+4858 people 4ea021f919 fix(opencode): finish status installation with invalid TUI settings (#25031) ...

* fix(opencode): keep server plugin installation independent of invalid TUI config

* test(opencode): format invalid TUI installation control

* Add host-owned OpenCode and Devin managed account profiles (#24636)

* Add host-owned OpenCode and Devin account profiles

* Manage OpenCode and Devin profiles in account Settings

* Expose registered account roots to host transcript readers

* Clarify managed profile provider flags

* Retain isolated Electron home in browser sidecars

* Restore inherited account environment and preserve cleanup retries

* Check relay environment values before merging

* Consolidate managed account type imports

* fix(accounts): use existing localized provider names

Align the new Japanese account copy with the existing catalog repair policy.

* Keep managed account baselines private to the execution host

* Align account enrollment help with accepted providers and flags

* Show the active System account in managed profile lists

* Document the validated Linux managed account scope

* Fix managed account removal, credential audits, and runtime bundling

* Quarantine removed accounts and audit captured credential snapshots

* Continue Antigravity IDE and 2.0 history in new CLI conversations (#24692)

* feat(antigravity): bridge IDE history into new CLI conversations

* fix(antigravity): preserve fresh-launch model and environment for IDE references

* fix(antigravity): forward IDE history opt-in through desktop IPC

* fix(antigravity): rebuild remote IDE reference startup on its host

* fix(antigravity): register IDE continuation action labels

* fix(antigravity): confine IDE references and bound metadata reads

* fix(antigravity): localize IDE continuation badges

* Preserve scanner service cache assertions and refresh Antigravity opening metadata

* Preserve Antigravity opening joins and target folder runtime authority

* fix(opencode): retry timed-out SSH plugin updates (#24666)

Preserve bounded retry behavior and the current-main status-envelope fields.

Original-PR: #24124
Reviewed-source: 103144f9c5

Co-authored-by: Justas Brazauskas <brazauskasjustas@gmail.com>

* fix(opencode): keep Go credentials private and resolve backend keys (#24615)

Preserve the complete credential storage, migration, IPC, Settings and rate-limit refresh change alongside standalone GLM plans, current-main database diagnostics and the reviewed unknown-backend environment correction. Keep native discovery cancellation third and selected environment fourth.

Original-topic-commit: 7903f1cddb
Original-topic-commit: 588117b5cf
Original-topic-commit: dedd4f8c86
Original-topic-commit: 6645dae104
Original-topic-commit: a25b80c02c1af7830b0e6a65e72d965b3ad98276
Restacked-from: a25b80c02c1af7830b0e6a65e72d965b3ad98276
Restacked-onto: b032867021

Co-authored-by: kespineira <kespineira@users.noreply.github.com>
Co-authored-by: kevimux <kevimux@users.noreply.github.com>
Reported-by: pullfrog
Reviewed-full-source: 849fe093073f4c1606bd65d79a0c725d955d0d1f
Native-helper-source: 80dbe23237


Reviewed-full-current-source: 36acb57d44adb3d378c0289c8c15f7da0fda214c

* Skip store notifications when refreshed work items are unchanged (#24530)

An identical forced refresh previously returned an empty Zustand patch, creating a new root state and notifying every subscriber. It now returns the current state after renewing the existing freshness timestamp. Real row or metadata changes still publish once.

* Read project activity once while sorting the sidebar (#24549)

Reuse the existing pure recent rank once per actual entry tuple within each synchronous sort; discard the lookup after sorting.

* Skip impossible link and tag matches in docs search results (#24646)

* Skip impossible HTML matches when formatting docs search results

After the existing link-removal phase, skip the unchanged HTML regex only when the current string has no closing delimiter.

* Skip impossible link and tag matches in docs search results

Skip the five unchanged link regexes when their protected-code input lacks ]; skip the unchanged HTML regex when its post-link input lacks >.

* Decode complete transcript lines without copying their bytes (#24546)

Decode each single owned Buffer synchronously; preserve concatenation for multipart records and remove a redundant tail array copy.

* Clear unused speech worker timers after errors and exit (#24924)

Call the existing idle timer cleanup when private current-worker state is cleared; keep current-worker guards, transcript/error order, deliberate warmth and stop deadline unchanged.

* Reuse documentation search excerpts during result navigation (#24574)

Memoize the existing pure excerpt renderer by complete raw text for the current result array, retaining original rendering as a miss fallback.

* Append subagent handoffs without recopying their message list (#24672)

Append each message reference to its private call-local scope list; retain the final merge and existing ordering.

* Cancel plugin retry timers when their fetch is replaced (#24700)

Reuse the existing timer clear block immediately after the fetch generation advances; preserve live retries and all state publications.

* Avoid rebuilding visible file paths for single-row selection (#24743)

Skip the visible path array only for keyboard replacement selection; the existing replacement branch never reads that order.

* Reuse prepared reads while searching saved agent conversations (#24535)

Use schema-complete session columns to enable the existing bounded statement cache without caching result rows.

* Reuse discovered mobile chunks during static traversal (#24774)

Iterate the existing live visited Set in FIFO discovery order instead of maintaining a second shifted queue; both actual callers own untouched esbuild JSON metadata.

* Skip unused image-size calculations in mobile web browser requests (#24941)

Use the existing mobile density budget only for mobile view; pass the existing constant to the existing assembler for web/default mode, where that argument is discarded. No cache, policy, request or native path changes.

* Skip diff analysis after a result has been discarded (#24711)

Move the unchanged pure render-limit calculation after both existing generation and section-token rejection fences.

* Skip late browser grab toasts after their surface closes (#24727)

Reuse the existing mounted-owner ref at the toast presentation entry while preserving all admitted extraction/screenshot and clipboard completion.

* Classify native chat waiting messages once (#24771)

Use the existing native filter traversal to populate a fresh waiting array, keeping the original predicate, policy, projection and output ordering while classifying each actual private slot once.

* Skip discarded Kanban pruning for an empty selection (#24782)

Guard only the existing open pruning branch when both its private selected Set and nullable anchor are empty; retain all original pre-guard projections and nonempty/anchor-only/public-helper behavior.

* Index discovered test files once during shard selection validation (#24532)

Verified selection plans previously validated each selected filename with a scan of all discovered files. One per-call Set now handles membership checks. Exact source/discovery checks, nonempty selection, fallback to all tests, shard balancing and manifests remain intact.

* Clear the watchdog benchmark heartbeat when CPU sampling fails (#24802)

Move the existing initial CPU observation and histogram enable inside the existing try/finally so their failures clear the sampler's owned heartbeat, preserving successful operation order and original errors.

* Reuse the parsed notification when leasing a push delivery (#24645)

* Reuse the parsed notification when leasing a push delivery

Pass the notification already parsed for the dismissal check into the existing private delivery builder.

* Reuse the parsed notification when leasing a push delivery

Pass the notification already parsed for the dismissal check into the existing private delivery builder.

* fix(accounts): canonicalize account removal to rm

Use account rm as the canonical removal command and keep account remove as an alias. Preserve the existing accounts.removeData RPC and the full original 63-path account component.

Original-Account-Source: cf71ae4cb6
Frozen-Account-Base: 08ee7ba9ef
Frozen-Integrated-Main: 53f9ea7839
Private validation source only; no ref or publication.

* Update README downloads badge

* Let retired-cache GC observations settle across the existing six-turn budget (#24967)

* Collect test-selection evidence when full unit tests fail (#24955)

* Collect advisory unit-selection evidence from failed full runs

* Trigger checks after retargeting the evidence fix to main

* Check each project repository once while filtering mobile cards (#24540)

Reuse exact raw source/slug matching decisions within one project filter call, preserving the matcher, membership, negative matches, output order and identity.

* Skip renderer callbacks after their request has ended (#24631)

Reuse the existing pending-request identity check before starting its deferred renderer callback.

* Decode single-piece saved-session tails without copying them (#24632)

Decode the existing owned Buffer directly when the EOF tail has one piece; preserve the existing concatenation for multiple pieces.

* Avoid irrelevant scroll-action searches in Mac snapshots (#24731)

Check the current pure action name before searching for vertical scroll actions in the same immutable array.

* Stop copying every retained browser page before attachment (#24553)

Find the first page/generation match directly in the existing insertion-ordered Map instead of copying all page references into an array.

* Reuse event byte sizes when relay watchers notify several clients (#24558)

Store the existing grouped numeric byte-size array on the already call-local watcher batch sizing object, for reuse by later chunking clients.

* Stop building unused import candidates in the test planner (#24611)

Replace the existing ordered extension map/find with a loop that forms each candidate only when its existing lookup is reached.

* docs(terminal): explain local macOS/Linux shell startup files

Document the actual login/interactive shell startup-file order and existing shell setup behavior.

Co-authored-by: brynnclaw <261708852+brynnclaw@users.noreply.github.com>
Co-authored-by: Neil <neil@stably.ai>

* fix(editor): recognize Ruby task and configuration files

Add Ruby task/configuration filenames to the existing generated language associations.

Co-authored-by: ggbdpq <ggbdpq@gmail.com>
Co-authored-by: Neil <neil@stably.ai>

* Speed up large Markdown Find and render oversized tables (#24948)

* Speed up large Markdown Find and render oversized tables

* Poll table preview geometry outside hidden renderers

* Preserve Markdown navigation through refresh and tab restoration

* Confirm Markdown restoration when refreshed content is ready

* Trace table refresh positions and update Unicode search reference

* Recognize queued measurement scrolls before restoring Markdown anchors

* Rebuild Markdown Find ranges after renderer components change

* fix(source-control): generate clean OpenCode messages locally and over SSH (#24613)

* fix(source-control): generate clean OpenCode answers on local and SSH hosts

Restack the original focused change onto current main, preserving every owned source and test blob and the merged CI contract and journal cleanup fixes.

Original-commit: 64bb15e3f2
fix(source-control): generate clean OpenCode answers on local and SSH hosts

Use configured models and JSON answer/error events, preserve run-first arguments, and handle the precise v2 variant rejection. Hydrate SSH execution-host PATH through the existing bounded login environment resolver before direct spawning.

Credits: andy-murr (PR #5197 SSH environment intent) and coelho-doti (PR #13065 argument-order intent).

Original-commit: 1b60ec5d11
fix(source-control): retry inline OpenCode model and variant options

Original-commit: 98fdecc7a3
Preserve OpenCode named errors without a data message

Restacked-from: 98fdecc7a3
Restacked-onto: f7b1f9d8be

* fix(ci): prevent concurrent pnpm refresh during mobile typechecks (#24776)

* fix(ci): run mobile typechecks without concurrent dependency refresh

* test(ci): check effective Linux E2E package list

* test(ci): preserve the mobile production compiler barrier

---------

Co-authored-by: Orca Integration Recovery <orca-validation@invalid.example>

* test(terminal): restore the live fish fixture prerequisites (#24947)

A restored pane waits for the initial status replay before subscribing to
PTY output. This fixture never settled that replay, so fish printed its
mode-2031 arm before the renderer connected. Its PTY API also omitted the
reset-input listener required by the serializer, aborting attachment.

Settle and dispose the existing startup-snapshot registration and provide
the same reset-listener mock used by the other PTY tests. The real fish
child-stdin assertions and timeouts remain unchanged. No production change.

* fix(shortcuts): defer TUI editing chords in terminal-first mode (#24640)

Restack the original focused change onto current main, preserving every owned source and test blob and the merged CI contract and journal cleanup fixes.

Original-commit: f707cde14a
fix(shortcuts): defer TUI editing chords in terminal-first mode

Original-commit: 0c6348e49e
docs(shortcuts): describe deferred preview terminal chords

Original-commit: be62c1b6c6
Align worktree history shortcut metadata with terminal conflict policy

Restacked-from: be62c1b6c6
Restacked-onto: f7b1f9d8be

* Register supervised Qoder China and Qwen Code (#24616)

* Add Qoder session history and search with real CLI coverage

* Allow the real Qoder marker file to end with a newline

* Keep Qoder tool output out of history previews and search

* Keep Qoder search pages readable by older clients

* Verify persisted Qoder history after a real generated and resumed task

* Negotiate Qoder filters before searching an older execution host

* Combine search client imports for the CI plugin gate

* Keep the relay search oracle aligned with legacy agent filtering

* Register supervised Qoder China and Qwen lifecycle integration

* Cover Qoder China mobile assets and mixed-host resume gates

* Verify Qoder provider tags against the older released wire parser

* Verify China and Qwen keep independent Windows hook scripts

* Verify Qoder registrations against the installed older Windows release

* test(qoder): align search capability contracts and pin old-host fencing

* fix(qoder): rank exact picker identities and command aliases first

* test(qoder): preserve the regional CLI shared icon expectation

Keep the full bundled-asset and no-remote-image checks, with an explicit
shared-logo basename for Qoder China. The map also works with older
catalog type unions.

* fix(qoder): align China catalog entry with fallback order

---------

Co-authored-by: Orca Integration Recovery <orca-validation@invalid.example>

* Use the measured pnpm lookup policy automatically in hosted root CI (#24951)

* Select lookup automatically for the measured hosted root-install profile

* Record hosted automatic-mode cold cache publication proof

* fix: bound remote generation setup and honor OpenCode option terminators

Count execution-host profile resolution inside the existing request deadline, cancel its waiter promptly, and pass only the remaining time to the child. Shared bounded profile probes keep their existing cache lifetime; the SSH transport margin is unchanged.

Read OpenCode output format from active final-argv options before -- so literal prompt arguments cannot select the JSON finalizer.

Fresh exact-source controls reproduce nine failures before; 139 related checks pass after, including primary/fallback delays, deadline boundaries, cancellation, parser metadata, and SSH lanes. Node typecheck and strict changed-file lint pass.

* Continue Antigravity IDE and 2.0 history in new CLI conversations (#24692)

* feat(antigravity): bridge IDE history into new CLI conversations

* fix(antigravity): preserve fresh-launch model and environment for IDE references

* fix(antigravity): forward IDE history opt-in through desktop IPC

* fix(antigravity): rebuild remote IDE reference startup on its host

* fix(antigravity): register IDE continuation action labels

* fix(antigravity): confine IDE references and bound metadata reads

* fix(antigravity): localize IDE continuation badges

* Preserve scanner service cache assertions and refresh Antigravity opening metadata

* Preserve Antigravity opening joins and target folder runtime authority

* fix(opencode): retry timed-out SSH plugin updates (#24666)

Preserve bounded retry behavior and the current-main status-envelope fields.

Original-PR: #24124
Reviewed-source: 103144f9c5

Co-authored-by: Justas Brazauskas <brazauskasjustas@gmail.com>

* fix(opencode): keep Go credentials private and resolve backend keys (#24615)

Preserve the complete credential storage, migration, IPC, Settings and rate-limit refresh change alongside standalone GLM plans, current-main database diagnostics and the reviewed unknown-backend environment correction. Keep native discovery cancellation third and selected environment fourth.

Original-topic-commit: 7903f1cddb
Original-topic-commit: 588117b5cf
Original-topic-commit: dedd4f8c86
Original-topic-commit: 6645dae104
Original-topic-commit: a25b80c02c1af7830b0e6a65e72d965b3ad98276
Restacked-from: a25b80c02c1af7830b0e6a65e72d965b3ad98276
Restacked-onto: b032867021

Co-authored-by: kespineira <kespineira@users.noreply.github.com>
Co-authored-by: kevimux <kevimux@users.noreply.github.com>
Reported-by: pullfrog
Reviewed-full-source: 849fe093073f4c1606bd65d79a0c725d955d0d1f
Native-helper-source: 80dbe23237


Reviewed-full-current-source: 36acb57d44adb3d378c0289c8c15f7da0fda214c

* fix(opencode): enforce deadline through executable startup

Keep synchronous Windows PATH resolution and child startup within the existing request budget.

---------

Co-authored-by: Orca Integration Recovery <orca-validation@invalid.example>
Co-authored-by: Justas Brazauskas <brazauskasjustas@gmail.com>

* fix(accounts): recover interrupted profile removal on startup

Scan private removal backups before quarantined cleanup, reuse the existing state schema to validate the exact UUID, and preserve registered or unmarked profiles. Mark account rm as destructive using the existing typo recovery policy. Retain the full original account component and public author ancestry.

Account-PR: 24636
Original-Account-Source: cf71ae4cb6
Reviewed-Public-Parent: 6abaf736e3
Frozen-Integrated-Main: 53f9ea7839
Private source only; no ref or publication.

* fix(persistence): preserve projectGroupOrder on restart for flat folder-scan groups

Retain saved project ranks when the project remains in its folder-scan group.

Co-authored-by: lurunzi <lurunzi@gmail.com>
Co-authored-by: Neil <neil@stably.ai>

* fix(runtime): reclaim temp files orphaned by interrupted mobile store writes

Reuse the existing stale temporary-file cleanup policy when each mobile store opens.

Co-authored-by: LDH1103 <ldh517525@gmail.com>
Co-authored-by: Neil <neil@stably.ai>

* fix(terminal): show actionable copy for missing folder workspace paths

Show the existing folder-path failure with concrete recovery instructions.

Co-authored-by: Wayn_Liu <wayntingliu@gmail.com>
Co-authored-by: Neil <neil@stably.ai>

* docs: reference local orca.yaml and .worktreeinclude

Document the existing workspace configuration, include/copy rules and sharing behavior.

Co-authored-by: Neil <neil@stably.ai>

* fix(orchestration): allow model selection for OMP workers

Pass an explicitly requested model through the existing OMP worker launch catalog.

Co-authored-by: Neil <neil@stably.ai>

* fix(cli): preserve primitive success results in JSON output

Check for an object before inspecting screenshot fields so primitive success results remain printable.

Related: https://github.com/stablyai/orca/pull/14735

Co-authored-by: Neil <neil@stably.ai>
Co-authored-by: VXNCXNX <VXNCXNX@users.noreply.github.com>

* Show loaded file changes before deleting a workspace

Expose up to ten already loaded changed paths without altering deletion counts, hydration or authorization.

Related: https://github.com/stablyai/orca/pull/22778

Co-authored-by: Neil <neil@stably.ai>
Co-authored-by: Michiel de Gooijer <mdgooijer@gmail.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(keybindings): record and match Option+digit shortcuts on macOS

Use the physical digit key for explicitly assigned Mac Option+digit shortcuts while retaining modifier checks.

Co-authored-by: marcuslannister <marcus@lannister.cc>
Co-authored-by: Neil <neil@stably.ai>

* fix(editor): don't throw closing a stale active file

Recover stale active-file selection within the owning workspace before choosing a surviving editor.

Related: https://github.com/stablyai/orca/pull/24715

Co-authored-by: m4air <m4air@m4airs-Air.localdomain>
Co-authored-by: Neil <neil@stably.ai>

* Fix Project Settings targeting for multiple local checkouts

Carry the selected checkout identity into the existing project Settings action.

Co-authored-by: fsmeier <1506919+fsmeier@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Neil <neil@stably.ai>

* feat(documents): open CSV and TSV files from the OS

Extend existing OS document associations and delivery to CSV/TSV, preserving restoration and authorization.

Co-authored-by: Neil <neil@stably.ai>

* feat(cli): link GitHub and GitLab items on worktree create and set

Expose and validate the existing workspace link fields, preserving omitted values and provider identity checks.

Related: https://github.com/stablyai/orca/pull/22609

Co-authored-by: marco song <marco.song@mvlchain.io>
Co-authored-by: SongMarco <20613630+SongMarco@users.noreply.github.com>
Co-authored-by: Luca Critelli <lucacri@gmail.com>
Co-authored-by: Neil <neil@stably.ai>

* feat(sidebar): include folder workspaces in keyboard navigation

Use the rendered sidebar row order and host identity when cycling through folder and Git workspaces.

Related: https://github.com/stablyai/orca/pull/10555

Co-authored-by: Neil <neil@stably.ai>
Co-authored-by: JeongUk Park <jeongph.dev@gmail.com>

* fix(build): turn off MSBuild file tracking for Windows native rebuilds

Default Windows native rebuilds to TrackFileAccess=false while preserving explicit caller preferences.

Co-authored-by: B1nh M1nh <43268322+b1nhm1nh@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Neil <neil@stably.ai>

* Fix Ctrl+M in Linux and Windows terminals

Keep the Minimize menu item but disable its accelerator registration on Linux and Windows.

Co-authored-by: Zhichang Yu <yuzhichang@gmail.com>
Co-authored-by: Neil <neil@stably.ai>
Co-authored-by: Ahmed Nagy <ahmednagy25t@gmail.com>

Related contribution: https://github.com/stablyai/orca/pull/24143

* fix(startup): read a nushell login PATH from $env.PATH

Use Nushell login command syntax and preserve the actual login PATH value.

Related: https://github.com/stablyai/orca/pull/22677

Co-authored-by: Kh05ifr4nD <meandSSH0219@gmail.com>
Co-authored-by: Neil <neil@stably.ai>

* fix(cursor): run local hooks through sh for non-POSIX login shells

Keep POSIX hook syntax inside one quoted sh command so the login shell can invoke it safely.

Related: https://github.com/stablyai/orca/pull/22663

Co-authored-by: Kh05ifr4nD <meandSSH0219@gmail.com>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Neil <neil@stably.ai>

* Fix word wrap for both panes in side-by-side diffs

Forward wrapping to both diff panes through the existing editor option path and clean up listeners.

Co-authored-by: Wooseong Kim <innocarpe@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Neil <neil@stably.ai>

* fix(monaco): highlight Svelte block closers inside markup

Return Svelte block closers to the existing markup tokenizer state.

Co-authored-by: Wooseong Kim <innocarpe@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Neil <neil@stably.ai>

* fix(repos): keep active clone dialog open on outside clicks

Ignore accidental outside dismissal only while cloning; Escape, Close and Back still cancel.

Related: https://github.com/stablyai/orca/pull/24581, https://github.com/stablyai/orca/pull/23430

Co-authored-by: Neil <neil@stably.ai>
Co-authored-by: Nawapat Buakoet <nawapat.b@covest.finance>

* fix(editor): keep chat visible after closing Markdown tabs

Count remaining unified chat tabs before clearing workspace selection during editor close.

Related: https://github.com/stablyai/orca/pull/24273, https://github.com/stablyai/orca/pull/23760

Co-authored-by: Neil <neil@stably.ai>
Co-authored-by: Wooseong Kim <innocarpe@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>

* Honor typed starting numbers in chat ordered lists

Forward the existing parsed ordered-list start attribute to both chat Markdown renderers.

Related: https://github.com/stablyai/orca/pull/19765

Co-authored-by: Neil <neil@stably.ai>
Co-authored-by: Frederic Barthelemy <git@fbartho.com>

* Normalize base source once while checking changed-code diagnostics (#24557)

Reuse the exact existing moved-code matching algorithm with run-local lazy normalization of immutable base source blocks across diagnostics.

* Support real OpenCode sessions in native Chat (#24647)

* Use bounded OpenCode context for vault session continuation

OpenCode database and synthetic row paths are not text transcripts. Use the
vault preview or captured pane context, preserving actual transcript paths
containing a hash and supporting both OpenCode lanes and Windows paths.

Adapted the intent of #11859 and extended it to actual installed v2 vault rows.

Co-authored-by: mrcha033 <mrcha033@users.noreply.github.com>

* Read real OpenCode sessions in terminal-backed native Chat

Reuse the bounded AI Vault SQLite worker for v1 and v2 session pages and live updates. Keep terminal input as the real execution path and pace OpenCode Stop through its two-Escape interrupt.

Co-authored-by: xodmd45-ctrl <xodmd45-ctrl@users.noreply.github.com>

* fix(opencode): publish approval cards for permission requests

* Send OpenCode native approval through its Enter selector

* Resolve mobile Chat readability for folder workspaces

* Bound OpenCode part batches and preserve v2 image attachments

* Prefer live migrated OpenCode sessions over legacy copies

* Consolidate mobile Chat eligibility test imports

* Consolidate OpenCode SQLite protocol type imports

* Update native chat settings contract for both OpenCode agents

* fix(native-chat): reconcile bounded OpenCode transcript reads

* fix(native-chat): dispatch OpenCode questions safely

* fix(native-chat): keep native discovery and transcript windows current

* feat(accounts): link standalone GLM Coding Plans (#24618)

* feat(accounts): link standalone GLM Coding Plans

Adapt the reviewed GLM accounts contribution to current main, retain Antigravity behavior, guard late credential results, expose storage protection, and redact quota errors.

Co-authored-by: Luchong <lu740528977@gmail.com>

* fix(accounts): retain GLM credential results during quota refresh

* fix(accounts): make GLM credential editing desktop-only

* fix(accounts): mirror the host GLM site in paired clients

* fix(accounts): report unknown GLM host details and split web settings tests

Apply the independently reviewed Accounts correction from697284a without the v2 adapter commits. Preserve the saved-key store and serialized write behavior.

* fix(zcode): ship required GLM account translation entries

* chore: record GLM reconciliation hook validation

* chore: validate installed GLM commit hooks

* test: complete GLM account fixtures and web API inventory

---------

Co-authored-by: Luchong <lu740528977@gmail.com>

* fix(native-chat): route transcript requests through shared SQLite worker

* fix(ci): prevent concurrent pnpm refresh during mobile typechecks (#24776)

* fix(ci): run mobile typechecks without concurrent dependency refresh

* test(ci): check effective Linux E2E package list

* test(ci): preserve the mobile production compiler barrier

---------

Co-authored-by: Orca Integration Recovery <orca-validation@invalid.example>

* test(terminal): restore the live fish fixture prerequisites (#24947)

A restored pane waits for the initial status replay before subscribing to
PTY output. This fixture never settled that replay, so fish printed its
mode-2031 arm before the renderer connected. Its PTY API also omitted the
reset-input listener required by the serializer, aborting attachment.

Settle and dispose the existing startup-snapshot registration and provide
the same reset-listener mock used by the other PTY tests. The real fish
child-stdin assertions and timeouts remain unchanged. No production change.

* fix(shortcuts): defer TUI editing chords in terminal-first mode (#24640)

Restack the original focused change onto current main, preserving every owned source and test blob and the merged CI contract and journal cleanup fixes.

Original-commit: f707cde14a
fix(shortcuts): defer TUI editing chords in terminal-first mode

Original-commit: 0c6348e49e
docs(shortcuts): describe deferred preview terminal chords

Original-commit: be62c1b6c6
Align worktree history shortcut metadata with terminal conflict policy

Restacked-from: be62c1b6c6
Restacked-onto: f7b1f9d8be

* Register supervised Qoder China and Qwen Code (#24616)

* Add Qoder session history and search with real CLI coverage

* Allow the real Qoder marker file to end with a newline

* Keep Qoder tool output out of history previews and search

* Keep Qoder search pages readable by older clients

* Verify persisted Qoder history after a real generated and resumed task

* Negotiate Qoder filters before searching an older execution host

* Combine search client imports for the CI plugin gate

* Keep the relay search oracle aligned with legacy agent filtering

* Register supervised Qoder China and Qwen lifecycle integration

* Cover Qoder China mobile assets and mixed-host resume gates

* Verify Qoder provider tags against the older released wire parser

* Verify China and Qwen keep independent Windows hook scripts

* Verify Qoder registrations against the installed older Windows release

* test(qoder): align search capability contracts and pin old-host fencing

* fix(qoder): rank exact picker identities and command aliases first

* test(qoder): preserve the regional CLI shared icon expectation

Keep the full bundled-asset and no-remote-image checks, with an explicit
shared-logo basename for Qoder China. The map also works with older
catalog type unions.

* fix(qoder): align China catalog entry with fallback order

---------

Co-authored-by: Orca Integration Recovery <orca-validation@invalid.example>

* Use the measured pnpm lookup policy automatically in hosted root CI (#24951)

* Select lookup automatically for the measured hosted root-install profile

* Record hosted automatic-mode cold cache publication proof

* fix(native-chat): keep OpenCode history usable at read limits

Continue past failed database probes while preserving discovery cancellation.
Verify rows displaced by a capped tail before deciding whether to replace history.
Represent oversized v1/v2 rows with the existing omission text and stable cursors.

* Continue Antigravity IDE and 2.0 history in new CLI conversations (#24692)

* feat(antigravity): bridge IDE history into new CLI conversations

* fix(antigravity): preserve fresh-launch model and environment for IDE references

* fix(antigravity): forward IDE history opt-in through desktop IPC

* fix(antigravity): rebuild remote IDE reference startup on its host

* fix(antigravity): register IDE continuation action labels

* fix(antigravity): confine IDE references and bound metadata reads

* fix(antigravity): localize IDE continuation badges

* Preserve scanner service cache assertions and refresh Antigravity opening metadata

* Preserve Antigravity opening joins and target folder runtime authority

* fix(opencode): retry timed-out SSH plugin updates (#24666)

Preserve bounded retry behavior and the current-main status-envelope fields.

Original-PR: #24124
Reviewed-source: 103144f9c5

Co-authored-by: Justas Brazauskas <brazauskasjustas@gmail.com>

* fix(opencode): keep Go credentials private and resolve backend keys (#24615)

Preserve the complete credential storage, migration, IPC, Settings and rate-limit refresh change alongside standalone GLM plans, current-main database diagnostics and the reviewed unknown-backend environment correction. Keep native discovery cancellation third and selected environment fourth.

Original-topic-commit: 7903f1cddb
Original-topic-commit: 588117b5cf
Original-topic-commit: dedd4f8c86
Original-topic-commit: 6645dae104
Original-topic-commit: a25b80c02c1af7830b0e6a65e72d965b3ad98276
Restacked-from: a25b80c02c1af7830b0e6a65e72d965b3ad98276
Restacked-onto: b032867021

Co-authored-by: kespineira <kespineira@users.noreply.github.com>
Co-authored-by: kevimux <kevimux@users.noreply.github.com>
Reported-by: pullfrog
Reviewed-full-source: 849fe093073f4c1606bd65d79a0c725d955d0d1f
Native-helper-source: 80dbe23237


Reviewed-full-current-source: 36acb57d44adb3d378c0289c8c15f7da0fda214c

---------

Co-authored-by: mrcha033 <mrcha033@users.noreply.github.com>
Co-authored-by: xodmd45-ctrl <xodmd45-ctrl@users.noreply.github.com>
Co-authored-by: Luchong <lu740528977@gmail.com>
Co-authored-by: Orca Integration Recovery <orca-validation@invalid.example>
Co-authored-by: Justas Brazauskas <brazauskasjustas@gmail.com>

* fix(accounts): protect registered UUID case variants during removal recovery

Compare validated account UUID identities without changing stored IDs or filesystem paths. Protect registered originals and quarantines, including explicit retry variants, and accept equivalent UUID spelling in a valid removal backup. Retain the complete account component and published contributor ancestry.

Private source only; no index, ref, or public mutation.

* Drain removal fixture jobs before resetting and deleting their records (#24977)

* Reuse measured Electron preparation for current Terminal Perf refs (#24968)

* feat(jcode): add Jcode as a supported TUI agent with managed hooks

Ports PR #10521 onto current main: agent catalog, managed hook service,
agent-status listener, session resume, AI Vault parser, per-pane daemon
isolation, and Source Control AI support.

Co-authored-by: Neil <neil@stably.ai>

* feat(jcode): evidence-backed status pipeline (turn_start, live pre_tool, questions)

Co-authored-by: czzczz <chanzrz_zbf@foxmail.com>

* fix(jcode): register vault fixture, dynamic model discovery, script refresher

Co-authored-by: czzczz <chanzrz_zbf@foxmail.com>

* fix(jcode): keep finished-turn detail so completion notifications fire

Co-authored-by: czzczz <chanzrz_zbf@foxmail.com>

* fix(jcode): show the prompt in agent rows instead of jcode's repainting title

Co-authored-by: czzczz <chanzrz_zbf@foxmail.com>

* fix(jcode): pre-warm the per-pane daemon so a cold runtime dir cannot time out

Co-authored-by: czzczz <chanzrz_zbf@foxmail.com>

* fix(jcode): stop the OSC color skip from crashing every pane connect

Co-authored-by: czzczz <chanzrz_zbf@foxmail.com>

* refactor(jcode): fold three reverse-scan copies into one, reuse the shared hook POST

The jcode journal reader, the Claude transcript reader and the Command Code
transcript reader each carried their own copy of the same reverse chunked line
scan; they now share one tested helper. jcode's managed hook script drops its
hand-rolled curl for buildPosixAgentHookPostCommand, which also gains it the
raw-JSON transport and the --noproxy guard the bespoke copy was missing.

Co-authored-by: czzczz <chanzrz_zbf@foxmail.com>

* fix(jcode): narrow dynamic reads with predicates, pin the Windows hook shape

CI's anti-slop audit rejects Reflect.get: parse dynamic input into a named type
instead. Adds Windows script-shape tests too, since Windows is the platform this
change could not be exercised on directly.

Co-authored-by: czzczz <chanzrz_zbf@foxmail.com>

* test(jcode): measure the gate against the synchronous path, not the clock

The absolute 1s bound was the flake CI shard 3/8 hit: it is tight enough to catch
a synchronous gate on an idle laptop and too tight on a loaded runner. Measuring
the same POST both ways on the same machine makes the claim a ratio, which is what
the test is actually about. Mutation-checked: a synchronous gate reads 6120ms
against a 1517ms observer.

Co-authored-by: czzczz <chanzrz_zbf@foxmail.com>

* test(jcode): drop the wall-clock gate assertion

The bound was machine-speed sensitive and flaked on CI shard 3/8 at 1525ms. The
structural assertions (detached gate branch, foreground observer branch) and the
no-hang stdin drain cover the same contract without a timer.

Co-authored-by: czzczz <chanzrz_zbf@foxmail.com>

* fix(jcode): address CodeRabbit review on #22539

- Windows posted no payload at all: the shared builder reads `payload@-` from
  stdin, which the gate has already drained and observer hooks never receive, so
  every Windows event was dropped. Write the env var to a temp file and pipe it.
- The daemon pre-warm never fired for daemon-host spawns, which is the default
  local path; it now runs there too, and from the final env so the daemon gets the
  hook port and token.
- A failed runtime-dir mkdir took down every local terminal, jcode or not.
- removeJcodeManagedHooks matched the raw line, so a user hook whose comment
  mentioned the managed script was deleted; matching on Windows never worked.
- A managed entry left by a copied home or a platform switch is now repointed
  instead of being reported as user-owned forever.
- The OSC colour skip only checked launchAgent, so a command- or telemetry-named
  jcode pane still leaked the reply into its composer.
- `['hooks']` and a commented scalar are recognised, instead of appending a
  second [hooks] table that makes jcode reject the whole config.
- A failed tool's error is marked as tool output rather than agent prose.
- The vault keeps a session's stored name, counts its tokens, and skips
  background_task and [Scheduled task] turns.

Co-authored-by: czzczz <chanzrz_zbf@foxmail.com>

* fix(jcode): read the journal once per turn, key prompts by byte offset

The jcode prompt reader ran a synchronous bounded file scan plus a JSON
parse on every hook event. jcode blocks on pre_tool, so a turn that ran
four tools charged the user eight scans of latency it did not need — the
prompt cannot change inside a turn.

- Cache the journal read per pane, refreshed on the turn boundary that
  can change it. The cache holds the whole evidence record, since
  hasExplicitUserPrompt needs the transcript-evidence flag and not just
  the text, and it joins the existing pane-scoped lifecycle (close,
  rename, reset) rather than living in a module singleton.
- Key a journal prompt by its absolute byte offset instead of its
  region-local line index. The backward scan windows the file from EOF,
  so appending shifted every boundary and reminted the key for a prompt
  that never moved; a repeated turn_end then slipped past the same-hash
  dedupe as a second done event with duplicate telemetry.
- Pin the platform in the daemon pre-warm tests. The pre-warm is a no-op
  off POSIX, so the dedupe and retry cases would have passed vacuously
  on a Windows runner.

Co-authored-by: czzczz <chanzrz_zbf@foxmail.com>

* fix(jcode): quote the managed hook path, drop tools from patch prompts

Three fixes, all on paths this PR could not exercise locally.

The managed hook command was stored as a bare path. jcode tokenizes that
string shell-style before exec'ing it directly (parse_hook_command,
crates/jcode-terminal-launch/src/lib.rs): unquoted whitespace splits, and
every unquoted backslash is consumed as an escape. So on Windows
`C:\Users\me\.orca\agent-hooks\jcode-hook.cmd` reached exec as
`C:Usersme.orcaagent-hooksjcode-hook.cmd` and no hook fired at all, and a
POSIX home with a space split into two arguments. Store the path
single-quoted (verbatim, backslashes included), falling back to double
quotes for a path containing a single quote. Existing bare entries are
already repointed by the stale-key path, and getStatus accepts both forms
so the repair is not reported as a user-owned hook. The quoting helper was
previously dead code that only tests called; the three production sites
now use it. isJcodeManagedCommand also normalizes separators, since a
`/`-only needle never matched a Windows entry.

Commit-message generation feeds a staged patch to `jcode run` as the
prompt — attacker-influenced text — while jcode's default profile exposes
shell, read, write, and MCP. Pass `--tool-profile none`, which resolves to
an empty allowed-tool set in jcode's config (base_allowed_tools), matching
the read-only posture claude (plan) and codex (read-only) already take.

docs/reference/jcode-hook-events.md was never actually in this PR: the
repo ignores docs/** and tracks reference docs by allow-list only, so the
captured-payload evidence four source comments point at was silently
dropped. Allow-list it.

Co-authored-by: czzczz <chanzrz_zbf@foxmail.com>

* test(jcode): pin the tool-profile flag in the generation plan too

The argv assertion lives in two places; --tool-profile none only landed in
one, so the plan test still expected the unrestricted argv.

Co-authored-by: czzczz <chanzrz_zbf@foxmail.com>

* fix(text-generation): refuse an oversized argv prompt on Linux too

The pre-spawn size guard only ran on Windows. Linux caps a single argv
entry at MAX_ARG_STRLEN (32 pages, 128 KiB on a 4-KiB-page host) and
execve fails with E2BIG past it, so an agent that delivers the whole
prompt as one argument — jcode, and the other argv-delivery agents —
failed on a large staged diff with an error the user could not act on.

The cap is per-argument and in bytes, which is why it is not the Windows
line budget: 40k chars trips Windows and is nowhere near the Linux limit,
so folding them together would have refused prompts Linux runs fine.

Co-authored-by: czzczz <chanzrz_zbf@foxmail.com>

* fix(jcode): register in main's remote-installer guard, drop our duplicate

Rebasing onto 853 commits of main surfaced two things the earlier branch
had hidden.

main already owns a guard for the issue-#7253 bug class
(`remote-hook-service-registry-coverage.test.ts`). This branch had added a
second, near-identical one — a parallel implementation of a test that
already existed, which is what AGENTS.md's reuse rule is about. Deleted
ours and registered jcode in main's, which is the one that has kept pace
with every agent added since.

Also fixes a missing separator in the mobile icon map. `pnpm tc` does not
cover `mobile/`, so only the session-route closure suite caught it.

Co-authored-by: czzczz <chanzrz_zbf@foxmail.com>

* fix(jcode): decompose the four files jcode pushed over max-lines

Adding an agent tipped four modules past their line budget. AGENTS.md
forbids a `max-lines` disable or a per-file bump, so each is split on a
real seam rather than silenced:

- agent-catalog.tsx keeps `AgentIcon`, which 70+ files import, and the
  rows move out. The rows alone exceed the 300-line budget a `.ts` file
  gets, so they follow the primary/secondary split this repo already uses
  for commit-message agent specs.
- getAgentResumeArgv -> agent-resume-argv.ts, re-exported so the 18 call
  sites keep one import path.
- isDiscoverableSessionFile/pathSegments -> session-file-discovery.ts.
- remoteCodexSources -> remote-session-scanner-codex-sources.ts; Codex is
  the one remote agent with two CODEX_HOME roots.

Also:
- Records the readiness-census baseline jcode now needs. main added that
  gate while this branch was out; the fixture is the recorded 74-case
  matrix, not a hand-written one.
- Restores two entries a rebase resolution silently dropped from
  config/tsconfig.cli.json (gitlab/project-ref-parser,
  startup/shell-path-probe). Nothing to do with jcode; losing them was a
  conflict-resolution mistake on this branch.

Co-authored-by: czzczz <chanzrz_zbf@foxmail.com>

* feat(native-chat): open structured chats on the paired Orca server that owns the workspace (#24205)

* fix(native-chat): a host admits structured sessions by client capability, not its own chat setting

A host's experimentalStructuredNativeChat decided whether any paired client could reach
agentSession.* at all, and whether session.tabs.* showed it structured tabs. That setting is the
host user's own launch preference: whether a new agent opens as a chat or a terminal is decided by
whoever launches it. Using it as admission control meant a client whose own preference was
"structured chat" was refused on a host whose preference was "terminal", and chats opened while
the setting was on were withheld from mobile once it was turned off.

The gate now asks one thing: did the client advertise agent-session.structured.v1 (in-process
callers negotiate nothing and are always admitted). Tab projection and restore follow the same
rule. With the setting no longer gating anything, the separate cleanup gate (close, cancel,
unsubscribe, release), which existed only so those kept working after the setting was switched
off, is identical to the main gate and is folded into it. The settings listener that republished
tabs when the setting changed is removed, since projection no longer depends on it.

The host setting still picks the default for launches that start on the host itself
(agent.launch from mobile, orchestration worker-start).

* fix(native-chat): the desktop declares structured chat support to paired hosts

The desktop renderer advertised agent-session.structured.v1 (and the Claude, turn-item and
background-task capabilities that go with it) to its own main process but not to a paired Orca
server. The server therefore refused every agentSession.* call from the desktop and stripped
structured chat tabs out of the tab list it published to it, so a structured chat running on a
paired server never appeared on the desktop, even though the renderer already mirrors a host's
agent-session tabs and drives each one against the server that owns its workspace.

The same renderer reads structured chats on either host, so the remote Electron list now carries
the same structured-session capabilities as the local one, and the capability test pins that
nothing is advertised only locally.

* feat(native-chat): open structured chats on the paired server that owns the workspace

With the structured-chat default on, an agent launched in a workspace that lives on a paired Orca
server always opened as a terminal (or the terminal-backed chat view). Three things kept it off
the structured path: the launch check refused every host but this machine, a remote workspace was
handed to the host-published terminal path before the structured route was even considered, and
the structured launch pipeline sent create and every follow-up call to this machine's runtime.

A workspace's owning runtime is fixed, so the pipeline now derives it from the workspace instead
of assuming this machine (structured-agent-session-owner.ts, the same derivation the chat pane
already uses to read a session). The launch intent carries that target; the pre-create support
check, create, the publication check and fence read, the launch prompt send, held option picks,
the "focus this chat" marker, the placeholder tab's host, and tab close/purge all use it.

The launch check now accepts a paired server and asks that server's own capabilities (read from
the status the client already cached for it) rather than this machine's. An SSH workspace stays
terminal-backed: no Orca runtime runs there. The host still answers createSupport before
anything is created, so an older server that refuses shows the failure in the chat tab.

A chat the user closed before its create landed is now also retired on the paired server when it
publishes, as the local sync already does. Orchestration workers placed on another runtime are
unchanged: federation creates terminal agents only.

* fix(native-chat): negotiate client-chosen launch mode so released phones and old servers keep terminals

Hosts advertise agent-session.structured.client-launch-mode.v1: they admit
structured sessions by client capability alone. A remote client that does
not advertise it (phones released before agent.launch) asks createSupport
to pick the launch mode, so the host keeps answering that with its own
setting, exactly as before. Cleanup methods keep their own named gate so a
future admission condition cannot make close or cancel refusable.

* refactor(runtime): keep the Electron client capability list in its own module

protocol-version.ts is at its line budget; the list is what the desktop
advertises to paired hosts, not the host's own contract.

* fix(native-chat): the desktop declares it picks each launch mode itself

Paired hosts and the desktop's own main process then answer createSupport
by the workspace rather than by their own chat setting.

* fix(native-chat): pin each structured chat to the host it was launched on

- Route: a paired server opens a chat only when it advertises the
  client-chosen launch mode; an older server keeps its terminal. Its
  capabilities come from the store's host status, not the compatibility
  cache that is empty after boot or reconnect.
- A launch command override is this machine's: the route applies it only
  locally, and a host's createSupport refuses on its own override.
- The owning host is resolved once, from the same value the route used,
  and carried on the launch intent, its persisted record (legacy records
  load as local), the provisional tab and every mirrored chat tab. Close,
  purge, retry and reload read it instead of re-deriving it from a
  worktree id two hosts can share; an owner that cannot be named refuses.
- Cancellation tombstones record their host: only that host's
  authoritative inventory retires one, restored cleanup closes it there,
  and a paired host's tombstone expires after 30 days if it never answers.
- A paired server's frame settles launches it published, as the local
  inventory already does for this machine.

* fix(native-chat): a paired server that declines a chat opens its terminal instead

createSupport only reads, so both of its non-answers are settled before
anything is created:
- A paired server that answers it cannot run the chat (a WSL repo, a
  Claude account mismatch, its own launch command override) closes the
  chat tab and opens the terminal the route would have chosen, with a
  notice saying why. This machine's own decline stays a failed chat.
- A host that could not be asked closes the chat tab and leaves one
  failure toast, instead of a lingering "could not confirm" chat.

* test(native-chat): a provisional chat carries its launch's host and hands pre-create failures on

* chore(native-chat): justify the two type assertions this change's lines touch

* test(native-chat): state why each staged test fixture is cast

* fix(native-chat): chats that already exist keep showing whatever the chat setting says

The structured chat setting decides only what new agents open as. With it
off, this machine's structured chats used to be hidden while the host,
which no longer reads the setting, still reported them to the workspace
activation gate, so a workspace holding only a chat opened empty. The
local chat mirror and its startup restore now run whatever the setting
says, the continue-after-restart offer follows the chats that exist, and
the setting's copy says it applies to new agents.

* fix(native-chat): the browser client keeps its host terminal on paired servers

A browser client whose own preferences turn structured chat on took the
structured route for every paired-server workspace, but its handshake
never says it reads structured sessions, so the server refused the chat
and the user got a failed chat tab where a host terminal used to open.
The route for a paired host now also asks what this client advertises to
it: the desktop's list does, the browser client's does not. Its handshake
list is now a named constant the route reads, so the two cannot drift.

The chat setting's copy now says it runs on paired Orca servers too;
WSL and SSH hosts still use terminal chat.

* fix(native-chat): a retried launch a paired server declines opens its terminal too

A launch restored after a reload settles only through its Retry, so a
declining paired server left a failed chat there while a first launch got
the server's terminal and a notice. The chat's Retry now hands the same
pre-create failures to the same replacement, carrying the prompt the
launch had staged.

* refactor(native-chat): a paired host's cancelled-chat record ends on its 30-day TTL

The paired census re-read a host's whole inventory after every
authoritative frame to retire tombstones, and a tombstone restored after
a reload needed a second such frame, so in practice it retired nothing.
A tombstone guards a random session id and is inert once stale; the chat
is already closed on its host whenever a frame shows it. The census, its
trigger in the mirror layer and its cleanup are removed; the owner-scoped
tombstones, close-on-sight, the TTL and publication marking from frames
stay.

* fix(native-chat): a chat's pane and status read from the host recorded on its tab

The chat pane and its sidebar status still derived the host from the
workspace id, which two hosts can share; a paired chat in a non-active
same-id workspace was read from this machine. Both now read the owner
stamped on the tab, as close, purge and publication already do.

* fix(native-chat): "Resume in chat" follows the terminal resume's host rule

Agent Session History offered "Resume in chat" for a conversation
recorded on this machine into a paired server's workspace, where its
transcript does not exist. A chat now resumes a conversation only on the
host that recorded it, as the terminal resume does, and that host is the
one asked whether it can resume history.

* fix(native-chat): the chat setting says older paired servers keep terminal chat

* test(native-chat): pin that a host advertises the client-chosen launch mode

* fix(native-chat): mirror this machine's chats only where it holds them

Round 1 ran the local chat mirror for everyone so existing chats show
whatever the setting says. That gave every desktop a permanent
session-tabs listener, which turns on the runtime's phone replication
paths, plus two full session-tab censuses at startup, and made the
browser client mirror its remote host a second time.

The runtime now says whether it holds structured chats: its structured
host is built only when saved chats were restored at startup or a client
created one here, and it announces the moment one is built. The mirror,
the startup restore and the continue-after-restart offer run only when
the setting launches chats or the host holds some, and never in the
browser client. A chat a paired client creates here with the setting off
still appears at once. The chat behaviour settings show wherever chats
exist, and the setting's copy says it picks what new agents open as. The
toggle-off teardown this made dead is removed.

* test(native-chat): route a paired-server launch over the capability lists both sides really advertise

* test(native-chat): record install listeners without a cast

* fix(native-chat): a paired server admits a chat before any of it exists here

The desktop opened a paired server's chat tab, launch record, queued
prompt and focus intent before asking the server, so a "no" needed a
replacement that undid and redid all of it, and every piece it missed
was a bug: the workspace deselected, the caller told "failed" while a
terminal ran its prompt, the caller's arguments and other queued prompts
lost, and a create whose reply was lost treated as never sent.

A paired launch now asks the server first and commits nothing until it
answers. Admitted opens the chat as before. Declined runs the caller's
own launch as the server's terminal, with the existing notice (a resume
fails instead, having no terminal equivalent). Unreachable opens nothing
and names the server in one toast. The new-tab launcher reports the
host's surface for paired workspaces, as it did before paired chats,
with the prompt delivery of whichever surface got the prompt. The
replacement and its error classes are gone, and the probe inside a
launch is back to its old meaning: a "no" is a failed chat with Retry,
and no answer leaves "Could not confirm" with Retry and the prompt kept,
here as on this machine.

* fix(native-chat): mirror this machine's chats only once it holds one, not once its host is built

Session history, resume preparation, terminal resume commands and replay-safe phone launches all
build the structured host for users who never had a chat, which turned on the chat mirror and the
structured-only settings rows until the next restart. The signal is now derived from the host's
records (or a records file still owed its import) and pushed when the first chat is restored or
created. A throwing listener no longer fails the install that fired it.

* fix(native-chat): a fork's reveal never seeds a terminal beside the surface the launcher opens

Forking into a paired-server workspace revealed it as if nothing would open there, so the reveal
created a blank host terminal beside the forked chat (and beside a forked agent terminal on main).
The launcher always opens the fork's surface itself, so the reveal now says so for every surface,
as the fix-checks launch already does.

* fix(native-chat): a declined direct launch keeps the caller's CLI args; an unreachable resume toasts once

When a paired server declines a "Fix checks" chat in a new workspace, the terminal that opens
instead now carries the recipe's saved CLI arguments, launch platform and launch source, as the
terminal route did. "Resume in chat" to a server that cannot be reached showed the admission's
"Could not reach" toast and the vault's generic one; the admission marks its failure notified and
the vault adds nothing.

* fix(native-chat): a declined background create opens its terminal without switching workspaces

Since #23974 a worktree create the user moved away from must not pull them onto the new
workspace. When a paired server declined that create's chat, the fallback terminal opened as a new
agent tab, whose host create selects the workspace. The create now opens its own agent terminal the
way main's background branch does: in place from the request's startup plan (so its CLI args carry),
without selecting the workspace. A create the user is still watching keeps the new-tab fallback.

* test(native-chat): name the launch's host in main's new outbox fence test

Main's new staging-failure test calls settleStructuredAgentLaunchPrompt without the target this PR
made required; it is a local launch, as in the sibling tests.

* fix(native-chat): a paired server's new chat shows no model until the server reports the one it started

A chat on a paired server starts with the server's saved model and options, but the picker showed
this desktop's saved selection (or the catalog default) until the server reported a model, and a
pick made in that window was remembered on the server under that guessed model. A paired launch
now carries no desktop seed, and until the server reports its model the picker names no model and
takes no picks. Local chats are unchanged.

* test(native-chat): seed the paired repo without a cast

The repo literal already satisfies Repo, so the changed-lines cast gate has nothing to excuse.

* feat(native-chat): createSupport reports the saved selection a new chat on this host starts with

A chat on a paired server starts with the server's saved model and options, which the desktop could
not read, so its picker showed a guess. createSupport's answer, which the desktop already waits for
before a paired launch, now also carries that seed as a new optional field (older clients ignore it).
Create and createSupport read it through one resolver so they cannot drift.

* fix(native-chat): a paired server's new chat shows the selection the server will start it with

The paired server now names its saved model and options in the admission answer the desktop
already waits for. That seed goes into the launch intent and its persisted record, so the picker
shows the server's model at once, stays pickable like a local chat, and remembers picks on the
server under that model; a reload shows the same. The locked picker remains only for a server too
old to name a seed.

Also moves host admission and launch-outcome tracking into their own modules: the latest main
merge left structured-agent-session-launch.ts over the max-lines limit.

* test(native-chat): expect the launch intent's new seed argument in exact-call assertions

* refactor(protocol): move the Electron remote client capability list into its own module

Merging main left protocol-version.ts one line over the max-lines limit on this branch. The list of
capabilities the desktop advertises to a paired host moves, unchanged, into
electron-remote-runtime-client-capabilities.ts, the module the next PR in the stack already uses
for it; importers point there.

* fix(native-chat): a paired chat with no saved server model is pickable; Retry shows the server's current seed

A server whose user never saved a chat model sends no seed, and the desktop showed a locked,
model-only picker for it, although that is the common case: no server that can admit a paired chat
predates the seed field. Such a chat now behaves like a local chat with no saved model: the CLI
default, pickable. The lock and its snapshot helper are gone.

Retry kept the first admission's seed while the create probe, which already runs on every attempt,
reported the server's current one and dropped it. The probe's seed now replaces a paired launch's
seed and the picker's, so a retried chat shows what its create will run.

* test(cross-version): stub the launch seed resolver createSupport now reads

* test(protocol): pin the desktop capability divergence against what a paired server receives

Every paired transport sends the shared remote base plus the Electron list, so the
divergence test now compares that union with the renderer's local list instead of
the declared Electron list. A capability added only to the shared base can no
longer slip past it. The two base-only capabilities it surfaced are recorded:
skills.install-result.v2 has no local caller; the authoritative-inventory label is
read by the local tabs sync but dropped by main, and is marked unsettled.

The turn-item and both background-task-stop capabilities were already sent through
the shared base, so the Electron list no longer repeats them. The wire set is
unchanged; this PR's real change on the wire is structured.v1, the Claude
structured capability and the client launch-mode capability.

* fix(native-chat): the desktop tells its own host it picks each launch mode, so retrying an existing chat works with the setting off

* docs(native-chat): name the real exit for the released-phone createSupport rule

* fix(native-chat): the route reads the capabilities a paired host actually receives

The renderer decided whether a paired host would admit a chat from the desktop's Electron list, but
every desktop transport sends that list plus the shared remote base. They agreed only because the
route's checks happened to sit in both. The route input is now built with the same
remoteRuntimeClientCapabilities the transports use (the browser client already sends its list as is),
and a test pins each against the real handshake.

* test(cross-version): a released client still gets the host-setting createSupport answer; a launch-mode client gets supported plus the seed

* test(native-chat): let main's …

* Read OpenCode Go usage from the selected account (#24770)

* Add host-owned OpenCode and Devin account profiles

* Manage OpenCode and Devin profiles in account Settings

* Expose registered account roots to host transcript readers

* Clarify managed profile provider flags

* Retain isolated Electron home in browser sidecars

* Restore inherited account environment and preserve cleanup retries

* Check relay environment values before merging

* fix(opencode): keep saved Go credentials in main-owned storage

Co-authored-by: kespineira <kespineira@users.noreply.github.com>
Co-authored-by: kevimux <kevimux@users.noreply.github.com>

* test(opencode): retain legacy key ownership across worker writes

* Consolidate managed account type imports

* test(accounts): provide OpenCode credential API in lifetime fixture

* fix(accounts): use existing localized provider names

Align the new Japanese account copy with the existing catalog repair policy.

* Keep managed account baselines private to the execution host

* Align account enrollment help with accepted providers and flags

* Show the active System account in managed profile lists

* Document the validated Linux managed account scope

* fix(opencode): resolve Go keys by execution backend

* Use host-private account restoration for OpenCode Go usage

* Fix managed account removal, credential audits, and runtime bundling

* Quarantine removed accounts and audit captured credential snapshots

* fix(accounts): canonicalize account removal to rm

Use account rm as the canonical removal command and keep account remove as an alias. Preserve the existing accounts.removeData RPC and the full original 63-path account component.

Original-Account-Source: cf71ae4cb6
Frozen-Account-Base: 08ee7ba9ef
Frozen-Integrated-Main: 53f9ea7839
Private validation source only; no ref or publication.

* fix(accounts): recover interrupted profile removal on startup

Scan private removal backups before quarantined cleanup, reuse the existing state schema to validate the exact UUID, and preserve registered or unmarked profiles. Mark account rm as destructive using the existing typo recovery policy. Retain the full original account component and public author ancestry.

Account-PR: 24636
Original-Account-Source: cf71ae4cb6
Reviewed-Public-Parent: 6abaf736e3
Frozen-Integrated-Main: 53f9ea7839
Private source only; no ref or publication.

* fix(accounts): protect registered UUID case variants during removal recovery

Compare validated account UUID identities without changing stored IDs or filesystem paths. Protect registered originals and quarantines, including explicit retry variants, and accept equivalent UUID spelling in a valid removal backup. Retain the complete account component and published contributor ancestry.

Private source only; no index, ref, or public mutation.

* Allow cold node-pty setup on Windows ARM CI

Keep a ten-minute bound only for node-pty rebuilt on Windows ARM CI hosts; all other node-gyp calls retain five minutes. Cold setup in two completed ARM jobs left compilation less than a minute.

---------

Co-authored-by: kespineira <kespineira@users.noreply.github.com>
Co-authored-by: kevimux <kevimux@users.noreply.github.com>

* fix(source-control): default Codex to GPT-5.6 Terra low (#24495)

* fix(source-control): use Codex's configured model by default

Source Control AI pinned Codex to gpt-5.5, which Codex retires on
2026-10-14. Any path that still resolves to that slug would then break
commit-message and PR-field generation.

Follow the Antigravity precedent (#21606): add a "Config default" entry
for Codex, make it the default, and omit --model when it is selected so
`codex exec` uses the model from the user's Codex config or Codex's own
default. Explicit model choices still pass --model.

Older remote servers would still build `--model default`, so advertise
git.codex-configured-model.v1 and have clients refuse the sentinel for
servers without it, the same way Antigravity is gated. The gate now
covers both agents, and the two capabilities live in their own module
because protocol-version.ts is at the max-lines limit.

Written with AI assistance (Claude Code).

Fixes #24481

* fix(source-control): honor -m, repo overrides and low effort for Codex

Review on #24495 found three gaps in the configured-model change.
The remote compatibility gate only recognized --model, so a recipe
passing Codex's -m short flag was rejected on older servers. The gate
also ignored the repository's per-operation model override that the
server applies. And moving Codex to Config default dropped the low
reasoning effort the pinned model used, which would change cost and
latency for users whose Codex config sets a higher effort.

* fix(source-control): match gate repo and model checks to the server

Repo ids can repeat across hosts, so the configured-model gate now reads
the repo row for the worktree's host instead of the first id match, the
same row the server applies. A recipe passing --model default or
-m default no longer counts as an explicit model, since an older server
still forwards that sentinel to the Codex CLI.

* fix(source-control): read only the worktree host's repo row in the gate

A worktree that names its own host must not fall back to the runtime
host's repo row, since the server applies the row for the worktree's
host. Also cover an environment id that needs URL encoding.

* fix(source-control): default Codex to GPT-5.6 Terra low

---------

Co-authored-by: Jinjing <6427696+AmethystLiang@users.noreply.github.com>

* fix(jira): search plain text when task search input isn't JQL (#22900)

* fix(jira): search plain text when task search input isn't JQL

The Tasks page sent every keystroke to Jira as JQL, so partial input like `s` failed with a 400.

Fixes #22265

* fix(jira): keep the JQL rejection reason behind Details

For plain-text searches the reason ("Field 'login' does not exist") reads as noise.

* fix(jira): announce the text-match notice to screen readers

* fix(jira): skip the wildcard when the last search word has punctuation

Jira doesn't split a wildcard term into words, so `login,*` or `c#*` matched nothing.

* fix(jira): keep the text-match live region mounted between searches

* fix(jira): keep hyphenated words like sign-in as plain text search

* fix(jira): retry a key-shaped search as text when no issue matches

Input like `utf-8`, `sha-256` or `covid-19` matches the issue-key shape, so
the search ran only `key = "UTF-8"` and showed an empty list with nothing to
explain why. An empty exact-key lookup is a wrong guess, not an answer, so
retry it as a text search — the same let-Jira-decide-then-fall-back rule the
JQL path already uses.

Splits the key and text builders apart so the caller can tell which branch it
took; `buildJiraTextSearchJql` keeps its combined behaviour for the
smart-workspace caller.

* fix(jira): let Jira's answer settle key, JQL and text guesses

Bare `in`/`is` sent everyday phrases down the JQL path, a key lookup Jira
answered with 400 never reached the text search, and a missing key like
PROJ-1401 was read as a lost connection. One key-or-text search now serves
the Tasks page and new-workspace, and one parser reads the status prefix.

* fix(jira): require IN to be its own word before a function name

`input (raw)` and `init()` read as `in` + a function call and took the JQL path.

---------

Co-authored-by: Neil <neil@stably.ai>

* Wait for OpenCode worker input before the first dispatch (#24601)

* fix: wait for OpenCode worker composer before first dispatch

Reuse captured composer readiness on local and paired execution hosts and revoke launching-shell paste anchors.

Co-authored-by: Brennan Benson <79079362+brennanb2025@users.noreply.github.com>

* chore(deps): update reviewed dependencies across Orca (#24561)

* chore(deps): update reviewed desktop dependencies and tooling

* chore(deps): update compatible mobile packages and Fastlane

* chore(deps): update cloud transports and enforce release age

* chore(deps): patch documentation dependencies and record review

* chore: remove dependency review reports

* test(linear): smoke-load resolved SDK through CommonJS loader

* fix(deps): keep native rebuilds from reinstalling addon dependencies

* fix(native): invoke installed node-gyp directly for Node rebuilds

* test(cloud): exclude observer probes from row-lock timing budget

* test(mobile): preserve the CSS writer receiver in viewport spy

* test(native): remove obsolete batch-shim fixture exception

* Stream native rebuild output through the process wrapper

* fix(antigravity): discover global skills from the CLI configuration root (#24592)

* fix(antigravity): keep quota probes free and visible without Gemini OAuth (#24593)

Consolidates the reviewed version and visibility work from #24283 with the probe gating and structured error classification from #24296. Reject unsuccessful version probes, preserve diagnostic precedence, and assert that real quota reads start no model turn.

Co-authored-by: Pablo Werlang <19828711+werlang@users.noreply.github.com>

* fix(antigravity): launch POSIX hooks through sh with bounded JSON stdin (#24596)

* fix(antigravity): make POSIX hooks executable with bounded JSON stdin

* test(antigravity): decode SSH hook shell command before assertions

* test: check plugin fixture worktree cleanup (#24779)

* fix(files): ignore nested generated directories on macOS (#24620)

* fix(files): ignore nested generated directories on macOS

* fix(files): filter generated filenames containing newlines

* Check E2E packages where the installer reads them (#24785)

* test: check plugin fixture worktree cleanup

* test: check E2E installer package environment

* Preserve Mermaid exports through mobile bundling (#24661)

* Update README downloads badge

* ci: skip installed glibc tools in SSH host qualification (#24733)

* Fix terminal width cutoff on wide panes (#24687)

* fix(terminal): let wide panes use up to 1024 columns

Adapt the wider viewport limit proposed in #16578 to the current runtime, shared RPC schemas, and preview sizing.

Co-authored-by: innocarpe <innocarpe@users.noreply.github.com>

* test(terminal): wait for probe output after command echo

* test(terminal): align RPC boundary with wider viewport limit

---------

Co-authored-by: innocarpe <innocarpe@users.noreply.github.com>

* fix(runtime): report an open agent question as blocked to tui-idle waits (#24533)

OpenCode's question tool (and Pi/OMP ask tools and custom modals) put the
hook row in a waiting state but paint no dialog wording the blocked-text
layer knows, so the hook lane read the wait as pending and the tui-idle
wait timed out with no blocked reason. For agents whose hooks are
authoritative, a wait the hook reports with no recognised dialog text now
blocks with the existing agent-interactive-prompt reason, unless input
reached the pane after the row (it may have answered the question before
the next hook arrived).

* fix(runtime): read Codex's 'Implement this plan?' menu as blocked (#24536)

After a Plan-mode turn Codex shows a menu that owns the keyboard, but its Stop
hook has already fired, so a tui-idle wait settled ready and sent text into the
menu. A Codex blocked text anchor now names it an interactive prompt while its
key row ends the tail, written against a new 0.160.0 capture that is replayed
by the readiness census.

* feat: live updates for agent state rules (#24387)

Orca downloads a newer agent-state-rules.json from a fixed GitHub release (stable or next channel), validates it like the bundled rules, and applies it without a restart; a local override wins over the download, which wins over the bundled rules. A hand-started workflow from main is the only publisher; merging publishes nothing.

* fix(runtime): settle Codex tui-idle waits on its hook done, leaving working rows to the rules (#24541)

A headless orca serve has no window to write the Codex ready title, so a
Codex tui-idle wait settled only after three quiet seconds. Rule files gain
profile.hooks: "turn-end": a fresh hook done settles the wait, while a
working or permission row leaves the decision to the rules, so a Codex
whose Esc posts no event (before its Interrupt hook) cannot hang the wait.
Codex moves from identity-only to turn-end.

* fix(release): accept the build identity as a later declarator in the minified telemetry check (#24219)

* fix(windows): link the CLI launcher's C runtime statically so orca.exe runs without the VC++ Redistributable (#24484)

* feat(terminal): point an old terminal's Codex shared-server banner at a new terminal (STA-9051) (#24501)

* feat(terminal): point an old terminal's shared-server banner at a new terminal

A terminal opened before the update that added Orca's codex wrapper is
still served by an older terminal daemon, so a typed codex there joins
Codex's shared server. The banner now says why and offers a new terminal
instead of the global Fix, which changes Codex settings and stops a
server other sessions use.

Detection reads the owning daemon's protocol from the router's in-memory
session map, only after a pane is already found on the shared server.

Refs #24217, STA-9051

* refactor(terminal): simplify the old-terminal banner after review

- Open new terminal now works from Activity, which shows panes from
  worktrees that are not active: it activates the tab's worktree first.
- Inline the legacy-daemon check in the IPC handler over the existing
  getLegacyDaemonAdapters instead of a new routing export.
- One banner frame with the variant chosen inline; the Fix dialog is a
  sibling rather than a children slot.
- Rename CodexSharedServerJoin to CodexSharedServerStatus.

* fix(terminal): open the new terminal in the pane's own workspace, and only promise it where it helps

Open new terminal now always goes through the folder-aware workspace activation
(returning early when that fails) and reveals the floating panel for floating
panes, so folder workspaces and panes viewed from Activity open in the right place.

The old-terminal variant now shows only when the shell Codex was typed into gets
Orca's codex function from this build: zsh, bash and PowerShell from protocol 37,
fish from 39, cmd.exe never. The shell is the parent of the Codex process in the
process table the shared-server check already reads.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* feat(terminal): explain an old terminal's shared server in Learn more

Old-tab banner says Orca now gives each Codex its own server, and gains a
Learn more dialog: why it matters, why this terminal still shares, Open new
terminal, and a quieter way into the existing Turn off / Stop server steps.

* fix(terminal): shorten the old-terminal Learn more copy to one line

* fix(terminal): drop the global fix from the old-terminal dialog

A new terminal already runs Codex on its own server there, so turning off sharing everywhere only changes settings outside Orca and can end other sessions.

* fix(terminal): treat fish without config as a shell Orca does not wrap

* fix(terminal): return focus when a Codex shared-server dialog closes

Both banner dialogs are controlled with no Radix trigger, so Esc or X left
focus on document.body. Capture the active surface when the banner opens a
dialog and restore it on close via useModalReturnFocus; Open new terminal
skips the restore so the new terminal keeps focus.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(terminal): return focus when no new terminal opens, and keep an open banner dialog when Codex ends

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(filesystem): deleting a workspace no longer fails as 'outside allowed directories' in WSL-runtime C:\ projects (#24242)

* fix(filesystem): list a WSL-runtime repo's worktrees through its distro when authorizing paths

* fix(filesystem): record the Git that listed each repo's roots and keep WSL roots under repair

- Registration now takes the distro the caller listed through (create and catalog scan pass theirs)
  instead of re-resolving the runtime, so a stale-routed listing is relisted on the next miss.
- A runtime awaiting repair no longer counts as a routing change, so its last WSL listing stays
  authorized instead of being replaced by a host-Git listing.
- The routing check runs after each awaited refresh, so a runtime switch during an in-flight
  rebuild is corrected in the same request.
- Reuse the shared distro helper for listing; move the drift check into the relist policy and
  refresh the root set once per batch.

* test(worktrees): fail the host-Git scan registration test when nothing registers

* fix: dismiss Codex account prompt and return focus to terminal (#24683)

* chore(worktree): include create timing and spare outcome in the workspace create events (#24483)

* chore(worktree): include create timing and spare outcome in the workspace-created event

The workspace_created and workspace_create_failed events gain optional,
numbers-and-enums-only fields built from what the create already measured:
total and per-phase durations, the prepared-checkout hit/miss and miss
reason, the execution host (local/WSL/SSH), a worktree count bucket, how
many other creates were in flight, whether the repo has a post-checkout
hook (file existence only, probed after the create returns), and for a
failure the phase it died in plus elapsed time. No new git process runs;
consent and opt-out are unchanged.

* fix(worktree): attribute failed_phase by error, label WSL-path repos, skip the hook check with telemetry off

- failed_phase now names the outermost timed step the thrown error (or its cause) left, so a
  caught failure or a concurrent sibling step can no longer be misattributed; the old-relay SSH
  error keeps its cause so it still reads as git_worktree_add.
- execution_host follows the same rule Git routing uses, so a \\wsl.localhost repo reads wsl.
- The post-checkout hook check does not read the repo when telemetry is disabled.
- Privacy page mentions the miss reason code and the failed step.

* test(worktree): pin the old-relay SSH add error to git_worktree_add through its cause

* fix(worktree): name the create event field sets for their role, and type the old-relay test's caught error

* fix(worktree): send create events from runtime creates and record what the spare checkout did

Runtime creates (CLI, agents, phone app, paired clients, orchestration, server
automations) reuse prepared checkouts like the app's own creates, but recorded
no timing and sent no events. Both entry points now start one shared sender
(workspace-create-telemetry.ts), so every create sends exactly one event with
the same fields, plus create_entry_point (app | runtime).

Spare-checkout fields:
- concurrent_preparations: peak prepared-checkout builds and background
  discards running during the create, excluding the one it used; the window
  closes before the create's own re-arm starts.
- prepared_checkout_claim / prepared_checkout_discard phases, so on a miss
  git_worktree_add minus the prepared_checkout_* phases is the plain checkout.
- prepared_checkout_reset (none | base_moved | retargeted) replaces the
  retargeted flag; prepared_checkout_origin (prefetch | rearm) on hits.
- workspace_create_failed carries the spare outcome and its wait.
- repo_index_size_bucket from one stat of .git/index in the existing
  post-create probe (telemetry on, local/WSL only, 2 s cap).

* fix(worktree): add spare build and idle time, the re-arm prefetch origin, and a tracked-file count

- prepared_checkout_build_ms / prepared_checkout_idle_ms on hits: from arming
  the spare to ready, and how long it sat ready before the claim (0 when the
  create waited). readyAt is recorded in the pool's existing ready handler.
- prepared_checkout_origin gains rearm_then_prefetch: an automatic re-arm that
  the dialog prefetch then asked for too, so rearm means the re-arm alone.
- repo_file_count_bucket replaces the index byte size: the entry count from
  the 12-byte index header, which is the same in every index version; left
  out for a split or sparse index.
- The shared sender never lets a failed send change the create's result or
  error; it logs instead and still ends the create's concurrency membership.
- Tests pin the runtime SSH create's timing hand-off and the throwing-send
  cases on both entry points.

* fix(worktree): leave out the file count under any sparse checkout and time spare builds monotonically

- The repo probe also reads .git/config.worktree, where git sparse-checkout
  --sparse-index writes index.sparse, and omits the file count whenever
  sparse checkout or a sparse index is on in either file, with Git's boolean
  spellings. core.hooksPath there is honoured too.
- prepared_checkout_build_ms / _idle_ms use performance.now(), like every
  other duration; the build is timed from its own start (buildStartedAt).
- The origin field comment names all three values.

* fix(worktree): keep the spare's build time on its first build and count worktrees by lock reason

- prepared_checkout_build_ms runs from the first build's start (including any
  wait for the base fetch it is built on) to its first ready; a later tip
  refresh no longer restarts it, though it still counts as new preparation
  work. prepared_checkout_idle_ms runs from the latest ready (build or
  refresh) to the claim.
- The worktree count reads each .git/worktrees entry's locked file and leaves
  out entries whose lock reason names an Orca preparation, the way the
  listing does, instead of subtracting this process's spares. That covers
  spares from other processes, crash leftovers and spares being discarded,
  and cannot run one low while a spare's admin dir does not exist yet. It has
  its own 1.5 s cap inside the probe.

* ci(release): skip the orcad template for tags that predate it (#24872)

A patch cut from a base older than #24155 has no orcad template source, so
the template job could never pass and every desktop build waited on it.

* fix(runtime): make OMP 18.4 workers ready at the composer and keep stale-title clears display-only (#24295)

* fix(runtime): require OMP composer readiness and retain native title evidence

Keep timer-cleared titles in display state so they cannot settle an idle wait.
Require OMP's captured empty composer through the existing screen rules.

Co-authored-by: drakeo338 <paranoyouz@gmail.com>
Co-authored-by: Ahmed Nagy <ahmednagy25t@gmail.com>
Co-authored-by: Saurav M Hiremath <sauravhiremath@gmail.com>

* fix(runtime): veto OMP setup by screen and keep stale-title clears display-only

OMP readiness no longer keys on the composer's thinking-effort hint, which
OMP retires after a few uses, hides once a conversation exists and renders
differently per glyph preset. Instead OMP's setup wizard (alternate screen or
a "Setup step N of M" heading) vetoes every ready lane, and OMP's own `π >`
idle title counts for the omp identity only once quiet on a screen read clear
of the wizard. Other OMP titles take the general path again, so the
post-turn `π -` title, 17.x `π:` titles and first-party blocked waits behave
as before. `π >` is no longer explicit idle for every identity.

The stale-working-title clear stays out of runtime records, but display
readers (worktree ps, phone status and titles, terminal list, orchestration
pointer, orphan adoption and reveal titles) now project records through one
display accessor, so they show the cleared title and status again. A title
the clear retired can no longer prove an agent is present; the foreground
process must, which also covers the retained Cursor spinner, so the Cursor
presence exclusion is reverted.

* test(runtime): refuse OMP's setup splash, which has no step heading

* test(runtime): a genuine title retires the stale-working display clear

* test(daemon): record the OMP captures' inherited shrink cursor divergence

* fix(runtime): settle OMP's idle title by its age and veto only its alternate screen

OMP 18.4 re-asserts bracketed paste every second once a terminal answers
its capability probe, which Orca's terminals do, so an idle OMP pane never
went quiet and worker-start still timed out. Its idle title now counts once
it has stood a quiescence window on a screen read clear of the setup wizard.
The setup veto now reads only the alternate screen: the wizard always runs
there, and a 'Setup step N of M' line on the normal screen is an answer's
own text.

* test(runtime): let the grid ingest OMP repaints before the title arrives

* fix(runtime): keep the stale-title clear's process checks and record lifetime

The stale-working-title clear is now an explicit branch instead of an early
return. It still skips the title/status evidence readiness and delivery
read, but runs what main's clear did to re-derive process state: the
foreground-agent re-probe (an exited shell-launched agent loses its
identity) and the completion-race exit check behind a fresh done hook. The
clear now lives on the PTY record, so an SSH relay drop and reattach keeps
the cleared display, and a recreated tracker is seeded from it. Presence
judges a retired title by display ordering, so a renderer pane title older
than the clear cannot prove an exited agent is present.

* docs(runtime): OMP's idle title precedes its setup wizard by its startup tail

* fix(runtime): judge agent presence by the displayed title while a clear stands

A title the stale-working timer cleared used to veto presence outright, so a
live, busy agent whose title cleared to its name and whose process Orca
cannot recognize read as no agent, and the answer depended on whether a
renderer pane was mounted. Presence and the terminal status query now read
the display projection instead, which is what they read before the clear
stopped rewriting records: a cwd spinner still clears to a neutral title, so
the foreground process decides for an agent that exited behind it.

* fix(runtime): compare a cleared title with a lifecycle cleared the same way

The stale-working clear no longer writes the prompt lifecycle, so the
blocked-dialog check compared a cleared title with a still-working
lifecycle and ignored a trust or update dialog painted after the clear:
terminal status said idle, the guarded send accepted, and a mounted pane's
interactive wait went missing. The lifecycle is now projected through the
clear wherever a cleared title is compared with it, as main recorded both
together. That comparison can only report a wait on the user, never idle.

* fix(runtime): verify prompts behind a stale-title clear against main's baseline

Prompt-submission verification read the native lifecycle while a
stale-working clear stood, so output after a swallowed Enter behind a stuck
spinner counted as delivered, and the first spinner frame after the clear
was not a new turn. Its starting status now goes through the same lifecycle
projection the blocked-dialog check uses, and the first working status
after a clear starts a new turn, as main recorded it. Readiness, delivery
and the mailbox still read only native evidence. The blocked-dialog comment
now says the projected pair reproduces main's verdict.

---------

Co-authored-by: drakeo338 <paranoyouz@gmail.com>
Co-authored-by: Ahmed Nagy <ahmednagy25t@gmail.com>
Co-authored-by: Saurav M Hiremath <sauravhiremath@gmail.com>

* fix(files): replace watcher subscriptions after native failure (#24723)

* fix(files): retry failed shallow directory watch replacement (#24724)

* fix(git): share pending watcher reconciliation startup (#24725)

* Skip dependency installation for known headless build inputs (#24716)

* ci: defer headless dependency installation until graph analysis is needed

* docs: align headless CI rollout with platform and cache policy

* test: isolate headless detector output from the parent CI step

* Reuse pnpm verification records in Alpine CI (#24817)

* ci: reuse pnpm verification records in Alpine builders

* ci: qualify consumers of the verification restore action

* ci: match Linux verification cache archive paths

* ci: overlap ARM SSH setup and independent observation waits (#24714)

* fix(tests): run watcher crash harness against current code (#24705)

* fix(tests): resolve watcher crash harness from repository root

* fix(tests): rebuild watcher crash harness from current sources

* fix(tests): register watcher interruption callback as an owner hook

* ci(release): publish after a skipped orcad template (#24882)

* ci(release): publish after a skipped orcad template

#24872 skips orcad-template for tags that predate it, but a skipped ancestor
skips every job that keeps the implicit success(), so publish-release and the
post-release jobs never ran for v1.4.219.

* test: brace-free filter in the orcad downstream contract

* Reduce scheduled CI cache warming to every six hours (#24881)

* Reduce scheduled CI cache warming to every six hours

* Document cache warmer recovery interval and measured tradeoff

* ci: move ARM Mac qualification to macOS 15 (#24760)

* fix(files): keep watcher streams across SSH reconnect races (#24722)

* fix(files): preserve watcher streams across SSH reconnect races

* fix(types): prune the checked SSH watcher from suppression baseline

* fix(files): retain established SSH watches through connection loss

* fix(files): fence initial SSH watcher callbacks across reconnects

* test(files): verify guarded SSH watch errors and cancellation

* Keep large Markdown previews responsive (#24880)

* Keep large Markdown previews responsive

* Fix large preview review navigation and Find budgets

* Initialize preview scroll caches once and check viewport visibility

* Restore large previews after loaded rows are measured

* Refresh loaded Markdown rows after viewport changes

* Keep Markdown revisions visible and reuse bounded search text

* Skip slower Windows root package-store restores in CI (#24885)

* Skip slower Windows root package-store restores in CI

* Update reviewed mobile dependency-store cache expression

* Advance full shutdown deadlines in mocked Codex connection tests (#24893)

* fix(plugins): restore development watchers after folder replacement (#24726)

* fix(files): retry failed shallow directory watch replacement

* fix(plugins): restore development watchers after folder replacement

* fix(plugins): reconcile root bindings without broad macOS watches

* fix(plugins): avoid source-watch restarts on Windows child edits

* fix(plugins): preserve full filesystem identity precision

* fix(plugins): recognize root replacement on inode-less volumes

* fix(plugins): preserve availability without reliable directory identity

* fix(watchers): preserve directory identities when checking replacements (#24888)

* fix(watchers): preserve full directory identity precision

* fix(watchers): recover replaced inode-less directories

* fix(watchers): decline unavailable creation-time identities

* fix(opencode-go): don't fall back to OPENCODE_API_KEY when the credential database is unreadable (#24605)

* fix(opencode-go): stop before OPENCODE_API_KEY when the credential database is unreadable

An unreadable OpenCode credential database read as 'no key', so the Go key
resolver fell through to OPENCODE_API_KEY, which OpenCode shares with its
Zen provider and can show usage for the wrong key. The database read now
reports unreadable separately; with an env key set the resolver stops, and
the usage fetch uses a configured cookie or shows a readable error.

* fix(opencode-go): treat a denied credential-database listing as unreadable, not missing

* fix(codex): skip the hook trust RPC when Codex's session index is unreadable (#24604)

An unreadable state DB was read as 'not busy', so the trust grant ran a
short app-server RPC that can refresh an abandoned backfill lease. A
tri-state pending check lets the trust grant take its fallback, while other
callers keep their existing boolean behaviour.

* refactor(cursor): move the desktop-login read onto the shared foreign SQLite reader worker (#24603)

* refactor(sqlite): rename the OpenCode SQLite worker entry to foreign-sqlite-reader (STA-9122)

The worker thread that reads OpenCode's database off the main thread is about
to read other apps' databases too, so its entry is renamed to what it is:
src/main/foreign-sqlite-readers/foreign-sqlite-reader-entry.ts, built as
out/main/foreign-sqlite-reader-entry.js.

Why now: #24572 fixed the Cursor focus freeze with a second, dedicated
worker. Rather than grow one worker per foreign app, the next commit moves
Cursor onto this entry and deletes that worker. This commit is the rename
only; #24572's cursor-desktop-profile-worker-entry lines stay until then.

It moves out of ai-vault/ into a new foreign-sqlite-readers/ module because
it will no longer be session-scanner code; the module will own the readers,
their dispatch, protocol and main-process client.

The entry still routes only OpenCode kinds in this commit. The OpenCode
dispatch, protocol and process entry stay in ai-vault/ and stay OpenCode-only,
because the SSH/WSL relay reader bundles them (build-relay.mjs).

Every reference is updated: electron.vite.config.ts input key, knip entry,
the plain-node entry guard and its test, the asarUnpack list (the scanner
service still spawns this entry under ELECTRON_RUN_AS_NODE), and the
electron-builder test that reads the filename. The filename and the
beside-or-one-up (Rollup chunks) lookup now live in
foreign-sqlite-reader-entry-path.ts, which the OpenCode spawn reuses, plus an
Electron-main resolver that uses the packaged app.asar path.

* fix(cursor): move the desktop-login read from its dedicated worker onto the foreign SQLite reader (STA-9122)

#24572 fixed the Cursor focus freeze (#24360) with a dedicated worker
(rate-limits/cursor-desktop-profile-worker*.ts). Orca already runs OpenCode's
database reads on a worker, and more foreign-app SQLite reads are coming, so
keeping one worker per app means one entry, build input, asarUnpack line, knip
entry and guard line each. This keeps one pattern instead: Cursor's
state.vscdb read runs on the shared foreign SQLite reader entry, and the
dedicated worker, its entry and its config lines are deleted.

What moves:
- The read itself is a pure cursorProfile reader in
  foreign-sqlite-readers/readers/ (was rate-limits/cursor-desktop-state-db.ts),
  run only on the worker. A separate dispatch owns the new kinds and refuses
  an unknown kind. The entry routes OpenCode kinds to the untouched OpenCode
  dispatch, so the relay's OpenCode reader stays byte-identical.
- ForeignSqliteReaderClient gives each reader its own WorkerThreadRequestQueue
  lane (own lazily started, idle-torn-down thread; one shared factory) with
  in-flight dedupe per database path. Any failure resolves to the reader's
  existing failure value and never falls back to the main thread.

Kept from #24572, so every reader gets them:
- Await worker retirement before respawning. Worker.terminate() cannot
  interrupt a native SQLite call (e.g. a WAL-index rebuild), so the old
  thread lives on until that call returns; respawning at once stacked a new
  thread on the same work for every timed-out read (#24572 measured three
  live workers). This belongs in the shared host, which fire-and-forgot
  terminate(): LazyWorkerThreadHost now takes awaitRetirement and refuses to
  spawn until the terminated worker settles, and the queue fails calls closed
  meanwhile. Opt-in, because pure-JS clients (session scanner abort, port
  scan) respawn right after an abort. A rejected terminate() also ends
  retirement, so it cannot latch the reader off (raised in #24572's review).
- 10 s Cursor timeout, 2 consecutive deaths, a queue cap of 8.
- #24572's worker tests, rewritten against the shared client: responsive
  caller plus coalesced probes, unavailable worker without path leaks,
  stalled-worker recovery, no respawn before retirement, dispose settles.

Tests: reader, dispatch, client (timeout, 10 s default, crash, malformed,
unavailable without a main-thread read, dedupe, queue cap, own thread per
reader), queue retirement (stalled and rejected terminate), import boundary,
and an event-loop test reading a ~50 MB WAL with no -shm on a real worker.

* test(sqlite): walk the reader import boundary with the shared source-tree scan (STA-9122)

* fix(sqlite): key reader dedupe on a caller-supplied key, not the path alone (STA-9122)

* Skip slower root package-store restores in Linux PR jobs (#24896)

* test: advance mocked Claude child exit deadlines (#24897)

* perf(tab-bar): a change to one tab no longer re-renders every tab (#24261)

With many tabs open, a change to any one tab (a retitle, an agent finishing, a tab switch, a git status write, or a browser tab update on SSH and web clients) re-rendered every tab in the strip, so the strip stuttered. Each tab is now a memoized row that re-renders only when its own values change, with stable handlers, a stable drag id list and stable drag sensor options. Editor tabs get their own git status, and mirrored browser tabs keep their page-id list while the ids don't change.

Part of #24241: opening, closing or reordering a tab still re-renders every tab once.

* Reuse the headless detector compiler without installing full dependencies (#24895)

* Reuse the headless detector compiler without full dependency setup

* Keep optional compiler-cache saves from failing cache warming

* Stop mocked renderer imports from qualifying headless CI (#24902)

* Decouple headless running-work tests from the renderer

* Keep the shared running-work probe contract documented

* fix(opencode): read the binder's session store off the main thread; ship the reader worker in orcad (#24638)

* fix(opencode): read the binder's session store on the foreign SQLite reader worker (STA-9122)

Before: the OpenCode session binder listed new sessions from opencode.db with
node:sqlite on the main thread every 60 s (and on SessionStart kicks), so a
large or contended store could stall the app the same way Cursor's did.

After: the read is a pure openCodeBinderSessions reader in
foreign-sqlite-readers/readers/, run only on the worker. The binder's
correlation, pane snapshot and process sweep stay where they were.

- The binder round awaits listSessions and re-checks its generation right
  after, so a stop() during the read discards the round before it touches the
  unbound map or the watermark.
- The client's in-flight dedupe key now includes the cursor, so a stale round
  from before a restart cannot hand its rows to the restarted round.
- Idle teardown is per reader. The binder lane keeps its thread for 120 s,
  longer than its 60 s poll, so the thread is not respawned every round.
- A timeout, crash, malformed reply or unstartable worker resolves to [] (no
  sessions), the value the old read already returned on failure.
- An absent store still reads as [] without a log line, and a permission or
  corrupt-file failure still logs (kept from #24577, now in the reader: it
  stats the path and throws anything but ENOENT/ENOTDIR to the client's log).
- The binder lane inherits #24572's limits from the shared lane: no respawn
  until a timed-out worker has exited, 2 consecutive deaths, a queue cap of
  8. Its timeout stays 60 s, matching its poll.
- dispatch switches on the destructured kind, so a new kind without a case
  still fails to compile.

orcad: the hook server runs there too, so orcad now ships
foreign-sqlite-reader-entry.js beside orcad.js (ORCAD_ARTIFACTS, built as an
orcad child). build-orcad runs a smoke check that starts the built worker
under the build's Node and under the pinned runtime, and does a real binder
read on a fixture DB, a Cursor read of a missing file and an OpenCode history
list. The OpenCode history scanner uses the same entry and was bundled into
orcad without it, so on orcad it always failed closed; it can now run.

Tests: reader (cursor, same-ms ids, OpenCode 2 rows, missing then created,
corrupt, inaccessible directory), retirement gate for the binder lane, dispatch
routing, client lane (rows, failure -> [], dedupe per cursor, own thread, idle
teardown default and override), binder loop with an async listSessions
(failure -> [], stop during the read), orcad path resolution through orcad's
host adapters, artifact list, and the smoke check against good, missing and
non-reading entries.

* test(opencode): cover the binder read deadline with fake timers and name the failure test accurately (STA-9122)

* refactor(ai-vault): delete the unused session-scanner worker thread (#24607)

* refactor(ai-vault): delete the unused session-scanner worker thread

Production always scans through the forked session-scanner service process;
the worker thread was reachable only under NODE_ENV=test or the
undocumented ORCA_AI_VAULT_SERVICE_PROCESS=0 switch, and nothing fell back
to it on service failure. Remove the thread (spawn, client, protocol, entry,
tests), its build entry, knip and plain-node-guard listings, and the
backend switch, so session-scanner-background always routes to the service.

- Move the scan options type to the service protocol as
  AiVaultServiceScanOptions.
- Tests now mock session-scanner-service-spawn, the seam production calls.
- Repoint the hot-path listing reliability gate from the worker-client test
  to the service-client test, which covers the same bounded-queue,
  cancellation, and fault-restart properties for the real executor.

STA-9122

* test(ai-vault): cover the service's Claude-vs-OMP subagent lister choice

Runs the real service entry and subagent reader, replacing only the two
per-agent listers, so a swapped lister choice fails.

STA-9122

* Skip slower root package-store restores in macOS PR jobs (#24908)

* Skip slower root package-store restores in macOS PR jobs

* Update the companion cache-policy contract

* Overlap independent Linux headless runtime builds (#24910)

* test(persistence): wait for real worker readiness before timeout (#24793)

* Add Qoder session history and search (#24614)

* Add Qoder session history and search with real CLI coverage

* Allow the real Qoder marker file to end with a newline

* Keep Qoder tool output out of history previews and search

* Keep Qoder search pages readable by older clients

* Verify persisted Qoder history after a real generated and resumed task

* Negotiate Qoder filters before searching an older execution host

* Combine search client imports for the CI plugin gate

* Keep the relay search oracle aligned with legacy agent filtering

* test(qoder): align search capability contracts and pin old-host fencing

* fix(cursor): show the primary model pool in compact usage (#24830)

* fix(cursor): show the primary usage pool without hiding exhaustion

Use Cursor's reported plan percentage when its base allowance disagrees.
Select Cursor Models for compact display while preserving maximum-pool
warning, overflow, sorting and collapse behavior.

Adopts the plan mapping and primary headline intent from PR23531.

Co-authored-by: DakaAlvarez <149860458+Dacadev97@users.noreply.github.com>

* fix(cursor): describe compact usage summaries

Correct the existing six translations and fallback to describe one summary per provider after the compact Cursor primary-pool adoption. Preserve quota mapping, selectors, alerts, sorting and detailed presentation.

Validated source patch: d2a233e5f90a2cf8ccfb02dfafe99e0e03add48d with normal installed commit hooks, localization catalogs and changed-code quality. Standalone copy uses a private index and preserves the reviewed candidate ancestry.

Co-authored-by: DakaAlvarez <149860458+Dacadev97@users.noreply.github.com>

---------

Co-authored-by: DakaAlvarez <149860458+Dacadev97@users.noreply.github.com>

* fix(native-chat): one ordered journal writer (#24127)

* fix(claude): settle a queued send the CLI withdrew from its own cancelled frame

Claude reports each uuid-stamped command's lifecycle (queued, started,
completed, cancelled). A send it withdraws from its queue gets `cancelled`
before the interrupt or cancel_async_message answer, so a lost or failed
answer no longer leaves that send pending: it settles as withdrawn, with the
same reason and words as the receipt path.

A command the CLI already started also ends `cancelled` when its turn is
interrupted or fails, so `cancelled` after `started` is not a withdrawal;
an echoed send has left the waiter lists and is never reached.

Tests replay real 2.1.280 captures, scrubbed.

* fix(claude): release a doubted send when the CLI reports its session idle

A Claude send whose write ended in doubt is recorded `unknown`, and a live
`unknown` reads as work still owed, so the chat showed Working until the
child exited. Claude sends `session_state_changed idle` only once its whole
queue has drained, so it can no longer be holding that send. The runtime now
routes that report to the host's existing release, the same one Codex's
thread-stopped report uses; it retires `unknown` only, never `pending`.

* fix(claude): keep a command's started mark when a redelivery re-emits queued; fixtures name msg_lifecycle_v1

* fix(claude): settle every terminal lifecycle state of a send the CLI never echoed

A send the CLI started, then cancelled before any echo, stayed pending: it may
already be in the conversation, so it is released as doubt (unknown, recovered),
never withdrawn and never re-sent. A late echo still accepts it.

The 2.1.280 schema has two more terminal states. `discarded` (the CLI ended its
session with the send still queued) settles as not delivered; `refused`
(declined before it queued) settles as not accepted by the provider. After
`started`, either one is doubt, as `cancelled` is.

The late-settlement path gains an `unknown` outcome, which the host records as
released doubt.

* fix(claude): release a send the CLI took but left unanswered when it goes idle

`session_state_changed idle` comes only once the CLI's queue has drained, so a
send it took that is still unanswered there got no echo and never will: a turn
that throws can leave `started` with no terminal state. Idle releases it as
doubt.

What proves the CLI took a send is its lifecycle frame. On a CLI that reports no
lifecycle, it is the send's place on stdin: one whose write finished before an
interrupt went out was read before the interrupt was, so the first idle after
that interrupt releases it too. A send armed ahead of the interrupt but written
after it is left alone, since the CLI may still run it.

* fix(native-chat): keep the idle sweep off a Claude child that holds a send

A Claude retrying a rate-limited request has taken the send but echoes nothing,
so no turn row exists yet and the sweep rested the child after the idle window,
turning the send into doubt. The adapter now reports whether the CLI holds a
send (lifecycle `queued` or `started`, not yet echoed or ended), derived from
the live waiters, and owed work counts it.

Nothing is stored: every held send leaves the live set on its echo, its
terminal lifecycle state, the CLI's idle, or the child's exit, so the hold ends
with the send.

* fix(claude): count only a started send at idle and as a held send

2.1.280's end-of-turn cleanup can report idle before it re-reads its queue, so
a send read in that window goes queued, idle, started. Releasing every taken
send at idle doubted that live send and dropped Working. Only a `started` send
is released at idle or keeps the child from the idle sweep; a `queued` one ends
by starting and echoing, by a terminal lifecycle frame, or with the child.

The stdin-order path for CLIs without lifecycle frames is removed: a doubted
send retired there disables content matching on CLIs that mint their own echo
ids, and no Orca failure called for it. Those CLIs keep the earlier behaviour.

Comments that said only a failed write or child exit ends a waiter, or that
idle comes only once the queue has drained, now say what ends one.

* docs(claude): say only what the CLI's lifecycle frames and idle actually prove

* fix(claude): hold the idle sweep while Claude has a send queued, not only started

The sweep rested a child whose CLI had queued a follow-up behind a turn, dropping
the send it had already taken. The hold now spans the CLI reporting it took the
send until its echo, a terminal lifecycle state, or the child's exit. The idle
release still covers only started sends: 2.1.280 can report idle before it
re-reads its queue.

* refactor(native-chat): give provider-proven late dispatch settlement its own module

* test(claude): pin a steer a Stop interrupts after it started as doubt, not withdrawn

* fix(native-chat): one ordered journal writer

Every journal write, streamed or direct, lands in the chat's one write queue
in the order it is issued, and has landed in the fold when its call returns
(except while an owed import is paid, when it lands in queue order). The event
sink stops being a queue ahead of it; journal-write coalescing, which moved a
replaced write to the tail, is removed; closing a sink no longer loses writes
already handed over. The ten flushStreamedEvents patches go. A streamed text
item takes its place at its first delta, so a direct write inside the
coalescing window never lands above text already streamed. A Stop interrupts
whatever its bookkeeping writes do.

* fix(native-chat): a failed Stop holds the lane until its queue pause lands

A Stop whose note write or stop call failed skipped the wait for its
withdrawal and pause, so the lane freed first; with writes queued behind
owed work, the drain could hand the waiting card to the agent after Stop.

* fix(native-chat): a journal write body is typed synchronous

The queue's ordering rule needs every write body to finish before it returns;
serialize still took a promise-returning body, so an await inside one would
let a later write land first. The body type now refuses a promise.

* fix(native-chat): a stream's first window snapshot always lands

The first-delta write counted as the growth checkpoint, so the window's
snapshot was skipped until 32 more characters arrived: a stream showed only
its first token, and a Codex reasoning row could sit as an empty aside. The
coalescer now marks the row-creating emit and the first snapshot after it,
and both providers' growth throttles write those.

* test(native-chat): streamed text rewritten in place above a Stop note

Covers a stream whose later deltas wait in the window across a Stop, for
Codex and Claude, and a Codex item completed inside the window.

* chore(native-chat): drop comments that still describe coalesced journal writes

* fix(native-chat): type the draft-table write body synchronous too

* fix(native-chat): an empty delta owes no streamed-text emit

The opening snapshot is forced past the growth rule, so an empty second
Codex delta rewrote the row with identical text. The coalescer now marks a
stream dirty only when a delta adds text.

* fix(native-chat): a mutation's open pays an owed import first

With the flushes gone, a Stop naming no turn, a goal set or a /clear could
read the fold while provider rows still waited behind a restore's owed
import: the Stop answered cancelled:false and interrupted nothing. The open
every mutation shares now waits for the import, as a reader's does; a failed
import is reported and never refuses the mutation.

* chore(native-chat): whenImported says mutations await it too

* test(native-chat): a Stop interrupts before its withdrawal or pause settles

Pins the order for a write that is held and then fails, for a Stop naming
its turn and one naming none.

* test(native-chat): a Stop ends a starting child before its withdrawal or pause settles

* test(native-chat): Stop order tests wait for the interrupt, not a 50 ms timer

* test(native-chat): settlement-order test reads rows through the journal row parser

Replaces a Reflect.get field walk, which the low-evidence audit rejects, with
parseJournalRow and the named row types.

* fix(native-chat): an empty delta still owes its emit, just not a forced one

The previous fix stopped an empty delta from marking the stream dirty, which
broke the pinned contract that an empty stream is snapshotted and flushed.
The coalescer now marks a stream dirty on every delta and tracks separately
whether its text changed since the last emit; only a changed snapshot is
the opening one that skips the growth throttle.

* revert(native-chat): drop the first-text row write from the delta coalescer

The immediate first-delta emit (and the opening flag and counters it needed)
had no Orca-observed failure behind it and added a journal commit per
streamed item. The coalescer, the Claude checkpoints and the tests that
pinned the first-text row go back to main's behaviour; the one ordered
writer, the sink hand-off and the Stop rules stay.

* fix(worktrees): list a folder once when git reports it twice (#24357)

When git lists the same folder twice (a leftover worktree registration that points at the main checkout), Orca's runtime listing turned each line into its own worktree with the same id, so `orca worktree current`, `active` and `branch:` failed with selector_ambiguous, and paired clients saw a duplicate row. The runtime scan now keeps git's first row per folder, the rule the desktop sidebar already uses. Separately, for a bare or separate-git-dir repo added through a linked worktree, the scan no longer relabels the main row with that worktree's folder (it relabels only when the folder's git dir is the common git dir), so the worktree keeps its own row and branch in the CLI and the sidebar. No extra git command runs.

Part of #23631: the "Profile state writer command timed out" toast in that issue has a separate cause.

* Bound AI Vault cache loading and keep atomic saves responsive (#24789)

* Bound AI Vault cache loading and cooperative atomic saves

Preserve schema 3 caches across compatible releases while limiting bytes, JSON structure, and newest unique rows. Keep in-process entries authoritative and retain a valid prior snapshot when the newest row cannot fit.

Credits @AmethystLiang for the original PR10708 cache bounds and cooperative persistence intent.

* Use checked cache JSON properties in cooperative serialization

Preserves lazy own-property access and all serializer bounds, yields and errors.

* Resolve explicitly configured command aliases for workers (#24648)

* feat(orchestration): resolve explicitly configured command aliases

* docs(orchestration): explain configured command aliases

* fix(orchestration): validate configured aliases with target shell grammar

* fix(orchestration): use actual shell and refuse assignment-only aliases

* test: preserve typed calls in configured worker target checks

Replace Reflect.apply with the existing typed prototype call pattern so the unchanged regression cases pass the anti-slop lint gate.

* Cancel the journal import test sampler before database teardown (#24767)

* fix(antigravity): bound Windows hook stdin on the owning runtime (#24622)

* fix(antigravity): bound Windows hook stdin before posting status

* fix(antigravity): publish Windows hook companion before core

* test(antigravity): name Windows hook payload cases explicitly

* Add verified native Antigravity Accounts on the owning runtime (#24691)

* Add verified native Antigravity accounts on the owning runtime

* Keep Antigravity usage tied to its observed native account

* Refuse oversized encrypted Antigravity snapshots before writing

* fix(antigravity): localize account heading and search terms

* fix(worktrees): a worktree delete git fails partway stays listed and can be retried (#23952)

* fix(worktrees): delete removed checkouts in git, not in Orca's file pool

Local worktree removal renamed the checkout into a sibling trash root and
deleted it in the background with a recursive fs.rm in the main process.
That queued one request per entry on libuv's shared 4-thread file pool, so
for minutes every other async fs call in the main process (the agent-session
store behind chat sends, file explorer reads) waited behind the delete.

`git worktree remove` now deletes the checkout inline in git's own process
again, so the card stays in its Deleting state for the length of the delete
while Orca's file pool stays free. No timeout applies to the call, so a
large delete is never killed halfway.

If git reports success but the path still exists (Git for Windows leaves
junctions and their parent directories in place), the leftover is deleted
with the existing removeHostTree; WSL checkouts stay with the distro.

Nothing creates trash any more: the scheduling queue, rename/restore
helpers and the trash_rename span are gone. The startup sweep stays to
drain entries older releases left behind, and now removes each emptied
trash root so the obligation ends.

* fix(worktrees): let Git delete Windows checkouts with long paths enabled

Removal now always runs Git's own recursive delete, and worktree creation
checks out with core.longpaths on Windows, so a deep checkout Orca created
could fail to delete with "Filename too long" (#6433). The Windows recovery
then finishes the delete but keeps the branch. Pass the same command-scoped
core.longpaths option to `git worktree remove` so Git can delete what it
created.

Also point the CI shard timing entry at the renamed real-git removal suite.

* fix(worktrees): keep an inherited GIT_ASK_YESNO out of the worktree delete

Git for Windows asks $GIT_ASK_YESNO whether to retry when a file stays
locked during a recursive delete. Orca's git env inherits the user's
environment, so an inherited value would run an arbitrary prompt program
in the middle of a removal. Drop it for the removal call only.

* perf(worktrees): run worktree deletes under their own limit, outside git admission

`git worktree remove` now deletes the whole checkout in Git's own process,
which takes 20-35 s on a large tree. It took a general git admission slot at
status tier for that whole time, and that cap is as small as two slots on a
machine with six or fewer cores, so two deletes blocked every status read.

Deletes now skip general admission and queue under their own limit of two
per host instead: two concurrent deletes already saturate one disk, and more
only slow each other down. Leftover cleanup runs inside the same slot.

* fix(worktrees): delete removed checkouts in the background and mark them removing

Since the checkout is deleted by `git worktree remove` in Git's own process,
a large delete takes 20-35 s. Answering the request only after that made web
and mobile (30 s), paired desktop (60/180 s) and the CLI (60 s) report a
failure for a delete that was still going, and mobile silently re-showed the
row.

The request now does everything that can refuse (lock, cleanliness, archive
hook, watcher/terminal gate, terminal stop, shared-link unlink), records the
removal in an in-memory table on the host and answers `removing: true`. The
delete, branch cleanup and metadata purge run after it in the same order as
before, and the watcher/terminal gate stays held until they finish.

- Listings mark rows in the table `removing` for clients that advertise
  `worktree.background-removal.v1` (the desktop renderer, paired desktop and
  web), and leave them out for everyone else (older clients, mobile, the
  CLI), which already dropped the row when the request answered.
- The outcome (removed, with any preserved branch, or the error) rides the
  existing worktrees-changed event as an optional field, sent after the row
  has left the table.
- A repeat delete while Git runs joins it. A create at the same path or with
  the same branch is refused with "Cleanup is pending; try again shortly";
  create's name search skips the path, so generated names move on.
- Nothing is persisted: after a quit or crash Git still lists the checkout
  and it can be deleted again. WSL checkouts still delete inline.
- `orca worktree rm` says the checkout is still being deleted.

* fix(worktrees): keep the existing Deleting card until the host's Git finishes

The host now answers a local worktree delete on acceptance and deletes in the
background. The renderer keeps the existing delete state set until the host
publishes how it ended:

- The delete that asked waits for the outcome on the worktrees-changed event
  (local IPC or the paired runtime's client event), then runs the same
  teardown, preserved-branch toast and card error an inline delete did. If
  that event is lost to a dropped connection, a listing that shows the row
  gone after it was marked removing finishes the wait, and one that shows it
  back without the marker fails it.
- Any other renderer (a reload, a paired desktop, web) sets the same delete
  state from the host's `removing` marker and clears it when the marker goes.
  A failure the host publishes lands on that card's existing error.
- Web advertises `worktree.background-removal.v1` so the host sends it the
  marker; paired desktop does through the Electron capability list.

No new component, style or state: the card reads the delete state it always
did. A host that predates this answers when done without `removing`, and the
renderer takes that as finished, as before.

* test(worktrees): type the removal harness and projection for the node typecheck

* fix(worktrees): don't fail a delete retry with an earlier attempt's buffered failure

A background removal's outcome that reached this renderer with no waiter (another client's
delete, a host-marked card, or one already settled from listings) was buffered for 60 s and
consumed by the next delete of the same workspace, so retrying a failed delete failed at once
with the old error while the host was deleting. Drop the buffered outcome before sending the
request; only an outcome that arrives after it can belong to it.

* fix(worktrees): let only a gap in host events settle a background delete from listings

Git unlists the checkout before the host deletes the branch, cleans the push target and purges
metadata, and the worktree-directory watcher refetches within 250 ms. The renderer read the
missing row as a finished delete, so the waiter resolved without the preserved branch (no
toast) and a failure in those last steps showed as success; the real outcome was then dropped.
The listing fallback exists only for a lost outcome event, so it now applies only after this
host's event stream had a gap: a new subscription or a replay after reconnect.

* perf(worktrees): let a bulk delete start each same-repo checkout delete once the host accepts the last

A bulk delete ran one worktree at a time per repo (#2259, for packed-refs and ref-lock races in
branch cleanup). With Git now deleting each checkout for 20-35 s before the request settles, N
worktrees in one repo took N times that. The renderer now queues same-repo deletes only until
the host accepts each one; a parent still waits for its nested children to finish. The host
serializes the branch cleanup step per repo itself, which also covers removals started by
different clients.

* test(worktrees): pin the host platform in the mocked removal suites so they pass on Windows

Removal now passes -c core.longpaths=true on Windows, so the exact-argv
assertions and command-keyed mocks never matched there (17 failures on a
Windows host). Pin darwin as the add-worktree suites already do, and drive
the one Windows-specific case through the same spy.

* test(worktrees): type the blocked git remove result instead of a broad object

The anti-slop static-analysis gate rejects `object` parameters.

* test(worktrees): clear the changed-code quality gate in the removal suites

Merge the duplicate node:fs import, build the mock child without a cast, read
worktrees:list rows through one typed helper, and give the remaining casts a SAFETY line.

* fix(worktrees): record each background delete durably and finish it after a quit or crash

A quit mid-delete left git to finish the checkout on its own while the branch
delete and metadata purge never ran; a crash left a normal-looking row. Each
accepted local removal now writes a record beside the profile state before git
starts, clears it on success or failure, and the host runs the same delete
again for any record left at startup, re-deriving what remains from git and
disk. An orderly quit stops the checkout delete without waiting for it.

* test(worktrees): type the interrupted-removal assertions for the node typecheck

* fix(worktrees): finish an interrupted delete that already removed the checkout's .git file

Quit stops git worktree remove mid-delete, and Git deletes the checkout's .git
file wherever it falls in directory order. Git then refuses the checkout
("validation failed ... .git does not exist") on every retry, so the startup
finish failed and the row could never be deleted from Orca. A registered
checkout this record owns that has lost its .git file now finishes like an
unregistered one: leftover files, prune, then the branch.

* fix(worktrees): let Git finish an interrupted delete, and never take a different checkout

A quit or crash that stops `git worktree remove` after it deleted the checkout's
.git file left a registered checkout Git refuses to remove. The previous fix
deleted that leftover inside Orca's process, which is the bulk delete this
change exists to avoid (and on Windows the leftover can be most of the
checkout). The startup finish now rewrites the missing .git file from Git's
own admin entry for that path and lets `git worktree remove --force` delete
it. `git worktree repair` is not used: it also re-points every other
registered path, including a checkout another repository now owns there.
Orca deletes the leftover itself only when no admin entry claims the path.

The startup finish forces, so it now leaves the path alone when the checkout
there is not the one recorded: a registered worktree on a different branch or
head, or a `.git` at a path Git already unregistered. The record is dropped and
the card shows why.

The record write before Git starts is now bounded (2 s, logged when exceeded)
so a stalled disk cannot hold the delete, and the outcome is published before
the record's clear reaches disk.

* test(worktrees): compare worktree paths by value and tear down with Windows lock retries

Git prints forward slashes in `git worktree list` on Windows, so the real-Git
removal suites never found a joined path there: positive checks failed and
negative ones passed without proving anything. They now compare Git's parsed
rows by value. Teardown uses the shared retrying removeTree, since Windows can
hold the deleted checkout busy for a moment after Git exits. Adds a
relative-path worktree case for the .git restore (skipped before Git 2.48).

* fix(worktrees): reply to a worktree delete when it has finished, not on a broadcast event

A current client's delete request now waits for the host's background delete and gets its real
result (removed, a preserved branch, or the error) as the reply, the way it did before the delete
moved off the request. A request that arrives while the delete runs joins it and gets the same
result. Every other view keeps reading the host's `removing` marker: the row leaving means the
delete finished, and the row listed again without the marker shows "The delete did not finish.
Try again." on a card that view had marked Deleting. A request whose reply is lost (a timeout or a
dropped connection) settles the same way from a fresh listing instead of reporting a failure.

Clients without the background-removal capability (mobile, the CLI, older desktops) are still
answered on acceptance and have rows under removal left out of their listings.

This removes the outcome on worktreesChanged and everything it needed: the renderer's outcome
waiters, early-outcome buffer and TTL, per-host event-gap generations, the request pre-registration,
and the accept callback bulk delete used. Bulk delete runs same-repo deletes in parallel only on
this machine, whose host serializes branch cleanup per repo; SSH and paired hosts stay serialized.

* test(worktrees): type the pending-removal host id in the background-removal suite

* fix(worktrees): answer a delete request even when a concurrent removal of the same worktree replaced its record

The desktop app's removal and the runtime removal (CLI, paired clients) coalesce separately, so
both can be accepted for one worktree. The second replaced the first's record, and the first
delete then finished without resolving the request waiting on it, leaving the desktop card on
Deleting indefinitely. Each delete now settles the request it was started for.

* fix(worktrees): run same-repo removal archive hooks and teardown one at a time on the host

Local bulk delete now sends same-repo removals in parallel, so their archive hooks, terminal
teardown and preflight ran at once; a hook that writes refs can race the repo's ref locks
(#2259). The host now serializes each local removal up to acceptance per repo, for every
client; Git's checkout delete still runs in parallel under the delete limit.

* fix(runtime): keep waiting worktree deletes out of a host's foreground call slots

worktree.rm now replies only after Git deletes the checkout (up to minutes), so on paired
desktop and web each waiting delete held one of the host's 8 foreground call slots, and a
bulk delete queued listing refreshes and every other foreground call behind it. Deletes now
run in their own lane with the same bound; the 2-slot background lane stays for status polls.

* fix(worktrees): join a same-worktree delete accepted while a removal waited its repo turn

The desktop app and the runtime (CLI, paired clients, web) check for a running delete before
they queue for the repo's acceptance turn. A delete of the same worktree from the other path,
accepted while this one queued, was missed: this request re-ran the archive hook, stopped the
terminals again and started a second `git worktree remove` on the directory Git was deleting.
The queued acceptance now re-checks and joins the running delete.

* fix(worktrees): fence a resumed delete's checkout from startup, and drop rows a listing read before the delete finished

A delete a quit or crash interrupted took its terminal and file-watcher gate only when the resume
job ran, after the first window was shown; session restore could open a shell or watcher inside the
half-deleted checkout first, and on Windows that handle can fail the resumed git delete. Loading the
records now fences each recorded path, and the resumed job takes the fence over in the same tick it
takes its own gate.

A listing that read git's registration before a delete finished, and replied after the removal
record cleared, returned the row unmarked, so other views briefly showed "The delete did not
finish". Listings now capture the pending removals before reading git and leave out a row whose
delete finished successfully since; a row whose delete failed stays listed as before.

* test(worktrees): keep git's auto-maintenance out of the real-git removal suite

CI's Git 2.55 failed the file-pool test in teardown with ENOTEMPTY on the scratch repo's
objects/pack after the test body passed: the 3,000-file commit's detached auto-maintenance was
still writing a pack. The scratch repo now disables auto-maintenance and auto-gc.

* fix(worktrees): one archive-hook approval covers a same-repo bulk delete again

Local same-repo deletes now start together, so each queued its trust prompt with a state snapshot
taken before the first prompt was answered; approving the first still showed the same prompt once
per remaining worktree. The queued check now reads the store when its turn comes.

* fix(worktrees): a delete Git fails partway stays listed with its error; Delete retries it

`git worktree remove --force` drops the checkout's registration even when it
cannot delete a file (root-owned files, `chflags uchg`, a read-only Windows
directory). Orca lists workspaces from Git, so the row vanished after the error,
leaving the checkout, the branch and Orca's metadata with no way to retry.

- A background delete that fails with the checkout still on disk, unregistered,
  and still the removed checkout's own leftover keeps its durable removal record
  with the error (`failure`) instead of clearing it. Every other failure clears
  it as before.
- Local listings (desktop list/list-all/detected, runtime list/ps/detected)
  add a row for each such record, carrying `removalError`, and for a pending
  removal whose checkout Git no longer lists (shown as removing).
- Delete on that row (desktop IPC and runtime worktree.rm) runs the recorded
  removal again: terminal teardown, then the leftover, prune, branch and
  metadata, under the per-host delete limit.
- The record ends on a successful retry, when the checkout is gone (listing or
  startup)…

* fix(source-control): generate PR details before Create PR on a ready branch (#24215)

* fix(source-control): generate PR details before Create PR on a ready branch

A pushed branch sent the placeholder title and empty body, skipping the
configured agent. A run Create PR submits unreviewed keeps the user's base
and never unchecks Draft; the Generate button is unchanged.

Fixes #22824

* docs(source-control): correct the auto-submitted Draft rule comment

* fix(source-control): never let generated details uncheck the user's Draft box

A Draft choice made before generation started was unprotected: the form only
keeps fields whose revision changed *while* a run was in flight, so a user who
checked Draft and then pressed Generate had the box silently unchecked by an
agent answering draft:false, and the next Create PR opened a real review ready
for review. Scope the rule to provenance rather than the in-flight window, and
apply it to reviewed runs and the prepare-branch route too, so the agent can
still flag unfinished work but never reverts a choice the user made.

* test(source-control): cover the Checks panel create-after-run and superseded-run guards

Two guards had no failing test. The Checks panel's Create PR must still send
the finished run's details while its last render shows the run as generating,
or the click silently creates nothing. Each panel's generation must return no
result when a later run replaced its record (Stop, then Generate finishing
first), or the stopped Create PR click would open a PR with the later run's
details.

* fix(source-control): don't reveal a background-created PR over another worktree

Create PR on a ready branch now waits for the agent, and the create still
runs if the sidebar panel closes meanwhile. When it finished, it always
opened the sidebar on Checks (and, with "open after create", the review),
even if the user had switched to another worktree, so Checks showed the
wrong worktree and the in-app link route switched the user back.

Reveal only when the created review's worktree is still selected, or the
panel that made it is still mounted, like the prepare-branch route.

* fix(source-control): keep Create PR in flight while the agent writes the details

Create PR on a ready branch generated the details and then created nothing,
with no message. When generation ended, Source Control's eligibility check
restarted and cleared the eligibility, and the create, which reads the panel's
latest state, returned early on the missing eligibility.

The click now stays in flight from the start of generation until the create
settles, as the prepare-branch route does, so the eligibility check stays
paused and the create sees the eligibility it started with. Repeated clicks
are counted so one returning early can't release another's hold.

* refactor(source-control): move Checks created-review linking into its own hook

No behavior change. The Checks panel's create hook was at its 400-line limit;
linking and refreshing a created review now lives in
use-checks-panel-created-review.tsx, the way Source Control keeps it in
use-hosted-review-created.ts.

* fix(source-control): finish a started Create PR run for the branch that was clicked

After the agent wrote the details, Create PR read the panel's latest render
instead of the state at the click. That render could have lost its
eligibility (no PR, no message), still show the run as generating, or show
another branch. Navigating away also gave opposite results: with the sidebar
open on another worktree no PR was made; after closing the sidebar first, it
was.

The create now always runs through the click's own closure, so a started run
creates the PR for the branch that was clicked wherever the user went, like
the prepare-branch route. The reveal still follows the selected worktree, so
nothing opens over another one. The "still shows generating" bypass in both
panels is gone: the click's render never shows the run as generating.

In the Checks panel, a create that returns after the panel moved to another
worktree still links the review to the clicked worktree and clears its
push-first flag; only the in-flight state, errors, opening the review and
the GitLab checks refresh are skipped, since they belong to whatever the
panel now shows. The same holds for a push-first create whose push finishes
after the panel moved.

* refactor(source-control): let the Create PR click own the in-flight flag

The click already kept Source Control's create in flight through generation,
so the composer stays up instead of closing while eligibility refreshes. It
did that with a per-worktree click counter next to the create's own in-flight
flag, so a repeated click that returned early couldn't end the first click's
hold.

Now the click owns the existing in-flight flag from the click until the
create settles, and a click while it is set is refused before it touches
anything, so the counter is gone. This matches what the UI already did:
Create PR is disabled while the flag is set, including after Stop until the
stopped request returns. Since the create now runs through the click's own
state, the hold only keeps the UI steady; it no longer decides whether a PR
is created.

* docs(source-control): describe the created-review foreground check by what callers pass

The Checks panel now passes whether it still shows the created review, not
just whether it is mounted, since a mounted panel can have moved to another
worktree while the create ran.

* fix(source-control): release Create PR as soon as Stop lands

After Stop, Source Control's Create PR stayed disabled ("Creating...") until
the stopped generation request came back. Locally that is under a second, but
when the cancel can't reach a slow or disconnected remote host the button
stayed disabled for up to the 75 s generation timeout.

A generation run's outcome now settles as soon as its record stops running,
not when the request returns. Stop marks the record canceled at once, so the
click that started the run ends, its hold is released, and the next click
submits the form as shown. The stopped request still finishes in the
background; its late result is dropped because the record is no longer
running for that request. Both panels' generation handlers use the same
helper.

---------

Co-authored-by: Neil <neil@stably.ai>
Co-authored-by: Brennan Benson <79079362+brennanb2025@users.noreply.github.com>

* test(windows): record native PTY stress lifecycle milestones (#25042)

<!-- orca-pr-loc -->
<!-- Programmatic LoC summary. Do not edit by hand; rewritten on every commit. -->

| | Files | Added | Deleted | Net |
| :--- | ---: | ---: | ---: | ---: |
| Test | 3 | $\color{#1a7f37}{\Huge{\mathbf{+}}}$​562 | 0 | $\color{#1a7f37}{\Huge{\mathbf{+}}}$​562 |
| Prod | 3 | $\color{#1a7f37}{\Huge{\mathbf{+}}}$​220 | $\color{#cf222e}{\Huge{\mathbf{−}}}$​8 | $\color{#1a7f37}{\Huge{\mathbf{+}}}$​212 |

<!-- /orca-pr-loc -->

## ELI5

When a Windows terminal stress test stalls, its old log cannot show where progress stopped. The test now records bounded, sanitized lifecycle observations while keeping its existing assertions and timing.

## What Changed

Record output-pipe, worker, native exit and final callback milestones, pending state at existing deadlines, and hashes of the native addon and supporting files. Redact addresses and credentials even when terminal controls split them, preserving the controls’ positions. Keep the warmup survivor and newest 31 terminals; later terminals still receive observers, with omissions counted explicitly.

## Why

[The failing Windows job](https://github.com/stablyai/orca/actions/runs/37132029374/job/111229398978) had one silent terminal and at least one undrained callback. [An earlier passing job](https://github.com/stablyai/orca/actions/runs/37123932023/job/111205697896) used the same source and cache inputs, without a loaded-addon hash. Observing the existing objects supplies missing evidence while preserving the gate. Bounded recent records retain the terminals most likely involved in a late failure.

## Linked Issue

Diagnosis of PR #25031’s Windows package failure. This does not claim that the original native cause is fixed or close a product issue.

## Visual Proof

N/A — test diagnostics add no UI or interaction change.

## Testing

- [ ] I manually tested these changes locally
- [x] Automated tests added and updated for reproduced failures.

On the exact previous public helper, eight strengthened controls failed and six passed. They reproduce control-interrupted credential/address leaks, omitted late terminals after eleven rounds, and late state after the 256-milestone cap. Independent review then reproduced six more leaks on the intermediate helper: usernames and hostnames masked only part of an email. The final helper runs the existing same-length redactor a second time, removing the remaining private domain while preserving controls and OSC framing. Those six before failures and two credential-overlap controls are retained. Corrected source `5037c18a898571539ade5336ad3519cd49dfb20f` passes 171 controls in four files, including all 22 observer controls, the full anti-slop audit, syntax/AST checks, focused format/lint and six quality gates.

The previous public source `f1ed3e97` passed [actual Windows CI](https://github.com/stablyai/orca/actions/runs/37136703455/job/111243011997): 25 PTYs, eight rounds, 526 passing tests and 26 skips, plus build, package and smoke checks. Its native stress/driver files remain exact here; the observer has changed, so fresh normal CI must qualify this complete correction. The passing run’s native-addon hash differs from the historical failure; it does not explain that failure.

## Review

All four original diagnostic files and public ancestry are retained; all 32,245 paths outside that scope match main `786a040b`. The existing redactor runs twice on a view with presentation controls removed, then controls are restored at their original positions; OSC framing separates title payloads from adjacent text. Logs retain 32 terminal records and 256 milestones with explicit omissions, plus bounded final snapshots. The native callback preserves its receiver, arguments, result and thrown errors; error observers preserve unhandled-error behavior.

Native assertions, deadlines, input, concurrency and cleanup remain exact. Companion hashes identify file bytes rather than in-memory DLL equivalence. The original native cause remains unverifiable.

## Agent skill upstream boundary

- [x] Not applicable; no upstream skill source is copied.

## Notes

No production API, RPC, native patch, dependency or workflow change. Blank output or a missing callback is not evidence of process death. Local helper controls run on macOS; the existing real ConPTY gate exercises Windows.

## Checklist

- [x] Small, focused diagnostic scope.
- [x] Explained the before/after, mechanism and choice.
- [x] Visual proof N/A with reason.
- [x] Self-reviewed privacy, bounded logging and callback behavior.
- [x] Cross-platform and remote impact considered.
- [ ] Fresh normal Windows CI must qualify the complete corrected source.

* fix(cli): orca file open opens PDFs and other binaries like the File Explorer (#24445)

files.open gated the desktop open on the mobile preview list, so PDFs returned opened:false with ok:true. The host now opens every file after an existence/directory check; the CLI fails loudly if an older host still declines.

Fixes #24328. Builds on #24331.

Co-authored-by: Wooseong Kim <innocarpe@gmail.com>

* Update README downloads badge

* refactor(orchestration): one function for six agent-to-agent sends (#24901)

* refactor(orchestration): send every agent-to-agent message through one sendAgentTurn

The structured mail-pointer lane and the structured worker preamble each carried a copy of
"send, wait out a pending start, read the verdict", and four dispatch-preamble sites typed into
a terminal directly. They now share sendAgentTurn, built from host.send, the host's settlement
waiter and sendTerminalAgentPrompt. Every caller passes delivery 'now', so nothing sends
differently; a source-scan ratchet keeps new direct sends out.

* test(orchestration): the refusal fixture uses a real wire refusal code

* fix(orchestration): derive the send fingerprint inside sendAgentTurn and pair target with turn

A `queue` send was refused by a real host: callers supplied a fingerprint over the body alone
while the host digests body and delivery. sendAgentTurn now builds the envelope with the
composer's own builder (extracted to structured-agent-session-send-mutation), so the fingerprint
is always over exactly the fields sent; `now` sends keep the identical digest. The pointer lane
no longer carries a fingerprint it cannot get right.

sendAgentTurn takes one argument, a union on kind that carries its own target and turn, resolved
by an exhaustive switch; the terminal turn names its purpose (only the dispatch preamble today)
instead of every terminal send inheriting the task lead line. A queued outcome keeps the host's
draft receipt, and both structured callers read outcomes exhaustively with unchanged behaviour.

The boundary test now also fences direct structured host sends, catches optional, bracket and
bound member uses, has a planted-offender self-test, and pins the unmoved terminal mail pointer
and agent-teams tmux senders.

* fix(orchestration): keep federation.ts under max-lines and satisfy prefer-template in the send ratchet

* fix(orchestration): wait on the answered submission's id when a replayed queue turn was already handed off

* test(orchestration): pin that a replayed queue turn waits for its hand-off to settle

* test(orchestration): fence call-result and cast host sends, raw terminal writes and the launch-prompt helpers

* feat(native-chat): a Claude subagent waiting on a permission prompt reads as waiting (#22634)

* feat(native-chat): a Claude subagent waiting on a permission prompt reads as waiting

A subagent's permission request reaches the parent session's callback naming the
subagent that asked (agent_id) and the tool call it gates (tool_use_id). The
pending request is recorded with the asking agent. On every drain the child-work
producer re-derives which children a pending request blocks and hands that set
to the Claude child decoder, the one owner of each child's live edges: a blocked
child reads waiting on every live edge it reports, and a child that starts or
stops waiting is a live edge of its own. Answering, denying or cancelling the
request returns the child to its prior live state; nothing is stored beyond the
pending requests.

A live task_updated carrying an error now reaches the record as the child's last
message, without an ending or a new state.

The replay test drives a scrubbed capture of the real CLI (foreground allow,
deny, interrupt, background allow, and the main agent's own request) through the
real adapter into the host's child records.

* test(native-chat): a subagent's request names it before its tool call is read

* docs(agent-status): a subagent asking for approval waits in every lane; the parent row keeps the session's own attention

* test(native-chat): hand canUseTool the asking agent without widening the helper's cast

* test(native-chat): an interrupted Claude subagent settles cancelled, not failed

A captured interrupt shows the spawn call's error result ("The user doesn't want
to proceed…") arriving before the subagent's own `task_updated {status: killed}`.
The spawn result ends nothing (the child ends only on its own terminal frame), so
the child stays live until its `killed` status settles it cancelled. A genuine
failure, captured with the subagent on a model that does not exist, sends its
`failed` status before the error result and still ends failed. Both captures now
replay through the real adapter into the host's records.

* fix(native-chat): a Claude subagent's prompt makes the parent row wait, not block

A subagent's pending prompt made the whole session `attention`, which reads as
the main agent's own `blocked` and outranks the fold's waiting arm, so the
parent row read blocked where a CLI Claude parent reads waiting. The main
agent's state now reads only its own pending prompts.

- A Claude prompt row carries the linkage of the agent that raised it: the one
  the permission request names, or the owner of the tool call it gates. The
  same join decides which child reads waiting, so the two cannot disagree.
- The status summary projects the session's own status from root prompts only;
  every other reader (delivery gates, teardown, restart) still asks whether
  anyone is waiting on a human.
- An answer keeps the prompt row's linkage by the journal's own rule: a
  revision that names no producer keeps the row's existing one.
- The child-tool queries gain the prompt's producer, so a prompt row and a
  child record answer "which agent" from the same join.

* test(native-chat): say which ids the permission capture scrubs and which are its own

* fix(native-chat): the status clock dates attention by the session's own asks only

The session's status is now `attention` only for its own pending prompt, so the
clock's fallback to a subagent's ask could no longer be reached, and it read the
journal by a different rule than the status it dates. Both now read root prompts.

The journal also stamps a Codex subagent's prompt with its thread (#22532), so a
Codex child's approval is that child's wait in the Codex lane too. Two tests
written for the earlier rule are updated: a subagent's ask leaves a running
session `working` on its turn's clock, and a Codex child's answered approval
leaves the settled parent's Activity row done with nothing unread.

* fix(native-chat): a completion still says the user is asked when a subagent asks

The turn-completion feed marked a completion `awaitingUser` from the status
summary's `attention`. The status now means the session's own agent is waiting,
so a subagent's pending approval stopped reaching the completion. The projection
now also says whether anyone is waiting on the user, as the delivery gates,
teardown and restart ask it, and the completion reads that.

The waiting-subagent replay answers its prompt with the adapter's current
response shape.

* revert(native-chat): a live Claude task's error stays out of the child's last message

No capture shows a live task_updated carrying an error, and it is unrelated to
a subagent waiting on a permission request; it leaves this PR.

* test(native-chat): settle the Claude session's startup before replaying a subagent's request

A startup frame drained child work during the first await, so answering a
request freed the child even with the answer's own republish removed.

* fix(native-chat): a Claude subagent's prompt row names it as its other rows do

The prompt row stamped only the asking agent's id, so a nested subagent's
request lost the agent that spawned it, its spawn call and its run. It now takes
the linkage the asker's own rows take: the gated tool call's, when that names the
same agent, else the one resolved through the agent's spawn call. The provider's
agent id stays the asker's id.

* fix(native-chat): a subagent's request makes the parent row wait without a child record

The parent row learned that a subagent needed the user only from that
subagent's child record, so a request no record carried (a Codex child the host
never registered, a Claude task past the live cap) left the row working or done
while the approval card sat in the chat.

"Someone in this session must answer" is now one derived session fact. The
projection names two facts instead of a mode flag: the main agent's own status
(attention only for its own request) and structuredAgentSessionAwaitsUser (any
pending prompt). The status summary publishes the second as an optional
awaitsUser, and the shared fold reads it: the main agent's own ask is blocked,
otherwise awaitsUser or a waiting child record makes the row wait. Every caller
picks the fact it means: the completion edge's awaitingUser and the delivery
gate read awaitsUser; the quit snapshot folds the same two inputs the sidebar
does.

* fix(native-chat): a client that predates awaitsUser still reads a subagent's request as attention

A status summary's status is now the main agent's own, so a client built before
the split would read a subagent's request as working (or idle) and fold it with
code that has no awaitsUser input. Clients advertise
agent-session.status-awaits-user.v1; at agentSession.subscribeStatus the host
sends any client that does not the pre-split summary: attention whenever
awaitsUser is set, without the main agent's own tool line, verdict and clock.
The feed and every in-process reader keep the canonical summary. Transitional,
like the turn-item downgrade.

* test(native-chat): a Codex subagent's approval makes its settled parent's Activity row wait

The test pinned the parent row done while a Codex child asked, through a harness
that fed no child records, so it proved nothing about the ask. It now drives the
ask twice through the real host status store: with no child record (the
session's awaitsUser alone) and with the child's own record waiting from
thread/status/changed. Both read waiting with needsAttention while the ask is
open, then done with nothing unread.

* docs(agent-status): a subagent's request reaches the parent row through awaitsUser in every structured lane

The store reference said a Codex child's request still read as the main agent's
blocked and that only the Codex hook lane fed a waiting child. Both structured
lanes stamp the asking child and feed child records, and awaitsUser carries the
request when no record does. The liveness comment goes back to main's: a child's
blocked is a failed task on an older host's legacy rows.

* fix(native-chat): the restart dialog still headlines a subagent's pending approval

The quit snapshot now records the main agent's own state, so a subagent asking
while the main agent worked recorded `working` and the dialog said "Was
mid-reply" where it used to say "Waiting for your approval". The headline now
comes from the snapshot's pending prompt, whoever raised it, with the existing
copy; `state` stays the main agent's own.

* test(orchestration): a subagent's pending approval holds structured mail delivery

Scoping the delivery gate to the main agent's own request left every gate test
green; a subagent's request now has its own case.

* fix(native-chat): a subagent's request is dated by when it was raised, on every client

Since the summary's clock became the main agent's own, nothing dated a wait
that only a subagent's request held: a pre-split client was sent attention with
no clock, where the old host dated it by the subagent's prompt, and a new
client's waiting row fell back to the time it first saw it, so after a reload a
request the user had already read could read unread again.

The session fact is now when someone started being asked: awaitsUserSince, the
oldest pending prompt whoever raised it, and its presence is what awaitsUser
meant. A row waiting on someone else's request takes that as its clock; the
downgrade for a client without the capability dates its attention by it, which
is what the old host published. A cross-version test pinned to the last
pre-split release runs the same journals through that release's projection and
through this one plus the downgrade, and compares the whole summary. The Codex
end-to-end test also reads the host's own status row, and keeps a read ask read
through a later row and a reload.

* test(native-chat): the pre-split parity check compares only the fields the split owns

An additive summary field is safe for old clients, so comparing whole summaries
against the pinned release would redden on one. The wire comment now says how
the downgrade dates attention: the main agent's own oldest ask, else
awaitsUserSince.

* test(runtime): an aged host-held working summary states that nobody is asked

The test built its working summary by overriding the status of a published
approval summary, which still carried awaitsUserSince, so the row correctly
read waiting. It now drops the request as its scenario says.

* fix(native-chat): the chat's subagent block says waiting when the strip does

While a Claude subagent's request was open, the sidebar and the composer strip
read waiting but the subagent block in the chat history a few pixels above
still read "Kicked off 1 subagent working": it shows the journal's roster
state, and the journal records no wait.

The structured chat now hands its transcript the subagents the strip shows
waiting, read from the host's child records through the strip's own row model
and matched by the provider id the roster names each one by. A running entry
the host says is waiting reads waiting in the group row, its entry and its
section head, with the strip's word and the question colour; it reads the
journal's state again as soon as the host stops reporting the wait.

* fix(native-chat): a collapsed subagent group shows a wait beside a failed sibling

A failed sibling took the group row's one alert slot, so a group with a waiting,
a working and a failed child read "1 working +1 failed" and hid the wait; it
now reads "1 working +1 waiting +1 failed". The waiting set keeps its identity
while a child frame changes no wait, so the transcript's subagent rows do not
re-render on every frame, and the test of a wait ending now updates one mounted
row instead of remounting it.

* refactor(claude): one needs-input state on the parent; the asking subagent alone reads waiting

Drop the split of the main agent's own status from a session-wide "someone must
answer" fact: awaitsUserSince, the agent-session.status-awaits-user.v1
capability and its old-client downgrade, and every reader change that only
consumed them (fold, equality, ingest, delivery gate, turn-completion feed,
quit snapshot, resume headline, status clock, status bridge, attention
dispatch) go back to main. The parent row again reads one needs-input state
for a pending request whoever asked, dated as before.

Kept: a request's owner recorded once on its prompt row with full producer
linkage; the asking subagent's own record reads waiting, re-derived on every
update; the chat history's subagent block reads that same state; an answered
subagent request stays in its subagent's group.

A subagent now waits only on a request the user can still answer (its card
open, no answer underway), and the adapter frees it before the host records an
answer or dismissal. So a waiting child record always sits beside the pending
card, and main's fold never reads the parent as waiting on it: no window after
an answer, and no ~3 s wait after a card dismissed by Stop.

* fix(claude): a subagent waits only beside its committed card

A subagent's wait was pushed to the host as soon as its request arrived,
while the request's card row reached the journal at least a microtask later.
So every subagent request published the parent row as waiting before
blocked (the main agent's own fold reads a waiting child that way), and
Activity got an extra unread "waiting" event that main never shows.

The card is now the one record of an open request. The translator records
the asker on the card once (its row's linkage) and counts the card open only
after the sink confirms its rows landed, then publishes the wait; anything
that closes the card (an answer underway, a dismissal handed to the host,
Claude's own withdrawal, the session's end) frees the subagent first. So
every publish that shows a subagent waiting also shows its pending card, and
the parent reads one needs-input state, exactly as on main.

This retires the registry's view of pending requests (unclaimed(), the
asking-child join) and the translator's holdsOpen. The prompt row's linkage
takes one rule: the agent the provider names, else the gated call's owner.

The parent-row proof now runs through the real deferred sink, durable
journal and status feed, publishing as production does, and checks at every
publish that waiting subagents have pending cards and that the parent row
matches a host fed no waits.

* fix(native-chat): a closed sink's dropped writes never read as landed

The sink's written() resolved ok when the sink was closed with writes still
queued, so a subagent's prompt card could count as open with no row in the
journal. written() now reports a close that dropped writes admitted so far
as not landed; drained() and lifecycleBarrier() keep reading a closed sink as
settled.

Tests: a card never opens when its sink closes first; a card Claude withdraws
while the sink holds the cancelled row back closes at once; two subagents
asking at once, and the main agent asking beside a subagent, keep the parent
row as before with each waiting subagent beside its own card; a process that
dies mid-request leaves no subagent waiting.

* fix(claude): a withdrawn subagent request frees its child before its card closes

Main's sink now hands each write to the journal as it is submitted, and an idle journal commits it
and runs the publication at once. Claude's own withdrawal of a subagent's request therefore closed
the card and published the parent row before the child's wait was freed, so one publish showed the
subagent waiting beside no pending card (fg-interrupt replay). The child's wait now also requires the
request to still be open in the registry, and a withdrawal republishes child work before the
journal takes the close.

* refactor(claude): trim subagent request waiting to the common pattern and its essential tests

The chat history no longer marks a subagent block as waiting: the approval card itself carries the
request, and the asking subagent's row in the sidebar and composer strip reads waiting, as before.
NativeChatWaitingSubagentsProvider, native-chat-waiting-subagents.ts and their renderer changes go.

A subagent waits while its request is still open and unanswered in the prompt registry and its card
has landed in the journal. The registry check also covers a withdrawal under backpressure, so the
card list no longer filters pending cancellations itself.

Tests: one integration file replays the captured CLI frames through the real adapter, sink, journal
and status feed (renamed claude-subagent-permission-request.test.ts), with the asking subagent's
state timeline, attribution, nested linkage and a card write that waits for the journal. The
producer-harness waiting test, the redundant prompt-card cases, the harness reducer swap and three
unused captures (deny, interrupt, main agent, failed subagent) are removed.

* test(claude): pin the parent row's dating when a subagent asked first, and narrow the oracle's claim

* Turn desktop notifications on or off per machine (#24518)

* feat(notifications): turn desktop notifications on or off per machine

Settings > Notifications lists each machine (this computer, SSH targets,
paired Orca servers) with a switch. Muted machines are stored as an
opt-out list so a newly added machine still notifies. Both notification
senders now name the machine a workspace runs on, and main skips the
desktop banner for a muted one; phone push is unchanged.

* fix(notifications): preserve phone alerts and honor machine mutes

* fix(notifications): bind machine mutes to configured sources

* fix(notifications): preserve chat completion subscriptions on owner collisions

* fix(notifications): reduce machine settings clutter with a collapsed section

* fix(notifications): align machine disclosure with settings rows

* fix(notifications): clarify machine switches affect only this computer

* Replace agentArgsOverride with unified removeAgentArgs approach (#25091)

Consolidate override detection and removal into removeAgentArgs. Previously
catalogs used agentArgsOverride to detect conflicts and removeAgentArgs to
strip them; now removeAgentArgs handles both by returning stripped tokens.
This eliminates redundancy and makes the intent clearer.

Enhance removeAgentArgOption with optional value filtering to support
selective removal for complex cases like Codex config overrides.

* Reuse the ancestor path while building mobile agent rows (#24539)

Preserve traversal order and cycle guards using one call-local path Set rather than a copy at every depth.

* Release waiting terminal output when a mobile subscription fails to start (#24547)

Dispose the failed subscription record’s existing terminal backlog before rethrowing its original start error.

* Avoid cloning terminal agent owners twice in relay listings (#24713)

Capture the existing host-age expression at its original time, then use one call-local fresh owner clone for the length check and unchanged published owner field.

* Reuse prepared statements for internal worker checks (#24573)

* Reuse prepared statements across remaining Dispatch reads

Use the existing schema-pinned complete Dispatch column list in five remaining read modules, enabling the existing bounded statement cache without caching results.

* Reuse prepared statements for internal worker checks

Use the existing complete Dispatch projection only for named-field internal consumers; preserve original wildcard reads for every returned failure snapshot.

* Prepare the Activity search query once per filtered list (#24701)

Lazily reuse a call-local search predicate inside the existing filter, using unchanged byte gates and thread text cache.

* Avoid rescanning shared chunks in the plain Node build guard (#24717)

Reuse a successful unchanged chunk-code verdict within one synchronous writeBundle invocation, while preserving traversal and current-code reads.

* Select the latest stable release tag in one pass (#24786)

Replace filter/sort/last selection with one traversal using the unchanged stable-tag regex and numeric comparator; update on equality to preserve the original last spelling. Both actual callers own private plain arrays from Git stdout.

* Avoid Resource Manager renders when terminal removal leaves its inventory unchanged (#24806)

Reuse the private React state object only when the existing single/bulk removal helper returns its identical inventory; keep every lifecycle revision, tombstone, known-ID write and helper invocation in its original order.

* Cancel pending cursor updates when a terminal closes (#24566)

Reuse the existing pane frame tracker to cancel owned deferred focus-class updates during existing cleanup, with disposed guards against late or reentrant delivery.

* Cancel terminal preview fit frames when the preview closes (#24712)

Reuse the existing frame tracker inside the box-fit owner, suppress scheduling after disposal and cancel pending frames at the start of the existing effect cleanup.

* Skip new mobile toast work after feedback owner cleanup (#24759)

Reuse the existing mobile mounted-ref lifecycle pattern at toast presentation entry, preserving admitted clipboard outcomes and every live animation/sequence/timer operation.

* Release terminal side effects after they have been delivered (#24548)

Clear consumed queue slots after successful apply or overflow carry; preserve backing identity across reentrant callbacks and report actual retained slots.

* Release relay handshake timers when a host connection leaves (#24554)

Reuse the existing first-frame finish pattern to remove stage-owned timer/message/close callbacks on receipt, timeout or close; check OPEN after successful async assignment verification.

* Avoid restarting error timers on closed mobile relay links (#24567)

Check the existing irreversible closed flag before scheduling the existing missing-close fallback timer.

* Avoid restarting error timers after mobile relay pairing closes (#24568)

Check the existing pairing owner closed flag before allocating its missing-close fallback alarm.

* Delete unreturned clipboard cache files after a failed write (#24599)

Reuse existing provider-copy best-effort deletion for a newly created clipboard cache file whose write fails before its URI reaches the caller.

* Skip new legacy file inventories after mobile search cleanup (#24792)

Reuse the existing mobile mounted-ref pattern only at legacy fallback entry after completed passive cleanup; preserve admitted work and all live search/authority/cache paths. Correct only the strict fully-unmounted inventory scenario and its sole golden.

* Discard completed AI Vault cancellation IDs in the relay worker (#24820)

Use the relay child's existing pending Set to ignore cancellation IDs after completion, matching its desktop sibling; preserve admitted active/queued cancellation and the bundled private sender's complete replies, errors and ownership.

* Release completed updater setup timers and callbacks (#24920)

Cancel completed deferred updater fallbacks in finally, clear only their exact pending callback, and drop the captured timer before the original fallback guard runs; preserve destroyed admission, successor ownership and updater/quit callbacks.

* Check daemon idle state without building discarded terminal inventories (#24749)

* Check daemon idle state without building discarded terminal inventories

Replace the private idle predicate full public inventory with a host-local full pass that still reads every Session liveness getter in original order.

* Remove overwritten daemon test mock assignment

* fix(updater): keep macOS Orca open when background instances block updates (#24952)

* fix(updater): guard macOS installs against running app instances

* fix(updater): match native app blockers and preserve quit lifecycle

* fix(updater): keep ordinary macOS quit on Squirrel's install-on-exit path

Converting every quit with a staged update into quitAndInstall made Cmd+Q
relaunch Orca, refused the quit when background instances existed, and
hijacked app.relaunch()+app.quit() restart flows (profile switch, admin
restart) into an update install racing the relaunched old app. Only
Update & Restart runs the running-instance preflight now; the
quit-without-install allowance is no longer reachable and is removed.

* fix(updater): preserve quit intent through macOS staging

* test(native-chat): explicitly model legacy published tab ownership

---------

Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Co-authored-by: m4air <m4air@m4airs-Air.localdomain>

* Cancel review animations and timers when the Markdown preview closes (#24644)

Track review frames with the existing frame owner, remember all active pulse timers, and retire only the detached preview generation.

* docs: update Android APK links to 0.0.52 (#25107)

* Preserve saved account credentials after enrollment errors (#25059)

* fix: preserve registered account credentials after enrollment errors

* refactor(orchestration): one function for six agent-to-agent sends (#24901)

* refactor(orchestration): send every agent-to-agent message through one sendAgentTurn

The structured mail-pointer lane and the structured worker preamble each carried a copy of
"send, wait out a pending start, read the verdict", and four dispatch-preamble sites typed into
a terminal directly. They now share sendAgentTurn, built from host.send, the host's settlement
waiter and sendTerminalAgentPrompt. Every caller passes delivery 'now', so nothing sends
differently; a source-scan ratchet keeps new direct sends out.

* test(orchestration): the refusal fixture uses a real wire refusal code

* fix(orchestration): derive the send fingerprint inside sendAgentTurn and pair target with turn

A `queue` send was refused by a real host: callers supplied a fingerprint over the body alone
while the host digests body and delivery. sendAgentTurn now builds the envelope with the
composer's own builder (extracted to structured-agent-session-send-mutation), so the fingerprint
is always over exactly the fields sent; `now` sends keep the identical digest. The pointer lane
no longer carries a fingerprint it cannot get right.

sendAgentTurn takes one argument, a union on kind that carries its own target and turn, resolved
by an exhaustive switch; the terminal turn names its purpose (only the dispatch preamble today)
instead of every terminal send inheriting the task lead line. A queued outcome keeps the host's
draft receipt, and both structured callers read outcomes exhaustively with unchanged behaviour.

The boundary test now also fences direct structured host sends, catches optional, bracket and
bound member uses, has a planted-offender self-test, and pins the unmoved terminal mail pointer
and agent-teams tmux senders.

* fix(orchestration): keep federation.ts under max-lines and satisfy prefer-template in the send ratchet

* fix(orchestration): wait on the answered submission's id when a replayed queue turn was already handed off

* test(orchestration): pin that a replayed queue turn waits for its hand-off to settle

* test(orchestration): fence call-result and cast host sends, raw terminal writes and the launch-prompt helpers

* feat(native-chat): a Claude subagent waiting on a permission prompt reads as waiting (#22634)

* feat(native-chat): a Claude subagent waiting on a permission prompt reads as waiting

A subagent's permission request reaches the parent session's callback naming the
subagent that asked (agent_id) and the tool call it gates (tool_use_id). The
pending request is recorded with the asking agent. On every drain the child-work
producer re-derives which children a pending request blocks and hands that set
to the Claude child decoder, the one owner of each child's live edges: a blocked
child reads waiting on every live edge it reports, and a child that starts or
stops waiting is a live edge of its own. Answering, denying or cancelling the
request returns the child to its prior live state; nothing is stored beyond the
pending requests.

A live task_updated carrying an error now reaches the record as the child's last
message, without an ending or a new state.

The replay test drives a scrubbed capture of the real CLI (foreground allow,
deny, interrupt, background allow, and the main agent's own request) through the
real adapter into the host's child records.

* test(native-chat): a subagent's request names it before its tool call is read

* docs(agent-status): a subagent asking for approval waits in every lane; the parent row keeps the session's own attention

* test(native-chat): hand canUseTool the asking agent without widening the helper's cast

* test(native-chat): an interrupted Claude subagent settles cancelled, not failed

A captured interrupt shows the spawn call's error result ("The user doesn't want
to proceed…") arriving before the subagent's own `task_updated {status: killed}`.
The spawn result ends nothing (the child ends only on its own terminal frame), so
the child stays live until its `killed` status settles it cancelled. A genuine
failure, captured with the subagent on a model that does not exist, sends its
`failed` status before the error result and still ends failed. Both captures now
replay through the real adapter into the host's records.

* fix(native-chat): a Claude subagent's prompt makes the parent row wait, not block

A subagent's pending prompt made the whole session `attention`, which reads as
the main agent's own `blocked` and outranks the fold's waiting arm, so the
parent row read blocked where a CLI Claude parent reads waiting. The main
agent's state now reads only its own pending prompts.

- A Claude prompt row carries the linkage of the agent that raised it: the one
  the permission request names, or the owner of the tool call it gates. The
  same join decides which child reads waiting, so the two cannot disagree.
- The status summary projects the session's own status from root prompts only;
  every other reader (delivery gates, teardown, restart) still asks whether
  anyone is waiting on a human.
- An answer keeps the prompt row's linkage by the journal's own rule: a
  revision that names no producer keeps the row's existing one.
- The child-tool queries gain the prompt's producer, so a prompt row and a
  child record answer "which agent" from the same join.

* test(native-chat): say which ids the permission capture scrubs and which are its own

* fix(native-chat): the status clock dates attention by the session's own asks only

The session's status is now `attention` only for its own pending prompt, so the
clock's fallback to a subagent's ask could no longer be reached, and it read the
journal by a different rule than the status it dates. Both now read root prompts.

The journal also stamps a Codex subagent's prompt with its thread (#22532), so a
Codex child's approval is that child's wait in the Codex lane too. Two tests
written for the earlier rule are updated: a subagent's ask leaves a running
session `working` on its turn's clock, and a Codex child's answered approval
leaves the settled parent's Activity row done with nothing unread.

* fix(native-chat): a completion still says the user is asked when a subagent asks

The turn-completion feed marked a completion `awaitingUser` from the status
summary's `attention`. The status now means the session's own agent is waiting,
so a subagent's pending approval stopped reaching the completion. The projection
now also says whether anyone is waiting on the user, as the delivery gates,
teardown and restart ask it, and the completion reads that.

The waiting-subagent replay answers its prompt with the adapter's current
response shape.

* revert(native-chat): a live Claude task's error stays out of the child's last message

No capture shows a live task_updated carrying an error, and it is unrelated to
a subagent waiting on a permission request; it leaves this PR.

* test(native-chat): settle the Claude session's startup before replaying a subagent's request

A startup frame drained child work during the first await, so answering a
request freed the child even with the answer's own republish removed.

* fix(native-chat): a Claude subagent's prompt row names it as its other rows do

The prompt row stamped only the asking agent's id, so a nested subagent's
request lost the agent that spawned it, its spawn call and its run. It now takes
the linkage the asker's own rows take: the gated tool call's, when that names the
same agent, else the one resolved through the agent's spawn call. The provider's
agent id stays the asker's id.

* fix(native-chat): a subagent's request makes the parent row wait without a child record

The parent row learned that a subagent needed the user only from that
subagent's child record, so a request no record carried (a Codex child the host
never registered, a Claude task past the live cap) left the row working or done
while the approval card sat in the chat.

"Someone in this session must answer" is now one derived session fact. The
projection names two facts instead of a mode flag: the main agent's own status
(attention only for its own request) and structuredAgentSessionAwaitsUser (any
pending prompt). The status summary publishes the second as an optional
awaitsUser, and the shared fold reads it: the main agent's own ask is blocked,
otherwise awaitsUser or a waiting child record makes the row wait. Every caller
picks the fact it means: the completion edge's awaitingUser and the delivery
gate read awaitsUser; the quit snapshot folds the same two inputs the sidebar
does.

* fix(native-chat): a client that predates awaitsUser still reads a subagent's request as attention

A status summary's status is now the main agent's own, so a client built before
the split would read a subagent's request as working (or idle) and fold it with
code that has no awaitsUser input. Clients advertise
agent-session.status-awaits-user.v1; at agentSession.subscribeStatus the host
sends any client that does not the pre-split summary: attention whenever
awaitsUser is set, without the main agent's own tool line, verdict and clock.
The feed and every in-process reader keep the canonical summary. Transitional,
like the turn-item downgrade.

* test(native-chat): a Codex subagent's approval makes its settled parent's Activity row wait

The test pinned the parent row done while a Codex child asked, through a harness
that fed no child records, so it proved nothing about the ask. It now drives the
ask twice through the real host status store: with no child record (the
session's awaitsUser alone) and with the child's own record waiting from
thread/status/changed. Both read waiting with needsAttention while the ask is
open, then done with nothing unread.

* docs(agent-status): a subagent's request reaches the parent row through awaitsUser in every structured lane

The store reference said a Codex child's request still read as the main agent's
blocked and that only the Codex hook lane fed a waiting child. Both structured
lanes stamp the asking child and feed child records, and awaitsUser carries the
request when no record does. The liveness comment goes back to main's: a child's
blocked is a failed task on an older host's legacy rows.

* fix(native-chat): the restart dialog still headlines a subagent's pending approval

The quit snapshot now records the main agent's own state, so a subagent asking
while the main agent worked recorded `working` and the dialog said "Was
mid-reply" where it used to say "Waiting for your approval". The headline now
comes from the snapshot's pending prompt, whoever raised it, with the existing
copy; `state` stays the main agent's own.

* test(orchestration): a subagent's pending approval holds structured mail delivery

Scoping the delivery gate to the main agent's own request left every gate test
green; a subagent's request now has its own case.

* fix(native-chat): a subagent's request is dated by when it was raised, on every client

Since the summary's clock became the main agent's own, nothing dated a wait
that only a subagent's request held: a pre-split client was sent attention with
no clock, where the old host dated it by the subagent's prompt, and a new
client's waiting row fell back to the time it first saw it, so after a reload a
request the user had already read could read unread again.

The session fact is now when someone started being asked: awaitsUserSince, the
oldest pending prompt whoever raised it, and its presence is what awaitsUser
meant. A row waiting on someone else's request takes that as its clock; the
downgrade for a client without the capability dates its attention by it, which
is what the old host published. A cross-version test pinned to the last
pre-split release runs the same journals through that release's projection and
through this one plus the downgrade, and compares the whole summary. The Codex
end-to-end test also reads the host's own status row, and keeps a read ask read
through a later row and a reload.

* test(native-chat): the pre-split parity check compares only the fields the split owns

An additive summary field is safe for old clients, so comparing whole summaries
against the pinned release would redden on one. The wire comment now says how
the downgrade dates attention: the main agent's own oldest ask, else
awaitsUserSince.

* test(runtime): an aged host-held working summary states that nobody is asked

The test built its working summary by overriding the status of a published
approval summary, which still carried awaitsUserSince, so the row correctly
read waiting. It now drops the request as its scenario says.

* fix(native-chat): the chat's subagent block says waiting when the strip does

While a Claude subagent's request was open, the sidebar and the composer strip
read waiting but the subagent block in the chat history a few pixels above
still read "Kicked off 1 subagent working": it shows the journal's roster
state, and the journal records no wait.

The structured chat now hands its transcript the subagents the strip shows
waiting, read from the host's child records through the strip's own row model
and matched by the provider id the roster names each one by. A running entry
the host says is waiting reads waiting in the group row, its entry and its
section head, with the strip's word and the question colour; it reads the
journal's state again as soon as the host stops reporting the wait.

* fix(native-chat): a collapsed subagent group shows a wait beside a failed sibling

A failed sibling took the group row's one alert slot, so a group with a waiting,
a working and a failed child read "1 working +1 failed" and hid the wait; it
now reads "1 working +1 waiting +1 failed". The waiting set keeps its identity
while a child frame changes no wait, so the transcript's subagent rows do not
re-render on every frame, and the test of a wait ending now updates one mounted
row instead of remounting it.

* refactor(claude): one needs-input state on the parent; the asking subagent alone reads waiting

Drop the split of the main agent's own status from a session-wide "someone must
answer" fact: awaitsUserSince, the agent-session.status-awaits-user.v1
capability and its old-client downgrade, and every reader change that only
consumed them (fold, equality, ingest, delivery gate, turn-completion feed,
quit snapshot, resume headline, status clock, status bridge, attention
dispatch) go back to main. The parent row again reads one needs-input state
for a pending request whoever asked, dated as before.

Kept: a request's owner recorded once on its prompt row with full producer
linkage; the asking subagent's own record reads waiting, re-derived on every
update; the chat history's subagent block reads that same state; an answered
subagent request stays in its subagent's group.

A subagent now waits only on a request the user can still answer (its card
open, no answer underway), and the adapter frees it before the host records an
answer or dismissal. So a waiting child record always sits beside the pending
card, and main's fold never reads the parent as waiting on it: no window after
an answer, and no ~3 s wait after a card dismissed by Stop.

* fix(claude): a subagent waits only beside its committed card

A subagent's wait was pushed to the host as soon as its request arrived,
while the request's card row reached the journal at least a microtask later.
So every subagent request published the parent row as waiting before
blocked (the main agent's own fold reads a waiting child that way), and
Activity got an extra unread "waiting" event that main never shows.

The card is now the one record of an open request. The translator records
the asker on the card once (its row's linkage) and counts the card open only
after the sink confirms its rows landed, then publishes the wait; anything
that closes the card (an answer underway, a dismissal handed to the host,
Claude's own withdrawal, the session's end) frees the subagent first. So
every publish that shows a subagent waiting also shows its pending card, and
the parent reads one needs-input state, exactly as on main.

This retires the registry's view of pending requests (unclaimed(), the
asking-child join) and the translator's holdsOpen. The prompt row's linkage
takes one rule: the agent the provider names, else the gated call's owner.

The parent-row proof now runs through the real deferred sink, durable
journal and status feed, publishing as production does, and checks at every
publish that waiting subagents have pending cards and that the parent row
matches a host fed no waits.

* fix(native-chat): a closed sink's dropped writes never read as landed

The sink's written() resolved ok when the sink was closed with writes still
queued, so a subagent's prompt card could count as open with no row in the
journal. written() now reports a close that dropped writes admitted so far
as not landed; drained() and lifecycleBarrier() keep reading a closed sink as
settled.

Tests: a card never opens when its sink closes first; a card Claude withdraws
while the sink holds the cancelled row back closes at once; two subagents
asking at once, and the main agent asking beside a subagent, keep the parent
row as before with each waiting subagent beside its own card; a process that
dies mid-request leaves no subagent waiting.

* fix(claude): a withdrawn subagent request frees its child before its card closes

Main's sink now hands each write to the journal as it is submitted, and an idle journal commits it
and runs the publication at once. Claude's own withdrawal of a subagent's request therefore closed
the card and published the parent row before the child's wait was freed, so one publish showed the
subagent waiting beside no pending card (fg-interrupt replay). The child's wait now also requires the
request to still be open in the registry, and a withdrawal republishes child work before the
journal takes the close.

* refactor(claude): trim subagent request waiting to the common pattern and its essential tests

The chat history no longer marks a subagent block as waiting: the approval card itself carries the
request, and the asking subagent's row in the sidebar and composer strip reads waiting, as before.
NativeChatWaitingSubagentsProvider, native-chat-waiting-subagents.ts and their renderer changes go.

A subagent waits while its request is still open and unanswered in the prompt registry and its card
has landed in the journal. The registry check also covers a withdrawal under backpressure, so the
card list no longer filters pending cancellations itself.

Tests: one integration file replays the captured CLI frames through the real adapter, sink, journal
and status feed (renamed claude-subagent-permission-request.test.ts), with the asking subagent's
state timeline, attribution, nested linkage and a card write that waits for the journal. The
producer-harness waiting test, the redundant prompt-card cases, the harness reducer swap and three
unused captures (deny, interrupt, main agent, failed subagent) are removed.

* test(claude): pin the parent row's dating when a subagent asked first, and narrow the oracle's claim

* Turn desktop notifications on or off per machine (#24518)

* feat(notifications): turn desktop notifications on or off per machine

Settings > Notifications lists each machine (this computer, SSH targets,
paired Orca servers) with a switch. Muted machines are stored as an
opt-out list so a newly added machine still notifies. Both notification
senders now name the machine a workspace runs on, and main skips the
desktop banner for a muted one; phone push is unchanged.

* fix(notifications): preserve phone alerts and honor machine mutes

* fix(notifications): bind machine mutes to configured sources

* fix(notifications): preserve chat completion subscriptions on owner collisions

* fix(notifications): reduce machine settings clutter with a collapsed section

* fix(notifications): align machine disclosure with settings rows

* fix(notifications): clarify machine switches affect only this computer

* fix: distinguish unavailable account metadata from absence

* fix(accounts): resolve registered profile UUIDs consistently

* Replace agentArgsOverride with unified removeAgentArgs approach (#25091)

Consolidate override detection and removal into removeAgentArgs. Previously
catalogs used agentArgsOverride to detect conflicts and removeAgentArgs to
strip them; now removeAgentArgs handles both by returning stripped tokens.
This eliminates redundancy and makes the intent clearer.

Enhance removeAgentArgOption with optional value filtering to support
selective removal for complex cases like Codex config overrides.

* Reuse the ancestor path while building mobile agent rows (#24539)

Preserve traversal order and cycle guards using one call-local path Set rather than a copy at every depth.

* Release waiting terminal output when a mobile subscription fails to start (#24547)

Dispose the failed subscription record’s existing terminal backlog before rethrowing its original start error.

* Avoid cloning terminal agent owners twice in relay listings (#24713)

Capture the existing host-age expression at its original time, then use one call-local fresh owner clone for the length check and unchanged published owner field.

* Reuse prepared statements for internal worker checks (#24573)

* Reuse prepared statements across remaining Dispatch reads

Use the existing schema-pinned complete Dispatch column list in five remaining read modules, enabling the existing bounded statement cache without caching results.

* Reuse prepared statements for internal worker checks

Use the existing complete Dispatch projection only for named-field internal consumers; preserve original wildcard reads for every returned failure snapshot.

* Prepare the Activity search query once per filtered list (#24701)

Lazily reuse a call-local search predicate inside the existing filter, using unchanged byte gates and thread text cache.

* Avoid rescanning shared chunks in the plain Node build guard (#24717)

Reuse a successful unchanged chunk-code verdict within one synchronous writeBundle invocation, while preserving traversal and current-code reads.

* Select the latest stable release tag in one pass (#24786)

Replace filter/sort/last selection with one traversal using the unchanged stable-tag regex and numeric comparator; update on equality to preserve the original last spelling. Both actual callers own private plain arrays from Git stdout.

* Avoid Resource Manager renders when terminal removal leaves its inventory unchanged (#24806)

Reuse the private React state object only when the existing single/bulk removal helper returns its identical inventory; keep every lifecycle revision, tombstone, known-ID write and helper invocation in its original order.

* Cancel pending cursor updates when a terminal closes (#24566)

Reuse the existing pane frame tracker to cancel owned deferred focus-class updates during existing cleanup, with disposed guards against late or reentrant delivery.

* Cancel terminal preview fit frames when the preview closes (#24712)

Reuse the existing frame tracker inside the box-fit owner, suppress scheduling after disposal and cancel pending frames at the start of the existing effect cleanup.

* Skip new mobile toast work after feedback owner cleanup (#24759)

Reuse the existing mobile mounted-ref lifecycle pattern at toast presentation entry, preserving admitted clipboard outcomes and every live animation/sequence/timer operation.

* Release terminal side effects after they have been delivered (#24548)

Clear consumed queue slots after successful apply or overflow carry; preserve backing identity across reentrant callbacks and report actual retained slots.

* Release relay handshake timers when a host connection leaves (#24554)

Reuse the existing first-frame finish pattern to remove stage-owned timer/message/close callbacks on receipt, timeout or close; check OPEN after successful async assignment verification.

* Avoid restarting error timers on closed mobile relay links (#24567)

Check the existing irreversible closed flag before scheduling the existing missing-close fallback timer.

* Avoid restarting error timers after mobile relay pairing closes (#24568)

Check the existing pairing owner closed flag before allocating its missing-close fallback alarm.

* Delete unreturned clipboard cache files after a failed write (#24599)

Reuse existing provider-copy best-effort deletion for a newly created clipboard cache file whose write fails before its URI reaches the caller.

* Skip new legacy file inventories after mobile search cleanup (#24792)

Reuse the existing mobile mounted-ref pattern only at legacy fallback entry after completed passive cleanup; preserve admitted work and all live search/authority/cache paths. Correct only the strict fully-unmounted inventory scenario and its sole golden.

* Discard completed AI Vault cancellation IDs in the relay worker (#24820)

Use the relay child's existing pending Set to ignore cancellation IDs after completion, matching its desktop sibling; preserve admitted active/queued cancellation and the bundled private sender's complete replies, errors and ownership.

* Release completed updater setup timers and callbacks (#24920)

Cancel completed deferred updater fallbacks in finally, clear only their exact pending callback, and drop the captured timer before the original fallback guard runs; preserve destroyed admission, successor ownership and updater/quit callbacks.

* Check daemon idle state without building discarded terminal inventories (#24749)

* Check daemon idle state without building discarded terminal inventories

Replace the private idle predicate full public inventory with a host-local full pass that still reads every Session liveness getter in original order.

* Remove overwritten daemon test mock assignment

* Preserve the registered account ID when selecting UUID variants

* fix(updater): keep macOS Orca open when background instances block updates (#24952)

* fix(updater): guard macOS installs against running app instances

* fix(updater): match native app blockers and preserve quit lifecycle

* fix(updater): keep ordinary macOS quit on Squirrel's install-on-exit path

Converting every quit with a staged update into quitAndInstall made Cmd+Q
relaunch Orca, refused the quit when background instances existed, and
hijacked app.relaunch()+app.quit() restart flows (profile switch, admin
restart) into an update install racing the relaunched old app. Only
Update & Restart runs the running-instance preflight now; the
quit-without-install allowance is no longer reachable and is removed.

* fix(updater): preserve quit intent through macOS staging

* test(native-chat): explicitly model legacy published tab ownership

---------

Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Co-authored-by: m4air <m4air@m4airs-Air.localdomain>

---------

Co-authored-by: Brennan Benson <79079362+brennanb2025@users.noreply.github.com>
Co-authored-by: Jinjing <6427696+AmethystLiang@users.noreply.github.com>
Co-authored-by: OrcaWin <alpha-eng@stably.ai>
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Co-authored-by: m4air <m4air@m4airs-Air.localdomain>

* Fix admission of later Claude question rows and real CLI verification (#25119)

* Attempt each Claude question journal row after admission refusal

* test: read Claude auth JSON after configuration warnings

* test: retain the explicitly opted-in Claude profile without disabling write guards

* test: use the probed Claude profile in real handshake sessions

* Keep the real Claude signed-out control free of environment credentials

---------

Co-authored-by: Orca QA <orca-qa@users.noreply.github.com>

* fix: remove managed profiles through canonical UUID paths (#25126)

---------

Signed-off-by: Neil <neil@stably.ai>
Co-authored-by: Justas Brazauskas <brazauskasjustas@gmail.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Brynn Bendixen <brynnbendixen@gmail.com>
Co-authored-by: brynnclaw <261708852+brynnclaw@users.noreply.github.com>
Co-authored-by: 小七 <ggbdpq@gmail.com>
Co-authored-by: Orca Integration Recovery <orca-validation@invalid.example>
Co-authored-by: lurunzi <lurunzi@gmail.com>
Co-authored-by: Dongho Lee <126236161+LDH1103@users.noreply.github.com>
Co-authored-by: LDH1103 <ldh517525@gmail.com>
Co-authored-by: Wayn_Liu <115852642+Waynting@users.noreply.github.com>
Co-authored-by: Wayn_Liu <wayntingliu@gmail.com>
Co-authored-by: VXNCXNX <VXNCXNX@users.noreply.github.com>
Co-authored-by: Michiel de Gooijer <mdgooijer@gmail.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Marcus <70784566+marcuslannister@users.noreply.github.com>
Co-authored-by: marcuslannister <marcus@lannister.cc>
Co-authored-by: OrcaWin <alpha-eng@stably.ai>
Co-authored-by: m4air <m4air@m4airs-Air.localdomain>
Co-authored-by: Florian Schwarzmeier <1506919+fsmeier@users.noreply.github.com>
Co-authored-by: SongMarco <snubflow@gmail.com>
Co-authored-by: marco song <marco.song@mvlchain.io>
Co-authored-by: SongMarco <20613630+SongMarco@users.noreply.github.com>
Co-authored-by: Luca Critelli <lucacri@gmail.com>
Co-authored-by: JeongUk Park <jeongph.dev@gmail.com>
Co-authored-by: B1nh M1nh <43268322+b1nhm1nh@users.noreply.github.com>
Co-authored-by: Zhichang Yu <yuzhichang@gmail.com>
Co-authored-by: Kh05ifr4nD <meandSSH0219@gmail.com>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Wooseong Kim <2222333+innocarpe@users.noreply.github.com>
Co-authored-by: Wooseong Kim <innocarpe@gmail.com>
Co-authored-by: Nawapat Buakoet <nawapat.b@covest.finance>
Co-authored-by: Frederic Barthelemy <git@fbartho.com>
Co-authored-by: mrcha033 <mrcha033@users.noreply.github.com>
Co-authored-by: xodmd45-ctrl <xodmd45-ctrl@users.noreply.github.com>
Co-authored-by: Luchong <lu740528977@gmail.com>
Co-authored-by: czzczz <chanzrz_zbf@foxmail.com>
Co-authored-by: Brennan Benson <79079362+brennanb2025@users.noreply.github.com>
Co-authored-by: razshlomo <razshlomo@users.noreply.github.com>
Co-authored-by: Raz Shlomo <12373339+razshlomo@users.noreply.github.com>
Co-authored-by: Ahmed Nagy <ahmednagy25t@gmail.com>
Co-authored-by: KAPUIST <thsxornjs12@gmail.com>
Co-authored-by: JianJia2018 <39438074+JianJia2018@users.noreply.github.com>
Co-authored-by: kespineira <kespineira@users.noreply.github.com>
Co-authored-by: kevimux <kevimux@users.noreply.github.com>
Co-authored-by: Jinjing <6427696+AmethystLiang@users.noreply.github.com>
Co-authored-by: Kelvin Amoaba <97001695+AmoabaKelvin@users.noreply.github.com>
Co-authored-by: Pablo Werlang <19828711+werlang@users.noreply.github.com>
Co-authored-by: innocarpe <innocarpe@users.noreply.github.com>
Co-authored-by: Jinwoo Hong <73622457+Jinwoo-H@users.noreply.github.com>
Co-authored-by: drakeo338 <paranoyouz@gmail.com>
Co-authored-by: Saurav M Hiremath <sauravhiremath@gmail.com>
Co-authored-by: DakaAlvarez <149860458+Dacadev97@users.noreply.github.com>
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Co-authored-by: Orca Worker <orca-worker@localhost>
Co-authored-by: m4air <m4air@Mac.localdomain>
Co-authored-by: griffinmartin <griffinmartin@users.noreply.github.com>
Co-authored-by: Orca QA <orca-qa@users.noreply.github.com>

2026-10-03 20:32:06 -07:00

..

cli

fix(cli): orca file open opens PDFs and other binaries like the File Explorer (#24445)

2026-10-03 11:57:24 -07:00

main

fix(opencode): finish status installation with invalid TUI settings (#25031)

2026-10-03 20:32:06 -07:00

preload

fix(git): preserve SSH review context and worktree ownership (#24945)

2026-10-03 05:24:35 -07:00

relay

Discard completed AI Vault cancellation IDs in the relay worker (#24820)

2026-10-03 15:47:20 -07:00

renderer

Recognize Build terminals and preserve Windows confirmation key routing

2026-10-03 20:23:33 -07:00

shared

fix(opencode): finish status installation with invalid TUI settings (#25031)

2026-10-03 20:32:06 -07:00

types

…