through a shell command or a search path are recorded too, so the group is no longer empty when a mock was only opened that way. The unplanned "Eigener Dialog" button is gone.

### Changed

- The Session isolation dialog follows the decided design: two independent choices — file isolation On/Off and Shared checkout or Isolated native tree — applied with one Apply button. The Git worktree option is removed from this dialog. A file-isolation choice made here stays local to the session, survives the next turn and a resume, and the automatic escalation after repeated conflicts is no longer undone by the settings on the next turn.

## [0.2.241] - 2026-09-26

### Fixed

- A worker that stopped without reporting a final status no longer stays "running" in the chat's worker list. While any worker reads as running, the list re-checks the session's actual workers every 20 seconds; a row that no live worker backs any more for over 15 seconds is marked interrupted instead of spinning forever.
- Stopping a foreground shell command that was still preparing no longer leaves its terminal reading "running" forever. A cancel that lands before any process launched now settles the operation at once and reports that no command ran and no write-back is owed.
- Sessions in a repository with a submodule or nested checkout start again. A tracked path that is a directory on disk was walked into and every session was refused with "new file requires explicit source authorization"; it is now treated as one local entry, as Git does.
- A broken session can be deleted or archived again: removal now waits out the failing re-open the Desktop triggers instead of giving up after 4 seconds.
- Isolation checks recognize the same checkout root however its path is spelled, and same-session private edits are serialized from admission to commit.
- Native shell interpreters keep their non-interactive defaults.
- The Desktop chat no longer shows the answer-guard's hidden `<!-- answers: … -->` marker in replies or when copying them.

### Added

- Conversation indexing controls: Settings › Index has global defaults for chats and Assistant sessions, and each Assistant group can override them (Default / On / Off). Off pauses future indexing only; already indexed content stays searchable. A failed save is shown instead of leaving a wrong label.
- Repository Info gains section tabs and lists the sources of every session of the repository, HTML mocks first, each openable.

### Changed

- Add project is one source-first dialog: choose Local folder or Clone a repository, and the clone form takes the URL (with a GitHub picker) and target folder and previews the resulting path instead of chaining three prompts.

## [0.2.240] - 2026-09-26

### Fixed

- Reverting a file to exactly its original content no longer wedges `edit finish`. A version with no remaining difference is left out of the delivery plan, but the check that runs after publication still demanded to find it published — unsatisfiable once another session had moved the file on, so the same refusal repeated forever, and the second failure tripped the per-call circuit breaker on the very call the message told you to repeat.
- The per-call circuit breaker no longer keeps a ready private edit unit blocked after the unit's state has changed. `edit finish` carries no argument that could show the breaker its precondition was gone, so an edit that changed the unit was invisible to it; a changing edit result now dates the recorded failures to the unit state they happened under and clears exactly that tool's debt. Inspections alone clear nothing, and identical calls on an unchanged unit still trip the breaker.
- Attaching the terminal to an agent session works again. Every session except the one that was focused when the terminal opened was refused — so New Session always failed, and Active Session failed after switching chats — and the refusal was swallowed before it could be shown. A refused attach now reports itself.
- PowerShell commands run on machines whose PowerShell 7 comes from the Microsoft Store. The interpreter selected there is a Store alias that Windows refuses to open as a file, so every owned PowerShell launch was refused; the alias is now resolved to the program it names. The rule never to fall back to the older Windows PowerShell is unchanged.

### Changed

- Preparing a new session shows its progress where the logo normally sits — centered in the middle area together with any queued message, instead of pinned to the top.

### Added

- A handoff to a worker can carry a picture from the conversation. Name the message the picture came with and the real image travels with the task, so a worker asked to match a mock sees the mock itself rather than a description of it. A selection that matches no message, or names an image that message does not carry, fails with a clear reason instead of quietly sending the task without it.

## [0.2.239] - 2026-09-26

### Fixed

- Stale-read replacements now verify the touched lines against bounded, hash-matched read evidence instead of trusting a still-unique partial anchor. A sibling changing `tally=` to `tally=A` can no longer make another writer silently produce `tally=CA`; missing evidence requests a re-read and multi-edit batches refuse atomically.
- Parallel edit publication detects a moving parent head while observing checkout preimages and re-prepares the candidate. Recovery can recognize bytes from a still-pending accepted sibling journal without accepting arbitrary historical snapshots or unexplained external rollbacks.
- The private-hold guard now recognizes durable batch admission and persistence, with negative tests for missing safety checks. Added retained build-slot regressions for exclusive reuse, source refresh and abandoned-slot quarantine.

## [0.2.238] - 2026-09-26

### Fixed

- Manual full-tree isolation works: "Isolate session" no longer always defers with "native process/file leases have not drained" (it demanded a retired generation, which a live session never has), and Fold back, Status preview and conflict Resolve no longer refuse themselves with "publication is busy" by holding the checkout locks their own delivery needs.
- Several sessions or workers finishing edits on the same file at once all land: `edit finish` waits briefly for a concurrent publication instead of failing once with "checkout publication is busy", a sibling delivery landing mid-finish is re-prepared instead of refused, and a writer's replayed edit chain survives reanchoring so four concurrent appenders to one file all keep their lines.
- Two writers editing the SAME line no longer merge silently into a line nobody wrote (`tally=A` + `tally=C` became `tally=CA`); the overlap is reported as a conflict.
- A publication-busy refusal no longer trips the per-call circuit breaker and strands a ready private edit unit.
- A waiter that re-parks behind a file claim reports its true queue position instead of counting the writers behind it.
- Private edits no longer refuse with "private edits must use their pinned publication root" when the same checkout arrives once as `\\?\C:\...` and once as `C:\...`.

- Existing sessions survive tool definition changes between builds: a stale on-demand activation record no longer blocks the whole request with "schema changed for …". Unchanged repeated activations keep the earliest receipt position (stable prefix); a definition drift promotes the tool to the request root with its current definition on native dialects (Pi-style checkpoint: historical positional additions omitted, ordinary calls/results untouched), while the dispatcher treatment anchors at the reconciliation refresh — one fresh hidden schema-load receipt, persisted before encoding, idempotent across restart, retry and undo. Removed tools and revoked grants still never replay, and the dispatcher still fails closed without a clean receipt. Deterministic local construction blocks ("provider call blocked: …") no longer consume the empty-error retry budget when surfaced as error messages.

## [0.2.237] - 2026-09-25

### Fixed

- A session can commit the files its own worker wrote: the git staging guard no longer treats a parent's workers (or their workers) as another live session, while sibling and unrelated sessions still block.
- A worker whose retry replaced it, or whose session was closed while its process was already gone, now releases its file claims instead of blocking commits until the app restarts.
- `session` `send_input` to a busy worker no longer fails with "Zugriff verweigert (os error 5)", and it can no longer drop the worker's newest transcript entries.
- Private file edits no longer fail with os error 32 when a build holds an ignored artifact such as `build/debug/.ninja_log` open; untracked and ignored files are never opened by the capture.

### Added

- Repository Info shows every session's open and in-progress goals, including sessions that are not loaded, and can start an Overseer session that closes orphaned goals and supervises completion.

## [0.2.236] - 2026-09-25

### Fixed

- Background shells whose result the agent already read (wait, output or kill) no longer replay their exit as a separate turn after the run ends; any exit notices still pending arrive together in one turn.
- A "workspace busy" refusal no longer counts toward the repeated-failure breaker, so the retry that would succeed once the other session finishes is no longer blocked.
- A file claim stops blocking only when its owning process is proven gone; a claim whose owner cannot be checked is kept and reported as unverified instead of being dropped.
- Parallel private file edits publish per writer and in durable batches, and a terminated worker releases its claims.

### Added

- Sessions can archive or restore another session (`session` action `archive`); the sidebar updates when that happens.

## [0.2.235] - 2026-09-24

### Fixed

- Ordinary shell commands no longer queue behind another session's long workspace mutation: they wait briefly, then run on the session's last captured source and say so, instead of failing with "workspace busy" while other sessions wait to finish their edits. Explicit verification still requires the current source.

## [0.2.234] - 2026-09-24

### Fixed

- Foreground shell timeouts include preparation and completion work, with bounded cancellation cleanup.
- The Windows Bash tool discovers Git Bash utilities such as awk when they are absent from the app's inherited PATH.
- Long interactive prompts retain their logical line across terminal wraps and cursor updates instead of acquiring extra line breaks and padding.

## [0.2.233] - 2026-09-24

### Fixed

- Background Cargo startup uses the cancellable shell-start lifecycle instead of the synchronous tool watchdog.
- Shell startup cleanup preserves cancellation and timeout error classifications.

## [0.2.232] - 2026-09-24

### Fixed

- PowerShell commands flush formatted object output before returning their exit status, including PTY capture.
- Cancelled shell preparation cannot launch a job later without returning its handle; background and interactive startup retain cancellation ownership through handle delivery.
- Background Cargo commands preserve shell argument boundaries and target directories containing spaces or apostrophes.
- Edits retain their successful mutation receipt when post-edit verification cannot prepare or launch; retry guidance distinguishes invalid inputs from changing runtime state.
- Contract-review refusals identify the source whose revision or complete read needs refreshing.

## [0.2.231] - 2026-09-24

### Fixed

- Shell preparation errors retain their underlying cause and failing path instead of showing only a source-capture label.
- Commands can capture a checkout after the last previously captured file in a source directory was deleted, without rejecting the remaining empty directory.
- PTY output capture answers Windows' cursor-position handshake so commands start and return output instead of stalling.

## [0.2.230] - 2026-09-24

### Fixed

- Commands validate ownership of their leased execution view, and mixed build/edit command chains use the same isolated workspace view.
- Managed command write-back preserves concurrent file edits and deletions atomically; repeated baseline and foreign-claim conflicts escalate to private file mode.
- Background commands and interactive terminals report completion only after their file changes have finished writing back.
- Windows commands wait for child processes in their job to finish, so a parent exiting early cannot discard descendant output or file changes.
- Ordinary commands read current shared source files when isolation is off, instead of reusing a stale captured version.
- Isolated Cargo builds retain their dedicated output directory across command views, and Windows batch launchers work through the managed shell path.

## [0.2.229] - 2026-09-24

### Fixed

- Worker budget parking no longer queues a replacement steering message that restarts the cancelled run while resume waits for the actor.
- Interactive agent terminals answer the ConPTY cursor handshake instead of staying alive with zero output after input.
- Own-shell command lookup preserves platform PATH entries; piped wc output contains the bare count instead of an unsolicited stdin label.

## [0.2.228] - 2026-09-24

### Added

- Der Intercom-Tab ist im gehosteten Desktop-Chat eingebunden.
- Worker erhalten bei automatischer Orientierung ein aufgabenbezogenes Index-Paket, sofern der übergebene Kontext noch kein passendes, aktuelles Index-Ergebnis enthält.

### Fixed

- Befehle einer Sitzung mit unveröffentlichten privaten Dateiversionen laufen in einer materialisierten Arbeitsansicht statt über das Abfangen einzelner Kindprozesse. Builds wie `pixi run build` und Desktop-Terminals brechen dadurch nicht mehr still ab. Das Zurückschreiben vergleicht Inhalte per Hash, erkennt Konflikte mit zwischenzeitlichen Sitzungsänderungen und behält die Ansicht bei unvollständiger Übernahme.
- Worktree erzwungen entfernen: fehlgeschlagene Zeilen blockieren den Bereich nicht mehr, und die Übersteuerung greift tatsächlich.
- Steuerungsnachrichten halten einen laufenden Lauf nicht mehr an; das Modell liest sie im nächsten Schritt.

## [0.2.227] - 2026-09-24

### Added

- On-Demand-Werkzeuge behalten den eigenen Aktivierungsschritt und laden Schemas an dessen Position im Verlauf. Provider- und Modell-Einstellungen wählen die native Schemaübertragung oder den konstanten Dispatcher sowie HTTP oder unterstützte Responses-WebSockets.
- Geänderte Laufzeitbeobachtungen werden dauerhaft am aktuellen Gesprächsende gespeichert; unveränderte Beobachtungen erzeugen keine weitere Nachricht.
- Aufgabenbezogene Worker-Namen und gemeinsame Intercom-Kanäle je Sitzung und Projekt mit adressierten Nachrichten, Ziel-/Aufgabenübersicht, Verantwortlichen und dem wirksamen vorhandenen Datei-Isolationsmodus. Delegationen reservieren die Aufgabenverantwortung vor dem Worker-Start.

### Fixed

- Desktop-Pending-Nachrichten werden pro Sitzung in Eingabereihenfolge zugestellt und bearbeitet. Chip-Aktionen erhalten fremde und maschinelle Queue-Einträge und können einen bereits verbrauchten Steer nicht erneut einreihen.
- Korrekte NT-Dateiöffnungswerte im Shell-Broker; private Datei-Claims werden erst nach erfolgreicher Anlage ihrer Speicherdatei sichtbar.
- Leere Projekte erhalten eine Snapshot-Historie, damit die ersten privaten Dateiänderungen veröffentlicht werden können; Wiederholungen duplizieren diese Historie nicht.
- Cargo verwendet im eigenen Terminal standardmäßig den verwalteten Ausgabeordner außerhalb der Quelldateien; bestehende Verzeichnisse werden bei Erstellungsproben korrekt erkannt.
- Subagent-Brief-Felder erreichen den Handler; Researcher behalten die lesende Websuche, und aktive Sitzungen mit eigenem Speicherort behalten ihre Workspace-Zuordnung.

- Eigene Shell standardmäßig für verwaltete Windows-Bash-Aufrufe; PowerShell, Hintergrundjobs und interaktive native ConPTY-Terminals verwenden dieselbe Dateikoordination. Desktop-Terminals bleiben an den beim Öffnen ausgewählten Chat gebunden und unterstützen Resize, Eingabe und Prozessbaum-Abbruch.
- Die Regelprüfung erkennt mehrteilige Lock-Dateinamen und Verzeichnisse ohne Dateiendung in echten Quelltextliteralen.
- Externe Programme verwenden unter Windows x64 die vorhandenen Datei-Claims und privaten Versionen; bei ausgeschalteter Sandbox wird kein AppContainer erzwungen.
- Gewöhnliche und kanonische Windows-Pfade verwenden bei Shell-Umleitungen dieselbe Session-Dateischicht.

## [0.2.226] - 2026-09-23

### Fixed

- Statusfragen und Tempo-Hinweise setzen den Schutz gegen ergebnislose Untersuchungsschleifen nicht mehr zurück. Unbeantwortete Nachrichten lösen nach wiederholten stillen Tool-Runden eine sichtbare Antwortphase aus; das gilt auch für spätere Nachrichten und Textblöcke.
- Neue Steuerungsnachrichten unterbrechen die laufende Modellgenerierung und werden im selben Lauf zugestellt, ohne einen Benutzer-Stop vorzutäuschen.
- Erledigte Goals bleiben bei einer Fortsetzung erhalten, solange noch offene Arbeit vorhanden ist. Die explizite Aufräumaktion bleibt verfügbar.
- Der Release enthält den vorhandenen Fix für kollidierende Idempotenzschlüssel beim Veröffentlichen privater Dateiänderungen sowie manuelle Kompaktierung während eines laufenden Turns.

### Added

- Zoom in Dialogen: HTML-Vorschau, SVG, Bilder, Markdown und Text zoomen jetzt gleich mit Ctrl+Mausrad (Trackpad-Pinch ebenfalls), Ctrl++/Ctrl+- und Ctrl+0; der Inhaltspunkt unter dem Mauszeiger bleibt dabei stehen. Der Zoompegel wird pro Dialogfenster gemerkt und als Chip im Dialogkopf angezeigt (Klick setzt auf 100 % zurück). Bisher war in keinem Dialog zoomen möglich — das WebView bringt kein Browser-Zoom mit. Der Bildbetrachter behält sein eigenes, bildverankertes Zoom.

## [0.2.225] - 2026-09-23

### Fixed

- Die Plan-Karte mit Accept/Revise/Decline erscheint nur noch für echte Pläne aus dem Plan-Tool. Bisher reichte eine Überschrift mit dem Wort „Plan“ in einer normalen Antwort (etwa „Why the Token Plan failed“), und die Karte tauchte auch im Yolo-Modus auf, obwohl es keinen Plan gab.
- Verwaiste Index-Dateien werden jetzt automatisch entfernt: liegengebliebene Schreibdateien abgebrochener Indexläufe, überholte Sicherungen und aufbewahrte Kopien beschädigter Indexe (nach 24 Stunden). Bisher räumte nur ein manueller Reset sie weg, und sie konnten Hunderte MB belegen. Ein nie baubarer Index für das Home-Verzeichnis wird beim Sessionstart gelöscht statt nur gemeldet.

## [0.2.224] - 2026-09-23

### Fixed

- Der systemweite Snip-Shortcut funktioniert jetzt auch, wenn Unary minimiert, im Tray oder verdeckt ist; ein solcher Druck blockiert spätere Snips nicht mehr.

## [0.2.223] - 2026-09-23

### Added

- Eigene eingebaute Shell für das bash-Tool (opt-in über `UNARY_OWN_SHELL=1`): bash-kompatibel mit Pipes, Umleitungen, Schleifen, Funktionen und den üblichen Werkzeugen (`grep`, `sed -i`, `find`, `cp`, `mv` …) im Prozess. Dateiänderungen laufen vor dem Schreiben durch die Session-Dateischicht; externe Programme werden im geschützten Modus abgelehnt statt ungeschützt ausgeführt.

### Fixed

- Der Aufrufgraph des Index verknüpft Namen nur noch innerhalb derselben Sprachfamilie und nie mit fremdem Vendor-Code: TypeScript-`Record<…>` und jedes `.iter()` galten bisher als Aufrufe einer einzigen Rust-Datei und verfälschten `callers_of`, `impact`, Hubs und die Repository-Karte. Wirkt ab dem nächsten Reindex.
- Die Repository-Karte wird nur noch an Zeilengrenzen gekürzt und sagt das im Text, Vendor-Code steht zuletzt, die Signale `c`/`commits`/`touched` sind erklärt, und `repo_map` lädt nicht mehr den ganzen Index, nur um die Karte zu lesen.
- Worker bekommen ihr Index-Evidence-Paket wieder: es nutzt den warmen Index der Session statt jedes Mal den kompletten Index neu zu laden (5–7,5 s gegen eine 200-ms-Frist), liefert fertige Abschnitte auch bei Fristablauf und protokolliert jedes Ergebnis.
- Das Index-Modell im Cache wird bei jedem Start gegen seinen SHA-256-Stempel geprüft: eine abgeschnittene oder beschädigte Datei wird gelöscht und neu geladen, statt den Modell-Prozess bei jedem Start abstürzen zu lassen. Der Cache liegt pro Modell-Revision, sodass eine neue Revision nie die alten Gewichte weiterverwendet; abgebrochene Downloads und alte Revisionen werden aufgeräumt.
- Bash-Befehle kommen unter Windows byte-genau an: Git Bash hat beim Einlesen der Kommandozeile `\\` zu `\` gemacht, auch in einfachen Anführungszeichen und Heredocs, sodass Python-Escapes, Regex und UNC-Pfade still verändert wurden. Befehle, die das bisher mit zusätzlichen Backslashes ausgeglichen haben, brauchen diese nicht mehr.

## [0.2.222] - 2026-09-23

### Added

- Xiaomi MiMo-V2.6-Pro-UltraSpeed und MiMo-V2.6-Flash im Modellkatalog.
- GPT-6 Sol und GPT-6 Luna für OpenAI (API) und OpenAI Codex, mit Reasoning-Stufe `max`.

### Fixed

- Suchen und Abfragen, die Build-Werkzeuge nur erwähnen (`rg`, `grep`, `echo`), werden im Repository-Ledger nicht mehr als fehlgeschlagene Builds ("Build RED") gemeldet; bestehende Fehleinträge verschwinden.
- Angeforderte Audits und Fragen werden nicht mehr fälschlich durch Schleifenwächter gestoppt: Anhänge und Nachrichten während eines Laufs werden korrekt eingeordnet, gebündelte Lesezugriffe zählen als Fortschritt, und Lese-Hinweise entfallen bei Audits.
- Vertragshinweise beim Lesen einer Datei nennen nur noch deklarierte Verträge ohne Abschlusspflicht.

## [0.2.221] - 2026-09-22

### Added

- Claude Opus 5.5 im Anthropic-Modellkatalog mit aktuellen Preisen, dauerhaft adaptivem Thinking und Fast-Modus.
- Xiaomi MiMo-V2.6-Pro im Modellkatalog und MiMo-V2.5-TTS als VOICE-Ausgabeanbieter mit neun Stimmen.

### Fixed

- Automatische Account-Wechsel bleiben auch bei weiteren Anfragen und nach Neustarts erhalten; eine veraltete Account-Auswahl der Worker-Verteilung setzt den Fehlerwechsel nicht mehr zurück.

- Aktuelle ausdrückliche Nutzerentscheidungen ersetzen widersprechende alte Projektvorgaben im betroffenen Umfang. Quellrevisionen, genaue Spec-Zuordnung, persistierte Leseabdeckung und begründete Vertragsprüfungen machen offene Dokumentationsfragen sichtbar; ersetzte Erinnerungen werden nicht erneut als aktive Evidenz projiziert.

- Worker untersuchen neue Hinweise ohne starre Drei-Versuche-Grenze. Aufträge können Gesprächskontext, unabhängige Recherche und benötigte Edit-/Testfähigkeiten ausdrücklich wählen; reine Diagnosen blockieren keine Schreibzuständigkeit mehr.
- Suchergebnisse zeigen ihren Geltungsbereich; Worker-Nachrichten unterscheiden Zustellung von umgesetzter Korrektur.

- Nach dem Abschluss privater Dateiänderungen bleiben Build und Tests im gemeinsamen Checkout. Parallele Veröffentlichungen werden als Diagnosekontext gemeldet, ändern aber weder den echten Exit-Code noch lösen sie automatische Wiederholungen aus.
- Eine ausdrücklich angeforderte isolierte Nachprüfung erfasst den aktuellen Quellstand, ohne den Ausgangsstand anderer Sessions zu verändern. Unveröffentlichte eigene Änderungen werden weiterhin im privaten Testbaum geprüft.
- `edit inspect` akzeptiert jetzt einen `verify`-Befehl für eine gezielte private Nachprüfung nach der Fehleranalyse.

## [0.2.220] - 2026-09-22

### Fixed

- Nach dem Abschluss privater Dateiänderungen bleiben Build und Tests im gemeinsamen Checkout. Parallele Veröffentlichungen werden als Diagnosekontext gemeldet, ändern aber weder den echten Exit-Code noch lösen sie automatische Wiederholungen aus.
- Eine ausdrücklich angeforderte isolierte Nachprüfung erfasst den aktuellen Quellstand, ohne den Ausgangsstand anderer Sessions zu verändern. Unveröffentlichte eigene Änderungen werden weiterhin im privaten Testbaum geprüft.
- `edit inspect` akzeptiert einen `verify`-Befehl für eine gezielte private Nachprüfung nach der Fehleranalyse.

## [0.2.219] - 2026-09-22

### Fixed

- Dateiisolation beginnt bei nativer Writer-Anbindung mit dem ersten Edit. Zusammengehörige Änderungen bleiben bis zum geprüften Abschluss privat; Konflikte gehen zur Auflösung an denselben Agenten. Ein Wechsel auf OFF veröffentlicht keine offene Edit-Einheit.
- Prüfungen verwenden den aktuellen privaten Dateistand. Veraltete private Beobachtungen nach fremden Änderungen und die Bearbeitung von Rust-Methoden werden korrekt behandelt.
- Gemeinsame Auswahlfelder akzeptieren unveränderliche Optionslisten, sodass das Security-Panel den Desktop-Typcheck nicht mehr blockiert.

## [0.2.218] - 2026-09-22

### Added

- Eine schreibgeschützte `diagnose`-Rolle für Warum-Fragen: Sie untersucht und berichtet, ohne dabei Dateien zu verändern.

### Fixed

- **Eine Datei, die es nur im Arbeitsverzeichnis gab, kollidierte beim Übernehmen mit niemandem.** Wer eine Datei per Shell-Befehl, Generator oder externem Editor anlegte und sie anschließend bearbeiten ließ, bekam eine Konfliktmeldung für einen Pfad, den keine zweite Sitzung je berührt hatte — und konnte ihn nicht aufösen, weil es gar keine Gegenseite gab. Dahinter stand ein falscher Vergleichsmaßstab: Als Ausgangsstand wurden die Bytes von der Festplatte herangezogen, während der erfasste Stand diesen Pfad überhaupt nicht kannte. Genau das liest die Übernahme dann als Ändern-gegen-Löschen. Die Erstberührung eines solchen Pfades gilt jetzt als Hinzufügen. Das verengt ausschließlich den Ausgangsstand, nicht die Prüfung: Ein bereits erfasster Pfad behält seinen Vergleich, und eine echte Kollision zweier Sitzungen auf demselben Pfad meldet weiterhin einen Konflikt.

- **Im Fenster „Worker issues“ führte kein Knopf etwas aus, der mehr als einen Text übergab.** „Remove“, „Remove all“, „Automatic“ und „Via session“ endeten wirkungslos mit der allgemeinen Meldung, dass etwas schiefgegangen sei; allein „Inspect“ funktionierte, weil es nur eine Zeichenkette weiterreicht. Ein abgetrenntes Dialogfenster schickt jeden Knopfdruck an das Hauptfenster zurück, und die dabei übergebenen Daten lagen in einer Form vor, die diesen Weg nicht überstehen kann. Die Daten werden jetzt vor dem Übergeben in eine übertragbare Form gebracht — auf beiden Seiten der Verbindung.

- **Ein Worker-Abschluss blockierte länger als nötig und konnte Ausführungsergebnisse verlieren.** Abschlüsse bleiben jetzt reaktionsschnell und behalten ihr Ergebnis, auch wenn kein Artefakt entsteht oder eine Entscheidung des übergeordneten Laufs erst nach dessen Antwortfrist eintrifft.

- **Ein gewechselter Anbieter-Zugang überlebte den Neustart der App nicht** und fiel still auf den vorherigen zurück. Die Wahl wird jetzt dauerhaft gehalten.

- **Automatisches Umbenennen und automatisches Commit** nutzen wieder ein eigenes Modell mit eigener Voreinstellung für den Denkaufwand, statt die Einstellung des laufenden Gesprächs zu erben.

- **Laufende Worker sind wieder kompakt im Blick:** Der Hauptagent sieht die Live-Aufsicht direkt, Worker-Ergebnisse sind ehrlich abrufbar statt als zugestellt behauptet, ein Abschluss ohne Dateiübernahme wartet nicht mehr auf beschäftigte Eltern, spätere Rückfragen an den übergeordneten Lauf bleiben auch nach dessen Antwortfrist erreichbar, und bereits abgeschlossene Worker werden nicht mehr nachträglich wegen Budget abgebrochen.

- **Terminal-Ankündigungen beim Anhängen an eine Sitzung** erscheinen jetzt als System-Notiz statt als vermeintliche Nutzeräußerung.

- **Fehlgeschlagenes Speichern der Einstellungen blieb unsichtbar.** Scheitert unter Windows das Ersetzen der Datei, werden der Fehler und die beteiligten Ressourcennutzer jetzt protokolliert, statt nur auf der Konsole zu landen. Das macht den Fall auffindbar; die Ursache der ursprünglichen Dateisperre ist damit noch nicht behoben.

## [0.2.217] - 2026-09-22

- Complete Grok 4.7 in the generated model catalog and xAI provider pack, with Responses transport, four reasoning levels and the documented context ceiling.

- Keep parked workers recoverable, remove implicit integration builds, renew inherited budgets from real results, share input-bound check receipts, and keep worker briefs focused.

- Session opening overlaps runtime and plugin preparation while retaining registration ownership. Stop keeps unresolved runs and shell handles visible, and repeated confirmed shell stops remain idempotent.
- File queues preserve FIFO after owner release. Smart File Isolation keeps private versions across mode changes; imports and index reads now participate in the shared freshness contract, with read observations retained on resume.
- Large Git status scans use a separate repository lock. Explorer updates avoid quadratic subtree removal, preserve unchanged pages and stop reloading known empty folders.
- MAIN checkout registration reuses its verified physical owner across path spellings instead of refusing a new session with a generation-root conflict.

- Codex Voice uses the CLI conversation and delegation flow: natural answers, one task handoff, scoped results, persistent push-to-talk calls and interruptible WebSocket playback. Tools and private reasoning stay out of spoken output.

- Workers inherit the active parent profile across permission modes and resume. Built-in profiles keep short handoffs without automatic reviewer or formatting-repair rounds; the parent coordinates focused, integration, or release checks according to the selected profile.

### Changed

- Bei unbekannten, unsicheren oder möglicherweise veralteten Fakten nutzt der Assistent zuerst Websearch. Liefert die Suche keine brauchbaren Belege, folgt bei aktivem Research-Automatikmodus Deep Research; verbleibende Wissenslücken werden ausdrücklich benannt.

## [0.2.216] - 2026-09-21

### Fixed

- **In einer isolierten Sitzung scheiterte jeder Befehl weiterhin — zuletzt an der Umgebung, mit der er gestartet wurde.** Die Meldung lautete nur noch „ungültiger Variablenname" und nannte weder die Variable noch den Grund. Drei Dinge standen dahinter, und jedes einzelne hätte genügt. Windows reicht bei jeder Aufzählung der Umgebung interne Einträge mit, deren Name mit einem Gleichheitszeichen beginnt — für die Prüfung ein ungültiger Name, und sie stecken in jeder geerbten Umgebung. Zugleich setzte Unary selbst Variablen wie das Heimat- und das Temporärverzeichnis, die dem Arbeitsbereich gehören: Wer sie mitliefert, wird abgewiesen — gesetzt wurden sie danach ohnehin, die eigene Angabe war also keine Absicherung, sondern die garantierte Ablehnung jedes Befehls. Und weil Windows Variablennamen unabhängig von der Großschreibung behandelt, die interne Liste aber nicht, galten zwei Schreibweisen derselben Variablen als Dopplung. Die Umgebung wird jetzt an einer einzigen Stelle auf das Zulässige reduziert — und diese Stelle liegt auf dem Weg jedes Starts, ob Terminal, Hintergrundbefehl oder Build.

## [0.2.215] - 2026-09-21

### Fixed

- **Ein an den Chat angehängtes Terminal führte nichts aus, was der Assistent hineinschrieb.** Die Befehle erschienen, das Terminal wirkte bereit — und blieb untätig, weil jede Zeile in Windows' Eingabe-Fortsetzung landete statt abgeschickt zu werden. Unter Windows ist die Eingabetaste ein anderes Zeichen als ein Zeilenumbruch; geschickt wurde der Zeilenumbruch. Die Befehle stapelten sich unausgeführt, ohne Fehler, ohne Hinweis. Ein Zeilenumbruch wird jetzt auf das Zeichen abgebildet, das die Konsole tatsächlich als Eingabetaste versteht.

- **In einer isolierten Sitzung startete auch nach dem letzten Update noch immer kein Befehl — nur scheiterte er später.** Die vorige Version machte den Start überhaupt erst erreichbar, und dahinter wartete die nächste Hürde: Wo ein Programm außerhalb des Arbeitsbereichs liegen darf, war schlicht nichts hinterlegt. Damit konnte die Zulassung nur für ein Programm *innerhalb* des Arbeitsbereichs gelingen — und der Kommandozeilen-Interpreter liegt dort nie, sondern in seiner eigenen Installation. Diese Installationswurzel wird jetzt benannt, und weil eine Unix-ähnliche Toolchain ihre Nachbarordner mitbenutzt, eine Ebene weiter oben angesetzt. Liegt das Programm ohnehin im Projekt oder stammt es aus dem Arbeitsbereich selbst — ein projekteigenes Werkzeug, eine frisch gebaute Datei —, wird bewusst nichts hinterlegt: Es ist bereits bekannt, und eine zweite Angabe würde der ersten nur widersprechen und den Start erneut verhindern.

## [0.2.214] - 2026-09-21

### Fixed

- **In einer isolierten Sitzung ließ sich kein einziger Befehl mehr starten.** Kein Terminal, kein Build, kein Test — jeder Versuch endete mit einer Meldung über überlappende Bereiche, und zwar für die gesamte Dauer der Sitzung. Dahinter stand eine Schutzangabe, die nichts schützte: Die Verwaltungsdaten des Projekts wurden als „unantastbar" deklariert, obwohl sie im Projektordner selbst liegen und die Sitzung auf genau diesem Ordner arbeitet. Die Angabe wird nirgends durchgesetzt — sie nahm dem Befehl also keinen einzigen Zugriff, sie machte ihn nur unmöglich. Sie gilt jetzt dort, wo sie zutreffen kann: wenn die Verwaltungsdaten außerhalb des Arbeitsbereichs liegen. Liegen sie darin, ruht die Absicherung dort, wo sie ohnehin ruht — auf den Sperren des Betriebssystems und den Transaktionen der Datenbank — statt ein Versprechen zu geben, das an dieser Stelle niemand einlösen kann. Dass die Meldung dabei die unverständlichere von zwei möglichen war, lag an zwei unterschiedlichen Schreibweisen desselben Pfades, die verglichen wurden, ohne sie vorher auf dieselbe Form zu bringen.

- **Ältere Unterhaltungen ließen sich nach dem Update nicht mehr fortsetzen.** Sie meldeten beim Öffnen, die Zuständigkeit für ihren Arbeitsstand könne nicht übernommen werden — dauerhaft, bei jedem Versuch. Der Grund war kein Konflikt und keine fremde Sitzung: Der Besitznachweis solcher Arbeitsstände lebte früher nur im Arbeitsspeicher des Programms, das ihn ausgestellt hatte. Mit dessen Ende war das Geheimnis unwiederbringlich fort, und seitdem konnte kein späterer Start es mehr vorlegen. Es gibt jetzt einen eng gefassten Weg zurück: Der Nachweis wird neu ausgestellt, aber ausschließlich für einen Arbeitsstand, der noch nie weitergegeben wurde, sich in keinem Übergang befindet und an dem nachweislich niemand mehr arbeitet. Einer laufenden Sitzung etwas wegzunehmen bleibt weiterhin unmöglich.

### Added

- **Grok 4.7 ist da, ohne auf eine neue Version zu warten.** Das neue xAI-Flaggschiff (500k Kontext, Text und Bild, Reasoning bis `xhigh`) steht sofort zur Auswahl: Unary liest seine Modell-Listen aus Provider-Paketen, die neben der App liegen, und ein geändertes Paket greift bei der nächsten Anfrage — kein Update, kein Neustart. Der mitgelieferte Katalog führt das Modell ab dem nächsten Katalog-Lauf ebenfalls.

## [0.2.213] - 2026-09-21

### Fixed

- **Eine gehaltene Datei kam unter Windows nie im Checkout an.** Wenn eine Sitzung ihre eigene Fassung einer Datei zurückgibt, sichert Unary das Schreiben in den gemeinsamen Checkout mit einer Sperre ab — damit zwei Sitzungen nicht gleichzeitig dieselbe Datei schreiben. Diese Sperrdatei wurde nur zum Anhängen geöffnet, und auf ein solches Handle verweigert Windows das Sperren grundsätzlich. Das Ergebnis war nicht etwa „ich warte“, sondern „Zugriff verweigert“: Jede echte Übertragung brach ab, und eine Sitzung, die tatsächlich gerade schrieb, sah aus wie ein harter Fehler statt wie ein Grund, kurz zu warten — genau der einzige Fall, für den diese Sperre existiert.

- **Die private Dateiansicht überlebt jetzt einen Neustart der Sitzung.** Welche Dateien eine Sitzung in ihrer eigenen Fassung hält, stand ausschließlich im Arbeitsspeicher. Nach einem Neustart war dieses Wissen weg: Die Sitzung las wieder den gemeinsamen Stand, obwohl ihre eigene, noch nicht übergebene Fassung unverändert gespeichert war — und eine in der eigenen Fassung gelöschte Datei tauchte mit fremdem Inhalt wieder auf. Diese Zuordnung wird jetzt dauerhaft gespeichert, beim Wiederaufnehmen geprüft zurückgeholt und dabei niemals über bereits vorhandenes, neueres Wissen geschrieben.

- **Die Wiederaufnahme nimmt keiner laufenden Sitzung die Zuständigkeit weg.** Wer eine Datei-Fassung besitzt, wird jetzt dauerhaft nachgewiesen und zusätzlich durch eine Sperre des Betriebssystems abgesichert. Eine Sitzung kann ihre eigene Zuständigkeit beliebig oft bestätigen, eine zweite Sitzung erhält sie nicht — und ein beschädigter oder fehlender Nachweis führt zu einer klaren Verweigerung statt dazu, dass eine fremde Fassung übernommen wird.

- **Eine abgelehnte Befehlsausführung nennt endlich ihren Grund.** Solange die Verwaltungsdaten eines Projekts im Projektordner selbst liegen, darf dort kein isolierter Befehl starten — sonst könnte er genau die Daten verändern, die festhalten, wem der Projektstand gehört. Gemeldet wurde dafür aber ein Satz über die Zuweisung des privaten Laufzeitbereichs, der an dieser Stelle längst nicht mehr die Ursache war; zwei Sitzungen haben daraufhin an der falschen Stelle gesucht. Die Meldung benennt jetzt den Ort der Verwaltungsdaten, den Projektordner und den nächsten Schritt.

### Added

- **Private Laufzeitordner werden wieder aufgeräumt.** Jede isolierte Sitzung bekommt einen eigenen Bereich für Zwischenstände und Build-Ergebnisse. Früher lag er innerhalb der Projektverwaltung und verschwand mit ihr; seit er außerhalb liegt, blieb er unbegrenzt liegen — ein nach einem Absturz zurückgelassener Ordner war von einem aktiv genutzten nicht zu unterscheiden. Jeder Bereich führt jetzt einen Besitznachweis mit Projekt, Generation, Prozess und Ablauf, der bei jeder Nutzung erneuert wird. Aufgeräumt wird nur mit Beweis: Der Anspruch muss abgelaufen sein UND es darf nachweislich keine lebende Sitzung mehr daran arbeiten. Ein Bereich ohne Nachweis wird gemeldet und bewusst nicht gelöscht — schlimmer als ein liegengebliebener Ordner ist ein Aufräumer, der einen löscht, den er nicht verstanden hat.

## [0.2.212] - 2026-09-21

### Fixed

- **Die Datei-Isolation ist wieder eingeschaltet.** Sie war vollständig gebaut und vollständig getestet — und stand trotzdem auf „aus", weil ihr Schalter im September hart auf Nein gesetzt wurde, als das Starten einer Sitzung mit eingeschalteter Kette noch scheiterte. Die fünf Ursachen dahinter sind behoben, der Schalter war es nicht. Er fragt jetzt wieder die Maschine selbst, ob sie die Prozessgrenze halten kann, statt eine Konstante zu lesen: Wo sie es kann, arbeitet jede Sitzung auf ihrer eigenen Fassung jeder Datei; wo sie es nicht kann, bleibt alles wie bisher, statt eine Absicherung zu versprechen, die dort niemand einlösen könnte.

- **Die Arbeitsverlauf-Abfrage suchte an einer Stelle, an die nie jemand schrieb.** Für ein und dasselbe Verzeichnis gab es zwei Adressen: Die Aktivierung legte es als `workspace-registry` im Unary-Ordner an, die Verlaufsabfrage suchte es unter `agent/workspace-registry`. Jede Hälfte war für sich stimmig, deshalb sah nichts falsch aus — die Antwort lautete nur immer „dieser Checkout ist nicht registriert". Auf einem Rechner, der noch nie eine Sitzung isoliert hatte, kam stattdessen ein nackter Systemfehler zurück, der weder Ursache noch Abhilfe nannte, weil das Verzeichnis geöffnet wurde, ohne es vorher anzulegen. Beide Hälften nennen jetzt denselben Ort, und wer ihn öffnet, legt ihn zuerst an.

### Removed

- Der alte Isolationsweg, der jeder Sitzung eine vollständige Kopie des Repositorys anlegte, ist entfernt. Seit der Umstellung auf Datei-Isolation rief ihn nichts mehr auf: Für gewöhnliche Änderungen genügt die einzelne Datei, und ein Test- oder Build-Lauf bekommt seinen eigenen Baum auf einem anderen, schlankeren Weg — was die Vollkopie auf genau die zwei Fälle beschränkt, die sie wirklich brauchen.

## [0.2.211] - 2026-09-21

### Fixed

- **Handing a decision back to you works again.** When the assistant needs your go — you asked to see a plan first, or told it to stop — it is supposed to say so once and end its turn there. That one instruction was refused with "valid actions: get, create, update, clear", even though everything behind it existed: the pause, its storage across restarts, the release on your next message, and the tests covering all of it. A single list in the middle of the chain had never been told about it. The visible consequence was worse than a missing shortcut: with the proper instrument unavailable, the only way left to pause was to mark every open item as externally blocked, one call at a time — which is exactly the clumsy behaviour that instruction was built to replace, and which you saw as the run nudging itself onward instead of stopping.

## [0.2.210] - 2026-09-21

### Fixed

- **Per-file isolation can actually be switched on now.** The feature that gives each session its own copy of a file it edits — instead of everyone writing into one shared checkout — refused to start at all when enabled. Five separate faults sat behind one another, so each repair only revealed the next: the import ran against a workspace record that did not exist yet; the same disk identity was written down twice in different notations (one decimal with a prefix, one hexadecimal) and then compared as text, so a folder was mistaken for a foreign one while path and file index matched perfectly; there was no way for the component that had just performed the setup to say so, so the final step rejected the very session it had been told to build; a single untracked file anywhere in the project made the whole adoption fail, which in a real checkout is the normal state rather than an edge case; and forking, starting a new session or resuming all went through the old path that refuses a session prepared this way. All five are fixed, and the full test suite now behaves identically whether the feature is on or off.
- **Whole-project isolation now recognises the commands that need it.** Only test and build runs get their own copy of the project; everything else edits in place. But the check that decided this looked at the FIRST WORD of a command only — so `cd some/dir && cargo test`, `pwsh -File ./scripts/verify.ps1` and a test run captured into a variable were all treated as ordinary work and ran in the shared project, which is precisely what the feature exists to prevent. Commands are now read in full, while a command that merely MENTIONS a test — searching for the words `cargo build`, for instance — still correctly stages nothing.
- **A missing import stopped the whole project from compiling for checks.** One test module reached for a catalog lookup through a shortcut that does not reach the crate root, so an entire crate's test build failed on a single unresolved name and took the workspace-wide check down with it.

## [0.2.209] - 2026-09-21

### Fixed

- **Two sessions writing the same file finally notice each other.** Everything that protects a shared file — taking turns, the waiting queue, keeping your own copy — starts with a writer announcing which file it is about to touch. The most common write in the app, replacing text inside a file, announced nothing. That did not merely leave this one path unguarded: because nobody was ever announced, every OTHER writer looked for a conflict, found none, and wrote too. Four agents appending to one file produced twenty-four lines with not a single pause, and nothing anywhere recorded the file as being in use. Every path that changes a file now announces it before the first byte reaches the disk, so the queue and the private copy do what they promise.
- **The waiting queue no longer freezes when several writers wait at once.** Two pieces of internal bookkeeping were locked in opposite orders by the writer that waits and by the check that looks for waiting cycles, and each waiting writer asked "is someone else holding this?" ten times a second — every time taking an exclusive lock on the very file that records who holds what. With a handful of writers in play everything came to a standstill: nothing crashed, nothing timed out, it simply stopped. That question is now asked on a slow beat, and the cycle check no longer holds one lock while reaching for others.
- **A contested patch answers like a contested edit.** Applying a patch to a file another session was holding replied in its own words, while every other write had already moved to "you are next in the queue". It gives that same answer now, so the reply no longer depends on which action you happened to use.

### Changed

- **The retired per-session worktree switch is gone from the code.** Nothing in the app had set it for a long time, so the check enforcing it returned immediately every single time — yet it was the ONLY conflict check one write path performed, which made that path look guarded while it could never act. What remains is the one switch that is actually read on every write: Smart File Isolation.

## [0.2.208] - 2026-09-20

### Added

- **Chats can make pictures and clips now.** A chat could research anything on the web but could not produce a single image, because the tool boundary that keeps chats away from your machine also happened to fence off image and video generation. Those two are not machine access — they call a provider and write the one file you asked for — so they are allowed in a chat now, while reading, editing, running and desktop control stay out exactly as before. Ask a chat for a picture and you get one; if no image or video model is set up, it says so and points at the setting instead of quietly producing nothing.

### Fixed

- **A run you paused stays paused.** When the agent hands a decision back to you, the whole task board waits for your answer instead of the agent marking every line "blocked" — but that pause was only ever held in memory, while the unfinished work it applied to was written to disk. Restarting the app, or anything that picked the session up again without you having said a word, brought the open work back *without* the pause and the agent continued with exactly what you had stopped. The pause is now stored with the work it belongs to, and lifting it is stored the same way, so it neither disappears on its own nor comes back after you have answered.
- **Reopening a session no longer empties its task board.** On the first message of a turn the app notes which request the tasks belong to, and that note was written before the saved board had been read back — which made the saved board unreachable for the rest of the session: the panel came up empty and the agent lost the list of what it still owed you. The board is read first now.

### Changed

- **The agent can see what its own task updates do to the run.** Every time it changes the board it is now told, in the same breath, what that leaves the run doing — continue, wait for you, or stop — so it stops having to guess at the rule and spend a turn finding out it guessed wrong.

## [0.2.207] - 2026-09-20

### Added

- **The editor's colours are yours now.** Syntax highlighting was fixed in the app and could not be changed at all. Drop a theme file at `.unary/editor-theme.json` in a project — or next to your global settings, for every project — and every editor picks it up; a reload restyles editors that are already open. The palette that ships is written in exactly the same format, so it doubles as the example, and a file with a typo in it falls back to that palette with one warning naming the file instead of leaving you in front of a blank editor. A copy to start from is `shared/editor/editor-theme.example.json`.
- **The IDE window can be opened again.** The button in the title bar had been sitting there disabled while the desktop app was the only surface being built. It opens the IDE on your focused project again.

### Fixed

- **Opening a file twice opens one window, not two.** Every window the app detaches — the file preview, an aux panel, a panel pulled out of the dock — was addressed by a name containing the current time, which means every request was a *new* window. A double-click on a file in the explorer therefore started two windows racing each other to load the same file, and asking for a panel you had already pulled out buried the one you were looking at behind a fresh copy. Each of these now has one window that comes forward and switches to what you asked for, and two fast clicks can no longer slip past that.
- **A file opened from a second project edits that project's file.** Editor tabs resolved their file against whichever project the window had focused, so a tab holding a file from somewhere else showed the right text while its git history, blame, code intelligence and breakpoints belonged to a different file entirely — and saving wrote where the tab thought it was. Every tab now carries its own project.
- **Renaming a file no longer leaves its tab writing to the old path.** Moving or renaming a file from the desktop explorer updated nothing that was open: the tab kept the dead path, and the next save recreated the file under its old name. Tabs follow the file now, which is what the IDE already did — the desktop app simply had nobody listening.
- **The app stops repeating one warning every five seconds.** Two parts of the app wrote the list of running editors, each through the same temporary file, so one could be holding what the other was about to replace — and when a virus scanner or backup agent held that list for a moment, one of them gave up on the first attempt. The write is one shared, patient operation now, and a failure names the file it could not write, is reported once, and is followed by a line when it recovers.

## [0.2.206] - 2026-09-20

### Changed

- **The embedding weights now come from Unary's own release storage.** Semantic search used to fetch its model from Hugging Face on first use, so a third-party host decided whether indexing worked at all. The weights are published beside the installers instead, compressed to roughly a third of their size, checked against a digest when they arrive, and fetched only when they are not already on disk — Hugging Face stays as a last-resort fallback. They are deliberately not packed into the installer itself: that would put 372 MB into every download of every release, while the copy in your model cache survives updates and is paid once.

### Fixed

- **Semantic search no longer breaks because of a Hugging Face login you forgot you had.** The bundled embedding runtime downloads one public model and needs no account at all — but it was quietly sending whatever Hugging Face token happened to sit on the machine, and a stale one made the download fail with "401". The error then told you to check an API key under Settings → Plugins → Providers, which has nothing to do with this runtime. The download now asks for no credentials, tries the machine's token only if the anonymous attempt fails, and the message names the real cause and the real fix.

## [0.2.205] - 2026-09-20

### Changed

- **Orchestrator is a switch now, not a mode.** Turning it on used to replace the mode you were in — Ask, Plan or Yolo went away and "Orchestrator" took the slot, so you could not orchestrate *and* be asked before things happen. It is now a switch beside the mode: the mode still decides how much runs without your approval, the switch decides who does the work. A session saved under the old pseudo-mode reopens as Yolo with the switch on, which is what it always meant.

### Fixed

- **A desktop task no longer carries the whole desktop with it.** Handing native UI work to Pilot's controller attached a list of every open window — 28 of them on a normal working machine — even when the task already said exactly which window it meant. That list is written into the task itself, so it was paid for in full on every single handoff, and there were dozens per session. A task that names its target now gets that target's windows; only a task that names nothing still gets the full desktop, and a named target that has since disappeared falls back to the full list rather than claiming there are no windows at all.
- **A long wait ends when the desktop actually moves.** Waiting for an app to start meant sleeping a guessed number of seconds and sleeping all of them, even when the app had appeared after two. Long waits now watch for the desktop to change and return the moment it does. A desktop where genuinely nothing happens is still waited out in full — that is the case the guess existed for, and cutting it short would have been worse than the wait.
- **Clicking twice in the same window stopped paying the same setup twice.** Every click, keystroke and drag re-checked whether the target window would accept input and re-ran the whole focus handover from scratch — including burning a full two-second timeout again on a window that had just refused to come forward. Those checks are now remembered per window for a short time, so a sequence of actions in one window costs the setup once instead of once each.
- **Pilot records what an observation actually cost.** Its trace captured every action with its timing, but not the one number that dominates: whether a step shipped a full screenshot to the model or was recognised as unchanged and dropped. An investigation into Pilot's speed could measure how often it looked at the screen, never how much that looking cost, and had to infer the rest. The cost is now written down per observation.
- **A delegating agent is no longer told it can write files.** The very first line of its instructions promised an assistant that edits code and writes files, and a later section told it to commit its own work — both untrue for a session whose whole point is that it hands changes to workers. It read the promise, tried, and was refused by its own tools. Both statements now follow what the session can actually do.
- **A refused command says which part was refused.** An inspection-only shell rejected read-only commands over a construct the agent could not see — the message named neither the cause nor, in a delegating session, the truth that no approval exists for it, and instead suggested asking for one. The refusal now names the construct it tripped on and offers the two things that do work: rewrite it read-only, or hand it to a worker.

## [0.2.204] - 2026-09-20

### Changed

- **A shipped installer is never handed out uncompressed.** The previous release added a switch that let the installer store its payload instead of packing it — the fastest possible packaging step. It also left everything we ship sitting in the installer as plain readable text, and it saved about twenty seconds. That switch is gone, and the floor is now normal compression: the option only ever moves toward *more* packing, never less. For the same reason the shipped interface bundle now states explicitly that it carries no source maps, instead of leaving that to a default one stray flag could flip.

## [0.2.203] - 2026-09-20

### Changed

- **Release builds are substantially faster, at a fraction of a percent of runtime.** Every release paid for one whole-program optimization pass over the entire codebase — a phase no cache can reuse, and one that cannot start before it has seen all of it. It bought a low single-digit percentage in an app that spends its time waiting on the network, the disk and you. That pass is gone, the work now spreads across far more parallel units, and a repeated release from the same tree reuses what it already built instead of starting over. Crash evidence is untouched: the same line tables ship, and the same safety nets stay in place.
- **The installer's compressor is a deliberate choice instead of a fixed cost.** The Windows installer is packed by a compressor that runs on exactly one thread and offers no option to use more, so a maximally compressed installer is minutes of a single core no matter how large the machine — which is why it is opt-in rather than the default. The release script keeps the fast setting, `-Lzma` asks for the small shipping installer, and `-NoCompression` skips packing entirely while you are iterating on packaging itself.

## [0.2.202] - 2026-09-20

### Fixed

- **A compaction handover no longer presents restored history as new work.** Every turn of a session carries a position, and that position decides which turns belong to the part already summarized and which are still open. It was assigned as "how many turns have been stored so far" — the same thing only while turns arrive in the order they happened. Replaying the durable journal breaks exactly that: it writes a turn that HAPPENED earlier but ARRIVES later, the counter filed it last, and the boundary then read a piece of restored history as work done after the cut, so a handover could show you turns you had already seen as if they were new. Positions now follow when a turn actually happened, and the turns behind an inserted one move up accordingly.

## [0.2.201] - 2026-09-20

### Changed

- **Parallel work no longer runs in Git worktrees.** A session or worker that needs to write gets its own private workspace instead: real files, materialized by a native volume clone where the filesystem offers one and by an independent copy otherwise, tracked by Unary's own content store and catalog rather than by Git plumbing. Git keeps exactly the jobs it is good at — history, remotes, and the branches you create yourself — and it is no longer asked to isolate concurrent agents, which is where the old automatic worktrees kept getting stuck: a tree that could not be folded back left work stranded behind a merge conflict nobody asked for. Manual worktrees you create stay untouched and keep their own vocabulary.
- **Finishing work is now "Smart Merge to Parent".** A private workspace submits a candidate that is built against the parent's ACTUAL current state; if the parent moved in the meantime, the candidate is prepared again instead of applied blindly. A conflict hands over the complete package — base, your version, the target version and the affected paths — so the later contributor resolves it with full context, and there is no partial apply. Every integration leaves a receipt naming its base, contributors and what was deliberately excluded, and repeating it is idempotent rather than a second effect.
- **Interrupted work is recoverable instead of lost.** Crashes, takeovers and cancelled workers leave a visible waiting state with its own recovery step; a coordinator takeover fences the old owner and quarantines generations whose processes may still be alive rather than killing them, and unknown liveness always keeps the data.
- **A test or a build run sees a consistent tree, automatically.** Everything else works file by file, but a test that reads a checkout other sessions are part-way through writing tells you nothing about your own change — so a run recognised as a test or a build gets a private copy of the whole tree for its duration, and nobody has to ask for it. The copy costs the size of the CHANGE, not the size of the repository, and the build cache deliberately stays outside it: an isolated run that recompiled the world every time would be isolation nobody could afford. Commands that are neither a test nor a build run exactly where they always did, because the classifier only reacts to what stands at the START of the command — a miss costs nothing, while guessing would stage a whole tree for an `echo` that merely mentions a test.

### Fixed

- **A queued message could vanish from the pending list before it appeared in the chat.** Messages you send while the agent is busy sit in the composer as chips until the turn picks them up, and a chip was allowed to disappear on the strength of the event alone. Two things went wrong. The event that says "this one has been delivered" was matched to a chip by its timestamp, without comparing the text — and the harness re-stamps a message's timestamp when it merges queued messages or wraps one in its delivery envelope, so the wrong chip was cleared: it vanished with no message of its own, while the message that did arrive belonged to a different chip. Second, the safety valve that stops a chip wedging the composer forever simply deleted it after five seconds, so a message whose row was merely slow existed in neither place — not in the queue, not in the chat, not in the input box. A chip is now tied to its message by id or by the text itself, a timestamp decides only between candidates it cannot be confused with, and the valve hands the message back to be shown in the transcript instead of dropping it. The Desktop chat had none of this waiting logic at all and cleared chips the moment an event arrived; it now runs the same rules as the Code panel, from one shared implementation.

- **Language servers, MCP servers and the Python worker started outside the process-admission contract.** Unary decides per session which parts of the filesystem a child process may touch, and that decision is made before the process exists — a process that starts around it is simply unscoped. Three kinds did: local MCP servers, language servers started by the LSP manager, and the CPython worker behind `py()`. Each now passes the same gate the shell already passed, and a start that cannot be scoped is refused before a child exists rather than repaired afterwards; where no host installs the gate, nothing changes. One of these was found only because the worker who built the seam proved its own work inert — the function that installs it was private to its crate, so the compiler reported it as never used, and the seam would have shipped looking complete while being unreachable.

- **The native embedding lane padded batches with the wrong token.** Two lanes implement the same Granite embedding model, and short texts in a batch get padded out to the longest one. The Candle lane padded with the token the model itself declares for that purpose; the native lane read the same setting out of the model's configuration and then never passed it on, so it padded with token 0 instead. Both lanes went on producing embeddings and neither reported anything wrong — they simply filled the unused positions of a batch with different content. The native lane now uses the model's own padding token. The giveaway was a compiler warning saying the configuration field was never read: the value had been parsed, which made the setting look delivered, and the warning had been sitting in the build output long enough to be scenery.

- **The Hub sign-in error pointed at a settings entry that does not exist.** When you were not signed in, the message told you to open "Settings › Account" — but that entry is labelled Profile, so anyone scanning the sidebar for Account found nothing. Found by checking the rest of that class rather than treating the web-search case below as a one-off: all forty settings paths in the Rust sources were compared against the actual section list. That produced this one genuine fix, and retired one suspicion that turned out to be correct already.

- **The web-search fallback notice pointed at a settings page that does not exist, and named the wrong providers.** When no search service is configured, web search falls back to the managed browser and tells you so. That notice sent people to "Settings › Web search" — which is not a destination anyone can reach, because Web search is a section inside Knowledge & Research — and it listed five of the six providers you can actually configure, silently dropping Perplexity. Both came out of following a single dead-code warning: a second copy of the same notice had been left behind when the surface that rendered it was removed, nothing called it, and its three green tests kept it looking healthy. The two copies had drifted in opposite directions, each correct exactly where the other was wrong, so consolidating them nearly shipped the drift instead of removing it. The dead copy is gone, the surviving one names the real page and all six providers, and the configuration hint now lists all five environment variables the web layer actually reads instead of three.

- **A worker that has spent its token budget is paused reliably, instead of by luck.** The watchdog decides a budget from the child session's cumulative usage, and reading that usage began with a single non-blocking attempt at the child's own live state — so whenever the worker happened to be holding it, the reading came back as "usage unknown". Unknown usage is deliberately never treated as a breach, because inventing one would pause workers over a missing file. The result was that a worker which had spent its entire token budget simply kept running, and whether the ceiling bound at all depended on a race nobody could see. It was intermittent, which is how it passed a full test run and then failed on its own — and why looking for the change that broke it found nothing: no change did. The reading now retries briefly, reports a busy child as BUSY rather than as missing data — one is worth retrying, the other is not, and collapsing them into a single answer was the whole defect — and, once it has seen that session even once, falls back to the durable transcript, which is an ordinary file and not behind that lock. Without the live portion the fallback can only under-report, so a ceiling may bind one sweep late, but it binds.
- **A session renamed while it runs keeps that name after it closes.** An automatic rename is staged first and written to the session file at the actor's next safe point. Closing the session removes the actor, so the name survived only in an in-memory overlay that could never reconcile against a file it had never reached: the list showed the new name for the rest of the session, and the next start showed the old one. Closing now writes a pending rename through the same path a rename of an idle session uses. If that write fails the overlay is kept rather than dropped, because a list that still tells the truth beats a silent loss, and a close never fails over a title.
- **The sidebar's "Needs attention" group can be collapsed.** Its header drew the caret every collapsible subgroup draws but carried none of their wiring — no button role, no click or keyboard handler, no expanded state — and its rows rendered unconditionally, so it read as expandable while clicking it did nothing. It now toggles through the same persisted `expandedProjects` store as the branch, worktree and archived subgroups: open by default, held open while a sidebar search runs, caret following the state, rows behind the guard, and the same Collapse/Expand in the group's menu. A collapsed header keeps the warning dot and the count, so folding a running, queued or conflicted fold-back out of view stays an explicit user choice.

## [0.2.200] - 2026-09-18

### Fixed

- **A project's sessions are visible again after a branch switch.** The Desktop sidebar routes a project's sessions through three renderers — worktree rows, branch groups, and the flat list — and two of them could hand a session to nobody. `groupSessionsByBranch` dropped every pair whose isolated workspace looked like a managed or foreign worktree, which is correct only while a worktree row actually renders it: once that tree was folded back and removed, no row claimed the session, and because it still matched its project by cwd it never reached the unassigned bucket either. On top of that the flat list and the branch grouping read different sets — the flat list took `activeSessions`, the grouping also took the worktree sessions no row had claimed — while `groupByBranch` requires MORE THAN ONE branch group. A checkout that moved to a new branch therefore marked every session recorded on the old branch as detached (`isDetachedFromMainCheckout` compares the recorded branch against the live one), collapsed them into a single group, and rendered neither that group nor those rows: the project kept its count badge and showed almost nothing. Sessions are now excluded from branch grouping only by the ids the caller says it already renders, and one `planSessionRendering` decision names the renderer AND hands it the exact set, so a pair cannot fall between the flat list and the branch groups.

## [0.2.199] - 2026-09-17

### Fixed

- **Worktree return now explains why it is blocked instead of silently giving up.** Recovery tools can reconnect a session to its own recoverable worktree while preserving files and commits. Ambiguous cases remain safely blocked rather than being reported as a successful return.
- **File claims now follow completed turns and workspace changes promptly.** Older operations no longer overwrite newer ownership or erase the next writer's claim. Uncommitted changes remain tracked separately from active editing.
- **Chat shows a compact return-status box instead of a perpetual “Folding back…” label.** It displays reported blocking files, owners, conflicts and next steps. Queued returns are not labelled as running, and failed refreshes leave the last report explicitly unconfirmed.

## [0.2.198] - 2026-09-16

### Fixed

- **A branch switch reaches the session list at once instead of up to a minute later.** The Desktop sidebar groups a project's sessions by the branch its checkout is on, but the workspace state relay had exactly one branch-shaped kind — `branch-deleted` — so the most common branch event there is, a plain switch, had no producer at all and the 60 s poll owned the branch subgroup alone. Every checkout path in `crates/unc_rust_workspace/src/git.rs` now reports `HeadChanged` with the new branch, and a switch made OUTSIDE Unary (a terminal `git checkout`) is picked up by a new host HEAD watcher: one non-recursive watch of each registered project's git directory — never the HEAD file itself, whose inode git replaces on every switch — reading the branch out of HEAD rather than spawning `git`, reporting only when the branch actually moved. The sidebar registers the projects it shows (bounded at 64, idempotent, advisory: an uninstallable watch leaves the poll as the backstop), and because the event carries its root, the refresh re-reads ONE project's branch instead of spawning `git branch` per project.
- **The first workspace-state event of a burst no longer waits out its own debounce.** `workspace-state:changed` was purely trailing at 250 ms, so a fold-back, a worktree removal or a branch switch spent a quarter second doing nothing before the sidebar even started a worktree measurement that costs seconds. The channel now has a leading edge with a 250 ms throttle, so the burst still collapses to one additional refresh and the first one is immediate.

## [0.2.197] - 2026-09-16

### Fixed

- **A session working in its own worktree no longer asks the user to fold back work the harness is already folding.** The only fold-back text a session ever saw is resolved from a path on every catalog rebuild and may not spawn `git`, so it was state-blind — and it named the user-facing fold-back control beside the automatic path. A session that had finished its work read that as the user's next step and reported the fold-back as homework, while the pending row for its tree was `waiting-owners`/`automatic-transient` with the watcher armed, retrying by itself, waiting only for a foreign live session to release the files it still held. Nothing was stranded and nothing was owed. The live row now rides in the per-request `[Session facts v1]` tail — state, classification, watcher, attempt count, blocking file count and owner — where `waiting-user` is the only state that names a person, `waiting-agent` points at the machine assignment, and an unreadable store reads as "nothing pending", never as "a person must act". The path-derived note no longer mentions the user-facing control at all, because a model that never hears about it cannot hand it over as a task. The pending-fold-back store is read only for a cwd that already has the managed worktree shape, so an ordinary session pays no per-request disk access.

## [0.2.196] - 2026-09-16

### Changed

- **Codex Voice is a voice device now, not a second agent.** It was ported from the Codex CLI with the CLI's own architecture: the realtime model was the conversational front, it owned a `background_agent` tool and a delegation channel, and it handed work to the coding session while it kept talking. That shape shipped a second prompt source — the composer already sends the finished transcript as a user message, so one utterance could reach the session twice — and it made the model's own reasoning and progress chatter audible. The realtime session now only transcribes and reads out. Speech in becomes an ordinary user message to the session model; the only thing ever spoken is that session's finished answer, markdown-stripped, at the real end of the turn — never reasoning, never tool output. No protocol version advertises a tool, a `tool_choice` or a `delegation` target any more, v2 sets `turn_detection.create_response: false`, and a speech gate drops every output audio frame and output transcript that no read-aloud request asked for — including on the WebRTC route, where the composer bar mutes the remote audio element unless that gate is open. The `delegationAckFiller` setting is gone; `speakReplies` now means "read the session's answer aloud", and dictation stays input-only.

### Fixed

- **A spoken reply no longer depends on a system notice arriving.** The desktop live feed's end-of-turn speech hook sat in the `system_notice` branch, so a voice plugin read an answer out only when a notice happened to land with that turn and most turns stayed silent. It runs on `agent_end` now, for both voice plugins.

## [0.2.195] - 2026-09-16

### Fixed

- **A worker that had already finished no longer buys a supervision decision turn.** The stall watchdog treated an unreadable sample as a stalled worker: a contended child-snapshot lock — or no active handle at all, which is exactly what a worker writing its handoff looks like — left every evidence field at its default, and the ladder read those unmeasured zeros as "nothing happened", parked the worker, interrupted its active step and woke the parent for a decision. The report then printed `childProcessRunning=false, newOutputBytes=0, filesChanged=false` for measurements it never made, which reads exactly like a real stall. An unobserved sample now yields a check-in and can never nudge, park or terminate; the report prints `snapshotAvailable` and `unknown` for every unmeasured field; a worker whose completion slot is already claimed is skipped entirely; and a park interrupts the child only on loop evidence, because aborting the active step of a merely slow worker destroys the step it was in the middle of.
- **A supervision decision the parent can no longer act on is withdrawn instead of delivered.** Staleness was checked only where the decision notice was enqueued, never between that enqueue and the queue drain — so a worker that terminalized seconds after its park still woke the parent, for work whose patch was already applied. The queued wake is now retired by its deterministic id prefix as soon as the record stops awaiting a decision, and the resolution is stamped rather than deleted: resume no longer erases the decision it answered, a cancelled record records `Closed`, any other terminal state records `WorkerTerminal`, and every resolution word now has exactly one production writer.
- **File and commit progress are no longer invisible for a whole window after each commit.** The comparison was gated on a non-empty previous fingerprint, so the clean tree that follows a commit — the strongest progress evidence a worker can produce — reported "nothing changed", and a worker without a worktree reported it forever. An unknown baseline is now `unknown` instead of `false`, and a growing message stream with fresh activity withholds the first decision rung's park for a worker that is reading and planning rather than looping.

## [0.2.194] - 2026-09-14

### Fixed

- **A provider pack's own base URL survives the settings pass.** The pass that applies `settings.json` calls `set_provider_base_url(provider, None)` once per provider entry (`None` for an entry that carries only `enabled`/`visibleModelIds`), and `None` means "clear the override". A provider that lives in a pack has nothing to fall back to — it has no entry in the static built-in default table — so the clear wiped the URL the pack declares, on the provider config and on every model row, and the session-level setter did the same to the active model. Every later request then failed with `UnsupportedApi`, "provider … with api … is not implemented in Unary yet", while the dialect, the credential and the pack were all fine and only the endpoint was gone. Both setters now reset to the URL the provider was registered with (the pack's URL first, then the static built-in default), and a provider whose URL exists nowhere else keeps what it carries.
- **DeepSeek requests now match the provider's current wire contract.** The official endpoint receives `max_tokens`, native `low` effort stays low, and Strict Function Calling is no longer advertised on the non-beta endpoint. Thinking requests with tools omit the `tool_choice` field that DeepSeek rejects even when it is `"auto"`; reasoning-content replay remains intact. HTTP 402/422 and `finish_reason: "aborted"` now take their deterministic billing, request-body and cancellation paths instead of generic retries or errors.
- **A silent or keepalive-only provider stream no longer looks productive for five minutes.** Unary tracks bytes, complete SSE frames, valid JSON and semantic events separately, applies a bounded first-semantic-event deadline, and reports only sanitized counters and elapsed times. This makes an upstream inference stall distinguishable from authentication, request validation and a disconnected socket.

## [0.2.193] - 2026-09-14

### Fixed

- **A model from a provider pack can now actually run.** Listing a pack's provider was only half the path: five production sites still resolved the *model* through the compiled-in catalog, which by construction cannot contain a model that only a pack contributes. Selecting such a model therefore failed at execution time instead of at selection, with a misleading "provider … with api … is not implemented" error. Model resolution now uses the runtime registry, like the provider list already does.

## [0.2.192] - 2026-09-14

### Fixed

- **A provider added by a provider pack now exists in the app at all.** The provider list was built from the built-in, compiled-in catalog, which by construction cannot contain a provider that only a pack registers. A pack under `~/.unary/extensions/<extension>/providers/*.pack.json` loaded cleanly, registered its provider in the runtime registry, and resolved its models there — but the one list the settings dialog renders and every model picker filters never saw it, so a self-added provider was invisible everywhere at once with no error anywhere. That list now comes from the runtime registry.

## [0.2.191] - 2026-09-14

### Fixed

- **A provider you added yourself now appears in the model pickers.** A provider contributed by a provider pack (`~/.unary/extensions/<extension>/providers/*.pack.json`) starts life without an entry in `settings.json` — it cannot have one before it is first visible enough to be edited. The provider list treated that missing entry as "not configured" and overwrote the verdict the credential store had just given it, so a provider with a working API key was listed in Settings but absent from every model picker, the agent-mode model list, research and plugin model selection. A missing entry now means "enabled with defaults", exactly as it already did for local providers; switching a provider off explicitly still switches it off.
- **The Open IDE button no longer opens a half-finished second app.** It is disabled with a "Unary IDE — coming soon" tooltip while the Desktop app is the only surface under development.
- **The pinned chat header keeps naming the newest request while you read back through a long session.** Reading older history prepends pages and evicts the tail, so the loaded window stops holding the newest real human request while it still holds older user rows — and the header answered with the newest row the window happened to contain. At the top of a long conversation that is the FIRST request, which is what the header then showed instead of your last one. The authoritative anchor now owns the identity whenever that request sits outside the loaded window on either side, and only hands it back to a loaded row while the anchor itself is inside the window.
- **Saving settings can no longer lose the settings file.** The writer rendered the new state to a `.tmp` sibling and then, on Windows, DELETED the live `settings.json` before renaming the temp onto it — and deleted the temp too when that rename failed. A process terminated inside that window, or a rename the OS refused, left no settings file at all: the app then started on defaults, silently, with every provider, model-visibility and MCP entry gone and nothing in the log. The rename replaces the target directly now, transient sharing violations are retried, and a refused replace keeps both the previous file and the pending copy, naming the latter in the error. The same delete-then-rename shape is gone from `editors.json` and the IPC registry.

## [0.2.190] - 2026-09-13

### Fixed

- **Worker admission and execution use the same resolved account.** Account-specific rate limits are checked before a worker starts, credential binding is acknowledged by its session actor, and resume prepares against the actual child model rather than a potentially stale role default. Bound workers reject unrelated key overrides while retaining refreshed tokens for their admitted account.

### Added

- **A session can move itself into an existing worktree.** When a worktree already exists — created from the Worktrees panel, by an earlier turn, or by another session — the agent can now relocate the session into it (`session` action `switch_workspace`) instead of writing into that tree from the main checkout, where the system prompt names the wrong directory, `search` treats the gitignored `.unary` tree as out of scope and `git` keeps reporting the other checkout. The target must be a worktree of the session's own repository: a bare repository, another repository and a plain directory are refused with the session left exactly where it is. The next tool call already runs in the new tree; the working directory and system prompt follow at the turn boundary (one prompt-cache invalidation). A move back to the main checkout is the same action. The Worktrees panel offers the same move for the focused session, and a session moved into a branch-named tree keeps its Goals board instead of re-keying to an empty one.

### Changed

- **The live canary no longer requires a USD ceiling.** Cost limits apply only when explicitly supplied; missing cost measurements remain visible. This does not claim a completed 24-hour endurance test.

- **The chat header's worktree button now opens a dialog with a generated branch name.** One click on the branch icon opens "Create session worktree" with the managed `unary/session-<word>-<word>` name prefilled — two words instead of the session-id hex — and the name stays an ordinary editable input. The managed destination folder is shown read-only, Create stays disabled on an empty name, a collision with an existing local branch gets a numeric suffix, and a failure appears inside the dialog instead of a chat entry. The branch pill's inline form keeps its own path; both run the same create flow.

## [0.2.189] - 2026-09-13

### Fixed

- **Starting one session no longer blanks the whole "Needs attention" group.** A project the worktree overview could not measure — cut off for budget, or because `git worktree list` raced the fold-back that a session start sets in motion — arrived as `truncated` with empty rows, and both the Desktop sidebar and the Sessions dock treated that emptiness as the truth: every row of the project vanished, including trees nobody had touched, and nothing pulled them back until an unrelated refresh. One shared rule now folds every overview answer in (a truncated project keeps the rows it had), the sidebar retries a raced measurement a bounded number of times, the backend keeps its fallback snapshot across the invalidation a fold-back performs, and a failing listing reports Git's own reason instead of an empty list. The dock's notice also stops calling every truncation a time budget.

- **Close and maximize can no longer be pushed out of any dialog header.** The window controls now live in their own fixed track next to — never inside — the shrinkable action row, so a dialog whose action row is too wide clips that row instead of spilling over its own controls.

- **The image viewer's close button can no longer be pushed out of its header.** The dialog title grew to absorb the header's leftover space while the action row — zoom group, Regenerate region, Show in File Manager — asked for the same leftover with a zero flex basis, so the two split it and the actions box landed short of its own rigid content. The row spilled past the panel's right edge and the header's clip ate the window-control group first: a detached image window had no visible X, even though the button was in the DOM and focusable (measured live — the close button laid out at x=1908..1951 on a 1904px-wide window). The title now takes its own width and the actions row owns the leftover — the same rule and the same reason the file preview has carried since 2026-09-08.

- **The pinned chat header now always names the newest user message instead of the one before it.** The header resolved its identity from the scroll position — "the newest user row that has left through the top edge" — so it kept naming the previous request for as long as the newest one was still on screen, which is exactly the state right after sending, and it named an older message again whenever you scrolled back through the transcript. The identity is now the newest real human request (the newest loaded one, or the authoritative anchor when that request is outside the loaded window), and the measurement decides one thing only: whether that request's own row is on screen, in which case the header stays hidden rather than duplicating a message you can already read.

## [0.2.188] - 2026-09-12

### Changed

- **Maintenance release — no user-visible behavior change.** The runtime is the same code 0.2.187 shipped; the only committed difference is the documentation for how a compacted session is handed back, re-measured and corrected.

## [0.2.187] - 2026-09-11

### Fixed

- **An image produced by a tool now reaches the model on every dialect that can carry one.** A `read` of a PNG, a snip capture or a screenshot arrives as a tool-result block, and the OpenAI Responses family, Amazon Bedrock and LM Studio each discarded it with no error and no note — while the same picture attached to a chat message went through. That asymmetry is why the complaint reproduced on one provider and was impossible to reproduce on another. Responses now lifts the image into the user message that follows `function_call_output`, which has no image slot of its own, exactly as the Chat Completions dialect already did; Bedrock emits Converse image blocks in user content and inside tool results, where it previously accepted only text; LM Studio carries tool-result images the same way it already carried attached ones. Where an encoder genuinely cannot carry an image, it now leaves a note saying so instead of leaving nothing.

- **One predicate decides whether a model accepts images.** The wire encoders tested for the exact capability token `image`, while the harness's own capability check also accepted `images`, any token containing `image`, and any token containing `vision`. A model the picker advertised as vision-capable could therefore have every image dropped by the encoder in silence, with the two layers disagreeing in a way neither of them could see.

## [0.2.186] - 2026-09-11

### Fixed

- **The fold-back button works again after a mid-run move into a worktree.** When a session was relocated into its managed worktree while running, the chat's fold-back button could answer with a context-free "This worktree was not folded back." because the live session handle still pointed at the parent checkout. The button now passes the worktree it already knows, and the harness accepts it as the fold target — but only when that tree demonstrably belongs to the session. A genuine refusal now names its reason instead of showing an empty dialog.

- **DeepSeek keeps its wire format behind a custom base URL.** A DeepSeek gateway reached through a reseller or a self-hosted endpoint was recognised by hostname alone, so it silently fell back to plain OpenAI semantics — `store`, the developer role and a forced `reasoning_effort: low`. The dialect is now decided by the provider id as well, so a custom DeepSeek base URL keeps its thinking format and its reasoning-effort map.

- **DeepSeek thinking turns always replay `reasoning_content`.** DeepSeek's thinking mode requires the assistant's `reasoning_content` on every replayed turn; a tool loop whose turns carried only tool calls, or whose thinking signatures had been restored or compacted away, sent none and the request failed with a 400. Such a turn now falls back to an empty `reasoning_content`, the same way Kimi's does, so the loop continues instead of erroring.

- **A tool call refused for a misplaced field now names where that field belongs.** `additionalProperties: false` is the right contract, but its own sentence said only that `maxSources` was unexpected in `config` — next to a list of valid actions, which reads as though the action were wrong. A refusal now resolves the offending value's path and names its accepted home, e.g. `config.advanced.maxSources` for a `maxSources` sent directly under `config`; a name with no home there gets the properties its node does accept instead of a guess.

## [0.2.185] - 2026-09-11

### Fixed

- **The context meter's breakdown popover opens through the shared portal anchor.** It no longer clips inside the composer on narrow panels, and the popover-position policy guard covers the component again instead of exempting it.

### Changed

- **One click from the chat header to a session worktree.** A branch icon left of the auto-commit-and-push button opens the worktree form with the managed `unary/session-…` branch name prefilled, so moving a session into its own worktree no longer needs the detour through the branch pill.

- **The IDE chat toolbar's model picker refreshes the model registry when it opens.** Newly enabled providers and models appear in the picker without restarting the app, and the same refresh backs the chat's open-picker catalog.

- **Inco Kimi K3 and GLM 5.3 expose their reasoning-effort gradient.** Both models now offer low/high/max effort in the model picker instead of hiding the control; the effort mapping follows the gateway's documented gradient and is recorded as unverified in the provider notes until a live check confirms it.

- **Qwen Cloud Token Plan is a separate provider.** It now has its own `qwencloud-token-plan` identity, Qwen Cloud credential name, catalog and provider pack instead of being represented internally as an Alibaba provider. Existing saved Token Plan configuration migrates to the new identity; Qwen Cloud Pay-per-token remains an independent provider with its own key and billing realm.

- **A successfully folded-back session carries a "Folded back" state in the session lists.** The fold stamps the session header with the moment its worktree landed in the parent checkout; the sidebar row and the sessions panel show the state word from the worktree vocabulary until you open the session, and re-isolating the session retires the state. No color coding: the state is the word, like "Not folded back".

### Fixed

- **Folding one worktree no longer blanks the whole needs-attention group.** The overview builds its rows from `git worktree list`, which answers nothing while the fold removes a tree; that empty answer was cached and served as authoritative and erased every row of the group until the next refresh. A degraded measurement now serves the previous rows (or reports itself truncated so the sidebar keeps its state), and the sidebar keeps its attention counts on transient fetch errors.

## [0.2.184] - 2026-09-10

### Fixed

- **The image viewer opens as its own window with working edit and fullscreen controls.** The Desktop chat's image dialog stayed an in-app overlay even where the app can give a dialog a real window, and the detached viewer mounted the bare shared dialog without the region editor, so "Regenerate region" handed the composer a raw image source where it expected an edit draft instead of opening the editor. The chat path now detaches first — same window, remembered size and bridged actions as every other dialog, with the web build keeping the overlay — the detached window mounts the same wrapped viewer as the overlay, and the header's zoom presets and meta row collapse via container queries at narrow widths so the edit, reveal and fullscreen controls are never pushed out of the header.

- **A session worktree fold-back completes in the installed app instead of failing forever.** The final removal step ran `git worktree remove` with the app's inherited working directory, which is outside any repository — Git answered "not a git repository" on every attempt, the fold rolled everything back, and the retry loop preserved the same tree five times per sweep, which is also why the needs-attention row kept disappearing and returning. Removal is now rooted at the parent checkout, and a regression guard refuses bare un-rooted removal calls. A successful explicit fold-back also emits `harness-sessions-changed`, so the folded session immediately appears under its project instead of vanishing from the session lists until an unrelated reload.

### Added

- **Tool failures have individual audit evidence.** Memory Errors separates recorded calls from older summaries, preserves available worker and turn identity, and shows missing evidence and export limits instead of treating recurring errors as confirmed defects.

### Changed

- **A worktree fold-back that hits a conflict is resolved by the session that owns it, not parked for you.** Closing a session, quitting, the crash-recovery sweep and the retry watcher all used to leave a conflicted session worktree under "needs attention". Every path now hands the conflict to the owning session's agent: the session is resumed if it was closed, it rebases inside its worktree, and the fold runs again in the same turn. You see a conflict only when the agent has tried and explicitly reported which file it could not decide. A worktree you chose yourself is no longer listed as needing attention while it simply waits; its `Fold back` action stays available.

## [0.2.183] - 2026-09-10

### Fixed

- **Decision cards accept JSON-wrapped question arrays.** `request_user_input` now applies the existing schema-guided argument repair before its early hook validation, instead of rejecting valid questions before the executor can repair them. Invalid JSON and invalid card content remain rejected.

## [0.2.182] - 2026-09-10

### Fixed

- **Long-running workers remain supervised and actionable.** Watchdog checks follow the next deadline, stalled workers send durable decisions to the parent without bypassing a manual Stop, and a running subprocess alone no longer counts as progress. Stale acknowledgements cannot reopen a completed worker, and temporary monitoring failures retain a bounded retry schedule.

### Added

- **The checkout remembers what happened the last time something was built.** When several sessions work in one repository, a session that hit a build error had no way to know whether the error was new, since when it existed, or that a sibling had already seen it. Every `cargo_check` and every build or test command run through the shell now records its outcome and the exact error identities into one facts-only ledger per repository (shared across linked worktrees, keyed by commit and worktree so a divergent worktree's red build is never mistaken for the parent's). The agent reads it with `diagnostics` `action=repo_ledger` — red builds, open errors with first-seen session and commit, recently fixed ones — and gets a one-line brief at the start of a turn only when the ledger changed since it last looked. Nothing in it blocks anything, and nothing is injected into another session: reaching the session that saw the error stays an explicit `session` `send_input` or `finding`.

### Changed

- **The compaction handover is a set of pointers again, not a digest.** It used to inline every assistant answer, every reasoning digest and the Goals rows three times over — while the Goals board is durable and handed over anew at the start of every turn anyway. It now inlines every failed tool use, only the newest twelve successes, the newest reasoning digest and a count for the rest, and tells the model exactly where to look for anything missing (the segment ledger, the tool result by id, the repository ledger, the other live sessions). The handover ceiling dropped from 32 KiB to 12 KiB.

## [0.2.181] - 2026-09-09

### Fixed

- **A provider you switched off no longer offers its models.** Turning a provider off in Settings left its selected models in the picker, because every path that builds that list read the visible-model selection and ignored the enabled flag. Choosing one sent the request to the provider you had just disabled — and where two providers ship models with identical names, that arrived as a bare `Invalid API-key provided.` from the wrong account.

- **A pinned model names its provider.** The pinned block is the one place a row carries no provider header, so two providers offering the same model name were indistinguishable there.

## [0.2.180] - 2026-09-09

### Added

- **GPT Image 2.5 is selectable for image generation through the ChatGPT (Codex) account.** The Codex image path sent no model at all, so it could only ever produce whatever the backend happened to default to. The image slot now offers GPT Image 2.5 Sunburst, GPT Image 2.5 Flare and GPT Image 2 under the OpenAI Codex provider, and the chosen line is sent with every generate, edit and inpaint request. Leaving the slot on the plain account entry keeps the previous behaviour.

## [0.2.179] - 2026-09-09

### Fixed

- **A tool call that dies mid-flight no longer bricks the session for good.** When a call ended without ever producing a result, the transcript's last message advertised a `tool_use` nothing would ever answer, and the provider refused every later request in that session — on every account, permanently, with `tool_use ids were found without tool_result blocks immediately after`. The encoder now completes the pairing at the end of the transcript too, so a dead call reads as a failed step instead of a wall.

- **A pending chip disappears once its message is on screen.** The chip was released only when it carried an in-memory mark that a reload — or any append path that did not go through the resolver — never set, so it sat in the composer next to the very message it was waiting for. A transcript row stamped with the chip's id now releases it, because that row exists precisely because the chip was delivered.

## [0.2.178] - 2026-09-09

### Fixed

- **Folding a session back is a decision, not a dead end.** When the parent checkout carried its own uncommitted changes in the same files, the fold refused with git's raw error, the tree parked under "needs attention", and there was nothing to act on because the decision row had just been cleared. A conflict now always names the overlapping files and both ways out, and it is offered as `Keep your version` / `Overwrite from worktree` / `Later` instead of being handed to the agent as a rebase that cannot clear a blocker living in the parent's working tree.

- **A session is no longer a foreign owner of its own work.** The fold-back asked both file-owner lookups with an empty session id, and a claim exists for every write — so every changed path answered with the folding session's own id, the tree waited for "another session" that was itself, and no other session had to be running for it to happen. This affected every session, not one tree.

- **A leftover worktree no longer folds what the parent already has.** A fold that applied its changes and then failed to remove its tree left a leftover whose content was already in the parent; git still rejected the patch because the surrounding code had moved on, so every retry read as a content conflict. Paths that are byte-identical to the parent are now dropped before the patch is applied.

- **Automatic fold-back retries converge instead of repeating forever.** The watcher re-armed the same tree every 30 seconds without a ceiling, and its salvage path rebuilt the row from the snapshot the retry started with — discarding the blockers the refusal had just recorded. Retries now stop after five identical failures and hand over a decision that still names its files.

- **A per-file fold-back decision can no longer delete unfolded commits.** `Keep your version` and `Overwrite from worktree` cover uncommitted changes only, but the path continued into worktree removal and a forced branch delete, destroying every commit the session had made. It now refuses while those commits exist and names both remedies.

- **Shell commands containing an umlaut, an accent or an emoji no longer crash the tool call.** The Windows `nul` redirect rewriter copied the command byte by byte through a string slice, which split the first multi-byte character and panicked the whole call with "is not a char boundary" — with or without a redirect.

- **The composer's Fold back button keeps its size.** With no goals, no changes and no lane to show, the status strip renders nothing and the fixed-size chip inherited the strip's growth rule, stretching into a full-width bar over the composer.

## [0.2.177] - 2026-09-08

### Changed

- **The composer's context gauge opens on click, not on hover.** Moving the pointer across the context fill ring used to unfold a 420px breakdown panel over the composer. The gauge is now a toggle: click opens it, clicking it again or anywhere outside closes it, and Escape closes it too.
- **The Memory panel's subtabs show their own content.** Review and MEMORY.md are content-connected 10px subtabs, and the MEMORY.md tab renders the complete document even when no entry maps to it.

## [0.2.176] - 2026-09-08

### Changed

- **Memory writes are immediate.** Agents can store durable repository or global knowledge directly without a review queue. Historical ungrounded candidates are not silently activated, and automatic reusable-artifact capture is opt-in.
- **Dream uses a model and a fixed interval.** A global app-owned timer classifies eligible memories across repositories with the selected model; session-end triggers are no longer needed. Failed classification preserves the previous conclusions, and Dream cannot overwrite a direct memory just because its title matches.
- **On-demand tools no longer preload their long usage guides.** MCP keeps compact tool summaries; the browser, desktop, research and memory families receive full guidance at activation. Exact names, action schemas and persistent safety boundaries remain available.

### Fixed

- Corrected the fold-back deadline watcher's weak session-registry type so the harness builds without a nested `Arc` mismatch.

## [0.2.175] - 2026-09-08

### Fixed

- **MCP tools no longer need a new session after install.** A catalog refresh (an MCP server installed, a plugin toggled, a settings write) that arrived while a turn was still running was dropped, so the new tools stayed invisible until the user opened a new session — which the install message then told them to do. The refusal is now recorded and settled at the next safe turn boundary: the tools are advertised from the next turn of the same session, and the install message says so.

- **Compaction markers say why they ran and what they freed.** A context compaction in the transcript names its trigger in plain words ("Auto-compacted", "Compacted on request", "Compacted after context overflow", "Compacted to continue the turn") and the token delta before → after. A failed compaction states that the context was left unchanged, and a marker never implies a delta it did not measure.

- **A removed file preview says why it is gone.** The browser runtime records a bounded reason for every removed preview session, so a viewer polling one that was already dropped gets the cause — a window-ownership reap versus ordinary cleanup — instead of a bare "was not found", and the reap can no longer close a session that was started after its window died.

### Changed

- **Yolo mode is permission, not minimalism.** The mode prompt said "make the smallest change that works", which read as a quality cap and produced quick fixes instead of production-ready work. It now frames the mode as pure permission — everything allowed, no gate or ceiling, plan and switch to execution — and asks for a complete, production-quality solution instead of the first patch that touches the symptom.

- **The Quick profile defers verification without lowering the bar.** The verification-policy line now states that it governs only when builds, tests and lints run, never the quality of the change itself: quality comes from how the work is built, not from running checks, and a Quick session ships the same production-ready solution while handing the verification commands to the user instead of running them.

## [0.2.174] - 2026-09-07

### Fixed

- **Pilot reports underlying controller failures instead of hiding them behind a generic stop message.** Provider errors retain bounded details, and interrupted controllers preserve available supervision evidence without claiming that the user pressed Stop.

### Changed

- **Pilot receives bounded desktop tasks rather than whole programming jobs.** Native API and batch artifact preparation stays with the main agent; desktop controllers are instructed to observe the screen before extended planning and work in short UI steps.

## [0.2.173] - 2026-09-07

### Added

- **Detached file previews open at the size they were designed for, and remember where you put them.** A preview opened in its own window derives its placement from the geometry you last gave that window, or from the view's own design size clamped into the monitor's work area, instead of an arbitrary default. Below 860px the secondary header actions collapse into the shared overflow menu while the primary toggles and the window controls stay flat, and a preview session now belongs to the window that started it — closing that window ends it, and a session the runtime no longer knows about shows a designed "session ended" state with Retry instead of a raw error.

### Fixed

- **A message you sent while the agent was idle no longer arrives after the next one.** A steering message that missed the run loop's drain point stayed queued until the following prompt started and was then appended *behind* that newer prompt — measured once at 75 seconds early, delivered 1 millisecond late. Queued messages are now flushed first, in the order the UI shows them, so the transcript matches the order you sent them in.

- **Terminals and shells no longer see the Windows verbatim path spelling.** A canonicalized root (`\\?\L:\Repo`) is correct for comparing two paths and wrong for everything else: PowerShell cannot map it to a drive and rendered the provider-qualified `Microsoft.PowerShell.Core\FileSystem::` prompt, and a terminal group keyed on it silently forked one repository into two — so the new tab was created in a group nothing rendered. The prefix is now stripped once per language at the ingestion points, project entries written by earlier builds are repaired on read, and comparison sites keep canonicalizing.

- **The "Plan ready" card disappears once you answer it.** Accepting or declining a plan the model wrote into chat submits a harness-authored reply rather than a user message, and the check that decides whether a plan is still awaiting an answer only stopped at user messages — so it walked straight past your decision and remounted the card on every state rebuild for the rest of the session.

- **Clicking a folder in the editor opens its entry file instead of claiming "Permission denied".** Windows reports a directory read as access-denied, which surfaced as a file-permission error for paths that were simply directories. Reading a directory now says so, and the editor retries the usual entry files (`index.ts`/`tsx`/`js`/`jsx`/`mjs`/`svelte`, `mod.rs`, `__init__.py`) and opens the one it finds.

### Changed

- **Both Qwen Cloud realms are one brand, and a realm mismatch says so.** The credits-based Token Plan and pay-per-token DashScope International are one product line bought in one place, but the surfaces named them so differently that a valid Token Plan key refused by the metered host read as a broken credential. They are now "Qwen Cloud (Token Plan)" and "Qwen Cloud (Pay-per-token)", listed next to each other in settings and the model picker, and a 401 on either host names the sibling provider as the likely realm. Provider ids are unchanged, so stored accounts and picked models keep working.

- **A session worktree now folds back entirely on its own.** After every turn, a session that was isolated into a managed worktree returns automatically: its commits are cherry-picked onto the branch recorded when the tree was created (switching the checkout back itself when it drifted, never forcing past uncommitted work), the session reopens in the parent checkout, and the worktree and its branch are removed — no button, no dialog. If the fold stops at a real conflict, the session's own agent is woken with a machine instruction to rebase its worktree onto that same recorded branch and resolve the conflicts there; the automatic fold retries the moment that turn ends. Blocked files are retried by the existing release hook the moment their claim drops, and the automatic return survives an app restart. A `Fold back` remains as one action for the times you want it immediately: it runs the same automatic fold and hands a conflict to the session the same way — the choice between "Automatically" and "Via session" is gone, because both ran the same fold and differed only in exactly that handover.

- **Machine-assigned instructions reach the model again.** Instructions the app composes on its own — a plan to execute, a fold-back conflict to resolve — were stored as a message type that was never projected to the model, so the model silently never saw a single one of them. They now ride as a distinct machine message the model reads verbatim, shown in the transcript as a system notice instead of an invented user entry.

- **Session titles are always generated by the session's own model.** The dedicated rename-model setting is gone end to end: auto-rename (and its refresh pass) runs on the model the session is already using, with no configured override and no fallback model in between. The settings UI keeps only the generation's thinking effort.

## [0.2.172] - 2026-09-07

### Added

- **A session folds back only into the branch it left.** Folding a session's work back into its parent checkout now checks the branch recorded when the session started: an automatic fold-back into a parent that switched branches in the meantime is refused, and the explicit path names the mismatch instead of silently writing into the wrong branch. The action also reads the same in every surface now — chat, menus and dialogs share one canonical label. Isolation can no longer nest inside a managed worktree, and sessions isolated at a write seam return to the parent checkout automatically.

- **Worktrees that already gave up their work no longer linger.** When a worker's patch is already applied in the parent, its worktree is recognized as equivalent and cleaned up on close instead of being preserved as "not folded back"; a worker preview shows the parent state before the close decides, a worktree that is still busy reports its path instead of failing anonymously, and a retried worker receives bounded context about what its predecessor changed.

- **The memory store keeps a locator where it used to keep your words.** Evidence that captured verbatim user quotes is scrubbed to a reference to the turn where the words were said, across every memory scope in a one-time migration — and new captures record the reference instead of the raw text.

### Fixed

- **A single question and answer now shows its span and token tally.** The seam between assistant turns only appeared from turn 2 onward, so the most common case — one prompt, one answer — showed neither elapsed time nor cost. The first turn now gets the same seam, measured from the moment the prompt was sent.

- **Parallel worker batches with a shared contract no longer fail on each other's code.** When one worker implements a new interface and its siblings consume it, each consumer's change was compile-checked alone and rejected for missing definitions that existed only in a sibling's unapplied patch. The check now overlays the whole batch, tests the union, and waits within a bound for a sibling that has not delivered yet.

- **A worker integration that goes sideways now says what it did and why.** A worker that broke files it does not own integrates with an explanatory note instead of being silently preserved when the parent workspace is dirty or unmeasurable, and a stalled integration flush reports what it is waiting on — retry state, completed writes, sidecar path — instead of hanging without a word.

- **On Macs, GPU embedding picks a real GPU.** The index's GPU lane could end up on the CPU when adapter selection failed on macOS; it now selects a physical Metal GPU and validates the adapter before running.

## [0.2.171] - 2026-09-07

### Added

- **A fallback vision model can take over when the configured one cannot serve an image.** Settings → AI Agent gains "Fallback vision model" (`visionFallbackModel`): it is tried whenever the primary slot fails — every account of its provider refused, the call errored or timed out, or its own configuration is broken — and a fallback-only selection carries the whole route, so a text-only active model still gets its images read. Pilot tool-result images inherit the same fallback, and a turn whose every attempt is blocked still fails closed, naming both reasons.

- **Deep Research now searches as broadly and as deeply as its depth setting claims.** A job no longer sends four fixed phrasings of one topic to a single page per provider. Query variants are generated per job — the exact phrase, the current and previous year, and the angles a research question actually needs (documentation, comparison and alternatives, limitations, measured performance, migration, how-to) plus sub-questions parsed from your instructions — and every provider is asked for more candidates than the final evidence budget and paginated, so host diversity has something to choose from. Standard depth issues 8 variants and over-gathers 2×; deep issues 12 and 3×.

- **Full text is fetched for the sources that deserve it instead of waiting on the model.** The search wave upgrades the highest-ranked snippet-only sources to full text on its own, preferring one source per independent host before taking a second from any host, and follows primary-source-looking links inside fetched pages. `followLinksDepth` was a setting every job stored and nothing read; it now bounds that chasing.

- **A research job can investigate rather than search once.** Up to two further waves derive their queries from terms that appear across at least two independent hosts and are not already in the topic, and every wave is deduped, host-diversified and domain-filtered by the same rules as the first.

- **Four controls for that breadth and depth**, in Settings → Research and in the per-run dialog: query variants, gather multiplier, fetch depth and expansion rounds. Each defaults to "auto — from depth", each shows the derived value, and each states what it costs in provider requests rather than only what it does.

- **The local half of a research run is held to the same contract as the web half.** `mode: both` is the default, so half of every run is workspace retrieval — it now issues the same generated variants, attaches real file content for its top-ranked hits instead of metadata alone, collapses near-duplicates through one identity, ranks instead of returning whatever came back first, and reports a missing or stale index as a distinct status with a reason.

- **Research quality is measured, not only reviewed.** A new offline evaluation scores a recorded job — citation precision and coverage, host independence, fetched ratio, provenance completeness, engine coverage, identity hygiene, contradiction preservation and cost per cited source — against golden fixtures, so a change to the ranking or the gate produces a number instead of an impression.

### Fixed

- **Vision helpers now rotate provider accounts the way a chat run does.** An image transcription used to resolve one API key and call the provider once, so a rate limit, auth failure or quota error surfaced as a dead vision model while the chat model rotated past the very same account. It now walks the provider's enabled accounts under the executor's own rotation policy, attributes the limit headers to the account that earned them, and names every account spent or skipped — plus how much rotation is left — in the note the model reads.

- **The search scraper tier no longer bypasses the SSRF protections the fetch tier already had.** Public-web scraping used a plain HTTP client that followed redirects with no per-hop address validation, so a redirect or DNS answer pointing at a private or link-local address could be fetched and parsed into evidence. Scraping now goes through the same protected client as page fetches: redirects validated per hop, every DNS answer checked, ambient proxies disabled.

- **A search that ran and found nothing is no longer reported as "skipped".** Engine status now distinguishes failed, zero-result and completed-with-warnings, keeps the provider name with each error instead of discarding it, and never produces an empty reason — so a silently failing search provider shows up in the job instead of looking like an engine you never enabled.

- **The evidence budget can no longer be filled entirely with unfetched snippets.** The rule that at most half of the sources may be snippets was enforced on one insertion path and not on the ones used when appending a source or persisting a search, so a job could finish with nothing but snippets and nothing to cite from.

- **A cited paragraph is no longer excused by the paragraph above it.** The finalization gate skipped headings without closing the current paragraph, so a cited claim, a heading and an uncited claim counted as one cited paragraph. Uncited paragraphs now fail the gate.

- **Job files are written atomically and job ids can no longer alias.** Saving a job used a fixed temporary filename with no sync, beside a correct atomic writer the same code already owned, and two different job ids that sanitize to the same string addressed the same file. Saves now go through the atomic writer, a non-canonical id is rejected before any path is built, the stored id is verified against the requested one, and job files that cannot be parsed are reported as `corruptJobs` instead of silently vanishing from the list.

- **A crashed finalization no longer imports the same report twice.** The intent to import is persisted before the side effect, a retry adopts the document the earlier attempt created, and the import carries a stable identity so the doc side updates the existing page rather than adding a duplicate.

- **Cached and revalidated content says so.** A 304 revalidation was recorded as the source's own status code, making a stale body look like a fresh non-200 response, and a persistent cache hit entered the evidence with no sign it came from cache. The origin status is recorded and a cache hit is annotated with when it was stored and when it expires.

- **Research evidence no longer names a browser engine that does not exist.** Persisted source notes said a page was "rendered with direct Playwright" and a health check reported Playwright as available unconditionally, for a runtime that was removed. Both now name the managed Chromium runtime and report whether it is really reachable.

- **An unterminated `<script>` or `<style>` no longer leaks into extracted text.** Excluded regions were stripped by patterns that required a closing tag, so a truncated page rendered its script body as evidence — a prompt-injection path into a source excerpt. Excluded regions are now removed as whole subtrees, to end of file when the close never arrives.

- **One citation contract, enforced identically in both places.** The research gate and the document import disagreed about what a valid citation is, so a report could pass the gate and then be refused at persistence, failing the job after it was already declared good. Both now reject a bracketed digit-and-separator span such as `[1,2]`, both treat other bracketed prose such as `[v2]` or `[note 1]` as text, and a shared case table is asserted row by row across the two crates.

- **A research status poll no longer stalls the rest of the turn.** Ten read-only research and docs actions were classified as reads but left off the parallel-safe list, which the scheduler treats as a barrier that also blocks every later hoist.

- **Settings number fields clamp what you type.** `min`/`max` attributes are advisory only, so a typed or pasted 99 or 0 was persisted and silently disagreed with the value in effect. The fields clamp in the handler and show the clamped number.

- **Two remaining SSRF edges are closed.** NAT64 addresses such as `64:ff9b::7f00:1` embed an IPv4 address and were passing the IPv6 filter, and the port policy was a silence rather than a decision; the embedded address is now checked with the full IPv4 policy and ports are an explicit allow-list.

- **The NotebookLM engine now says that it cannot see your NotebookLM settings.** It runs with defaults because Deep Research has no way to read the configuration you saved in the NotebookLM panel; if you use the consumer service nothing changes, but a Discovery Engine setup was being ignored without a word. The caveat now rides in the engine status instead of presenting as an unexplained authentication failure.

- **Building from source works again after the native-resolution region-crop change.** A new parameter was not passed at three call sites, which broke compilation for anyone building the workspace. Observation calls that only want window metadata now request no crop.

## [0.2.170] - 2026-09-06

### Fixed

- **Terminal Up/Down history is reliable and shared across every terminal tab.** The frontend no longer replays its own in-memory command list with counted backspaces — a mirror that silently switched to the shell's separate history after any Tab, cursor key or paste, and forgot everything on reload. Arrow keys now always reach the shell, whose history the integration scripts share live between all Unary terminals (PSReadLine incremental shared file, bash `history -a/-n`, zsh `share_history`).

## [0.2.169] - 2026-09-06

### Changed

- **Burn avoids unnecessary work when embedding mixed-length inputs.** The CPU groups similar lengths while preserving caller order; the GPU keeps its existing batch grouping. Both paths copy back only the CLS vectors instead of every token's hidden state. The model, FP32 precision and allocation limits remain unchanged.

### Fixed

- **Human Stop does not silently drain queued input.** Steering and follow-up messages remain held until an explicit prompt or resume; an authorized peer wake can also start a session whose history is still empty without losing its provenance.

- **Original/verifier cleanup cannot close the same actor twice concurrently.** Close operations for one parent session are serialized, and real cleanup failures still propagate.

- **Shared panels and session controls stay consistent across Desktop and IDE.** Operation progress remains visible, unknown Git sync status is not presented as a clean comparison, and detached panels receive the same sync-state information.

- **Older Memory evidence remains readable without weakening quote validation.** Legacy evidence references are normalized to the canonical format, and startup recovery completes privacy cleanup before publishing refreshed context.

- **The direct OpenAI GPT-6 Astra picker exposes its documented Low-through-Max ladder.** Unsupported Off/Minimal choices are not offered for that API model.

- **Concurrent embedding requests no longer starve behind repeated work on the same device.** CPU and GPU admission is FIFO, expired waiters leave the queue cleanly, and the existing request and queue deadlines remain unchanged.

- **The embedding sidecar exits when its parent is already gone.** The parent-process check now runs before loading the model, avoiding an orphaned model load during shutdown.

## [0.2.168] - 2026-09-05

### Changed

- **GPT-6 Astra is available as the new Codex model.** Unary now exposes `gpt-6-astra` through the ChatGPT-account Codex transport and the documented direct OpenAI Responses transport, with its 1.05M context, Low-through-Max reasoning ladder, image input, and provider-specific Fast-mode encoding.

- **The code index no longer has a size ceiling.** The snapshot used to be one pretty-printed JSON file carrying every chunk's source text (264 MiB for this repository, a fifth of it indentation) and refused to index above 512 MiB. It is now a compact metadata snapshot plus an on-demand chunk-content store: huge repositories index, a one-file refresh no longer rewrites hundreds of megabytes, and an open workspace keeps far less resident. Existing indexes migrate in place on first load without re-embedding.

- **Index memory is budgeted, not just counted.** Open workspaces are evicted from the index cache when their resident size exceeds `indexResidentBudgetMb` (default 2048), and Index status shows resident MB per workspace and per-stage times of the last reindex.

- **Index search runs its lanes in parallel** and reports per-lane timings (`laneTimingsMs`) so slow queries can be attributed.

### Fixed

- **macOS release builds include the desktop-only Tauri plugins and sign their updater bundle again.** The desktop plugin dependencies had accidentally fallen under the Windows-only Cargo table, and the shell release script populated an environment variable the Tauri CLI does not consume.

- **The embedding sidecar is honest about GPU trouble.** A recovery episode writes one log line instead of twelve panic records, the lowered attention budget survives a restart instead of being re-learned by panicking, truncated inputs and recovered refusals are reported to the host, a malformed request no longer kills the process, and a request that fails to *send* is counted as a death like one that fails to answer. A corrupt crash log is reported instead of silently re-enabling a GPU lane it had condemned; a request queued behind a stalled one waits two minutes, not thirty.

- **Abandoned index stores are named and reclaimable, and the home directory is refused as a workspace root.** A cancelled index of `C:\Users\<you>` had left 376 MB on disk with no way to see or remove it.

- **`code_context` and `repo_map` stay inside `budgetTokens`** including their structured payload, not only their prose.

- **The `index` tool description no longer advertises a LateOn MaxSim reranker** that was removed on 2026-09-02.

## [0.2.167] - 2026-09-04

### Changed

- **Memory recall no longer runs a reranker model.** The `Shared reranker model` plugin setting is removed; an existing value is ignored on load and dropped on save. Recall keeps its hybrid retrieval order without the per-recall reranker latency.

### Fixed

- **xAI video generation no longer reports Grok Imagine Video 1.5 as unavailable.** The authoritative xAI provider pack now keeps the current video model that the video tool resolves, instead of replacing the built-in catalog with a coding-and-image-only subset.

- **Index panel names a missing workspace instead of a missing plugin.** With no workspace open it reported "Plugin: Missing — the index plugin is not registered" and showed built-in defaults as if they were workspace settings; both now say "No workspace".

## [0.2.166] - 2026-09-03

### Fixed

- **Provider sign-in no longer ends on a refused localhost page when automatic return needs recovery.** The recovery flow now opens its localhost callback before sending you to the provider, finishes and saves the account when the browser returns, and still keeps paste-back available when a firewall or browser blocks local delivery.

## [0.2.165] - 2026-09-03

### Fixed

- **A session working in its own worktree now knows the way back.** Before,
  the only hint a moved session ever got was a one-off “continue in this
  directory”, so one session merged its branch into the shared checkout by
  hand and then told the user to close it and start a new one to work there.
  The system prompt now states on every request that the directory is a
  managed session worktree of the shared checkout, that its commits fold back
  automatically on close, that *Fold back to parent* continues the same
  session in the shared checkout, and that merging by hand or asking for a
  new session is wrong. The working-directory line also follows a session
  that is moved mid-turn instead of staying on the old path.

## [0.2.164] - 2026-09-03

### Fixed

- **“Fold back to parent” shows up for a session that was moved into a
  worktree after it started.** The composer decided from the branch recorded
  when the session was created — the parent’s branch, which the conversion
  never rewrote — so every converted session failed the “is this a managed
  session worktree” check while the header pill already showed its
  `unary/session-…` branch. The button now decides from the live branch of
  the session’s directory — the same branch the fold-back itself checks; the
  stored branch only stands in until the live one is known.

## [0.2.163] - 2026-09-03

### Fixed

- **Region capture works again on Windows.** Clicking the region snip button
  failed with *“Plugin tool `snip` has an unclassified action `snip_region`”*
  before the picker could open: the native capture was missing from both of the
  lists that decide which snip actions may save a capture, so it was refused
  outright. It is now handled like every other saving capture, and a slow
  selection — the picker stays open as long as you hold the mouse — no longer
  reports a failure after a minute while the capture still lands in the
  gallery.

## [0.2.162] - 2026-09-03

### Changed

- **Guardrail profiles: Quick is the default, Standard is now Normal, and
  Strict stopped paying for nothing.** Verification is something you switch
  ON by picking Normal or above; a fresh installation runs Quick. Strict used
  to verify after EVERY edit (a full context re-send per change), demand the
  full repository gate, and reject any ambiguous edit anchor although the
  repair only ever picks the one occurrence whose surroundings are unique —
  it now verifies once at the end at workspace depth and repairs like
  Cautious; what still sets it apart is rejecting a stale edit target and
  asking before recoverable git operations. Saved `Standard`/`Flow` submodes
  and settings files keep resolving to Normal. An installation without a
  `guardrails` block now resolves the default profile in the prompt layer,
  the fold-back check and the edit guard too — previously only the worker
  verifier did, so a default install would have run half a profile.

### Fixed

- **Worker patches no longer end in “conflict” because of their own parent.**
  A second round of workers in one long session was marked failed with
  *“is being written by ‹parent›”* although every patch applied cleanly: the
  parent's own attribution claims from round one were read as a live write
  lease for as long as its turn ran, and the resulting deferral was fed to the
  conflict resolver. Integration now acts for the parent session, so its own
  claims never defer it; a deferral is reported as a deferral and the patch
  waits for the next safe boundary instead of failing. A “worker had
  conflict” badge whose patch was meanwhile applied by hand now clears on its
  own at the next boundary.

## [0.2.161] - 2026-09-03

### Changed

- **Branches panel: three scopes instead of nine sections.** Source Control ›
  Branches now shows Branches (Local, Remote, Pull requests, Stashes), Workers
  (every worker folder under `.unary/worktrees/` in one list, grouped by the
  session that owns it, with running / finished / leftover / conflicts as tags)
  and Snapshots (grouped by session, orphaned ones as their own group). The
  twelve header buttons are gone; every action lives in the ⋮ menu of the
  thing it acts on. **Clean up…** replaces the old remove-all chains with one
  dialog: pick a policy for finished work (merge into main first, or remove),
  tick sessions or single folders, optionally delete their branches and prune
  orphaned snapshots, read the consequences once in a summary block, and
  confirm once. **Start worktree review** — the session that walks through
  every worker folder and proposes what to do — is reachable from the head
  menu, the Workers bar, every session group and the Clean up dialog.

## [0.2.160] - 2026-09-02

### Added

- **Text panel with learning writing styles (Assistant mode).** The
  Desktop Assistant gets a "Text" side panel — a small editor next to the
  chat with grammar, rewrite, shorten, improve and tone actions on the
  selection. Styles carry a name and the user's own description, explicit
  rules, every approved version as an example, and traits derived from
  examples and corrections that the user confirms or rejects.

## [0.2.159] - 2026-09-02

### Changed

- **Memory captures what matters and stays silent about the rest.** The
  agent of the running session is now the primary memory author: a user
  correction, standing preference or durable decision is stored at once as a
  cleanly rewritten rule with the user's exact words kept as evidence, and
  becomes trusted immediately. The old regex preference detector no longer
  promotes anything on its own — a hit becomes a one-shot hidden hint that the
  running agent decides in the same conversation — and the one-off task
  instructions it had written into `MEMORY.md` are demoted on the next start.
- **Session compaction is loud and lazy.** It runs after a session ends or
  sits idle for ten minutes, skips one-turn and worker sessions, falls back to
  the shared knowledge model when the memory model does not answer, and
  reports pending sessions and the last failure (with model and error) in the
  Memory panel status and in the injected memory index instead of failing
  silently.
- **One trust ladder for every automatic memory write.** Rollout distillation
  now produces structured candidates through the same grounding and
  auto-promotion path as compaction artifacts; nothing writes model prose into
  `MEMORY.md` directly.
- **No manual memory mechanics except delete.** Helpful/Not-helpful,
  promote/pin, refresh, compact, rebuild and retention buttons are gone;
  promotion, usefulness (from recall usage), retention, stale sweep,
  embeddings and health run automatically, throttled to once per day per
  repository.
- **Memory panel v2.** Five scopes — Ledger · Session · Repo · Global ·
  Errors — replace the per-scope side rail. Repo and Global show one entries
  surface with a compact Review · Trusted · All switch where Trusted renders
  the composed `MEMORY.md`; the file opens from the head menu. Session shows
  what was loaded and what this session created; Ledger is the session fact
  sheet (worked on, solved, errors fixed, files, commands), now derived
  deterministically even when no model pass ran.
- **RPC sessions end cleanly.** Transport EOF or error now runs the session
  shutdown, so plugins receive `SessionEnd` and repository leases are released.
- GLM-5.3-Flash now uses its exact 1,048,576-token native context window. The
  1,310,720-token figure shown by OpenRouter belongs only to its Cloudflare
  route and is not advertised for direct Z.AI requests.

## [0.2.158] - 2026-09-02

### Changed

- **Claude Fable 5.1 works through Anthropic subscription accounts again.**
  Unary now identifies its OAuth requests with the Claude Code 2.1.258 client
  version required by the model instead of the rejected 2.1.75 version.

- **Product code moved out of the host packages into `shared/`.** Everything
  under `unary-tauri/src/` except the host and IDE layout (`main.ts`,
  `Shell.svelte`, `App.svelte`, `shell/`, `surfaces/`, the window-opening
  `popout/` helpers, the palette UI, `remote/`, the IDE-only `Dock*Panel`
  wrappers and a few IDE-only components) now lives in `shared/app/` — including
  the palette registry, `openTauriPanelWindow.ts` and the four dock panels the
  Desktop also mounts (Browser, Sidebar, Extensions, Git) — and everything
  under `unary-desktop/src/` except its
  `main.ts`, `App.svelte` and `snipWindowControls.ts` now lives in
  `shared/desktop/`; the `@unary/app-src/*`, `@unary/app-api/*`,
  `@unary/app-keymap/*`, `@unary/app-themes/*` and `@unary/desktop/*` aliases
  keep their names and point at the new trees, and the dead git history/graph
  tabs, edge rail, session branch group, markdown performance monitor and
  tree-worker client were deleted.

### Changed

- **The index has three modes and one engine.** Settings › Index now offers
  Lexical only, CPU and GPU. The two model modes run IBM Granite Embedding 97M
  R2 on the bundled pure-Rust Burn sidecar, on the CPU or on the GPU through
  Vulkan, D3D12 or Metal. ONNX Runtime, Dawn, the 311M model, the Qwen
  models, the LateOn late-interaction stage, the static fallback model, the
  per-repository model override, the semantic-search switch and the remote
  embedding ceiling are removed. The sidecar ships as one self-contained
  executable with no runtime libraries beside it. A settings file from an
  earlier build keeps the lane it named: every former CPU preset becomes CPU,
  every former GPU preset becomes GPU. Existing indexes built with the 311M
  model are rebuilt on the next reindex.

### Added

- **A "Lexical only" index mode for machines that cannot carry a model.** It
  loads no embedding model and starts no sidecar; exact symbol search, BM25
  text, path trigrams, the symbol graph, repository map and context packing
  stay, and every search result says that natural-language intent search is
  unavailable. A repository cannot override it.
- **The code index understands 32 more languages.** CSS, JSON, YAML, TOML,
  SQL, Dart, Haskell, OCaml, Objective-C, HCL/Terraform, GraphQL, Julia, R,
  Erlang, Clojure, Nix, PowerShell, CMake, Make, Solidity, Elm, Gleam, Groovy,
  F#, GLSL, HLSL, Starlark/Bazel, Fortran, Verilog, VHDL, GDScript and D now get
  tree-sitter symbols, definition and call tags, symbol-graph edges and
  declaration-sized chunks instead of plain line windows. Bazel `BUILD`,
  `WORKSPACE` and `MODULE.bazel` files are recognised by name.
- **A search on an index without embeddings says so.** When a conceptual
  query planned a semantic lane the profile cannot run, the result now names
  the reason and the repair step in both its text and its structured
  `semanticError` field, instead of answering as a hybrid search with two
  silent empty lanes.

### Fixed

- **Large workspaces no longer take the interface down.** Flat workspace scans
  are bounded again on the host, the Desktop Files tab says when a listing is
  partial, and transcript row limits no longer grow with installed RAM. On
  a machine with a large `target/` folder the previous release scanned millions
  of entries per refresh until the renderer ran out of memory.
- **Opening a workspace with many sessions no longer re-applies plugin settings
  once per session.** A plugin that has already applied exactly the requested
  configuration is skipped, so one slow plugin at startup no longer makes every
  session wait on the full settings timeout.

## [0.2.157] - 2026-09-01

### Added

- **Claude Fable 5.1 is available through Anthropic.** The model includes its
  one-million-token context, 128K output limit, current pricing, always-on
  adaptive thinking, Low-through-Max effort controls, tool search, and safe
  handling when a conversation switches models.

### Fixed

- **Large workspaces no longer freeze the Explorer or disappear behind a fixed
  entry limit.** The shared IDE and Desktop tree now loads direct children in
  cancellable pages, renders the first page immediately, and fetches nested
  folders only when expanded.

## [0.2.156] - 2026-09-01

### Added

- **Tencent HY4 Preview is available through OpenRouter.** The model ships with
  its current text, reasoning, tool-use, context, output-limit, and pricing
  metadata in the seeded OpenRouter provider pack.
- **Unary now has a dedicated Android application target.** The mobile build
  uses its own lightweight entry, registers only the supported sign-in and
  remote-control commands, and validates the Android SDK and NDK explicitly.
- **The Automations workspace is available in both desktop surfaces.** Event
  triggers can be configured from the same shared panel instead of a host-only
  copy.
- **The Index graph shows details for the selected node, relation, or cluster.**
  The explanation stays in a dedicated lower section while the graph remains
  visible.
- **Assistant sessions can switch between Local and Remote at the top of the
  sidebar.** Remote hosts expose their connected, loading, empty, and error
  states without mixing remote sessions into local projects.
- **Managed worktree sessions can fold their completed work back from the
  composer.** The attention action closes the temporary session only after the
  fold succeeds and reports the resulting parent outcome.

### Changed

- **Model, mode, Fast, and effort controls now form one compact inline control
  family.** Fast remains an on/off action while model, mode, and effort keep
  their own menus.
- **GPU Fast indexing now uses Unary's native accelerated model runtime.** The
  legacy browser-style GPU path is gone, and a lost device is reported as an
  explicit GPU-lane failure instead of silently continuing on an unsafe path.
- **Destructive dialog actions are consistently red, and the remaining dialogs
  open as independent windows when the host supports it.**
- **Full Windows releases reuse substantially more completed build work without
  introducing a reduced release mode.** The canonical signed, inspected and
  R2-publishable path now has observable compiler caching, retry-safe build
  caches, bounded temporary Defender exclusions, reusable dependency installs
  and artifacts preserved across packaging. macOS and Linux also avoid the
  duplicate model-runtime build walk and report the same phase timings as
  Windows.

### Fixed

- **A failed local release build no longer burns a version number.** Build,
  gate, packaging, signing, and incomplete publication failures retry the same
  version safely; only a version visible in the live R2 channel is final.
- **Stopping a session is now a hard manual barrier.** Pending background exits,
  continuations, watches, worker completions, and queued input cannot wake a
  session again after the user presses Stop.
- **Anthropic sign-in recovers cleanly from concurrent refreshes and revoked
  credentials.** Refresh-token rotation is preserved across processes, and an
  invalid grant clears the unusable credential instead of failing repeatedly.
- **Pinned and archived chats keep their project context.** Pinned chats are
  grouped by project and then branch, archived chats stay in one nested
  subgroup, and branch groups no longer duplicate their own session rows.
- **A failed fold-back keeps the session available for recovery.** Unary no
  longer closes the worktree session before confirming that its work reached
  the parent.
- **Codex Voice connects through the supported path again and lets the active
  microphone be selected.**
- **Drawing and composer resize affordances are visible and stable again.** A
  buttonless pointer move no longer ends an annotation stroke, and both resize
  grip lines render at the intended contrast.

## [0.2.155] - 2026-09-01

### Fixed

- **Model-only image and video selectors no longer expose an inactive effort
  action.** When reasoning effort is unavailable, both desktop surfaces now omit
  that action instead of wiring an empty handler.

## [0.2.154] - 2026-09-01

### Fixed

- **Image and video model settings no longer block a desktop update.** Their
  model pickers now remain valid when reasoning-effort selection is unavailable.

## [0.2.153] - 2026-09-01

### Fixed

- **Every stored chat message is reachable again — and the first scroll no
  longer jumps.** Long sessions now show explicit controls for older and newer
  history, keep failed loads retryable,
  and preserve your reading position while a page is added. The pinned user
  message also remains authoritative when it is outside the first loaded page;
  clicking it fetches the containing page and lands on the original request
  instead of silently losing it. The IDE and Desktop app use the same production
  chat behavior, with normal-flow rows and a stable reading anchor preventing
  the first real scroll from leaping away from the content you were reading.

- **Long chats compact without forgetting or repeating earlier work.** The exact
  handoff boundary survives an interrupted save and is repaired when the chat is
  opened again, while a failed save cannot leave the live conversation ahead of
  its history on disk. Manual compaction now reports the right cause, shows one
  completion marker, and displays failures instead of hiding them in a log.

## [0.2.152] - 2026-08-31

### Fixed

- **The pinned last user message no longer disappears.** The line above the
  chat that keeps your last message in view could stay empty for a whole run,
  so the only way to see what you had asked was to scroll all the way back up
  to it. It gave up whenever it landed on a moment where the transcript was
  not measurable yet — a freshly opened session, a chat in a background tab —
  and nothing brought it back until the next manual scroll. It now shows your
  newest message whenever it cannot measure, instead of nothing, and it also
  notices a message leaving the top of the view when no scrolling was involved
  at all.

## [0.2.151] - 2026-08-31

### Added

- **Memory entries have categories.** Every remembered item is now filed under
  one of eleven categories — persona, business, rules, design, coding style,
  architecture, repo facts, workflow, pitfalls, environment, other. The
  category is shown and filterable in the memory panel, and the repository
  memory file is grouped by it instead of by entry kind.
- **Memory can be switched off for a single chat.** Using memory and learning
  from a session are now separate switches that apply to that session only,
  without touching your saved settings. A session that pulled in web or MCP
  content can also be excluded from learning entirely — previously, fetched
  page content could end up in a durable memory entry.
- **Existing memory from Claude Code and Codex CLI can be imported.** Their
  memory folders are read and brought in as reviewable candidates with their
  origin recorded, so switching tools no longer means starting from an empty
  memory. Nothing imported becomes active without review.
- **Rule files can be split up.** `AGENTS.md`/`CLAUDE.md`/`UNARY.md` now
  support `@path` imports, and `.unary/rules/*.md` files with a `paths:` glob
  load only when you are working in a matching part of the repository. Large
  rule sets no longer have to sit in the one file that is sent on every
  request.
- **Each distillation run leaves an audit trail.** A consolidation diff records
  what a run added, merged, replaced or retired, and why.

### Changed

- **Memory no longer fills the prompt.** Session start now injects only a
  compact index of what is remembered; the actual entries are loaded on demand
  when they are needed. Memory capacity is also one number now instead of a
  character cap and a token budget disagreeing with each other.
- **Searching dreamed conclusions understands paraphrases.** It used to match
  literal text only, so asking in different words found nothing.
- Consolidated the technical spine and reference units under `docs/spec/` with explicit ownership and plan provenance frontmatter.

## [0.2.150] - 2026-08-30

### Fixed

- **Checkboxes look the same everywhere, with no exceptions left.** Some
  checkboxes — among them the ones in the "Delete all chats" dialog — came out
  as the oversized grey system box instead of the Unary control. Every checkbox
  in the app now draws the same box, including the task-list markers inside
  chat messages.

## [0.2.149] - 2026-08-30

### Changed

- **A dirty file no longer moves a session into a worktree.** Session
  isolation now fires only while another live agent session is actually
  writing the same file. Leftover uncommitted changes, external editors and
  non-Unary tools never force a worktree anymore — git dirt is history, not
  activity — and the per-turn `git status` sweep that fed the old check is
  gone entirely.

### Added

- **The editor resolves file collisions at save and close.** When a file
  changed on disk while it was open — typically because an agent session
  wrote it — saving or closing now opens a conflict view that diffs the disk
  version against yours and asks for an explicit choice (save mine, take
  disk, or for a closing tab discard-and-close) instead of overwriting
  either side silently.

### Fixed

- **"Snip region" shows its drag-selection box again in the IDE app.** The
  overlay window's `/?snipOverlay=` URL matched none of Shell's window routes,
  so the transparent fullscreen windows rendered the Overview surface behind
  the auth gate instead of the selection overlay. Shell now routes snip
  overlay windows to a dedicated surface; both frontends share the one
  overlay-window factory that reports creation failures; mixed-DPI monitors
  are matched by physical position instead of enumeration order; a drag now
  captures the pointer, and a double press of the snip shortcut no longer
  spawns two overlay sets. On the backend, confirming with a mismatched frame
  id no longer destroys the still-valid frozen frame.

- **The pinned last user message shows up without scrolling first.** The
  sticky preview of the last user message resolved its position in a single
  animation frame; right after sending, or right after opening a session, the
  transcript rows had not been measured yet, so the frame concluded "nothing
  to pin" and nothing re-checked until you scrolled by hand. An unmeasured
  frame is now recognised as not-ready and retried.

- **A reopened compacted session shows its full history again.** The chat was
  seeded from the model's context window, which a compaction truncates, so the
  transcript started at the compaction marker; it now reads the session branch
  and shows everything that happened before it.

- **Escape closes the topmost settings overlay again.** The Copilot OAuth and
  hub settings dialogs had lost their modal flag, so the key fell through to
  the dialog underneath.

## [0.2.146] - 2026-08-29

### Changed

- **Long-running workers are judged by evidence instead of silence.** Active
  commands, output, token growth and repository changes now keep legitimate
  work alive for its measured stage duration, while repeated tool loops with no
  effect are surfaced early for a supervisor decision instead of waiting for a
  fixed timeout.
- **Worker limits are predictable.** Per-role and per-dispatch ceilings now use
  wall time and total tokens rather than an estimated dollar value, and a
  paused worker reports the exact total needed before it can resume.
- **Context-fill controls apply immediately.** Changing a session's automatic
  compaction threshold no longer waits behind the prompt that is currently
  running, and both desktop surfaces show the inherited value when no override
  is set.

### Added

- **Knowledge pages can be managed and improved from one panel.** The panel now
  includes its page list, category filters and editor workflow instead of
  splitting those actions across separate surfaces.
- **Source-control rows have complete action menus.** Branch deletion, stash
  actions and the remaining row operations are available from the compact
  overflow controls without stretching the card.

## [0.2.145] - 2026-08-29

### Changed

- **Sync leads the source-control card again.** The Sync button sat in the far
  right corner of the overview card, a full card width away from the ahead and
  behind counters that explain why you would press it. It now renders first in
  that row, with the counters following it. The branch dropdown above it no
  longer stretches across the whole card either: it is only as wide as the
  branch name, capped, and truncates with an ellipsis past that cap — so a
  short branch like `main` stops squeezing the upstream label down to
  `origin/m...`. The commit button now runs to the card's own edges, making the
  control that commits the widest one in the panel.

## [0.2.144] - 2026-08-29

### Changed

- **Your own messages read as one pill again.** A user message in the
  transcript now carries the same surface as the pinned message at the top of
  the panel: one rounded block with a chrome-level fill and a soft lift, the
  avatar chip sitting inline at the start of the text instead of a separate
  name row above it. The previous fill sat under two percent above the
  transcript background with no border, which is what made the block stop
  reading as a container at all.

- **The agent knows what to call you.** The name you configure now reaches the
  system prompt as a session fact, next to the date and the working directory,
  so it survives a custom prompt. A second setting overrides the name used
  there independently of the transcript label; with both blank, no name is
  sent at all. The transcript label itself now defaults to the neutral
  "User" — your account name is no longer borrowed for it, because that label
  is a role marker rather than an identity.

### Fixed

- **The transcript's navigation buttons stop lingering after a click.** Using
  one of them left the rail pinned open until focus happened to move somewhere
  else, because a plain mouse click was treated as a focus hold. Only keyboard
  focus holds the rail now, so a clicked rail fades on exactly the same timer
  as an untouched one.

## [0.2.143] - 2026-08-28

### Changed

- **One markdown renderer for both chat surfaces.** The desktop chat now
  renders through the same canonical pipeline as the IDE — background
  parsing, stable streaming tails, dialogs, and the file-reference menu —
  with thin per-surface adapters for opening and revealing files. Remote
  images from safe sources render on both surfaces now, links whose label
  is a file path resolve like the path itself, and line anchors survive
  the desktop's ordinary click into its preview.

- **The name on your own chat messages is yours to set.** A new setting
  controls the displayed user name; it defaults to your account's full
  display name and applies immediately on both surfaces.

### Fixed

- **Starting Unary twice no longer starts Unary twice.** The launch gate now
  claims its own boot mutex in the first milliseconds of startup instead of
  probing one the app only creates several seconds later, after the backend has
  booted. A second launch that arrived inside that window used to boot all the
  way through and then race the first one, leaving two live instances
  contending over the same data directory until one of them died. A second
  launch now hands over to the running window and waits out a first instance
  that is still booting, instead of dropping the launch after half a second.

- **A development run no longer hands itself to the installed app.** The gate
  used the release identifier in every build, so starting a dev build while the
  installed Unary was running quietly exited the dev process and raised the
  installed window. Each build profile now uses the identifier it was actually
  built with.

- **Crash reports stop inventing Rust panics.** The unclean-exit verdict read
  the modification time of `panic.log` — the same file it appends its own
  reports to — so practically every short-lived run was reported as a panic
  that never happened. The verdict is now read out of the panic records
  themselves; the newest real record in one observed log was 19 days old.

- **Chat scrolling is calmer across the board.** Streaming coalesces to one
  update per frame (an internal timer race meant it practically never did),
  finished blocks stop re-rendering while the tail streams, growing code
  fences reserve their height so late highlighting only paints color,
  images reserve their box before they load, running tool cards hold a
  stable height until they finish, and the native-flow transcript preserves
  reading position while older history loads. Edges feel defined: the
  transcript no longer chain-scrolls surrounding panels, the scrollbar's
  appearance doesn't shift content, and “Scroll to bottom” glides briefly
  and lands pinned.

- **Right-click works on file mentions in the desktop chat.** The
  file-reference menu (Open Preview / Open in Editor / Reveal in File
  Explorer / Copy Path) now opens on both surfaces from one shared
  component.

- **Session switches show the right scrollbar immediately.** Opening another
  session resets the native-flow transcript before the new history paints, so
  stale scroll geometry cannot leak between sessions.

## [0.2.142] - 2026-08-28

### Changed

- **The chat transcript no longer guesses.** The estimate-based virtualizer
  (place rows at a guessed height, reposition after measurement) is gone;
  every loaded row now renders in normal DOM flow, offscreen rows cost
  nothing via `content-visibility: auto`, and the browser's native scroll
  anchoring keeps the reading position while older history loads (with a
  manual fallback where WebKit lacks it). Opening a session lands exactly at
  the bottom on the first paint — no settle frames, no flicker, no jumping
  while scrolling back, and roughly a thousand lines of scroll machinery
  deleted.

- **One stylesheet owns the chat on both surfaces.** The Desktop app's
  hand-mirrored chat CSS layers could override the hosted panel and drift
  from the IDE (two live divergences shipped that way). Desktop-winning
  values were folded into the canonical `AgentChatPanel` styles, the mirror
  layers are scoped to the backend-less mock preview only, and a contract
  test fails on any new unscoped chat selector — a fix on one surface can no
  longer miss the other.

- **User messages carry your name.** The chat bubble's static uppercase
  "USER" marker (right-aligned) is now a left-aligned label showing the
  signed-in account's display name, seeded from the cached auth preview and
  updated on auth changes; it falls back to "User" when signed out.

## [0.2.141] - 2026-08-28

### Fixed

- **Stored sessions open again.** Opening any session died on
  `e.stopSessionOpenBottomFollow is not a function` and left the panel on
  the old view with a "something went wrong" warning. The scroll-follow
  rework had deleted that step everywhere except at the one place every
  session switch runs through, so no stored session could be reopened at
  all. The dead step is gone, and a new gate check now catches a chat
  panel and its controllers drifting apart before a build ships.

- **The snip overlay fails closed on overlay and attachment errors.**

- **A provider rejecting the request body no longer kills the turn
  silently.** A below-floor image embed (16×16 app icons against xAI's
  512-pixel minimum) came back as a 400 that was classified as a generic
  error: it burned the in-place retry budget, rotated an account out of the
  rotation and ended the turn while the model saw the conversation simply
  stop. Request-content 4xx now has its own error class, images below a
  provider's floor are guarded before they are sent, and the rejection
  reaches the agent as a quiet harness note instead of vanishing.

- **The chat Git workspace keeps its head and tab switch pinned.** The
  popover scrolled as one block, so the title row and the Changes/Graph
  switch slid out of view; head, view switch and status row are fixed now
  and only the view body scrolls.

### Changed

- **Unary has one cursor.** Every `not-allowed`, `wait`, `pointer` and
  `grab` cursor is gone from both frontends — disabled and busy states are
  expressed with opacity, fill and spinners, and the pointer the system
  draws is the pointer you keep.

- **Every dismiss button is the same button.** The nine remaining
  hand-rolled × glyphs (AuthGate, live status area, fix-suggestion banner,
  changes review, context transfer, worktree panel, hunk diff, memory
  entries, doc research) render the shared `CloseButton`, and the native
  system notification, which cannot reach the shared control, mirrors its
  exact shape.

- **Checkbox and radio hover is visible again.** The old hover wash sat at
  ~1.05:1 and collapsed entirely on the light theme; hover now uses the
  state-ladder fill plus a halo ring (~1.4:1), and checked boxes get hover
  feedback for the first time.

- **The global Index settings open without a workspace.** Embedding
  hardware, model and thread count are reachable before any project is
  opened.

## [0.2.139] - 2026-08-27

### Changed

- **App and installer icons carry the rounded mark with a small AI
  badge.** The full icon set returns to the rounded-corner artwork
  (white mark on the dark rounded plate) and the app/installer icons add
  the V1 AI badge — the in-app header and the tray stay clean.

### Fixed

- **Detached windows can be dragged again.** Detached Settings and dialog
  windows became unmovable once the WebView2 evergreen runtime activated
  its native draggable-regions feature: the native side swallowed the
  pointer gestures but cannot move wry's frameless windows. Dragging now
  runs through JS everywhere; `app-region: drag` is gone, `no-drag` stays
  as defense.

- **“Open credential vault” now reveals the dock window.** The forwarded
  open worked, but the panel appeared behind the covering detached
  Settings window and looked dead; cross-window opens now unminimize and
  focus the dock window.

- **A vertical mouse wheel scrolls compact horizontal rails.** The
  dominant wheel axis is translated into rail movement, so a conventional
  wheel reaches controls beyond a one-line rail without Shift; trackpads
  keep their genuine horizontal direction, and Ctrl/Cmd+wheel stays a
  zoom gesture.

- **The snip overlay is screen-reader friendly and its recorder buttons
  behave.** The region picker announces instructions and the live
  selection size to assistive tech, the shortcut recorder's Accept button
  is clickable only while a recording with a draft exists (it used to be
  clickable before any recording started), and snip toolbars mounted as
  their own Svelte roots reach the window controls again.

## [0.2.138] - 2026-08-27

### Changed

- **The IDE and the desktop app now share one chat composer.** The message
  box — attachments and their mark-up dialog, the drag-to-resize handle with
  remembered height, paste and drop handling, the plan-mode suggestion, and
  the keyboard behavior while a run is busy — is one component on both
  surfaces now, so a composer fix lands everywhere at once.

### Fixed

- **Streaming no longer trembles at the bottom, and follow truly sticks.**
  Both surfaces run one shared follow policy now: the per-frame tug-of-war
  between scroll writers is gone (including the one-pixel flips on fractional
  display scaling), “Scroll to bottom” enters a follow mode that holds
  through further streaming until you scroll away, and deliberate upward
  scrolling lets go immediately.

- **Opening a session always lands at the newest message.** Late-arriving
  content (images, long histories, code formatting) keeps the view pinned to
  the bottom until your first real scroll — on both surfaces — instead of
  stranding the view somewhere mid-history.

- **Notification banners are never cut off at the bottom.** The banner window
  sizes itself to the card’s real rendered height (the reply-style banner’s
  answer row used to lose its lower half), and bottom-corner banners rise to
  keep their margin.

- **The Goals panel no longer freezes for the length of a work stretch.** A
  2026-08-27 audit found runs whose board was created and then left untouched
  for 20–55 minutes of real work, every checkbox flipped in one batch at the
  end — and the only live push (`GoalsUpdated`) could silently drop under UI
  backpressure. The run loop now injects a bounded “Goals staleness guard”
  reminder every 8 workspace mutations that pass without a `goals` call,
  receipts teach per-transition updates (starting an item counts, not only
  completing it) and call out when nothing is left `in_progress`, the
  settle-side reconciliation checkpoint re-arms once when new mutations pile
  up after an ignored one, the goals push rides the lossless relay instead of
  a zero-timeout send, the goals strip shows “updated Xm ago” plus per-item
  in-progress ages during an active run, and the session debug bundle reports
  `goalsLiveness` numbers (longest stale stretch, nudges fired) so a
  regression is visible in the artifact users already paste.

- **The chat-plan "Plan ready" card no longer goes stale — and all three of
  its buttons now work.** Replying in chat to accept a plan the agent wrote
  ("implementiere das") left the Accept/Revise/Decline card mounted for the
  entire implementation run: mid-turn state pushes only patch streaming
  counters, so the pending-plan-reply flag computed before the run survived
  until turn end. A starting run now clears it and pushes the state change
  immediately. Accept/Revise follow-up prompts queue as follow-ups instead of
  being silently dropped while the session is busy, Decline now reports the
  decision to the model in the IDE (desktop parity), Revise/Decline no longer
  fail with "no active plan" on a plan that exists only as chat prose, and
  declining a chat plan keeps the current mode instead of switching the
  session into Plan mode.

## [0.2.137] - 2026-08-27

### Fixed

- **Reaching the very bottom is reliable again.** The jump-to-bottom settle
  loop stays alive while a running session keeps growing the transcript
  (height-driven budget instead of a flat 24-frame cap), growth without a
  scroll event now updates the jump button and can re-engage follow when the
  user rested on the bottom, and autoloaded newer pages re-pin or preserve
  the position instead of ending the jump mid-transcript.

## [0.2.136] - 2026-08-27

### Fixed

- **Orphaned worker and session worktrees now heal while the app runs.** Four
  leak paths under `.unary/worktrees/` are closed: a watchdog force-terminated
  worker now has its work captured and its tree reaped immediately (the wedged
  runner's settle path never runs, so nothing else ever cleaned it); preserved
  cancel trees age out on the same 72-hour window as unapplied integrations
  instead of living forever; the orphan sweep now repairs failed cleanups under
  a still-open session by trusting that session's own records about which
  trees are disposable; and both the worker GC and the session fold-back retry
  re-arm from every prompt (30-minute gate) instead of only at app start, so a
  deferred fold-back completes once the blocking edit is gone.
- **Snip works in the desktop app again.** 0.2.135 taught the IDE's snip
  toolbar to mint the operationId receipt the backend demands, but the
  desktop app's titlebar and footer toolbars still invoked the plugin tool
  without one, so every region/window/screen capture and gallery write died
  with "mutating plugin tool `snip` requires a registered operationId
  receipt". Both desktop entry points now ride the exact tracked-receipt
  wrapper the IDE uses (`executePluginToolWithMutationReceipt`); the
  plain-browser preview keeps its receipt-free mock no-op.

## [0.2.135] - 2026-08-27

### Added

- **Machine-started sessions can now open quietly.** A prompt composed by the
  app — the sidebar's worktree review, for example — can carry a display
  title; the chat then shows a small "Worktree review" line instead of a fake
  user bubble holding the full generated text. Dev mode still reveals the
  exact prompt the model received. The worktree review also ends by asking,
  in one decision card, what to execute — clean up, salvage, or finish what
  is still missing.

### Fixed

- **Snip region works again.** The operationId receipt gate classified the
  snip plugin's capture_frame/cancel_frame steps as mutating while the snip
  toolbar minted no receipt, so every snip click was rejected in 2 ms with
  "mutating plugin tool `snip` requires a registered operationId receipt" —
  and the error notice rendered off-screen inside the titlebar, so the button
  read as completely dead. The two frame steps are receipt-free reads now
  (they persist nothing), the persisting captures and gallery writes ride the
  tracked-receipt lifecycle, and a failing snip shows its error as a toast.
- **The newest message really is sharp now.** 0.2.134 fixed the mount path,
  but rich rows (markdown, code — every assistant reply) armed at a higher
  CSS specificity carrying blur(5px) than the two-class reveal override
  could beat, so they stayed blurred forever after revealing. The shared
  transcript carries the full-strength reveal overrides again.

## [0.2.134] - 2026-08-27

### Fixed

- **The newest chat message no longer sits blurry at the bottom.** A message
  appended while the viewport rests at the bottom used to stay in the reveal
  animation's pre-state (blur + fade) until the next scroll; rows mounting in
  the bottom band now appear sharp immediately, and streaming no longer
  flashes blur at the bottom edge. The reveal animation remains when
  scrolling back through history.

- **No more trembling just above the bottom while streaming.** Follow-mode
  stopped detaching on the browser's own scroll clamp (fired when content
  above re-measures smaller) and only lets go when a move actually leaves
  the bottom band — the rapid detach/re-engage flicker is gone. Scrollbar
  drags during streaming now detach follow on both surfaces.

- **“Scroll to bottom” really reaches the bottom.** The button used to jump
  to the estimated bottom once; rows below then realized their true height
  and the viewport stopped short. The jump now re-pins until the height is
  stable.

## [0.2.133] - 2026-08-27

### Changed

- **Both chat transcripts now run on one shared scroll engine.** The IDE and
  the desktop app mount the same virtualized transcript component, so height
  estimation with learned row priors, batched measurement, reading-anchor
  restore, history-load preservation, and the scroll flight recorder
  (Ctrl/Cmd+Alt+Shift+D) behave identically on both surfaces.

### Fixed

- **Chat scrolling is steadier.** The desktop app no longer prices every row
  at a fixed guess with a small look-ahead window; growth above the viewport
  is compensated immediately instead of landing as one visible jump after
  scrolling stops; corrections can no longer apply twice around history
  loads; and follow-the-bottom re-engages by scrolling to the bottom on
  Windows displays at fractional scaling (125%/150%).

## [0.2.132] - 2026-08-27

### Changed

- **Work can move between sessions without creating two owners.** Handing off a
  plan now records who may continue it, survives a restart, and notifies the
  receiving session when adoption completes.

### Fixed

- **Peer messages and plan transfers recover instead of disappearing.** Their
  delivery state is stored durably, interrupted transfers can retry or roll
  back, and completed work is not exposed before its shared state is ready.
- **Session instructions keep their real source.** User messages remain user
  messages across forwarding and renaming boundaries, so coordination hints
  cannot be mistaken for instructions merely because their text looks alike.
- **Workspace switches apply one coherent settings snapshot.** Reloading now
  follows the switch authority, while explicit user-selected paths remain
  distinct from paths contained by the active workspace.
- **The Windows terminal closes the complete command tree.** Child processes
  no longer survive when the host cannot assign a nested Windows Job Object.

## [0.2.131] - 2026-08-27

### Changed

- **Work can move between sessions without creating two owners.** Handing off a
  plan now records who may continue it, survives a restart, and notifies the
  receiving session when adoption completes.

### Fixed

- **Peer messages and plan transfers recover instead of disappearing.** Their
  delivery state is stored durably, interrupted transfers can retry or roll
  back, and completed work is not exposed before its shared state is ready.
- **Session instructions keep their real source.** User messages remain user
  messages across forwarding and renaming boundaries, so coordination hints
  cannot be mistaken for instructions merely because their text looks alike.
- **Workspace switches apply one coherent settings snapshot.** Reloading now
  follows the switch authority, while explicit user-selected paths remain
  distinct from paths contained by the active workspace.

## [0.2.130] - 2026-08-26

### Added

- **The Goals panel fills while the agent works, not after.** Sessions often
  ran a whole fix with the panel blank and then flashed a fully completed board
  at the end. The agent is now told at the start of every request when the
  board is still empty, and a run that keeps changing files without tracking
  anything gets one in-flight reminder to put the work on the board — so the
  checkboxes flip live while you watch. Explicit investigate/audit requests
  now count as trackable work too.

### Fixed

- **The scroll-to-bottom button hides when you are at the bottom.** Windows
  fractional display scaling (125 %/150 %) could leave the chat resting a
  fraction of a pixel short of the arithmetic bottom, so the button stayed
  visible while you were already reading the end. The button now tolerates
  that rounding; follow behaviour is unchanged.
- **Settings changes reach running sessions.** Changing the model or agent
  settings used to apply only to sessions started afterwards; live sessions
  now pick the change up. The "Global Settings" button also works from dialog
  windows, and the effort menu's cache warnings state token counts instead of
  dollar figures.
- **The SCM Sync button grays out when the branch is already in sync** instead
  of offering a push that would do nothing.
- **Worktree rows keep their actions in one context menu** instead of a strip
  of icon buttons per row.
- **A finished background shell no longer claims to still be running** in the
  shell supervision list.

## [0.2.129] - 2026-08-26

### Fixed

- **Your sessions follow the folder when you switch branches.** Checking out a
  different branch in place used to leave every earlier session of that folder
  stranded in a group named after the old branch - a group that stayed even
  after the branch itself was deleted. The Dock sessions panel and the Desktop
  sidebar now group a folder's sessions under the branch it is on right now,
  and the branch a session started on is still shown on the row itself. A
  checkout made in a terminal is picked up when the window comes back to the
  front, instead of waiting for the next refresh.
- **Windows shows the application icon on the taskbar again.** The window icon
  set at startup replaced the executable's own icon, so the taskbar button and
  the shortcut disagreed.
- **The trust period in Remote Control settings uses the themed dropdown**
  instead of the unthemed native one.

## [0.2.128] - 2026-08-25

### Fixed

- **Fast scrolling in the Desktop chat no longer loops back over the same
  messages.** When a row's real height is measured while you scroll, the
  correction to the scroll position is now applied in the same instant —
  previously it landed one beat later, so the view could re-render rows you
  had already passed for a frame. Quick scrolling measures rows every frame,
  which turned that one-frame stutter into the visible loop.

## [0.2.127] - 2026-08-25

### Added

- **Scroll diagnostics recorder.** Pressing Cmd/Ctrl+Shift+Alt+D in the chat
  copies the last ten seconds of scroll activity - every scroll event, every
  scroll write with its writer, and the virtual window mapping - so a
  remaining scroll defect can be diagnosed from evidence instead of theory.

## [0.2.126] - 2026-08-25

### Fixed

- **Fast scrolling in the Desktop chat view no longer repeats content.** The
  master-view chat now pays row-measurement shifts above your reading position
  back to the scroll position in the same frame, so flinging quickly through a
  long conversation can no longer re-map the same messages under your finger.
  Slow scrolling was already correct; this closes the fast path.

## [0.2.125] - 2026-08-24

### Fixed

- **Fast scrolling through a long chat no longer repeats the same content.**
  While you scroll, the transcript keeps its row-height predictions frozen and
  re-prices them only after the gesture settles, so the view can no longer
  re-map the same messages under your finger — the “scrolling in a loop”
  effect near the newest message is gone.

## [0.2.124] - 2026-08-24

### Fixed

- **Worker and remote transcripts no longer yank you back to the bottom.**
  The live worker-session view and the remote-session mirror now follow the
  same rule as the main chat: scrolling up takes the view immediately and
  keeps it, a queued refresh can no longer drag you back down, and following
  resumes only when you deliberately return to the newest message.

## [0.2.123] - 2026-08-24

### Fixed

- **The endless jump-back at the bottom is gone.** While the chat follows new
  output, the row-height corrector now stands down completely instead of
  fighting the follow writer frame-for-frame, and it resumes the moment you
  scroll away — so fast scrolling near the newest message no longer loops.

## [0.2.122] - 2026-08-23

### Added

- **Name your devices.** The desktop's Remote Control card and the iPhone's
  Settings gain a device-name field; the name is what paired devices show in
  their lists. The iPhone also now registers as platform `ios` instead of
  `unknown`.
- **Choose how long a pairing is trusted.** The desktop pairing card offers
  1 day / 7 days / 30 days / permanent; the choice travels inside the QR code
  and the scanned device is trusted for exactly that period. Trusted devices
  can be removed from the desktop list as before.

### Fixed

- **Fast scrolling no longer loops back to the newest message.** The follow
  writer now re-checks ownership at the moment it writes, the session-open
  follow yields to the first upward movement, and re-engaging follow requires
  deliberate downward travel — a single momentum jitter can never hand the
  view back to streaming output. Both frontends share the same contract.

- **Delayed reveal jumps respect the user.** Settings search reveals and
  MEMORY entry jumps cancel when you scroll first, and stale retries can no
  longer scroll a later view or session.

## [0.2.121] - 2026-08-23

### Fixed

- **Fast transcript scrolling no longer repeatedly restarts at the bottom.**
  Once a wheel, touch or keyboard gesture moves away from the newest message,
  follow mode stays latched off through subpixel Mac trackpad momentum. It can
  return only after explicit movement toward the bottom or the bottom button.

## [0.2.120] - 2026-08-23

### Changed

- **Worktrees appear once, with actions that match what they are.** The sidebar
  now keeps Session, Subagent, Personal and External worktrees in one clear
  hierarchy, while Running, Uncommitted and Integrated are visible states
  instead of duplicate groups.

### Fixed

- **Worktree review and clean-up are safer.** Review is available from both the
  sidebar and the full overview; Session trees fold back, Subagent branches
  merge, and Personal or External trees are never folded back automatically.
  Uncommitted files are shown and block unsafe integration, and removing a
  worktree clearly preserves its branch.

## [0.2.119] - 2026-08-23

### Added

- **Pair an iPhone by scanning the desktop's QR code.** The phone camera now
  opens above the settings drawer, takes the whole screen while scanning and
  returns cleanly when pairing is done.

### Fixed

- **Conversation scrolling stays where you put it.** A Mac trackpad's first
  fractional movement now takes control before live output can pull the view
  back down, and following resumes only once the native scroller has genuinely
  reached the newest message. Long chats, navigation jumps and rows that finish
  rendering no longer fight the reader.

- **High-resolution scrolling is consistent across the app.** Vertical gestures
  are no longer stolen by horizontal rails, event feeds preserve a reader's
  position, and browser and remote-control views coalesce momentum without
  replaying a delayed backlog after the gesture ends.

- **Sessions remain in the right sidebar group.** A session stays associated
  with its original branch, a parent remains active while its background workers
  run, and inactive-worktree cleanup no longer mistakes live worker folders for
  abandoned ones.

- **Background-worker results describe what actually happened.** Cost reaches
  the final handoff, missing local toolchains are reported as environment
  failures instead of patch conflicts, and several completion/reconciliation
  edge cases no longer strand finished work.

## [0.2.118] - 2026-08-22

### Fixed

- **Die Sidebar zeigt wieder deine ganze Chat-Historie.** Pro Projekt wurden
  nur die 10 neuesten Sessions angezeigt, alles ältere lag hinter einer
  „Load more“-Zeile — bei 23 Chats also 13 unsichtbare. Weil das meist nach
  einem Branchwechsel auffiel, sah es aus, als hätte git die Chats
  verschluckt. Die Voreinstellung steht jetzt auf 50 und wird bei
  bestehenden Installationen einmalig angehoben; ein selbst gesetzter Wert
  bleibt unberührt. Einstellbar unter Einstellungen › Sessions (0 = alle).

## [0.2.117] - 2026-08-22

### Fixed

- **Ältere Chats verschwinden nicht mehr aus der Sidebar.** Die Chatliste zeigt
  pro Projekt nur die 10 neuesten Sessions, der Rest liegt hinter „Load more“ —
  und diese Aufklappung wurde bei jedem Neuaufbau der Liste vergessen. Wer mehr
  als 10 Chats in einem Projekt hat, sah die älteren dadurch immer wieder
  verschwinden, oft direkt nach einem Branchwechsel oder einer neuen Session.
  Der aufgeklappte Zustand bleibt jetzt erhalten. Das Limit selbst steht in
  Einstellungen › Sessions.

## [0.2.116] - 2026-08-22

### Fixed

- **Switching git branch no longer makes your earlier chats disappear.**
  Sessions were filed under the branch they were started on, so a plain
  `git checkout` left every older chat in a group that then folded shut —
  starting one new session made the rest look deleted, and deleting it
  brought them back. Chats now belong to the folder they run in and follow
  it onto the new branch, and a branch group only sorts the list, it never
  hides anything.

## [0.2.115] - 2026-08-22

### Added

- **Clone a GitHub repository by picking it from a list**, with search,
  instead of pasting a URL.

- **The goals list in a chat window shows its items.** Opening it no
  longer leaves a blank card.

### Changed

- **Stable updates now ship the Mac and iPhone work** that lived off the
  main line: conversation scrolling, the jump-to-bottom button, remote
  sessions, and the memory cleanup from the last few releases.

## [0.2.114] - 2026-08-21

### Changed

- **Scrolling a long conversation behaves like it used to again.** Several
  attempts this week to smooth out the transcript ended up fighting the mouse:
  a short scroll up did not really leave auto-follow, so the next streaming
  token pulled the view back down. The scroll logic is restored to the state
  that felt right, while the genuine paging fix and the new bottom-button and
  hover-rail behaviour stay.

## [0.2.113] - 2026-08-21

### Fixed

- **Long conversations can reach their newest messages again.** If the app was
  re-hydrated or you switched sessions while a page of the conversation was
  being read, loading further pages stopped permanently for that chat. The
  newest messages then stayed out of view and the scroll-to-bottom button had
  nothing left to scroll to, which is what made it look dead while "go to top"
  kept working.

## [0.2.112] - 2026-08-21

### Fixed

- **The scroll-to-bottom button works again.** In long conversations the
  button — both the one in the hover rail and the one that stays visible while
  you are away from the bottom — could do nothing at all: it waited for older
  parts of the conversation to be paged back in, and when that could not
  finish it gave up without moving the view. It now always jumps to the end of
  what is loaded, and keeps pulling the newest messages in behind that.

## [0.2.111] - 2026-08-21

### Fixed

- **The chat transcript stays where you put it.** Scrolling near the bottom
  no longer springs back and forth while message rows finish rendering, and
  the navigation controls work again: jumping to the newest message, or
  between your own messages, now holds its landing position instead of being
  dragged away the moment the surrounding rows settle. Until now only "go to
  top" survived that tug-of-war, which is why every other button looked dead.

## [0.2.110] - 2026-08-21

### Fixed

- **Chat navigation, decided.** The jump-to-bottom button is now always
  visible whenever you are away from the newest message — no more hunting
  for it. The slim navigation rail with the prompt ticks is gone; the
  remaining rail appears only while you work the scrollbar with the pointer,
  and it no longer blocks clicks or scrolling when hidden. Scrolling gently
  near the bottom no longer jumps back and forth while message rows settle.

- **Remote sessions load reliably again.** Opening a session on a remote
  host could exceed the relay's message-size limit, so the controller never
  received the transcript, timed out and reloaded forever — stealing the
  input focus on every cycle. Large histories are now trimmed to fit the
  relay budget and paged back like any windowed session.

## [0.2.109] - 2026-08-21

### Fixed

- **The desktop chat finally scrolls the way it should.** Opening a
  conversation no longer flashes the oldest messages before jumping down,
  switching between conversations now starts at the newest message instead
  of inheriting the old scroll position, and jumping to the previous or next
  message stays anchored while older history loads in. Fast trackpad
  direction reversals no longer leave the list leaning the wrong way.

- **The assistant's memory stops hoarding junk.** Questions and problem
  reports were being filed away as permanent preferences, and routine tool
  rejections, probe answers and empty results were being remembered as known
  defects — some of it broadcast into every later prompt. Captures like
  these now wait: nothing reaches the global memory surface without being
  curated first or actually recurring, and uncurated one-offs expire on
  their own.

## [0.2.108] - 2026-08-20

### Fixed

- **No more endless password prompts from the Mac's keychain.** On macOS,
  running the project's own test suites made the system ask again and again
  for permission to read stored credentials — a fresh dialog after every
  rebuild that stole the keyboard focus mid-typing, and „Always Allow“ never
  stuck because each rebuilt test program counts as a new app. Test runs now
  use their own throwaway credential store and never touch the real keychain;
  the app itself keeps reading and storing credentials exactly as before.

## [0.2.107] - 2026-08-20

### Fixed

- **Isolated work sessions clean up after themselves again.** Sessions and
  background workers that run in their own copy of the repository could be
  left behind forever when returning their work hit a snag: an unimportant
  link or a regenerated lock file was enough to make every retry fail the
  same way, and a finished worker whose handoff nobody picked up kept its
  copy for good. Both now resolve themselves: the return merges around
  unrelated churn instead of failing on it, and an unpicked-up handoff is
  wrapped up after three days instead of never — while any work that exists
  nowhere else is kept safe on its branch.

## [0.2.106] - 2026-08-19

### Added

- **The Mac finally reports what the graphics chip is doing.** GPU load was
  simply blank on macOS while every other reading was live, so the one number
  that tells you whether a local model is actually using the hardware was
  missing. It now comes straight from the system and sits next to the others.
- **Preview an HTML file without leaving the app.** Opening one now renders it
  in the same managed browser the agent already uses, instead of handing you a
  wall of markup or pushing you into a separate window.

### Fixed

- **The chat's navigation rail stays reachable.** Moving the pointer toward the
  rail could make it slip away before you arrived, so clicking a marker turned
  into a chase. It now stays open while you approach it and while you are on
  it, fades between states instead of snapping, and tells a screen reader what
  it is and which message you are on.
- **A remote host no longer reconnects into its own ghost.** After a dropped
  connection the app could latch onto the dead half of the previous session and
  sit there looking connected while nothing arrived.
- **Background workers cannot stall past their deadline in silence.** The
  supervision that force-ends a wedged worker is now part of the same code path
  that waits for it, so a worker that stops responding is ended and reported
  instead of holding up everything behind it. This lands the completed version
  of work that was previously reverted here because it did not build on macOS.
- **The research panel no longer breaks on an incomplete configuration.** A
  missing engine list in the footer summary took the whole panel down instead
  of showing what it did know.

## [0.2.105] - 2026-08-19

### Fixed

- **Semantic code search works again on macOS.** The Mac app shipped with a
  local search-assistant program older than the app itself, so every attempt to
  understand your code by meaning failed the moment it started. Nothing said
  so: search quietly fell back to plain text matching and the status panel kept
  naming the model it could not actually use. Searching by intent now works,
  and if that helper ever fails to start again, the status panel says so
  outright instead of reporting everything as ready.
- **Workers can no longer hang silently past every deadline.** A worker
  frozen mid-finalization stayed `running` for 87 minutes while the main
  agent waited, because the supervision ladder was advisory in four
  compounding ways. A blocking `wait`/`wait_all` now sweeps the watchdog
  itself (on entry, per supervision interval, once at the deadline) instead
  of sleeping through worker deadlines, and its timeout carries actionable
  `stall_evidence` for every worker past its deadline or heartbeat grace
  while still non-terminal. Cooperative escalations that fail silently
  (wedged finalization slot, undeliverable interrupt) now increment a
  persisted failure counter; after three consecutive failures the sweep
  force-terminates the worker through the same cancel+abort path a manual
  close uses. The watchdog scheduler thread survives transient panics and
  resolution failures with 60s-backoff re-arm instead of dying permanently
  while its started flag stays true. Stalled heartbeats emit a
  `WorkerDeadlineGrace` event (300s cooldown) that previously had a
  formatter but no producer.

## [0.2.104] - 2026-08-18

### Fixed

- **Session-worktree work now flows back automatically — and is never destroyed
  at close.** An audit found work done in isolated session worktrees NEVER
  returned to the main checkout on its own (in this repository alone: 16
  stranded session branches with 27 unmerged commits), and worse, closing a
  session could silently DELETE its worktree, branch and every commit in it
  before the fold-back ran. Fold-back is now commit-preserving (cherry-pick
  with patch-id de-duplication instead of an all-or-nothing patch), conflict
  partial results are recorded and retryable, quitting the app folds every
  session back, a startup sweep recovers cold trees after a crash, and
  isolating a session moves its claimed files out of the shared checkout so a
  sibling session can no longer commit them as its own.

## [0.2.103] - 2026-08-18

### Added

- **Session templates carry the whole session setup.** A template now spans
  model, tools, worker roles and limits, so starting a recurring kind of work
  is one pick instead of a dozen settings. Both frontends ship a template
  editor and a template picker in the new-session dialog, including
  configurable time and budget limits per template.
- **Memory-pressure admission gate.** After the 2026-08-17 resource-exhaustion
  incident (96 % Windows commit charge, several parallel rustc children at
  6-8 GiB each), new compile jobs are refused while commit charge is critical
  instead of walking the machine past the commit wall. Incidents land in the
  run ledger with app-log and crash-dump plumbing to diagnose them.

### Changed

- **Remote screen capture is damage-tracked.** The remote-admin screen share
  stops doing a full PNG round trip per frame and sends only what changed,
  with ack-based flow control.
- **CPU use is recorded centrally.** Per-session CPU utilization is captured
  durably and one central CPU budget replaces scattered ad-hoc checks.

### Fixed

- **Orphaned session worktrees are visible and integrable.** The overview now
  lists worktrees whose session is gone, with an integrate-all action, and a
  batch of stranded sessions has been merged.
- **Worker terminal status stays terminal.** A finished worker no longer
  flips back to running (zombie revive), a worker that dies on a broken seed
  no longer takes the run with it, and goal reconciliation demands delivery
  evidence before checking an item off.
- **Self-memory writes no longer fail under lock contention.** Store locks are
  fair (ending load-dependent write failures), composition is shard-aware,
  and the legacy user-memory engine is consolidated.
- **The credential vault opens from the detached settings window.** Settings
  runs as its own native window, which previously left the vault unable to
  reach it; app-bound encrypted cookie stores are handled type-safely.
- **A question asked mid-run always gets a visible answer.** User turns that
  arrive while work is in flight are answered instead of being swallowed by
  the run loop.

## [0.2.102] - 2026-08-17

### Changed

- **Compaction handovers carry the model's reasoning — and got smaller.** The
  session ledger now captures a bounded thinking digest per assistant turn, so
  the deterministic handover preserves why decisions were made, not only what
  happened. At the same time the handover's ledger budget was halved (64 →
  32 KB), the user-intent section is byte-bounded (original request and latest
  intent always survive), and a session that withholds tool schemas gets one
  line naming the already-activated tools so the model does not re-activate
  them after a compaction. Anything omitted stays one recovery call away via
  the segment pointers.

### Removed

- **The dead "Compaction model" setting.** Compaction has rendered its handover
  deterministically from the session ledger since 0.2.77 and never calls a
  model, so the setting and its thinking-level dropdown configured machinery
  that could no longer run. The setting, the unreachable LLM-summarization
  code behind it, and background pre-compaction (which only existed to hide
  LLM latency) are gone; a blocked or unset model can no longer fail a
  compaction. Session consolidation commits the same deterministic summary and
  no longer asks you to pick a model.

## [0.2.101] - 2026-08-16

### Added

- **Search panel filter chips.** Results now offer type and folder facets built
  from the results themselves, so narrowing a large result set takes one click.
- **Settings search.** A search box in the settings dialog header jumps straight
  to the group that contains what you typed.

### Fixed

- Agent scopes load without an open workspace, and a control that stays locked
  now says why instead of failing silently.

## [0.2.100] - 2026-08-16

### Fixed

- **Terminal sessions survive a panel switch again.** Open shells are restored
  with their scrollback instead of disappearing when the terminal panel is
  remounted.
- **Sessions no longer get stuck on "New Session".** A lost race between the
  automatic rename and the first turn kept the old placeholder title.
- Notification banners stop stacking and overlapping; their accumulation is
  bounded.
- The chat scroll no longer jumps in the remaining cases found by the scroll
  audit, and toast sizing plus the dialog editors follow the shared panel
  design.
- Index settings are saved by the owning scope again (project vs. global), and
  the index graph actions are exposed under one canonical, singular name.

### Changed

- **Index panel redesigned** around index statistics and the code graph.
- Panel settings live in the three-dots overflow menu across the panels, for
  one consistent place to find them.

### Added

- Providers and their OAuth login can be driven by an extension through
  `unary.providers.controls`, applied fail-closed by the host.

## [0.2.99] - 2026-08-16

### Fixed

- **Interrupting the agent no longer throws away a running command.** A queued
  message used to kill a foreground shell that had been quiet for a few
  seconds. The process is now parked in the background instead: the agent reads
  your message first and can then re-attach, let it keep running, or stop it.
- **Panel buttons work again after a maximized layout is restored.** Opening
  Index, Memory & Dream and the other side panels used to do nothing when the
  saved layout had a maximized tab group; panels are now revealed.
- Failed turns show the plain red error box again, and the model's narration
  between tool calls is normal text instead of a bordered chip.
- GLM-5.3 requests match the current vendor documentation (reasoning effort
  levels and streaming tool-call wire surface).

### Changed

- **Memory panel split by purpose.** Repo scope now shows only what forms
  `MEMORY.md`; bugs, friction and ideas are gathered globally under Errors.
  Global entries use a full-width list, and every button in the panel uses the
  shared UI controls.
- Consistent panel design language across Index, file search, diff and review,
  git worktree, sessions, subagents, automations, browser, debug and remote
  panels — same headers, spacing and controls; the chat composer's active
  switch lost its filled box.
- File search gained tag-based filtering.

### Added

- Provider catalogs can ship as `unary.providers` extension packs, exported
  from the built-in catalog and verified field-for-field on load.
- The plugin creator generates a fuller working plugin scaffold.

## [0.2.98] - 2026-08-15

### Fixed

- **The agent's worker fleet no longer strands after a policy change.**
  Switching the session mode (or any other policy re-apply — a permission
  change, auto-compaction, picking another model) could permanently disable
  spawning subagents with a misleading "only available inside harness-backed
  agent sessions" error, even after conditions returned. The orchestration
  handle now survives every policy apply.

### Added

- **Memory learns across sessions in layers.** At session end the memory
  system now distills related records into higher-level knowledge blocks
  (scenario and persona layers) that are read first at the next session
  start — instead of every repo forever searching only its atomized records.
  One bounded model call per layer per day/week, refusal-safe.
- The agent's `harness_health` self-diagnosis now names the exact running
  build version, so a known-and-fixed defect can be told apart from one that
  is fixed in the tree but not yet in the installed build.

### Changed

- The fast-mode toggle now states honestly, for Codex models, that the
  endpoint accepts the priority tier but does not confirm it and may not
  apply it on every plan.

## [0.2.97] - 2026-08-15

### Added

- **The agent can now debug running code.** It can launch or attach to a
  program, set breakpoints — including conditional and log breakpoints — run
  until one hits, inspect the call stack and variables, evaluate expressions
  on the stopped frame, change variable values, break on exceptions and stop
  the program cleanly. This works in every window where chat runs, the IDE and
  the desktop app alike, with the debug adapters already configured for the
  debugger panel (Python and Go today).
- Extensions can now register their own dock panels and fill them with their
  own UI.
- New **UNARY Studio** theme family: dark and light, each with a glass
  variant.

### Changed

- Notification banners act in place: answer inline and trigger actions
  directly on the banner instead of switching to the session.
- In the desktop app, dialogs moved the close button to the far edge of the
  header.

## [0.2.96] - 2026-08-15

### Added

- The **Search** panel is now available in the desktop app too — the same
  "Search" tab the IDE dock has, with results opening in the file preview.
  Saved layouts are unaffected.

### Changed

- **Memory panel navigation redesigned.** The scope switch (Session, Repo,
  Global) stays on top; each scope's pages — MEMORY.md, Entries, Errors — moved
  from a pill row into a slim rail on the right edge, whose active item looks
  exactly like the active session in the session list. Status became a quiet
  footer line (online state and entry count), and bulk actions live in the
  panel's “…” menu instead of the navigation row. Errors are shown only in the
  Global scope.
- The Knowledge panel's “Manage docs / Improve and version” switch uses the
  same right-hand rail, so both panels navigate the same way.
- Chat rendering polish: assistant messages and tool-call views render richer
  content, run dividers mark where a new run begins, and hover fills follow
  rounded corners.

## [0.2.95] - 2026-08-15

### Added

- **Multi Expert mode**, off by default. A new session mode where every
  request is answered by a roster of models you configure instead of one: each
  expert answers independently, then a judge model compares the answers and
  delivers a single verdict - the winner, a synthesis, and where the experts
  disagreed. The settings block names the roster and the judge and warns
  plainly that a request bills every model on it, so token use multiplies by
  roughly the roster size. When the judge or every expert fails, the error
  becomes the answer - it never quietly falls back to a single model.

### Changed

- Settings groups model providers into visual sections and drops the favorites
  list; every settings gear is the same crisp cog icon, and window controls
  (minimize, maximize, close) are drawn as shapes instead of text glyphs, sharp
  at any size. Maximizing a dialog is opt-in.
- The worktree sidebar keeps one removal path: the row's delete x is gone, the
  kind menu can remove all worktrees at once, and its overview entry works.
- Research settings shows its per-run options with a Save bar, like the rest.
- One checkbox control everywhere - the same control in every panel.

### Fixed

- HTML file previews render like the file does in a browser. A page's own
  stylesheets, images and fonts used to load as broken links, so previews came
  out unstyled; scripts now run behind an explicit per-file toggle instead of
  being silently disabled.
- Renaming a session shows up in the session list immediately, while the
  rename is still being saved.
- The Memory panel's Global tabs read and compose user memory through the same
  path as the rest of the panel, and no longer show status and action columns
  that did nothing there.
- URA scripts get an honest host interface: paths are anchored where they are
  documented to be, return values match their documented shapes, and saved
  state actually persists.
- The branch dialog's MODE switch is only as wide as its content.

## [0.2.94] - 2026-08-15

### Added

- The Git panel has a **Leftover folders** section. A worker checkout whose Git
  registration is gone but whose directory is still on disk appeared in no panel
  at all — every other section walks `git worktree list`, so a folder Git had
  already forgotten was invisible while the sidebar kept warning about it. Six
  of them sat in this repository. Each row names the branch that still holds the
  work, its size on disk and its age, and removes with one button; a checkbox
  deletes that branch along with the folder. The remover re-checks three fences
  itself: inside this project's `.unary/worktrees/`, not registered with Git,
  and no running session working in it.
- Local branches can be cleaned up in bulk. The section header carries
  **Cleanup merged** and **Delete all but current**, the second of which keeps
  the current branch, the main branch and anything checked out in a worktree,
  acts on the branches the active filter is showing, and asks a second time
  before it force-deletes the ones Git refused as unmerged.

### Fixed

- The sidebar's leftover-worktree warning counts the folders that are really
  there. It counted only the removals that had RECORDED a failure, so a folder
  whose registration vanished quietly was warned about by nobody, and its
  tooltip claimed the work was already in the parent branch — true for a
  recorded cleanup failure, unverified for the rest. It now names where to
  remove them instead.

## [0.2.93] - 2026-08-14

### Added

- GLM-5.3 is selectable on the z.ai Coding Plan. It is the model the plan now
  serves to every tier — 1M tokens of context, up to 128K tokens of output, and
  the same two thinking depths as GLM-5.2, so High and Max keep working as
  before.

## [0.2.92] - 2026-08-14

### Changed

- The scroll-in animation feels soft again. Trading the blur for a short fade
  made it read as a pop, so the curve is long and decelerating again and a small
  blur is back — this time on a promoted layer and only for the first 200ms, so
  the GPU softens a texture it already has instead of redrawing the row on every
  frame. Ordinary trackpad and wheel scrolling keeps the animation now; only a
  real flick still skips it.

## [0.2.91] - 2026-08-14

### Fixed

- Scrolling back through a long chat no longer drifts under the cursor. A row
  that resolves above the viewport is compensated the moment it resolves rather
  than after the scroll settles, and the transcript now measures the row spacing
  the theme actually applies instead of assuming it — it assumed 16px while the
  themes ship 10px, so a few hundred rows of history were off by more than a
  screen.
- Code blocks and long messages start rendering before they reach the screen.
  Their preload distance was measured against the window instead of the chat
  scroller, which meant nothing was ever preloaded and code colorized under the
  cursor as it scrolled in.
- A message taller than about eleven screens could stay invisible: the scroll-in
  animation waited for a fraction of the row to be on screen that a row that
  tall can never reach.

### Changed

- The scroll-in animation keeps its look for a fraction of the cost: history
  rows fade and settle without the blur pass that re-rendered a full-width block
  on every frame, the blur stays on the live row at the end of the transcript,
  and nothing animates at all during a fast flick.
- The transcript keeps more rendered rows ahead of the direction you are
  scrolling and fewer behind it, so a flick meets fewer blank gaps for the same
  amount of work.
- Paging is symmetric again: scrolling back down to the end loads the newer
  messages by itself instead of waiting for a button, and jumping to a message
  stops walking there one screen at a time.

## [0.2.90] - 2026-08-13

### Fixed

- Built-in capabilities that come up after a session starts — the code index,
  Research, Memory and the browser — could be offered to the model and then
  refused the moment it asked for one. The list the model is shown and the gate
  that answers for it are now brought together at the one point both are read,
  so what is offered can always be loaded.
- A code search hit now carries a fingerprint of the symbol it found, so an edit
  can be applied straight from a search result without reading the file first.
  It previously carried a fingerprint of the whole file, which no edit ever
  accepts, and on Windows checkouts the line endings made it disagree a second
  time.

### Changed

- Index and Research say what they are: built in and switchable per workspace,
  rather than extensions that look installable.

## [0.2.89] - 2026-08-13

### Added

- Tools can be activated while a run is in progress: a withheld tool's schema
  arrives immediately instead of waiting for the next turn.
- The Assistant's library tabs are one surface again instead of two lists that
  drifted apart.

### Fixed

- A repeated identical call could lock a run in place. Goal ids now resolve from
  a unique prefix and the repeat breaker stops the loop instead of riding it.
- `activate_tool` accepts every tool name the capability map advertises, so a
  card the model was shown could no longer be refused when it asked for it.
- The sidebar's attention badge counts the rows it actually pulls.

## [0.2.88] - 2026-08-13

### Added

- The Git panel has a section of its own for subagent worktrees. Worker trees no
  longer hide among the checked-out branches, and the two cleanup actions moved
  out of a section they were never about into that one, as "Remove all" and
  "All projects".
- One list of every worktree nobody is working in right now, with the ones that
  never reached the parent branch first. Each row opens its session, folds the
  tree back or merges its branch, reviews the diff, or removes the folder.

### Changed

- The Git panel's buttons are round like every other control, and they now carry
  a meaning instead of one uniform look: filled creates, plain opens a dialog,
  and a filter tints itself while it narrows the list.
- Every worktree states what it is - conflict, not folded back, undetermined, or
  merely left over - and the list says the distribution once instead of leaving
  it to be counted. A leftover folder was previously counted on the project row
  but had no row of its own to act on.

### Fixed

- A project on a single branch listed all of its sessions twice, once flat and
  once inside the branch group.

## [0.2.87] - 2026-08-13

### Fixed

- Every build after 0.2.84 started and then froze. The sidebar's branch poll read
  and wrote the same reactive value inside one effect, so the effect re-ran
  itself without end, spawning a git process per project on every pass.

## [0.2.86] - 2026-08-13

### Added

- A `Tools` control sits beside the chat filter. It shows the catalog mode, the
  tool families callable right now and the ones still withheld, and loads a
  withheld family on request.

### Fixed

- Withholding a tool family only ever changed the prose. The prompt told the
  model to load those families first while their full schemas travelled in the
  request anyway, so the catalog saved nothing and the model spent turns opening
  a door it did not need. A withheld family now stays off the wire.
- A family activated during a session kept being listed as withheld in the
  prompt, so the model tried to activate what it already held.

## [0.2.85] - 2026-08-12

### Fixed

- Switching a checkout's branch in place now moves every session of that folder
  with it. The worktree list behind the grouping was read once per project and
  never again, so it kept naming the old branch and the sessions of one folder
  split across two branch groups - which a single directory can never be in.
- The per-worktree menu in the inactive list was invisible: the rule that
  reveals a row's actions on hover did not cover those rows, so the menu
  shipped unreachable.
- Fold-back is only offered where it can work. It applies to a session's own
  tree under `.unary/worktrees/session/`; on a worker tree it could only ever
  answer "not a Unary-managed session worktree". Worker trees now offer merging
  their branch, reviewing its diff, or removing the folder.
- When fold-back is unavailable in bulk, the menu says which trees block it and
  why, instead of restating the rule next to a count of zero.

### Changed

- A session that still owns a worker worktree nobody folded back is marked on
  the whole row, and its branch group carries the same count as the project.
- The project row states the session count and the attention count in one line;
  the separate chip that repeated it is gone.

## [0.2.84] - 2026-08-12

### Fixed

- Removing a worktree failed with "not a git repository" in the installed app.
  Git resolves the repository from the working directory, never from the path
  you hand it, and the app ran the command in its install folder - so every
  removal reported a failure and removed nothing.

### Changed

- Sessions are grouped by the branch their folder is actually on. A directory
  has one branch at a time, so switching in place now moves every session of
  that folder with it instead of leaving them filed under the branch they
  started on; the old branch stays visible as their origin.
- Every branch with sessions is its own named group. Previously one branch was
  a loose list and all others were groups, which required an answer to "which
  one is main?" that a project with several worktrees does not have.
- Worktrees appear exactly once, at their place. "Not folded back" is no longer
  a second list next to the worker trees but a marker on the tree itself, a
  count on the project row that is visible while the project is collapsed, and
  a filter that narrows the same tree.
- A worktree whose worker is still running no longer counts as "not folded
  back". Unfinished work is what a running worker is for.

### Added

- One list of every worktree nobody is working in right now, with the ones that
  never reached the parent branch first. Each row opens its session, folds the
  tree back, inspects the changes, or removes it.
- A new-session button on every branch group, starting the session in the
  folder that branch is checked out in.

## [0.2.83] - 2026-08-12

### Fixed

- A stale worktree could only be deleted after opening its dialog. The row now
  carries the same delete control, and using it no longer toggles the row.
- The hand cursor had reappeared on two desktop shell surfaces.

## [0.2.82] - 2026-08-12

### Fixed

- A worker was offered withheld tools it could not reach. The role filter
  removed the very entry that activates them, so the prompt advertised eleven
  tools and none of them could be loaded.

## [0.2.81] - 2026-08-12

### Fixed

- Turning Unary Flow off left it running. Every check compared against the
  tool's name while the code needed its family, so the switch matched nothing.
- Plugin tools were missing entirely from sessions that load schemas on demand.
  Memory, index, research, browser, desktop and NotebookLM reported that no such
  tool existed, while the prompt still carried full instructions for using them.
  0.2.80 made withholding effective and, in doing so, froze the catalog before
  the plugins had registered.

## [0.2.80] - 2026-08-12

### Fixed

- Turning Unary Flow off did not actually remove it. The switch compared against
  the tool's old name, so the tool stayed in the catalog and its prefix tokens
  were billed on every request of every session that had deliberately switched
  it off.

## [0.2.79] - 2026-08-12

**Withholding actually works now.** A session configured to load tool schemas on
demand kept the schemas off the wire, and then told the model nothing had been
withheld - so the thirteen tools it had held back were unreachable for the rest
of the session. The prompt listed them correctly; only the way back was missing.

### Fixed

- `activate_tool` answered "This session withholds no tools" in sessions that
  were withholding. The dispatcher never received the gate the session had
  built.
- The tool catalog setting showed "Withheld (default)" whether or not anything
  had ever been saved, so a configured session and an untouched one looked
  identical.
- The Flow toggle still called the tool by its old name. It is `ura` now, and it
  runs Lua in a persistent kernel rather than a declarative workflow language.

## [0.2.78] - 2026-08-11

**The agent picks the right tool noticeably more often.** Three tool families
carried descriptions that were accurate and still left the choice ambiguous, and
the measurement named every one of them. `index` had no per-action descriptions
at all. `list_windows` never said it returns nothing about what is inside a
window, so "look at this window and tell me what's on screen" went to the wrong
action. And memory's repo store and personal store both pointed at each other
without saying what actually separates them - a rule worded as a general habit
("I always …") is still a rule about the repository you are in.

Across the five measured models, action-choice accuracy now sits at 96-100 %,
with tool choice at 96-100 %. Every point came from clearer wording: nothing was
renamed, no action was removed, and no test was relaxed.

### Fixed

- The prompt told the model it could load a tool schema through a search this
  build does not ship, so it invented an activation call and lost two turns
  before giving up. It now names the mechanism that exists.
- Arguments a provider sent as JSON *text* were rejected even though the payload
  inside was correct, and the error repeated the schema instead of naming the
  argument. Both are fixed: the wrapper is undone before validation, and what is
  refused now says which argument is wrong and what shape it needs.
- Failed tool calls were filed as known bugs and replayed into every later
  session - including a guard correctly refusing a costly command, a
  read-before-edit precondition, and once a C++ header as the bug title.
- The same call failing twice in the same way is answered immediately instead of
  after five silent turns.
- NotebookLM shipped a full page of instructions even with no account connected.
  It now costs one line saying it is disconnected.
- An empty reasoning block could be sent to the provider, which is exactly what
  breaks conversation replay and prompt caching.
- Two diagnostic counters read as contradictions because their names implied a
  relationship they never had: a context remainder of 0 looked like "no context
  used", and a rendered-row count looked like more messages than exist.

- Asking to list an app's UI-atlas control ids took a screenshot instead of
  reading the atlas.
- A working preference for the current repository could be filed as a global
  fact about you, where it would follow you into unrelated projects.

## 0.2.77

**The tool catalog no longer ships whole on every request.** A session now
starts from a measured core and loads the rest of a tool's schema the moment it
decides to use one. Measured over a 52-turn session: the same tool-choice
accuracy, 1.9 points of action accuracy given up, and 54 % fewer prompt tokens.
Settings › AI Agent lets you choose: withheld (new default), eager (everything
up front, best accuracy), or fixed.

**Compaction stopped calling a model.** It is built from the session ledger
alone, so it costs milliseconds instead of seconds and works with no
summarisation model configured. The handover also stopped inlining tool output —
it cites the call id instead, so a compacted session no longer grows with the
work it did.

**Tested models and providers are grouped first**, with a badge whose tooltip
carries the actual measurement: tool-choice rate, action-choice rate, and the
date of the run. Nothing is hand-curated; a model is listed because it was
measured.

### Fixed

- Small purpose-built sessions (Pilot controller, assistant) could have started
  with an empty catalog under the new default.
- Flow sessions and worker milestone reporting kept their own tools, which the
  measured core did not know about.
