fix(media): preview host files and stream videos across clients - #9023
Conversation
Thread transfer impact✅ Thread transfer remains within every enforced ceiling.
Baseline: Scenario and decoded snapshot size10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.
Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed. |
ApprovabilityVerdict: Not approved Macroscope's review found this PR not approvable — This PR adds a substantial cross-platform media-preview capability and changes host-file authorization, signed asset serving, HTTP streaming, and default file-preview behavior. Its broad runtime and sensitive file-access scope, together with new static-analysis suppressions, warrants human review. You can add or adjust custom eligibility rules. Learn more. |
Reuse first-frame preparation on web and the native thumbnail cache on mobile. Keep streaming playback and use a shared filmstrip icon for video references. Portal media dialogs, preserve failed remote destinations, and refresh idle file previews without interrupting active playback.
|
@macroscope Please rerun the UI Consistency check on The PR now includes fresh before/after screenshots and Chromium/Android recordings of workspace chips opening the file viewer, with the raw Markdown source beside the rendered evidence. CI tests and typechecks pass; real Chromium playback uses the signed HTTP URL and returns 206 for byte ranges. |
|
Manual reviews triggered for commit All prior checks · these links stay valid even if you push more commits. |
|
Just FYI for future @mentions, I'm Rerun triggered. Results will be posted as check runs. |
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): keep queued message editing inside the queue panel * fix(web): keep queued messages in place while editing * fix(web): match composer actions to draft and modifier state * fix(web): keep composer shortcut tooltip stable on Mod * feat(web): summarize T3 orchestration actions * feat(mobile): port chat summaries and transitions to orchestration v2 Adapt grouped tool summaries and the floating working timer to V2 run, attempt, and queue state. Bring over the composer, keyboard, and disclosure transitions while retaining the V2 activity inspector and queue controls. Keep OV2 web composer and grouping behavior intact; share only the existing command label parser with mobile. * fix(chat): remove added tool summary status counts * fix(mobile): keep scroll bounds current after animations * fix: reconcile main's round-17 features after the rebase Restores main features dropped by the policy replay: #8569 theme wiring, settings search rework, #8803 workspace-mutation refresh (v2-adapted), video + image previews (web and mobile, v2-adapted), #8862 Expo glass, and the round's docs. Timeline thinking rows (#8984) stay on the v2 work-live system. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(web): port working and thinking timeline rows to orchestration v2 The v2 equivalents of main's #8984 and #8922: a "Working for ..." header anchors the active run, the trailing live tool row survives between actions in past tense instead of vanishing, and a shimmering Thinking row marks reasoning gaps. During workspace preparation the header shows "Setting up worktree..." (driven by the local dispatch flag or the v2 run's preparing status, so remote viewers see it too), the composer footer span is gone, and draft promotion waits until the run starts or startup fails instead of navigating mid-preparation. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: reconcile main's round-18 features after the rebase Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(server): evaluate automatic thread settlement in the v2 orchestrator Ports #8600's server-owned settlement to orchestration v2 instead of keeping client-side evaluation. A ThreadSettlementService sweep runs at startup, on auto-settle settings changes, and once per minute: it evaluates inactivity and merged or closed pull requests over v2 thread shells and dispatches the new guarded thread.auto-settle command, which rejects threads that changed after the sweep's snapshot or carry any explicit override, then reuses the orchestrator's settle lifecycle. With the server deciding, the clients drop their effectiveSettled evaluation and partition on the persisted settledOverride like main. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: reconcile main's round-19 features after the rebase Adopts main's round-19 features into the v2 stack: payload-budgeted orchestration replay (#8992), sidebar row subscription leases (#9052), tool group virtualization and scroll anchoring (#9106), repeated-command and browser-group presentation, inline assistant citations (#9146), per-cwd provider skills discovery (#8778), Claude composer skill dispatch (#9128), grok health probe and model negotiation (#9154), and the failed-tool thinking fallback (#9165). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): right-align the stash shoulder tab again Round 17 adopted main's #8850 ComposerBanner.Attachment (mx-auto plus the standalone drawer-inset width) without main's matching mounts, so the stash tab's ml-auto lost to the attachment's auto right margin and the tab centered over the composer. Column now spans its attachments like main does, the stash tab zeroes the right margin, and the stash menu keeps the full dock width. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): realign the composer and timeline with main The branch had drifted from main's composer and work-log design in ways unrelated to orchestration v2: a pre-revert "Working for" shoulder tab on the composer (main reverted #8693 and re-landed #8734 without it), an inline stash variant plus in-flow stash menu, expanded tool rows that hid their icons, an unmounted woke-thread banner, a composer scroll observer main never had, and a right-panel toggle that lost its showRightPanelControl gate so it rendered twice with the panel open. ChatComposer and its satellites now start from main's files with only the v2 delta re-applied (dispatch modes, queued-message editing, runtime request ids, response capability). Background tasks surface as a ChatView banner in main's backgroundLiveness shape instead of a composer tab. SimpleWorkEntryRow takes main's PlainWorkEntryRow body with the V2ItemInspector kept behind the expander. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: reconcile main's round-20 features after the rebase Renumbers the v2 migrations 044-052 to 045-053 behind main's 044_ClearAutomaticProjectModelDefaults, and adopts main's sticky new-thread selection (#9164), local-only worktree bases in the v2 launch path (#8751), the PR summary read for settlement (#9176), Claude per-cwd skills (#9210), the provider editor redesign with the branch's dedicated environment fields re-grafted (#8508), and the client half of continue-threads-across-restart (#9167). The server-side continuation markers stay unported: they live in the v1 session directory, and v2 recovery terminalizes running runs on restart, so the capability is withheld until the v2 runtime carries them. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(server): reduce v2 recovery and runtime resource usage Index event sequence lookups, recover only threads with pending work, and page projection verification and rebuilds. Bound provider event logging and omit turn histories when resuming Codex threads. Allow delegated thread identifiers through relay routes. Add focused regression coverage and document the performance constraints. * fix: reconcile main updates with orchestration v2 Retain main's composer, work-log, settings, mobile and performance changes through c8f77e0d441 while preserving v2 runs, queued messages, provider handoffs and durable history. Port native compaction and logout, asynchronous Codex questions, provider usage accounting, automatic settlement and PR refresh into the v2 services. Bound live event retention during replay and delivery, measure thread replay before decoding, and read checkpoint metadata without loading transcripts or patches. Keep main migrations through 047 and move the v2 migrations to 048–058. Preserve the existing branch history and the pre-rebase backup. Model: GPT-6. Harness: Codex. * fix(orchestration): stabilize Codex turn mapping and settlement - Preserve Codex turn identity while suppressing duplicate diff notifications - Optimize settlement projections and isolate thread visit handling - Add concurrency and regression coverage across server and mobile * fix(chat): match main timer and task placement Restore the completed work timer divider and text size from main. Keep todo-list progress in the composer and omit it from web and mobile timeline entries, including completed task lists. Verified pending, running, and completed task projection; 187 focused web tests and 35 mobile tests pass. Web and mobile typechecks pass. * fix(mobile): restore composer and timeline behavior from main Show Send when a running thread has draft content. Separate submission follow from first-message anchoring so later sends do not reserve extra blank space. Restore Android initial composer insets and iOS focus-aware dictation insets. Keep opening and final assistant replies visible around completed folds, anchoring Worked for at the first hidden item while preserving v2 relationship cards and execution-attempt behavior. Validation: 107 focused tests and the mobile typecheck pass. Formatting passes; scoped lint and React Doctor report warnings but no errors. No simulator run. * fix(orchestrator): Stop treating a wait timeout as a dead child (#7427) * fix(orchestrator): Show when a completed delegated child still has work (#4793) Co-authored-by: Julius Marminge <julius0216@outlook.com> * fix(orchestrator): Stop finished Codex turns from sitting on Waiting (#7105) Co-authored-by: Julius Marminge <julius0216@outlook.com> * chore: format v2 files flagged by CI * test(server): include the shell indexes migration * fix(mobile): pin the patched notifications dependency * fix(client-runtime): resolve work log source imports * test(server): refresh replay runtime instruction expectations * test(server): correlate OpenCode replay message identities * ci: run checks on v2 branch pushes * test(server): use Effect Vitest for Cursor provider checks * fix: reconcile main updates with orchestration v2 Retain main's changes while preserving v2 orchestration, queue/steer controls, composer-only tasks, timeline timers, and mobile scrolling fixes. Port opt-in restart continuation through durable v2 effects, with shutdown race guards, activation gating, retry deduplication, and native Codex resume. Use narrow projection reads for control effects and runtime-request replies. Surface Claude fallback notices without failing the turn or hiding the notice. Report missing workspace folders before provider startup. Carry over custom models and prices, bounded client caches and stream cleanup, lazy image loading, persistent changed-file trees and sidebar filters, Safari cookie import, theme fixes, POSIX file-link case, private-host favicon filtering, native provider update paths, and platform portability updates. Migration ids remain unchanged. Validated scoped typechecks and focused server, web, mobile, client-runtime, contracts, desktop, shared, SSH, script, and resource-monitor tests. Preserved all 347 original commits and checked the final tree against both saved tips. Model: GPT-6. Harness: Codex. * fix(server): explain fetch failures during worktree preparation Worktree preparation previously exposed only a generic fetch failure. Classify known authentication, network, repository access, and reference-lock errors using stable Git diagnostics, without retaining raw output or credentials. Unknown failures keep the existing generic message. Cover failure classification and redaction, a real missing local remote, and propagation into a failed prepared run without creating a worktree or running setup. The launch test waits for the persisted failure event. Validation: 38 focused tests, server typecheck, and scoped lint passed. * fix: reconcile upstream fixes with orchestration v2 Carry main's session refresh, provider maintenance, runtime diagnostics, composer focus, preview, usage, and mobile outbox fixes into the v2 branch. Keep queue/steer submission, composer-only task progress, v2 subagent cards, and LegendList scroll ownership. Project thread and shell events before transport buffering while retaining full durable history. Dismiss native questions when provider turns finish, with a transaction guard that preserves answers submitted concurrently. Port Claude limit notices and Codex file approval details to v2 adapters. Validated with focused server, web, mobile, client-runtime, shared, desktop, and marketing tests; affected package typechecks and scoped lint pass. All 349 branch commits retain their authors and messages. Migration files and the previous worktree-fetch, stash, panel, and mobile inset fixes remain unchanged. * fix(server): make project removal honor v2 threads Offline CLI and HTTP project removal dropped force and left native v2 threads behind. Move the nonempty-project guard and durable child cleanup into the shared project service, and forward force from CLI, HTTP, and WebSocket calls. Reuse the thread deletion planner and command lock, hydrate migrated history before attachment cleanup, and validate child receipts. Commit the project deletion after its children so failed cleanup can be retried safely. Validation covers CLI deletion with active and archived threads, missing workspaces, durable cleanup, partial retries, migrated attachments, receipt collisions, and concurrent thread updates. Scoped server tests, typecheck, and lint pass. Implemented with Codex (GPT-6). * fix(mobile): render generic message attachments (#9929) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: shivam <91240327+shivamhwp@users.noreply.github.com> * fix(mobile): use the archive eligibility guard when dispatching (#9930) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: shivam <91240327+shivamhwp@users.noreply.github.com> * fix(orchestration): persist linked pull requests (#8689) * feat(mcp): update thread metadata (#8690) * fix(server): keep old failures from waking snoozed V2 threads (#9903) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * refactor(shared): share model-selection command choice (#10577) * refactor(project): share create and update inputs (#10578) * refactor(server): share attachment message intake (#10580) * feat(mcp): expose thread organization commands (#10554) * feat(mcp): expose existing queued message commands (#10555) * feat(mcp): expose pending user questions (#10556) * feat(mcp): expose thread model selection (#10557) * feat(mcp): expose fork and merge-back commands (#10558) * feat(mcp): expose preview list and close (#10559) * feat(mcp): expose selected environment preferences (#10560) * feat(mcp): expose the existing thread search query (#10561) * feat(mcp): expose scheduled task run-now (#10562) * feat(mcp): expose project service operations (#10563) * feat(mcp): expose attachment upload and send (#10564) * feat(mcp): expose project thread launch service (#10565) * feat(mcp): expose branch-backed workspace discovery (#10566) * fix(orchestration): map late steering to follow-up turns - Re-route steering that races completion into idempotent follow-up dispatches - Preserve scheduled-task attribution and provider ownership history across clients * fix: reconcile main's round-24 features after the rebase Port main's pull-request discovery, active thread ordering, async question dismissal, settlement fixes, provider-session import, attachment context, and provider correctness changes into orchestration v2. Keep the branch's intentional composer and subagent behavior while adopting main's web and mobile fixes. Prevent headless setup terminals from hanging on the color probe, and move the v2 migration block to 050-061 after main claimed 048-049. * chore: remove accidentally committed audit artifacts * fix(ci): repair rebased checks and stop duplicate runs Restore the failed-before-start timer guard, align two server fixtures with the reconciled behavior, and remove dead files, exports, and dependencies surfaced by Knip. Drop the temporary branch push trigger now that the PR is mergeable, so each update runs the pull-request workflow once. * fix(web): port auto-balance updates to v2 chat Keep main's batch machine-update banner and update action while preserving the v2 runtime-based environment lock used by draft load balancing. * chore: format files exposed by CI * fix: reconcile main's round-26 updates after rebase Adopt TypeScript 7 and Effect rc.112 across orchestration v2, including the TaggedError API migration and updated Effect-aware tests. Restore main's composer-aware scroll-to-end clearance while retaining selected-model settings sync, preview recording transfer, image galleries, desktop context menus, and layout hit targets. Regenerate the lockfile on the upgraded dependency baseline. * fix(web): restore compact load-earlier control * perf(orchestration): bound v2 transport payloads Advertise bounded socket snapshots and authoritative dispatch validation, omit raw command output and inline file bodies at the wire boundary, and preserve compact status metadata across web and mobile. Add transport-budget coverage for snapshots, resume, commands, legacy import, and projection maintenance. * fix(web): preserve tool failures after output redaction * fix: restore sidebar behavior after v2 rebases Restore pinned-thread shelf classification, server-owned unread state, hidden-subagent-safe project ordering, guarded jump hints, draft upload cleanup, and active-provider archive guards across the current and legacy sidebars. Bring the surrounding current-main sidebar work forward as well: canonical project favicons, stable row layout, thread file drops, account-aware mobile provider badges, and deferred desktop keyring loading. * fix(server): consolidate V2 migrations and refine runtime recovery * fix(web): simplify timeline rows and preserve collapsed composer controls * fix(web): smooth composer transitions and group approval worklogs Keep collapsed model controls in a strip, contain transition overflow, and preserve timeline spacing. Render approval requests as regular grouped worklog entries. Implemented with GPT-6-Astra via Codex. * fix: reconcile main updates with orchestration v2 Adapt question attachments and Android push verification to V2 requests and shell events. Preserve composer transitions and compact worklogs while integrating upstream loading, navigation, and mobile changes. Release consumed application replay pages without retaining earlier batches. * fix(server): report OpenCode descendant stop failures * fix(server): abort external OpenCode sessions on release * fix(server): retain thread baseline diffs across root runs * fix(server): fail OpenCode turns on unexpected stream EOF * fix(server): bound OpenCode runtime request replies * fix(client): bound socket resets after cold HTTP failures * fix(server): query only due scheduled tasks during polling * fix(server): retain normalized OpenCode turn usage Accumulate owned step usage once and preserve partial or unavailable telemetry for failed, interrupted, or reconnected turns. * perf(server): scope ordinary control reads to their targets * fix(server): retry initial title generation after transient failures * fix(server): isolate Cursor metadata generation from workspaces * fix(server): preserve Claude Read image previews across clients * fix(web): preserve generic files when editing queued messages * test(server): assert tool output redaction before storage fidelity * test(web): cancel queued animation frames during worker cleanup (#10880) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * fix: reconcile main devices and pull requests with orchestration v2 * feat(providers): add Pi coding agent (#7211) Co-authored-by: Claude Fable 5 <noreply@anthropic.com> Co-authored-by: Mike Olson <mwolson@member.fsf.org> Co-authored-by: Julius Marminge <julius0216@outlook.com> * feat(providers): standardize ACP providers (#6461) Co-authored-by: Julius Marminge <julius0216@outlook.com> * feat: render background completions as typed notifications * fix(mcp): omit recursive screenshot metadata from tool inputs * fix(pi): use native forks and preserve rollback session identity * fix(pi): cap OpenRouter output budgets pending upstream fix * fix(web): show ACP sidebar icons and hold onboarding height while loading * feat(server): deliver delegated completions through a durable mailbox * fix(acp): support Devin terminals, questions, and native subagents * fix(server): find active turns when answering async questions * fix(web): populate sidebar ACP branding from environment settings * fix(acp): preserve native child messages and final summaries * fix(server): distinguish delegated task results from completed turns * test(server): align Codex delegation instruction assertion * test(server): align delegation fixtures with task result semantics * fix(server): classify Claude V2 structured terminal failures (#9897) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * fix(orchestration): exclude rolled-back work from bounded recovery (#8464) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Julius Marminge <julius0216@outlook.com> * test(orchestration): cover bounded V2 socket fallback paging (#9907) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * fix(server): wait for native Codex start before Stop (#10024) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Julius Marminge <julius0216@outlook.com> * fix(server): keep the native Grok default model (#10025) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Julius Marminge <julius0216@outlook.com> * fix(server): fail V2 turns when the OpenCode event stream ends (#9905) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * fix(web): preserve file attachments when editing queued runs (#9928) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * fix(server): preserve project mutation fields across transports (#9920) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * fix(mobile): throttle streaming thread visit updates (#9931) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * fix(orchestrator): preserve task-step elapsed time across restart (#10051) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Julius Marminge <julius0216@outlook.com> * fix(server): preserve Claude interruption status during steering * fix(server): wait for nested completion delivery before publishing results * test(server): verify background delivery with real providers * fix: keep working timers anchored to the active run * perf: page complete turns and bound timeline reconciliation * perf(client): narrow thread subscriptions and navigation updates * feat(mobile): manage queued messages in a dedicated sheet * fix(web): fold completed trailing background activity * fix: reconcile main settings and previews with orchestration v2 * perf(client): reconcile replay batching with orchestration v2 * fix: reconcile main Codex model selection and UI updates * fix: reconcile main context previews and rewind updates * fix(build): include protobuf and Connect license notices * fix(mobile): pin expo-audio so the release smoke patch stays in use (#11518) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix: reconcile main release and thread updates with V2 * fix: repair v2 CI after environment disable and dead exports ConnectionCatalogEntry gained a required enabled flag in #11478, but the threadShell harness never set it, so enabled-gated atoms filtered every environment out and two tests failed. Two dead exports also tripped knip. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(web): update notification tests for v2 thread shells and drop dead composer state ThreadNotificationCoordinator presents raw OrchestrationV2ThreadShell records, but its tests still fed the pre-v2 thread shape (session / latestTurn), which crashed presentThreadShell on missing DateTime fields. Rebuild the fixtures as v2 shells with pendingRuntimeRequest and run statuses, and remove the composerHasUnsentContent binding left unused by the compaction gating change. * fix(client-runtime): avoid Array#toSorted in thread lineage ordering * fix(server): reject partial output from failed Cursor runs (#11534) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(server): wake paused Cursor replay runs on mismatch (#11535) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(server): load V2 replay fixtures through the platform path service (#11566) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(server): report missing interrupt-and-restart capability for forced restarts (#11565) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(server): preserve Cursor directory and lint search results (#11533) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(server): roll back question attachment copies when respond preparation fails (#11557) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(server): restart the live session on model changes after dead records (#11505) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(server): replay launches with server-allocated thread IDs (#11508) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * fix(web): keep the active-run header with the prompt that started it on steer (#11828) * refactor(web): centralize provider instance icons (#11829) * feat(web): add the thread action menu and inline rename to the chat header (#11830) * fix(web): keep the preview mini-player clear of the inline thread details card (#11831) * fix: reconcile main snooze controls with orchestration v2 * fix(web): adapt registry icons to light and dark themes * fix: reconcile main worktree setup and title changes with v2 * fix(server): isolate V2 migrations from the V1 database * fix(ci): verify V2 branch pushes and remove unused helper * revert: restore existing CI push triggers * fix(web): retain server-side queuing on v2 after rebase * fix(test): account for optional encoded provider settings * fix(build): parse executable imports without matching generated source * fix(build): isolate executable parser from Vite config * fix(server): project legacy thread shells during import * fix(server): replay command events across persistence pages (#11499) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(orchestrator): report terminal runs after wait timeout (#11574) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(server): skip corrupt scheduled-task rows instead of stopping the scheduler (#11585) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(server): preserve due schedules across equivalent time formatting (#11590) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> * test(server): replay denied Claude writes through V2 (#11597) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(server): fork Codex threads at the native turn boundary (#11490) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(server): align MCP delegation support with live provider adapter registry (#11578) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * test(server): prove v1 to v2 cutover on a copied database and flag divergent migration ids (#11639) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(server): bound legacy thread projection requests (#10512) * fix: reconcile main thread updates and Git improvements with v2 * fix(mobile): restore permission registry concurrency protection * fix(web): confine composer glass transition to input surface - Move transition glass styling off the host wrapper - Add backdrop saturation to the main composer surface * fix(server): restore hub limits updates in V2 (#11963) Co-authored-by: Julius Marminge <jmarminge@gmail.com> * fix(mobile): reject preview builds from v1 source * fix(mobile): skip preview validation without release credentials * fix: stop retained background work after a turn settles * refactor: remove legacy token streaming * fix: format subagent task names across clients * fix: distinguish unsupported server connections * fix: reconcile main updates with V2 orchestration * fix(server): preserve PR links across V2 discovery and import * fix: reconcile main monograms and PR refresh with V2 * fix(web): reset thread scroll and ignore hydration as a new turn * fix(ci): pin patched Expo core during release resolution * fix(web): invert follow-up behavior with Mod+Enter * fix(web): show linked pull requests in thread details * fix(web): restore main thread-switch scrolling without layout resets * fix: reconcile main setup transitions with V2 threads * fix(mobile): keep cached thread list across relaunch The shared shell cache codec never overrode activityRunStartedAt and unsettledAt with DateTimeUtcFromString, so any snapshot holding a working or unsettled thread encoded fine but failed to decode on the next cold launch. The store discarded the whole row and the Home list stayed empty until the environment reconnected. Add the two overrides and extend the mobile cache round-trip test with a running, unsettled thread so the codec and the JSON overrides stay in sync. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * perf(client): coalesce persistent cache writes during streaming (#12109) * perf(server): suppress unchanged shell enrichment refreshes (#12110) * perf(mobile): skip unchanged thread row renders (#12116) * perf(mobile): yield to UI during shell cache encoding (#12117) * perf(client): narrow mobile and web environment subscriptions (#12126) * perf(mobile): ignore irrelevant config updates in thread lists (#12127) * perf(mobile): limit thread model options to its provider (#12128) * perf(client): stop scanning threads for unused shell timestamps (#12129) * feat(mobile): make the composer pill the hub for the running turn The pill above the composer only tracked queued messages, and a follow-up sent during a turn always queued because mobile hardcoded its dispatch mode. Steering meant sending the message and then promoting it from the queue sheet, and the turn's subagents were only reachable as transcript rows. The pill now carries an agents segment alongside the queue count, scoped to the running turn and hidden once it settles. Tapping either segment opens a sheet: agents lists the turn's subagents and opens a child thread, and the queue sheet is rebuilt on the native header with compact rows, swipe to remove, a context menu, and full editing that saves through queued-run.edit while keeping the message's place in line. Follow-ups become a choice. A Follow-ups settings screen picks queue or steer, the send button says which one it will do, and long-pressing it uses the other for a single message. On a hardware keyboard the Command chord does the same, so the composer text view now reports whether the submit was the alternate and names both chords for the iPad shortcut HUD. Steering travels as "auto" so a turn that ends mid-flight degrades to a queued run instead of bouncing the message back into the draft, and the button only offers Steer when the provider can actually steer. Web's dispatch resolver moves into client-runtime so both clients share it. The lineage banner is gone from the transcript, taking mobile's disconnect action with it; merge back to source now lives in the thread header's git menu. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: reconcile main composer and provider updates with V2 * test(server): restore Cursor usage coverage after V2 rebase * feat(web): add compact PR checks to the workspace card (#11981) * feat(web): show subagent details and history in workspace card (#12079) * fix(server): start V2 provider turn when checkpoint baseline capture fails (#12153) Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(server): retain durable checkpoint index fixes on v2 * fix(server): finalize v2 runs when checkpoint ref lookup fails * fix(server): reject v2 file restore in shared workspaces * fix(mobile): allow changing provider in a started thread (#12184) Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: reconcile main updates with V2 runtimes and timelines * fix: align V2 question and checkpoint timelines across clients * fix(mobile): guard question controls during answer submission * fix: adapt multi-model thread creation to V2 launches * fix: restore V2 worktree setup transitions across clients * fix(mobile): match web provider handoff dividers * feat(mobile): rebuild the Circe orb and add Circe Mesh sign-in The mobile app opened straight into Circe on a fresh install, so a user never saw the step that connects them through Circe Mesh. The voice orb was also the wrong object: a dark blob with fat translucent bands behind it, plus a soft square around its glow on some Android GPUs. Orb - One Skia canvas with six ordered layers: atmospheric glow, rear fibers, the sphere surface, the hull ring, fibers refracted inside the sphere, front fibers, then grain. Three fiber planes are what produce depth; nothing here is a real 3D render. - The body is an SkSL runtime effect that reconstructs a surface normal per pixel, so the sphere lights like an object instead of a flat radial gradient. The ring is modulated by angle, because a uniformly bright ring reads as neon rather than as light. - Fibers are silk filaments, not an audio waveform: 18 rear, 20 refracted interior, 5 front, each with deterministic per-strand variation and a gaussian envelope centred on the sphere. - Motion is split so the sphere feels heavy: it barely moves and the field carries the animation. One frame callback drives the scene, and audio level reaches the renderer as a shared value without re-rendering React. Sign-in - The signed-out gate waited forever on Clerk's isLoaded, so a device that could not reach Clerk skipped Welcome entirely and landed in the app. It now resolves to signed-out after three seconds, and a real stored session still resolves from the token cache without the network. - Adds the Welcome screen and its Clerk auth step, including the Circe Mesh onboarding request that already existed behind it. Theme - The orb follows the app theme. Light and dark share geometry and differ only in luminosity: dark leans on the rim and pulls the bloom back. Dev tooling - The orb gallery was gated on process.env.APP_VARIANT, which Expo never inlines into the bundle, so the route never registered. Now gated on __DEV__. It exposes every state, both appearances, three sizes, and three levels. Both soft-edge traps found here are silent: `opacity` on a large radial-filled shape and BlurMask each make Skia allocate a layer, which renders as a soft-edged square on some Android GPUs. Every soft edge in the orb is a gradient with its alpha baked in, so no layer is allocated. * fix(mobile): render the orb as a dark lens, not a lit copper sphere The previous orb was technically competent and visually wrong. It modelled a conventional lit solid sphere: the shader reconstructed a surface normal, applied directional light from the upper-left, and started the body gradient at a bright cream `hotColor`. That is a polished orange ball, which is what it rendered. The palette file in the same commit already said the middle of the sphere must read as near-black, so the code contradicted its own design. Replaced the rendering model rather than re-tuning colors. Dark base and a separate transparent shell - `OrbSurface` is split into `OrbBase` (opaque, `core`/`coreWarm`/`ember` only, no directional term) and `OrbShell` (transparent hull light). `hot` is now only ever a thin lip or a subsurface accent, never a fill. Interior fibers moved inside the glass - The refracted fiber plane now renders between the base and the shell. It previously rendered after an opaque sphere, so the strands could only look printed onto a surface. The ring is the shell, not a stroke - Removed the uniform 360-degree `Path` circle that painted over the shader's angular variation and cancelled it out. The shell pass carries an uneven profile built from three angular harmonics plus a travelling phase, and the only hard edge is a roughly one-pixel lip at the hull. Refraction instead of compression - Interior strands were squeezed by constant `x *= 0.91 / y *= 0.82`, which reads as a narrowed bundle. They now derive a lens depth from the horizontal position and use it both to pull the strand toward the optical axis and to shift its phase, so the fiber visibly bends as it enters the sphere. Fewer, quieter strands - 43 strands down to 20 (11 rear, 7 interior, 2 front), average alpha roughly halved, and one hero strand per plane instead of every fifth strand being equally prominent. The field should be perceived, not counted. Motion is time-based and far slower - The field advanced a fixed increment per rendered frame, so a 120 Hz device drifted twice as fast as a 60 Hz one and a full cycle took about half a second. It now advances by elapsed time; `fieldCycleSeconds` is 11 s at rest and 5 s while listening. The sphere itself only breathes. One state model instead of a dozen unused knobs - `waveAmp`, `massIntensity`, `glowResponse` and `strandAmplitudeScale` were varied by state and never read by the renderer. Replaced with six parameters that are all consumed, and a single microphone `energy` value that scales field amplitude, shell brightness, bloom, and core warmth together. Scene bounds - The welcome canvas was `size + 2 * size * 0.95`, about 487 dp tall for a 168 dp sphere, which opened a large gap between the copy, the orb, and the auth controls. The vertical padding is now independent of the fiber field at roughly 26% of the sphere. Grain is off at rest so the idle frame stays clean. Also syncs the stable web icon filenames from `assets/circe`, which `scripts/lib/circe-boot-assets.test.ts` asserts byte-for-byte and which the regenerated assets had left stale. Verified on a physical Android device in both appearances and in the orb gallery across states and sizes. * feat(mobile): adopt Circe design system v1 tokens and shell lighting Mobile was still carrying the pre-v1 palette: a cool blue-gray dark surface (`#16181b`), a cool `#0f1620` Circe canvas, an off-brand amber primary (`#96600a` / `#c99a2e`), and a serif stack that led with Times New Roman. The design system asks for warm layered near-black, warm ivory paper, restrained copper as the single brand accent, and an editorial serif. Tokens (global.css, mobileTheme.ts, regenerated uniwind themes) - Dark surfaces are now the layered warm near-black set: canvas `#0c0d0e`, surface `#121415`, raised `#191b1d`, hover `#1e2022`. Pure black is out. - Light surfaces are warm ivory: canvas `#fcf9f4`, surface `#fffdfa`, raised `#f6f0e9`, hover `#f2ebe4`. - Borders move to low-opacity warm rules: `rgba(56,43,35,.07/.13/.20)` on light, `rgba(255,255,255,.065/.10/.16)` on dark, replacing opaque beige borders. - Copper becomes the primary action token: `#a5482c` on light for legibility, `#e08a63` on dark. The brand accent is identical in both appearances. - Circe tokens gain the full v1 set: copper ramp, peach, semantic success, warning, danger and neutral, plus surface, surface-raised, and copy. Status colors are now semantic only rather than decorative. - The display serif drops Times New Roman, which the design system rules out, for a stack led by Iowan Old Style. Bundling Instrument Serif needs a native rebuild and is deliberately left as a separate change. Orb shell (§13) - The palette moves onto the v1 ramp: `#100e0d` core through `#6d3526` deep copper, `#e18a62` copper, `#ffd8bd` peach, `#fff4e9` hot lip. - The shell's angular profile is now three art-directed light lobes instead of a sum of harmonics: the strong warm regions sit upper-left and lower-left, and a narrow brilliant flare sits on the right edge. Each lobe drifts slowly. - The copper band starts around 81% of the radius, matching the design system's gradient stops, so the falloff is broad rather than a hairline. - Bloom becomes two passes matching the specified glow: a broad atmosphere that spills past the hull and a narrow warm glow hugging the shell. Welcome screen - Adopts the light onboarding treatment: warm ivory paper, near-black editorial ink, one burnt-copper phrase, a near-black primary CTA whose only brand cue is a restrained copper hairline, and low-opacity warm borders. Tests - `uses the Circe graphite palette as the default` asserted the old hexes and is replaced with the v1 invariants: light paper is warm (red leads blue), dark is a layered near-black that is neither pure black nor a colored slate, and copper is the same accent in both appearances. - The hard-coded variable count in the palette-role test is replaced with the presence of every Circe token, which is what the code actually depends on. * fix(mobile): give the orb volume and rebuild the field as one ribbon The previous pass over-corrected. Adding a dark base and a separate shell did fix the order, but nothing was left between them, so the sphere rendered as a near-uniform black disc under a hairline of light. Two causes, both structural. There was no volume layer - `OrbBase` stays in `#100e0d`-`#1a100c` and `OrbShell` only lights the hull, so the region in between had no light at all. - Adds `OrbVolume`, a transparent pass between the interior ribbon and the shell. It carries broad low-frequency copper across the outer 40-50% of the sphere plus two asymmetric lobes, a lower-body glow and a left-side light. Its alpha is capped at 0.42 and it never reaches white, so it reads as smoked glass rather than a second opaque sphere. - The shell's `pow(1 - z, 2.4)` falloff was the other half of the problem: it is near zero until the final pixels. Replaced with two explicit art-directed fields starting around 46% of the radius. This is brand artwork, not a physically correct rim term. The field was twenty independent sine waves - Every strand had its own frequency, amplitude, phase, offset and speed, which mathematically wants to become spaghetti however few strands remain. - `WaveField` is replaced by `RibbonField`: one shared centerline carrying a broad S-curve, with eight filaments as small offsets from it, so the group reads as a single piece of silk. Five faint atmosphere fibers keep their own trajectories at alpha 0.04-0.10. - Both centerline harmonics carry integer phase coefficients, so the curve returns to its exact starting shape after a phase revolution and the keyframe interpolation stays seamless. Refraction is now visible - The interior plane delays the shared centerline's phase by lens depth, grows its amplitude inside the glass, and pinches the bundle by up to 42% toward the optical axis at the centre. Interior filaments are roughly 1.5x more visible than before, so you can see the strands enter the object. - The front plane carries only the two highlighted filaments rather than a second full field. State wiring - `fieldAmplitude` was defined and tested but never read by the renderer, so tuning it did nothing. It now scales the path amplitude, and changing state rebuilds the interpolated frames. - Bloom is documented as microphone-responsive but `OrbGlow` was never passed the level. It now receives `energySV` and its three gradient fields genuinely respond. `alphaColor` is a worklet so the stops are built on the UI thread. - Adds a regression that fails if any `OrbStateParams` key has no consumer in a production renderer file. That is the class of bug this commit is fixing. Glow is now three separate fields rather than one: a broad peach atmosphere at about 1.55R that visibly lights the page around Circe, a medium warm bloom, and a localized shell aura. * feat(mobile): build a dedicated welcome hero illustration The welcome screen was a standard auth page with the product orb dropped into it. The orb was shared with the home and voice screens, so every attempt to make it a brand hero traded off against its job as a state indicator: it came out either too dark to be a focal point, or too luminous to read as "idle". The real problem was the abstraction, not the shader. Separates the two visual systems - `CirceOrb` stays the product orb: home, voice, listening, thinking, speaking, compact, interactive, stateful. - `CirceWelcomeHero` is a new, decorative brand illustration used only on the welcome and auth screens. It has no states, no audio input, and no interactivity, so it is free to be bright. - Both remain in the same canvas so the illustration is one composition rather than several widgets stacked in a column. Rebuilds the page composition - The hero now sits above the headline. It is full-bleed, cancelling the screen padding, so it reads as artwork rather than an inset widget. - Order is logo, hero, headline, subcopy, CTAs, divider, benefits, legal. The hero is one wide canvas, 300dp tall, with this layer order - atmosphere, so the page picks up warmth around the object - halo arcs, four flattened ellipses at very low alpha - rear ribbon fan - orb core, then the interior ribbon clipped and refracted through it - front filaments crossing over - dust motes The orb is luminous now, not a dark ball - A dedicated shader climbs warm brown, copper, then peach, holding the deep core to about 30% of the visible area rather than most of it. The previous product-orb treatment was near-black across the whole body, which is correct for a state indicator and wrong for a focal point. - The rim is modulated by three angular harmonics plus a travel phase, so it is never uniformly bright. A value hash adds faint grain so the volume is not a mathematically smooth disc. The ribbon is one flow field, not independent sine waves - A single master spline crosses the hero. Every filament is an offset from that curve, so the strands stay related and read as one piece of silk. - The bundle is tight where it passes the orb and opens toward the edges, which produces the left and right fans from a single construction. - Inside the glass the shared curve is phase-delayed, amplified and pinched toward the optical axis, so the fan visibly narrows as it passes through the object instead of merely being clipped by it. - Both harmonics carry integer phase coefficients, so the curve returns to its exact starting shape after a phase revolution and the loop stays seamless. Motion is slow drift only, driven by wall-clock time so it is identical at any refresh rate, and fully suppressed under reduced motion. Two things worth recording for the next pass. The first ribbon attempt opened the bundle from 8% to 123% of the orb radius within half a screen, which read as a bowtie starburst rather than a ribbon; the spread is now deliberately gentle. Second, the fallback for a driver where the runtime effect will not compile has to be its own component: `RadialGradient` and `Shader` both use hooks, so swapping them inside one component changes that component's hook order between renders. The hero is also surfaced in the development orb gallery, since the welcome route redirects as soon as a session exists and is otherwise hard to inspect. * refactor(mobile): rebuild the welcome hero as a woven ribbon over a lit sphere The hero looked wrong for structural reasons, not tuning reasons. The ribbon morphed its whole spline once per cycle, and the orb was a dark procedural sphere with the interior ribbon painted on top of it. Correctness - Removes geometry morphing entirely, which removes the class of bug rather than patching it. `ribbonPhase` was emitted in [0, 2*pi] while `usePathInterpolation` expects a [0, 1, 2, 3] input range, and `masterCurve` and the per-strand jitter carried half-phase coefficients, so the geometry at 2*pi did not equal the geometry at 0 and the loop had a real seam. - The illustration is a brand mark, not an audio waveform. The centreline is now frozen. Life comes from a highlight travelling along the ribbon and from a rigid 4dp drift over 12s, both implemented as slow out-and-back ramps, so there is no loop boundary to seam in the first place. - Fixes the compositing order. The glass shell is now painted after the clipped interior ribbon, so the strands genuinely sit inside the sphere instead of on top of it. The ribbon is now a woven surface - One art-directed Catmull-Rom centreline, and every strand is offset along that curve's own perpendicular rather than in raw Y. Parallelism is the point: the perpendicular separation between adjacent strands is exactly `|offsetA - offsetB| * halfWidth` at every sample, and the test asserts it. A Y-offset construction only holds where the curve is horizontal and drifts apart through every bend. - 24 filaments, ordinary 0.55-0.8dp, hero 0.9-1.15dp, glow at 2.8x core width and low alpha rather than a 6x fuzzy halo. - The bundle contracts around the sphere and fans toward both edges, which is what makes the mesh read as converging on the object. - Interior geometry is only built across the sphere plus a margin, since it is clipped to the sphere; building it across the full hero width tripled the stroked segment count for nothing. The orb is now two baked layers - `hero-orb-body.webp` and `hero-orb-glass.webp`, generated by `scripts/generate-circe-hero-assets.ts`. Radius-driven shader ramps read as concentric bands: they cannot express asymmetric directional lighting, a Fresnel rim or a specular lobe. The asset is shaded from the reconstructed sphere normal with a key and fill light, a directional terminator, a subsurface glow for internal illumination and limb darkening. - The glass face is nearly clear, carrying only the Fresnel rim and two specular lobes. A broad sheen across the face fogged the body into polished metal, which is the opposite of glass over warm copper. - Full-surface hash grain is gone. It read as dithering and broke up the volume; it is replaced by 20 discrete internal light motes. - `CirceOrb` remains fully procedural for product states. The interior strands are shifted hot and lifted slightly. At the same copper as the body they vanished into it entirely, which is how the first pass shipped with an invisible interior ribbon. Hero orb radius drops to 0.215 of the width, clamped to 76-88dp, so the mesh dominates the composition rather than the sphere. The gallery gains a Frozen/Live motion switch and defaults to frozen, so a still frame can be judged before motion is allowed to excuse anything. * perf(mobile): cut per-frame work in the welcome hero ribbon Reduces the cost of the woven surface. These are defensible reductions in work per frame; see the caveat below on what I could and could not verify. - Ordinary strands no longer carry a highlight gradient. Every one of the 72 filament instances used to create two animated derived values, so all of them re-evaluated a worklet and allocated a point on every frame. The gradient now lives in its own component used only by the four hero strands. - Stroke joins are miter rather than round. Skia emits join geometry at every vertex, and this ribbon is a densely sampled polyline, so round joins were generating thousands of join primitives. At this sampling density the two are visually identical. - Sampling drops from 26 to 14 steps per segment. Stroke geometry is generated per segment, so this is a direct cost driver. - The halo pass is limited to the strands meant to catch the light. Wide translucent strokes are pure fill rate and overdraw, and a halo on all 24 strands across three planes was the largest single contributor. Measurement caveat, recorded because it is easy to misread: the screen renders at the same frame time with the hero removed entirely, so this change is not demonstrably responsible for any measured improvement, and `dumpsys gfxinfo` on this device reports internally inconsistent numbers (455 frames over 12s is a 26ms average, while the same sample reports a 61ms median). Do not treat the hero as the performance owner for this screen without a cleaner instrument. * refactor(mobile): art-direct the hero orb assets and align the hero vocabulary The previous pass produced a planet. The body had a directional falloff down to 0.24 and a round specular, which reads as a sphere under a hard key light rather than as the reference's luminous object. The shading is now art-directed rather than physical. Body (`hero-orb-body.png`) - Ramp is deep brown through warm brown and copper to peach-copper, matching the reference palette rather than the previous darker set. - The directional falloff floor rises from 0.24 to 0.45, so the shadow side stays warm brown. That single number was responsible for the planet look. - Limb darkening drops to a mild term. The fresnel rim belongs to the shell layer and should not be doubled up here. - The suspended specks are baked in. Rendering them live was a second source of truth for something that never moves. Shell (`hero-orb-shell.png`) - One anisotropic highlight streak replaces the round specular. Studio lighting reads as an elongated streak; a round dot reads as a shiny ball. - The rim is biased so it is stronger top-left, top and right rather than uniform all the way round. - The outward bloom is much tighter. The first attempt kept near full strength across the entire image margin, which rendered as a solid opaque donut around the sphere. - The face stays at 0.012 alpha, verified from the exported alpha profile, so the shell adds a rim and a streak without flattening the body's depth. Both layers now place the sphere at 0.93 of the half-image, reserving margin for the bloom to extend past the silhouette. The generator and the components share `HERO_ASSET_SPHERE_SCALE`, because a mismatch here silently misaligns the rim against the body edge. Interior mesh brightness is reduced. Pushed harder it read as a glowing stripe cutting the sphere rather than as light travelling through glass. Files are renamed to the hero vocabulary: `HeroRibbonMesh`, `HeroHaloArcs`, `HeroAmbientParticles`, `HeroOrbShell`. The geometry module keeps its specific name rather than becoming `heroMath`, since it holds ribbon geometry and not general math. * refactor(mobile): replace the welcome hero renderer with the approved illustration The hero is now the approved reference artwork, supplied as a single transparent plate. Everything the previous passes built to approximate it is deleted. Why the replacement rather than another pass - `scripts/generate-circe-hero-assets.ts` computed a sphere normal, applied key/fill dot products, a directional shade term and limb darkening, then rasterized the result. That is cached shader maths carrying a PNG extension, so it inherited every limitation of the procedural sphere it replaced and read as a glossy planet. - The mesh was one Catmull-Rom centreline with filaments offset along its normal. That construction keeps strand ordering fixed for the whole length of the ribbon, so it can only ever draw a bent sheet of parallel strands. The reference has strands that cross and change depth, fans that differ left from right, and ribbon width that varies deliberately. Those relationships are the design, and deriving them independently then compositing them at runtime produced a belt around a ball. - The layer order was correct and the alpha profile was mathematically correct the whole time. Neither of those was the problem, which is exactly why implementation-level checks kept passing while the screenshot stayed wrong. Deleted: `scripts/generate-circe-hero-assets.ts`, `HeroRibbonMesh`, `HeroHaloArcs`, `HeroAmbientParticles`, `HeroOrbBody`, `HeroOrbShell`, `heroRibbonGeometry` and its test, `heroTokens`, and both generated orb layers. What replaces them - `apps/mobile/assets/circe/welcome-hero-base.png`, 1536x1024 with real transparency, rendered as one image. - `CirceWelcomeHero` is a plain React Native image at the plate's own aspect ratio, so the composition is never cropped or distorted. There is no Skia canvas left in the hero, because there is no longer any Skia content to compose with. - Static by design. The previous revision drifted the mesh 4dp every 13s and swept a highlight along it; for a brand illustration, movement should come from light and only after the still frame matches. There is no animation to approve yet. Layout moves the hero below the subcopy, which is where the reference puts it. Asset note: the plate is 2.2MB as PNG. It is committed exactly as supplied; converting to WebP would cut it to roughly 200KB with no visual change if that matters for bundle size. * feat(mobile): finish the welcome screen against the reference Composition and typography now match the approved reference, and the screen fits without scrolling. Authored type rather than a system stack - Bundles Instrument Serif from @expo-google-fonts. It matches the reference's high-contrast editorial serif with ball terminals, and it is what the design system already named for identity moments while noting it needed a rebuild. - Registered natively in app.config.ts so release builds pay no runtime cost, and also loaded at runtime in App.tsx so a dev client built before this change can still render it without a full native rebuild. - The headline sets `fontFamily` explicitly on both spans. It previously layered a serif class over AppText's font-sans, and the two were fighting. Real brand assets - The wordmark uses the approved `circe-mark.png` instead of a redrawn SVG ring. - The Google mark was a single blue shape: the canonical four-colour paths had the blue quadrant duplicated as a full outline, which painted over the other three. Replaced with the correct brand paths. The hero no longer wastes height - The supplied plate carried about 250px of fully transparent margin above and below the mesh, which at hero scale became ~60dp of dead space and pushed the whole sign-up screen into a scroll. The plate is cropped to its content bounds (margins only, no artwork removed) and re-encoded as WebP: 2.2MB to 540KB. - The viewport is now derived from the scaled plate rather than a fixed `245-260dp`, so the composition holds across widths instead of leaving a gap on tall screens and cropping on short ones. Layout - Hero sits directly under the wordmark and above the headline, as instructed. - Headline copy is now "Talk to every machine, / from anywhere.", which covers the voice and remote aspects in two balanced lines. The previous first line was long enough to spill onto a third. - Removed the benefit row. It read as filler rather than information, and it was the last thing keeping the page scrollable. - The account link is copper throughout, the primary CTA carries a copper hairline and a copper shadow cast, and the legal links are underlined and open the real Terms and Privacy URLs in the in-app browser. * feat(mobile): rebuild the Circe orb, live voice, and no-device state Orb: port the Web Threads field to SkSL (rear field, rim caustic, refraction), keep the idle lens clean, and tune appearance. Live voice: caption shows Circe only, one failure notice instead of two, mint only on an online node, and delegate corrections/quick actions without depending on the speech model's judgement. Weather/time: propose the deterministic lookup in the bounded grammar before the project guard so quick actions never fall through to a chat model. Home: replace unusable controls with an honest no-device state, and show a connecting placeholder on cold start instead of a false no-device claim. * fix(mobile): align the theme bridge test with the design system tokens The generator test still asserted the pre-design-system screen colors (#faf7f1 / #16181b) while the authored global.css and the generated bridge use the v1 warm-paper palette (#fcf9f4 / #0c0d0e). Regenerate the committed bridge (alpha normalized to 0.2) and assert the authored values. This unblocks the mobile PR after its rebase onto main. * test(circe): register merged upstream migrations 67-69 in the manifest tests The upstream orchestration V2 merge added three migrations above Circe's shipped 41-66 slots. MigrationsRemap and the V2 migration test still asserted a contiguous manifest ending at 66 and V2 at upstream's id 53, so both failed. Extend the expected manifest to 69 and assert the Circe-remapped V2 slot while keeping the schema and index checks intact. * fix(circe): reconcile web imports with the merged V2 client surfaces The orchestration V2 merge removed exports the web app still imported, so the web bundle failed to resolve ../T3Wordmark and two components referenced removed APIs at runtime. Point V2LifecycleRow at CirceWordmark, render provider rows through the centralized ProviderInstanceIcon instead of the removed PROVIDER_ICON_BY_PROVIDER map, and use the V2 useThreadProjection hook in place of the retired V1 useThread. * test(circe): exercise the message-context migration at its remapped slot The upstream ProjectionThreadMessageContext migration registers at 67 on the Circe line, but the test migrated to 51 and asserted id 51 (CirceFollowUpQueue), so it never exercised the guarded migration. Migrate to 66, keep the manual column, then apply 67 and assert migration 67 was recorded. * test(circe): replay in shared workspaces and guard foreign databases V2 file restore now requires an isolated worktree, but the replay and fork fixtures dispatch checkpoint.rollback in a shared workspace while asserting conversation rewind, so they are conversation-only (restoreFiles: false). Drop the upstream-numbered LegacyV1Cutover integration test and cover the real invariant instead: a database recording another product's history under a Circe migration id is refused with ForeignDatabaseError. * fix(circe): map the V2 thread runtime to desktop orb statuses The orb bridge read session.status and backgroundLiveness, which the V2 shell removed, so the desktop orb mis-rendered agent status and failed typecheck. Map preparing/queued/starting to starting, running to running, waiting to waiting, and pending background tasks to monitoring, and update the fixtures. * test(circe): read the foreign-database defect with the Effect 4 Cause API * fix(circe): apply V2 module deletions and reconcile ownership guards * chore(circe): re-key service tags and align sqlite runtime with V2 Service Context tags across apps/server/src still carried upstream `t3/...` keys while the package name makes the expected deterministic key `@absterrg0/circe/...`. Re-keyed all 56 declarations, including the nine orchestration-v2 services whose class-suffixed keys (e.g. `.../CommandPolicy/ CommandPolicyV2`) the prefix-only pass did not reach. Also: - Aligned persistence/Laye…
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix: port main fixes stranded by the v2 rewrite
- Port thread pinning (#5312) into the orchestration-v2 command pipeline:
thread.pin/unpin commands, thread.pinned/unpinned events, pinnedAt on the
v2 thread state and projected shells, promotion semantics (pin clears
settle/snooze, settle clears pin) matching the v1 decider, and client
pin/unpin operations in the v2 dispatch style.
- Port the regenerated-title context anchoring (#5365) into
ThreadTitleRegenerationService: pin the first user message ahead of the
retained tail when the digest is truncated.
- Re-apply the right-panel controls positioning from #5260 to the v2
ChatView title bar controls.
- Repair merge artifacts: committed conflict markers in BranchToolbar,
duplicate capability keys, duplicate CommandPalette import, v1 turn
naming in DiffPanel's focus-refresh effect, onSend signature merge.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(server): align migration expectations with renumbered ids
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(orchestrator): Surface waiting background work (#4378)
* fix(web): align git action progress button layout
- Match progress button spacing and single-line height to static git actions
* fix: repair conflict-marker artifacts from rebase auto-resolutions
rerere replayed stale resolutions during the rebase and committed nested
conflict markers in several files. Restore the branch-intended v2 shapes
and re-graft main's compatible additions (pending-card opacity comments,
theme-editor keybinding test, mobile scroll re-arm effects from #5566).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(server): renumber v2 migrations after main's 037_ProjectionTurnsKeysetIndex
Main owns migration numbering: 037_ProjectionTurnsKeysetIndex landed on
main (#5493), so the v2 migrations shift from 037-045 to 038-046.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix: port main fixes stranded by the v2 rewrite (round 2)
Native subagent observability (#5219), wired per its spec's v2 merge plan:
- getWorkflowScript RPC re-homed onto the v2 WS surface (contracts, rpc
group, ws handler, auth scope, client atom).
- AgentsPanel fed by the spec's mapper swap: projectedSubagentsToRuntime
maps orchestration-v2 subagent entities into the panel model;
deriveAgentPanelModel's v2Projection leg is now live and the v1 fold
never runs. Agents surface wired into ChatView + RightPanelTabs.
Other ports and reconciliations:
- Shell reconnect-loop fix (#5561) ported into the v2 shell sync
(same-session resubscribes resume from the in-memory cursor), with the
cursor-resume regression test adapted to v2 fixtures.
- Mobile end-follow latch (#5566) ported onto the v2 ThreadFeed.
- Claude ede_diagnostic interrupt classification (#5557) ported into
ClaudeAdapterV2 (aborted_tools/aborted_streaming => interrupted; CLI
telemetry never becomes the failure banner). #5559 needs no v2 port
(unknown system subtypes are already ignored).
- Plan sidebar removed from the v2 ChatView/ChatComposer per main's
plans-fold-into-chat rework (#5558); rightPanelStore stays at main's
surface set.
- SettingsPanels rebuilt as main's refactored version plus the branch's
composer-context setting; sidebar snooze respects the time format
(#4438 follow-through).
- v1-only leftovers deleted: zombie v1 adapters/ingestion/tests the v2
rewrite removes, the v1-bound transfer-budget CI harness (#5350, needs
a v2 rebuild), and main's v1 client pagination machinery (#5493 client
side; the 037 keyset migration is kept — server-side v2 windowing is a
follow-up).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* chore(web): prune plan-sidebar leftovers after the inline-plans rework
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): port the refined live-follow gesture gating to the v2 timeline
The rebase kept the LegendList 3.3.3 upgrade and patch from #5449 and the
mobile end-follow latch from #5566, but the v2 MessagesTimeline/ChatView
still carried the branch's blunt any-gesture-breaks-follow listeners.
Port main's #5566 web mechanics onto the v2 follow architecture:
- resolveTimelineIsAtEnd measures the 40px follow re-arm band from real
geometry (contentLength/scroll/scrollLength minus the composer inset),
keeping the isNearEnd fallback for older state shapes.
- Follow now breaks only on gestures that can actually leave the live
edge: upward wheel with overflowing content, touch drags that exited
the end band, scrollbar drags vs content clicks, and keyboard
navigation (PageUp/Home/ArrowUp) — previously keyboard scrolling never
broke follow and the next stream chunk yanked the view back down.
- Listener attach retries across frames so a thread switch cannot mount
the list without its opt-out listeners.
Deliberately not ported: #5449's shouldRestorePosition disclosure
anchoring and follow-gated maintainScrollAtEnd — the v2 timeline keeps
maintainScrollAtEnd={false} with its own follow scrolls and anchor
system; flipping that core is a separate change.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): let LegendList own end-follow and disclosure anchoring (#5449)
Complete the #5449 architecture on the v2 timeline, following the
LegendList author's direction to lean on the list's native mechanisms
instead of app-side scroll layers:
- maintainScrollAtEnd is enabled and owned by LegendList, gated off only
while the user reads history (liveFollowEnabled), while a sent turn
anchors near the top (anchoredEndSpace), or during the two-frame settle
of a fold toggle.
- maintainVisibleContentPosition compensates size changes natively
({data, size, shouldRestorePosition}); fold toggles anchor compensation
to the toggled row via a disclosure anchor key, so the trigger stays
under the pointer instead of the viewport chasing the end.
- ChatView's hand-rolled streaming follow (double-rAF scrollToEnd on
every data change) is gone; the app now only owns streaming
adjustments during anchored-end-space mode, mirroring main.
- timelineLiveFollowEnabled state mirrors the follow refs so the
render-visible gate switches native follow off when a gesture breaks
follow and back on when the viewport returns to the end band.
Timeline tests updated to assert the native-ownership invariants.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): show Git action success inline in panel
- Keep success feedback visible in the Git action control for 10 seconds
- Move the running elapsed timer into the panel menu slot
* fix: repair rerere-damaged files and reconcile main's round-3 features with v2
Post-rebase reconciliation sweep:
- Sidebar: main's folded Sidebar.tsx/Sidebar.logic.ts adapted to v2 shells
(latestRun/runtime naming, waiting status instead of monitoring), with
subagent-thread filtering and main's pinned-reorder helpers re-exported
- Pinned drag reorder (#5581) ported into v2: thread.pin orderKey +
thread.pin.reorder command, thread.pin-reordered event, Orchestrator fold,
ProjectionStore/Maintenance, client-runtime commands and shell mapping
- Project favicon (#4849-era) and defaultThreadEnvMode flowed through v2
contracts (OrchestrationProjectShell, application event payloads)
- ChatView: main's #5592 header props, pull-request right-panel surfaces,
liveAgentCount badge (#5745) wired into the v2 panel layout
- enableAssistantStreaming -> enableLegacyTokenStreaming rename applied to
v2 RunExecutionService and replay testkit
- Removed v1 zombies resurrected by the rebase (provider service/reaper/
ingestion + v1 layer tests, server.test.ts, integration harness)
- routeTree: main's tree + branch's /settings/scheduled-tasks route
- Misc marker-sweep syntax repairs (rpc.ts, entities.ts, localApi.test.ts,
rightPanelStore.test.ts, GitManager.test.ts, mobile model menu helpers)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* chore(server): renumber v2 migrations 038-046 to 041-049 after main's 038-040
Main released ProjectionThreadsPinOrderKey (038),
ProjectionProjectsDefaultThreadEnvMode (039) and
ProjectionProjectFaviconPath (040), so the branch-private v2 stack shifts
up by three. Registry ids were already 41-49; this renames the files and
identifiers to match and updates the ledger expectations and through-id
boundaries in the migration tests (released boundary 37 -> 40).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(server): port round-3 main fixes into the v2 orchestrator
- a6c9b41f90 (agents open pasted images): ClaudeAdapterV2 now grants the
attachments dir alongside cwd via additionalDirectories and appends
'[Attached ... is saved at: path]' lines to the turn text so tools can
dereference pasted images (pixels alone are not tool-readable).
- 5bb8c03664 (settle leaves monitors running): thread.settle now joins
archive/delete in the provider-session detach set, so PR monitors, dev
servers and subagent fleets stop when the user parks the thread. The
settle guard already rejects active runs, and serialized dispatch closes
the re-engage race the v1 fix handled with onlyIfSettled.
- e70cdb478d (Claude resume handshakes) and 2c7267ad43 (reaper vs live
background subagents) are already covered structurally in v2: results
are turn-scoped with explicit zero-turn handshake drops, and idle
release is pinned while background work is pending.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(web): restore main's right-panel migration expectations after the panel-visibility merge
The keep-both merge nested main's plan-surface migration test inside a
branch popover test and dropped the threadPanelVisibilityByThreadKey key
from the migration results. Restore main's test body and include the
branch's (empty) visibility map in the expected persisted shape.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(server): expect attachment saved-at lines in ClaudeAdapterV2 turn text
Follow-up to the #5757 port: start and steer turns now append the
'[Attached ... is saved at: path]' line, so the adapter fixtures assert it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): restore the branch's slim chat header
The round-3 reconciliation took main's ChatHeader wholesale and wired its
full prop set, resurrecting the scripts/open-in/git-actions cluster the
branch had deliberately relocated into the thread panel. Restore the
79-line slim header (project favicon + name + thread title) and its
minimal ChatView call. #5592's header actions stay a documented v2
follow-up, as decided in round 2.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(web): add pull request actions to thread details
- Add ready, merge, and conflict-resolution actions to the PR row
- Share pull request action and handoff logic with the detail panel
- Fix thread details scrolling and row alignment
* fix(mobile): port main's composer stabilization into the v2 thread screens
Round-4 reconciliation of #5986/#5988 with the v2 cutover files:
- PendingUserInputCard adopts main's collapsible overlay redesign with the
v2 RuntimeRequestId/responseCapability plumbing (dead provider processes
still read-only the card)
- ThreadFeed adopts the thread-feed-live-follow transition model, the
user-scroll settle window, momentum handoff, and env-scoped feed keys
while keeping the v2 nearListEnd layout gating
- ThreadDetailScreen hides (not unmounts) the composer while a user-input
request owns the slot; multi-select answers flow as arrays end to end
(threadActivity toggle/build helpers + tests, ThreadUserInputQuestion)
- ThreadComposer keeps the v2 canStopThread stop gate under main's
onEditorFocusChange rename; standalone stop reuses the shared renderer
- Restored the branch's steer stop/send tests alongside main's composer
test suite
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): reconcile main's round-5 features after the rebase
- PullRequestDetailPanel takes main's #6039 rework wholesale (reactions,
update-branch, auto-merge, in-place editing); the thread-details action
hooks stay in usePullRequestActions with label maps extended for the
new action variants
- CommandPalette #6330 provider subtitles adapted to the v2 shell
(session -> runtime for provider instance and display name)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(web): prioritize pull request row actions
- Add conflict, draft, failing-check, and merge action ranking
- Show check progress and pull request details in row tooltips
* fix(web): restore main's collapse chrome and tab-status keying on the PR panel
The round-5 ChatView reconciliation kept the round-3-era PullRequestDetailPanel
call, so the panel mounted without chromeVariant="collapse" — the #6039
scroll-condensing chrome never engaged and the description scrolled under a
full-height chrome. Restore main's call exactly: collapse chrome,
composerDraftTarget for same-thread hand-offs, the isThreadOwnPullRequest
context check, and tab statuses keyed by the active surface id via
updatePullRequestTabStatus instead of a key rebuilt from the status payload.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(orchestration): bound thread history and resume payloads
- Add paginated thread history with bounded snapshots and replay limits
- Trim oversized wire payloads and support progressive mobile history loading
* feat(contracts): track thread title regeneration
- Add optional title regeneration request and start time to thread shells
- Cover cache serialization on mobile and client runtime
* fix: reconcile main's round-6 features after the rebase
- ChatView: #5880 auto-settle-on-merge setting flows into effectiveSettled,
#5644 browser favicon project registration effect, activeProjectRef memo,
desktopByTabId on both RightPanelTabs mounts
- server: provide ServerSecretStore to the McpSessionRegistry's
ServerEnvironment layer (#6325 reads publish opt-in per descriptor)
- mobile: 3-way merged main's deltas into the v2 thread screens
(NewTaskDraftScreen keeps the branch title seed + main's environmentId,
threadListV2 keeps both new test suites, queries imports deduped)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): keep the titlebar layout controls fixed across right-panel toggles
Restores main's one-inset rule (#5226) that a rebase resolution had
overridden with a conditional right-2 offset, which made the controls jump
sideways whenever the right panel opened. Also restores the live-agent
count badge on the right-panel toggle (#5745) that the round-6 replay
dropped, and applies the same fixed-position rule to the pull requests
page: the toggle now stays mounted at one absolute inset in both states,
with a footprint spacer in the list header so the refresh button never
slides underneath it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): align titlebar clusters to one shared pixel inset
The right controls carry mr-px (main's border compensation for anchoring
inside the panel frame), which left the sidebar trigger one pixel closer
to its edge and the sheet-mode tab bar one pixel tighter than the closed
state. Mirror the pixel on the trigger and the sheet layout-controls slot
so all three read the same inset.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): size the titlebar layout-control icons like the sidebar trigger
The trigger's icon falls through to the Button default (size-4) while the
right cluster hard-coded size-3.5, so the two ends of the titlebar read a
pixel apart on every edge. All five layout-control icons now use size-4,
matching the trigger and the pull requests page's refresh icon.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* perf(server): keep shell snapshots bounded and active-only
- Omit transcript bodies from shell rows
- Query archived threads separately and stream compact resume metadata
* fix: reconcile main's round-8 features after the rebase
Re-applies the deltas that mid-stack blob reverts discarded, and merges
main's work into the v2-owned surfaces:
- keybindings: main's STATIC_KEYBINDING_COMMANDS rename plus both new
commands (rightPanel.toggleMaximized alongside threadPanel.toggle)
- OpenInPicker: main's remote-open/SSH routing and favorite-editor
shortcut layered onto the branch's panel/toolbar variants; the
extracted shouldShowOpenInPicker now takes remoteOpenMode
- ChatMarkdown: main's bare-filename resolver (#6297) ported into the
branch's module-level component factory, plus #4133 title-attribute
stripping on links and images
- ComposerPrimaryActions: main's #4781 model (stop stays reachable, send
joins it when Enter-to-send is unavailable) carrying the branch's
steering send button
- ComposerPendingUserInputPanel: main's collapsible redesign with the v2
RuntimeRequestId and responseCapability gate
- ChatComposer: main's oversized-prompt submission guard wrapping the
branch's dispatch-mode send
- preview shell: main's container-aware width clamp ported into the
branch's usePreviewPanelInlineSize hook
- MessagesTimeline/Sidebar: main's day-aware timestamps, code-font tool
bodies and provider accent badges on the v2 runtime shell
- index.css: main's @variant dark migration (#6381) replaces the branch's
standalone .dark block
- contracts: main's send-turn image mime allowlist re-homed to
chatAttachment.ts, where v2 keeps the other send-turn limits
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(server): reject replaying a command receipt across threads in v2
Ports v1's #5246 guard into the v2 dispatcher: a stored receipt only
proves that this exact command already ran for the thread it was recorded
against, so returning it for a command aimed at a different thread reports
success for work that never happened there. The check is extracted as
canReplayCommandReceipt so the rule is unit-testable, and reuse now fails
with OrchestratorCommandIdConflictError like the v1 path.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(mobile): surface prominent activity status and metadata
- Keep prominent activity rows visible with lifecycle status and provider metadata
- Move feed sizing logic into tested helpers and preserve native measurement for activity groups
* fix: reconcile main's round-9 features after the rebase
Re-applies the deltas that mid-stack blob reverts discarded, and merges
main's round-9 work into the v2-owned surfaces:
- settings: main's Integrations page (#7082) coexists with the branch's
Scheduled Tasks page in the path union, section labels, icons, and
search catalog
- contracts: main's preview appearance/zoom/viewport settings imports
restored beside the branch's modelSelection home for ModelSelection
- mobile: main's built-in themes (#6619) re-applied to the v2 thread
screens and work log (useThemeColor over hand-rolled color-scheme
ternaries)
- MessagesTimeline: main's #7157 cleanup adopted (toolCallExpandedBody
class name unexported, implementation-detail test dropped)
- ChangedFilesTree: main's styled tooltip (#7209) carrying the v2 runId
- pullRequestDetail tests: branch's row-action coverage renamed onto
main's buildAddSelectionToAgentHandoff (#6597)
- lint: migrated the six branch-owned native title tooltips that main's
new no-native-title-tooltip rule (#7209) flags to styled Tooltips
(GitActionsControl, QueuedRunsControl, TimelineSystemDivider,
MessagesTimeline intent badge and MCP tool logo)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(server): honor withheld agent browser access in the v2 runtime
Ports #7083 into the v2 session path, which replaced the v1
ProviderService where main's gate lives. Instead of withholding the whole
t3-code MCP credential — on this branch it also carries the thread
orchestration and worktree toolkits — the credential is minted without
the "preview" capability when enableAgentBrowserAccess is off, so every
preview tool call rejects while orchestration stays available.
ProviderSessionManager reads the setting at prepare time (deny on an
unreadable settings file, matching main), rotates a reused credential
whose capability set no longer reflects the setting, and the session
config now carries browserToolsAvailable so the Codex adapter keeps its
developer instructions truthful via main's parameterized instruction
builders instead of the removed constants.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): restore the titlebar sizing and timeline fade lost to main's style simplification
Main's #6381 deleted the shared .workspace-topbar and scroll-fade rules
from index.css after inlining them at main's own call sites, but this
branch's slim chat chrome still references both classes. The round-8
rebase took the deletion without migrating the branch call sites, so the
header collapsed to zero height — the breadcrumb sat on the window edge,
timeline rows scrolled unfaded through it, and the thread-details popover
anchored to the collapsed header.
Restores both as composable utilities in #6381's own style: a
workspace-topbar utility for the titlebar rows, and the branch's
chat-timeline-scroll-fade mask (soft ramp plus a full-height scrollbar
column). Also drops the duplicated media override and its dead
settings-page-scroll-fade selector.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* refactor(web): finish aligning the branch with main's style simplification
A follow-up sweep against #6381 found the branch still carrying the
pre-simplification forms it replaced, which my earlier fix had papered
over with a compat utility instead of finishing the migration:
- ChatView now uses main's inlined titlebar sizing and the
data-workspace-titlebar-controls hook on both control clusters. The
class-based markup was silently missing the themed-toggle bridge
(html[data-theme-id] [data-workspace-titlebar-controls] …), so custom
themes lost their titlebar accent in the thread view.
- The scroll-to-end pill becomes main's Button size="xs" variant="glass"
instead of a hand-rolled button recreating it.
- MessagesTimeline uses main's consolidated topbar-scroll-fade utility;
the byte-identical chat-timeline-scroll-fade copy and the
workspace-topbar compat utility are gone.
- The composer-glass dark rules move into nested @variant dark like
main's (the raw .dark duplicates could drift from the nested copies
they shadowed), including the branch-only queue strip.
- The pre-#6381 dialog-glass/dialog-backdrop/dropdown-glass class rules
and their .dark variants are deleted: the #6381 utilities plus
call-site shadow utilities own every declaration, and the stale
dropdown rule still had the saturate-less backdrop-filter. The dead
model-picker-surface dark rule goes with them.
index.css now has zero raw .dark selectors outside the variant
definitions, matching the doctrine in
.macroscope/check-run-agents/ui-consistency.md. Verified against the
emitted production CSS: dark variants compile to :is(.dark,.dark *) with
their @supports color-mix fallbacks intact.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(orchestration): show provider retries in the work log
- Complete retry items when provider activity resumes
- Keep retry progress visible across web and mobile clients
* fix: reconcile main's round-10 features after the rebase
Ten days of main (194 commits, 640 files) re-integrated with the v2
re-architecture. The headline mechanics:
- migrations: main added 041-043 (AuthSessionClientConnection,
ProjectionThreadLinkedPullRequest, ProjectionThreadsUnsettledAt), so the
v2 block renumbers 041-049 → 044-052 with the migration tests shifted to
match
- contracts: OrchestrationClientOrigin (#7774) and the origin metadata
field live in applicationEvent.ts and re-export through the legacy
path; ProviderApprovalOption + acceptAlways + mcp-elicitation (#8058)
land in providerPolicy.ts; OrchestrationDispatchCommandError (#8824)
added; the send-turn image-mime home stays chatAttachment.ts
- threadSettled: main's settle-once-on-merge semantics (#7454) and
un-settle re-anchor (#8231) hand-merged onto the v2 duck-typed shells
(latestRun ?? latestTurn reads); web callers pass the new
ChangeRequestSettleSource shape
- timeline anchoring: main's #7897 (follow-up sends no longer push to the
top) ported by scanning user rows only; the branch test now encodes the
new semantics, as does mobile's #7969 settled-pinned shelf behavior
- vcs: branch's deleteLocalBranch coexists with main's pruneWorktrees and
the #7674 submodule checkout tests
- ws: v2 RPC surface keeps its dispatch path; main's attachment upload
RPCs (#8048) and client-connection analytics recording are wired;
providerUploadFeedback (#7949) fails explicitly pending a v2 route
- approvals: main's option-driven approval buttons (#8058) render through
the v2 canRespond gate on web and mobile
- ChatView/ChatComposer/MessagesTimeline/Sidebar/session-logic/
threadActivity keep the branch's v2 architecture; main's v1-coupled
deltas to those files are recorded for follow-up rather than
force-fitted
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): stop mis-marking recovered and text-reported tool failures in the v2 work log
Ports main's #7999/#7893 failure policy onto the v2 turn-item work log:
output text that reports a failure (command not found, ENOENT, nonzero
exit markers) now flags a row even when the provider item completed
"successfully", while the rendered row judges only its displayed result —
a command that merely greps for failure strings stays calm. Success now
also requires the failure check to pass, so recovered failures no longer
get the blue check.
The server half of #7893 needs no port: CodexAdapterV2 already projects
item.status directly, so a failed item never masquerades as completed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(orchestration-v2): project linked pull requests on threads (#8160)
Main's thread↔PR linking never reached the v2 runtime: the client types
were optional stubs and the v2 server dropped the field, so linking a PR
on a v2 environment silently no-opped and #7454's settle-once logic could
never match the linked identity.
The link now flows end to end: thread.metadata.update carries an optional
linkedPullRequest (object to link, null to unlink), the orchestrator
folds it into thread state, and both shell builders project it — no
migration needed since v2 shells persist as payload JSON. The client
command sends the field and the shell mapper surfaces it, so the existing
web/mobile badge and settle plumbing light up on v2 threads.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(orchestration-v2): carry approval options and app names to the client (#8058)
Round 10 ported main's option-driven approval buttons, but v2 runtime
requests had no way to deliver the data — every approval rendered the
default button set, MCP app names never showed, and worse, the Codex
app-server's mcpServer/elicitation/request went entirely unhandled on the
v2 adapter, so ChatGPT-app access requests could never be answered.
The v2 approval_request turn item now carries optional appName and
options, the client derivation passes them into ThreadPendingApproval,
and CodexAdapterV2 handles mcpServer/elicitation/request end to end:
unsupported shapes decline immediately (mirroring the v1 runtime), and
supported ones surface a mcp-elicitation approval built from the shared
describeMcpElicitation/toMcpElicitationResponse helpers, so the persist
tiers (session / always) advertise exactly the choices the elicitation
can express.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(orchestration-v2): route Codex thread feedback uploads through v2 (#7949)
The round-10 rebase stubbed providerUploadFeedback to an explicit error
because its v1 ProviderService route died with the v2 rewrite. The route
now goes through the v2 runtime: session runtimes may expose an optional
uploadFeedback capability, the Codex adapter implements it against the
app-server's feedback/upload request, and the WS handler resolves the
thread's live provider session through ProviderSessionManagerV2 —
failing with a plain-language reason when no session has run, the session
is gone, or the driver has no feedback channel. This also un-blocks the
dormant mobile feedback UI.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(analytics): credit v2 threads and turns to the starting client (#7774)
Main records which client surface started each thread and turn; the v2
dispatch path replaced the v1 handler that did the recording, so v2
environments only reported connections. The v2 RPC layer now records
client.thread.started on thread launches (plus client.turn.requested when
the launch carries an initial message) and client.turn.requested on
message dispatches, using the connection's announced origin. Recording is
best-effort — attribution can never fail the user's command.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(grok): fail hung prompts on xAI rate-limit completions (#8358, partial)
Ports the rate-limit half of main's #8358 into the branch's reworked XAi
extension: a prompt_complete carrying stopReason rate_limit now fails the
hung prompt with the -32003 usage-limit error instead of settling it as a
normal end_turn, so the turn surfaces "usage limit reached" rather than
silently ending. The prompt-completion deferreds carry the error channel
end to end.
The exit_plan_mode approval gate from #8358 is NOT ported here: it needs
a v2 plan-flow design in AcpAdapterV2 (the v1 GrokAdapter it lived in is
gone) and is tracked separately.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(orchestration-v2): show live context usage in the meter (#8144)
The v2 context meter could only show token counts after a compaction had
already happened — v2 had no live usage plumbing at all, so main's
compaction-threshold UX was invisible on v2 threads. Provider turns now
carry an optional tokenUsage report: the Codex adapter maps the
app-server's thread/tokenUsage/updated notification (total breakdown +
model context window) onto the active provider turn, ChatView picks the
newest report out of the projection, and the meter prefers it over the
compaction fallback — so usage and remaining-context percentages update
while the turn runs.
Claude's v2 adapter does not report usage yet; its meter falls back to
compaction items as before.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): keep following the stream after returning to the live edge (#6519)
Ports main's anchor-release semantics onto the branch's timeline anchor
state: the scroll-to-end pill and a manual scroll back to the live edge
both drop the send-time anchored end space before re-enabling follow, and
the pill's scroll runs a frame later so the list measures without the
anchor space and lands on the true end. Without this the timeline could
settle into following-end with the anchor still installed — following
nothing.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(grok): capture exit_plan_mode into the v2 proposed-plan card (#8358)
Completes the deferred half of #8358: Grok's plan-approval gate now works
on the v2 runtime. The XAi extension regains main's exit-plan helpers
(request schemas, plan-markdown extraction, the abandoned-with-feedback
response, and the plan.md session-path sniffing), and the v2 ACP adapter
grows a captureProposedPlan primitive that emits a completed
proposed-plan artifact for the active turn — one plan id per turn, so
plan.md rewrites and the exit gate update a single card.
The Grok flavor wires both ends: tool calls that write plan.md under a
Grok session dir surface the plan while plan mode is still active, and
x.ai/exit_plan_mode (plus the underscore alias) captures the final plan —
request content first, then the sniffed plan.md contents, then the
empty-state placeholder — and abandons the native gate so the turn does
not hang, mirroring the Claude ExitPlanMode pattern.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): repaint the composer glass and strip the thread-panel popover chrome
Two post-rebase regressions from the round-10 index.css merge:
- The merge seam ate the closing brace of .chat-composer-glass, silently
nesting the entire composer-glass section (shell, host, context strip,
shoulder tab, banner cap) inside it as descendant rules that never
matched — the composer surface stopped painting and thread content
showed straight through the input. The brace is restored and the
compensating over-close removed; every composer selector emits at top
level again.
- The thread-details popover grew dropdown-glass card chrome around the
panel: round 9 deleted the legacy components-layer .dropdown-glass rule
in favor of the @utility, which the popover's border-0/bg-transparent
suppressors no longer outrank (the utility emits later in the layer).
The suppressors are now important variants, matching the !overflow
override already there.
Verified against the emitted production CSS: shell::before is top-level,
no descendant-of-glass selectors remain, and the important suppressors
(including [backdrop-filter:none]!) are emitted.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): adopt main's attached-composer surface contract so the glass survives shoulder tabs
The composer went frameless exactly when the stash or tasks badge was
showing: main's #7150 css hides the classic shell chrome whenever the
shell :has() a shoulder tab or top drawer and repaints the glass on
[data-chat-composer-main-surface] instead — an element the branch's
composer body never rendered, since it predates the drawer system while
the badges and css came through the rebase in main's new form.
The branch composer's frame div now carries the main-surface attribute
(with main's relative z-10 stacking) so attached mode paints background,
outline, and backdrop on it and the tab connects to the surface, and
ChatView applies chat-composer-glass-shell-attached while banner items
render in the drawer slot, matching main's externalComposerDrawerAttached
wiring. Without a tab or banner the attribute is inert and the classic
shell chrome paints as before.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): converge ChatComposer on main's drawer-era body
Round 10 restored the branch's pre-#7150 composer body while its
satellites (shoulder badges, banner drawers, glass css) arrived in
main's new form, and every seam between the two was a visible bug:
opaque/frameless composer, overflowing stash tab, detached stash menu.
Rebuilt ChatComposer via a reverse three-way merge (main's body as the
base, branch delta re-applied): dispatchMode send boundary, live-capable
approval gates, latestRun reads, and the v2 context-window meter stay;
everything else now matches main, including ComposerPrimaryActions and
the sendDisabledReason send gating. Attachment uploads stay off until
the v2 claim path lands.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): collapse settled tool runs behind main's summary toggles
The v2 timeline rendered every tool call as its own raw row; main's
tool-group collapsing (generated "Ran N commands and changed M files"
summaries, the live "Running <program>" pill for the active tool run,
and the "+N previous log entries" clamp for mixed groups) never made it
into the v2 row model.
Ported the work-live/work-toggle row kinds and group summarization into
the v2 derive, keyed on v2 item types (command_execution, file_change,
file_search, dynamic_tool, subagent) and runId lifecycle instead of v1
activities. Expanded groups keep the branch's richer per-entry detail
rows (V2ItemInspector) — only the collapsed presentation converges on
main.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): surface v2 todo-list plans as task progress
v2 already projected todo_list plan artifacts (deriveActivePlanState
existed with zero consumers), but nothing rendered them: todo_list turn
items showed as a bare "Updated tasks" work row, the composer Tasks
drawer never appeared, and the working row had no current-step label.
todo_list items now become inline turn-plan chips (mini step segments,
current step, N/M count, expandable step list) that fold with their
settled turn, ChatView derives the composer Tasks drawer progress and
steps from the running run's plan artifact, and the working row shows
"Working for Xs · <current step>" like main.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): show the command on collapsed tool rows, not its stdout
Collapsed command rows rendered "Command" plus the raw result JSON as
the preview; the command itself is the useful collapsed line, so it now
renders as the row text (whitespace-collapsed, truncated) with stdout
and the full payload behind the expander. Tool-like headings drop the
bold foreground for the muted secondary-label the summary rows use, and
the "+N previous tool calls" toggle loses its bold black label for the
same muted treatment.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): collapsed tool rows preview inputs for every tool type
Extends the command-row fix to the whole preview: file-change rows were
still leaking raw diff lines into the collapsed line. workEntryPreview
now resolves input-first — command, then touched-file paths, then
detail (which is input for the remaining types: search patterns,
reasoning text, error messages) — so outputs only appear behind the
expander.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix: reconcile main's round-11 features after the rebase
Round-11 rebase onto main (25 commits). Reverse-merged main's new work
into the v2 cutover files: #8395 muted ordinary-tool-failure treatment
(v2-adapted workEntrySignalsSevereFailure keyed on error items), #5931
sidebar project-filter combobox + #4c51 keyboard pin/settle with their
ChatView support graph, the auto-settle-mode migration through
threadSettled/threadListV2, #8235 file/unknown attachment schemas moved
into chatAttachment.ts with nullable attachment paths, #8481 client
analytics through the v2 ws layer, #8480 OpenCode server owner wired
into the driver beside the v2 orchestration adapter, and the mobile
semantic-theme migration applied to branch-only components.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(opencode): route child-session approvals through the v2 adapter
Ports the v2-applicable half of #8480 (the v1 adapter rewrite was not
carried; this branch's OpenCode path is OpenCodeAdapterV2). Permission
and question asks from child sessions — task subagents and their
descendants — were dropped because the adapter only looked up root
thread sessions. Related sessions now map back to the owning root
state (registered from task parts and session.created/updated parent
chains), and an ask that arrives before the relation is known resolves
it inline via session.get with a short forked backoff, then surfaces
the approval on the root turn. Replies already route by native request
id. Interrupts now tolerate the abort racing turn settlement instead
of failing the stop.
Covered by a new opencode_child_approval replay fixture where the
child asks for bash permission before the task part reveals the
relation.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix: reconcile main's round-12 features after the rebase
Small rebase onto main (4 commits). The composer stash-shortcut label
and the mobile start-task menu refactor merged onto the v2 composer
bodies, and main's new auto-settle list tests are ported to the v2
thread-list test file (latestRun/RunId shapes). The auto-settle
machinery itself already matched main from the round-11 reconcile.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix: restore main's automatic thread settling after the revert
Round-13 rebase onto main (2 commits). Main reverted the auto-settle
opt-in (#8596 undoing #8321), so the branch drops the ported
autoSettleMode machinery and returns to sidebarAutoSettleOnMerge with
settling-by-default, keeping only the v2 shell-shape delta in
threadSettled. The unpin confirmation (#7313) merges into the v2
thread-actions hook, and the mobile list tests re-sync to main's
reverted semantics in v2 shapes.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): restore the full-screen file-drop target over the chat column
Main's #6636 workspace file drop (drag anywhere over the chat column to
attach, with the dashed overlay) split across ChatView and the
composer. The round-10 rebase restored ChatView from the pre-#6636
backup wholesale and only the composer half was ever re-applied, so
the drop target, overlay, and drag-state plumbing vanished while
addDroppedFiles sat unused on the composer handle. Re-applies main's
ChatView half verbatim; the shared workspaceFileDrop module was
already identical to main.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(server): claim uploaded attachments at v2 dispatch
Closes the last gap from the rebase follow-up list (#8048/#8161 for
v2): pending uploads staged via the attachment upload URL flow were
never claimed by the v2 orchestrator, so the composer kept
supportsAttachmentUploads off and fell back to inline dataUrls with no
progress UI and no big-file support.
message.dispatch and thread.launch now claim pending refs at intake —
verify the staged file, copy it under a thread-scoped id (the pending
copy stays as the retry source), rewrite the refs, and release the
claimed copies if the dispatch fails. A launch carrying uploads
requires its thread id up front. The web composer reads the
attachmentUploads capability again like main, which lights up the
upload progress overlay, retry-on-failure, and PDF/ZIP attachments.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(web): show attachments on queued messages and edit them in the composer
Queued rows now render image thumbnails, and the pencil action loads the
queued message into the composer instead of an inline input: text and
stored attachments are editable (attachments removable, new images
addable), sending saves the queued run in place, and the user's own
draft is stashed untouched for the duration. queued-run.edit gains an
optional full-replacement attachments list end to end.
Built by Claude Fable 5 on Claude Code.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): drag-to-reorder queued messages and retire stale pending rows
Replaces the queue rows' up/down arrow buttons with a drag handle (arrow
keys still work on the focused handle). Also prunes optimistic queued
messages once the projection holds them: keying the prune on turn items
alone left a phantom clock row behind whenever a queued run was removed
or steered before it ever started.
Built by Claude Fable 5 on Claude Code.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix: reconcile main's round-14 features after the rebase
Round-14 rebase onto main (7 commits, mostly the Expo SDK 57 upgrade
plus the mobile glass restore and codex app-server buffering fix). The
mobile composer merged main's restored glass chrome onto the v2 body,
and the lockfile is regenerated from main's SDK-57 lock with the
branch's extra dependencies installed on top.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): dedupe the composer glass styles and align the chat column width
The rebase left a stale early copy of the composer glass-host, context-strip,
and shape() fallback rules that the identical later block always overrode.
The composer shell and queue/context strips also kept main's 48rem width while
the timeline moved to the 46rem content lane; they now share
--chat-content-max-width so one owner defines the chat column.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(mobile): replace remaining dark: utilities with adaptive semantic tokens
The v2 thread surfaces still styled borders, fills, and status text with
dark:/light: pairs, which do not follow registered custom themes and now fail
the no-mobile-uniwind-theme-escape-hatches lint. Convert them to adaptive
tokens, adding the missing amber/sky badge and neutral hairline/fill entries
to the theme generator.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(lint): allowlist the queue and relationships interop boundaries
ThreadQueueControl and ThreadRelationshipsBanner read theme variables only to
tint SymbolView icons and color native modal chrome, the same reviewed interop
pattern as the existing thread-feed entries.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(server): inject HostProcessPlatform into the Grok plan extractor
The plan.md path check read process.platform and process.env directly; thread
the host platform reference and the adapter's provider environment through
GrokAdapterV2Options instead.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* chore: retrigger ci
The pull_request workflows never fired for 6c3b84bbfc; only the
pull_request_target ones ran.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* chore: refresh macroscope ui-consistency check
Its findings were fixed in 9abca06b28 and the review threads are resolved;
the check only re-evaluates on push.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix: reconcile main's round-16 features after the rebase
Round-16 rebase onto main (5 commits, headlined by #8236 file
attachments in the client). Main's upload-aware send path — capability
probes, upload await/retry gating, uploaded-ref turn attachments with
dataUrl fallback, draft release on success — is woven into the v2
dispatch flow, timeline user rows render file attachments as download
links with the ChatView download handler, and the provider settings
editor cleanup keeps the branch's environment-field rows. The codex
feedback client flow stays unported, replay testkit configs gained
main's environmentThemesDir, and the rpc/settings/docs unions carry
both sides.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): keep failed tool items in the collapsed group summaries
An ordinary exit-code failure knocked its whole tool group out of the
"Ran N commands" summary (and the live pill mid-run) into the raw
"+N previous tool calls" clamp: v2's derived tone marked any
status=failed item as "error", which the grouping treats as a non-tool
row. That inverts v1's semantics and #8395's muted-failure rule — the
failed lifecycle status already carries the X marker and the summary's
includes-a-failure hint, so the tone override goes.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(server): keep Claude session approvals ephemeral
R03: Rescope Claude permission suggestions to the current session and add a whole-tool session rule when the SDK provides no suggestion.
* fix(orchestration): reanchor unsettled threads
Carry unsettledAt through v2 thread state and shells, stamp explicit and activity-driven reactivation, and clear the stamp when settling.
Finding: R11
Implemented by GPT-5.6 Sol in Codex.
* fix(server): observe pre-aborted Claude approvals
R15: Race Claude approval decisions with cancellation while checking already-aborted signals and cleaning up the listener on every exit.
* fix(server): include service launcher in bundle build
Build the service launcher after the clean CLI pack so publish and background-service installation receive the required artifact.
Audit: R01
* fix(orchestration): preserve legacy thread metadata
Import pin order, snooze state, unsettle time, and linked pull requests. Repair prior imports only where the stored v2 property is absent, so later v2 changes remain authoritative.
Finding: R08
Implemented by GPT-5.6 Sol in Codex.
* fix(web): honor disabled legacy plan mode
Force the effective composer interaction mode to build when the legacy plan setting is off, including threads and drafts with a saved plan mode.
Audit: R13
* fix(server): preserve Claude subagent models
R16: Apply authoritative assistant snapshot models to Claude subagents and buffer snapshots that arrive before task_started.
* fix(orchestration): honor migrated thread visibility in search
Use v2 ownership and lifecycle metadata when a legacy transcript belongs to a migrated thread. Legacy transcript rows remain searchable until lazy hydration finishes.
Finding: R09
Implemented by GPT-5.6 Sol in Codex.
* fix(orchestration): recreate missing worktrees before turns
Prune stale git worktree registration and recreate the saved branch at the saved path before provider startup. Recovery remains best effort so normal provider errors still report when repair is impossible.
Finding: R10
Implemented by GPT-5.6 Sol in Codex.
* fix(clients): restore Codex feedback submission
Intercept /feedback in web and mobile, show the upload result and feedback ID in the thread, and block duplicate submissions while an upload is active.
Audit: R12
* fix(server): preserve generic provider attachments
R02: Append persisted paths for every uploaded file on provider sends and steering while reserving native image payloads for supported images.
* fix(web): load workspace markdown images through assets
Classify markdown image sources and request environment-scoped asset URLs for workspace files while leaving ordinary web images direct.
Audit: R17
* fix(web): preserve Windows markdown paths
Normalize drive-path links and image sources before sanitization so file chips and signed workspace images receive usable paths.
Audit: R18
* fix(server): keep current provider context usage
R07: Project Codex last-turn and Claude assistant context usage, and retain the latest usage when terminal provider-turn updates omit it.
* fix(web): restore markdown file chip actions
Keep ordinary file-chip clicks in the in-app preview while restoring modifier-click editor opening, configured editor labels, remote gating, and reveal-in-file-manager actions.
Audit: R19
* fix(web): scope markdown actions to their environment
Use the owning thread or pull request environment for editor, shell, and remote-open actions instead of whichever environment is active. Add a multi-environment regression test for the action hooks.
Audit: R20
Implemented by GPT-5.6 Sol with Codex.
* fix(protocol): reject incompatible orchestration peers
Advertise and validate an explicit orchestration protocol before clients open RPC sessions. Announce the same protocol on WebSocket upgrades so hosts reject older clients before request decoding while preserving existing auth and relay parameters.
Audit: D03
Implemented by GPT-5.6 Sol with Codex.
* docs: explain legacy thread migration
Document which thread metadata and transcript data migrate, which runtime history does not, and how the fresh provider continuation uses the latest 32,000 characters. Add a safe read-only recovery procedure without claiming an export API.
Audit: D02
Written by GPT-5.6 Sol with Codex.
* docs: state portable handoff limits
Explain the eligible timeline items, whitespace-normalized 240-character prefixes, omitted tail risk, and practical preparation for provider or fork handoffs. Distinguish this rule from the legacy import's 32,000-character transcript suffix.
Audit: D04
Written by GPT-5.6 Sol with Codex.
* chore(repo): remove tracked audit scratch files
Remove obsolete implementation plans and the probe write marker so temporary work artifacts no longer ship with the repository.
Audit: H01
Implemented by GPT-5.6 Sol with Codex.
* fix(server): guard OpenCode prompt admission races
R14: Hold idle completion through prompt admission, reconcile status only for the current admission generation, and invalidate admission before abort.
* fix(server): restore Claude structured questions
Project AskUserQuestion as a structured user-input runtime request and return keyed answers to the Claude SDK instead of routing the tool through generic approval.
Finding: R04
Model: GPT-5.6 Sol via Codex
* fix(server): project Claude plans and todos
Translate TodoWrite and ExitPlanMode tool input into canonical todo-list and proposed-plan artifacts so every client can render Claude planning state.
Finding: R05
Model: GPT-5.6 Sol via Codex
* perf(orchestration): bound history reads in SQL
Load at most one turn-item page per thread in a fork lineage before decoding, keyed by the stable history cursor. Restrict message, plan, and handoff reads to that page plus live actionable state so cold opens and older-page requests no longer decode complete historical tables.
Finding: P01
Implemented by GPT-5.6 Sol in Codex.
* perf(orchestration): bound complete thread snapshots
Budget the serialized bounded projection after retaining live control state. Cap historical control arrays and large plan or handoff details only on the bounded route; the full thread-detail route remains available for complete text.
Finding: P02
Implemented by GPT-5.6 Sol in Codex.
* fix(server): restore Claude resume compaction
Pass the automatic compaction window to Claude and route resume-return dialogs through structured user input so users can compact, continue, or permanently dismiss the prompt.
Finding: R06
Model: GPT-5.6 Sol via Codex
* fix(server): allow protocol negotiation in CORS
Permit the canonical orchestration protocol header in browser API preflights so cross-origin web and desktop clients can negotiate compatibility while retaining authorization and DPoP headers.
Finding: D03
Model: GPT-5.6 Sol via Codex
* fix(server): preserve provider usage in persisted turns
Merge terminal provider updates with stored context usage before replacing the SQLite payload. Keep newer usage reports authoritative and verify the persisted projection after reload.
Finding: R07 follow-up
Model: GPT-5.6 Sol via Codex
* fix(server): preserve Claude planning lifecycle
Keep typed plan and todo records distinct from generic tool events, activate captured plans, and supersede older planning state within the owning thread. Ignore nested todo snapshots for the parent and retain identity across duplicate SDK messages.
Finding: R05 follow-up
Model: GPT-5.6 Sol via Codex
* fix(server): normalize Claude question answers
R04 follow-up
Convert client multi-select answer arrays to the comma-separated string shape required by the pinned Claude SDK while preserving single-select strings.
Implemented by GPT-5.6 Sol via Codex.
* fix(server): correlate OpenCode prompt admission
Stale cached user and status events could admit and complete a newly submitted OpenCode prompt. Generate the native message ID before submission and only advance admission when that exact message is observed.
Finding: R14
Implemented by GPT-5.6 Sol with Codex.
* fix(clients): anchor feedback in conversation order
R12 follow-up
Insert persistent feedback blocks by their timestamp within the canonical timeline while preserving projected row order. Keep real optimistic sends appended and suppress duplicate local messages already committed by the server.
Implemented by GPT-5.6 Sol via Codex.
* fix(web): retain markdown workspace ownership
R20 follow-up
Give inspector reasoning markdown its projected source thread and retain the explicit environment fallback for proposed plans without a thread reference. Workspace links and images now resolve through their owning environment after removal of the active-environment fallback.
Implemented by GPT-5.6 Sol via Codex.
* fix(orchestration): page history through its true end
Read the inclusive cursor, a full history page, and a look-behind row so older history does not terminate after one page.
Finding: P01 pagination termination
Model: GPT-5.6 Sol via Codex
* fix(server): cancel pending OpenCode prompts safely
Cancel pending SDK requests before aborting the native session. Preserve per-admission cancellation state and treat stopped initial prompts as interruption instead of provider failure.
Finding: R14 prompt cancellation
Model: GPT-5.6 Sol via Codex
* fix(orchestration): retain nested fork history when paging
Keep the original cursor owner through ancestor traversal and preserve the history budget across empty intermediate forks. Verify exact paged history against the complete nested projection.
Finding: P01 nested lineage
Model: GPT-5.6 Sol via Codex
* fix(server): recover OpenCode status reconciliation
Retain pending admission after transient status failures and use one generation-owned retry worker. Ignore stale timers and duplicate evidence so older prompts cannot finish newer steering.
Finding: R14 status reconciliation
Model: GPT-5.6 Sol via Codex
* fix(orchestration): select visible history before limiting SQL
Keep hidden local and inherited rows from consuming history pages. Preserve stop-request dependencies, source-run cutoffs, and imported history while loading related metadata from the selected cohort and using indexed watermark lookups.
Finding: P01 bounded history visibility
Model: GPT-5.6 Sol via Codex
* fix(web): port composer activity and grouping to orchestration v2
* fix(web): align queue headers and prevent stash overlap
* fix(web): share the outline for joined composer tabs
* fix(web): keep stash separate from the composer activity column
* refactor(web): use shared banner rows for queued messages
* fix(web): keep queued message editing inside the queue panel
* fix(web): keep queued messages in place while editing
* fix(web): match composer actions to draft and modifier state
* fix(web): keep composer shortcut tooltip stable on Mod
* feat(web): summarize T3 orchestration actions
* feat(mobile): port chat summaries and transitions to orchestration v2
Adapt grouped tool summaries and the floating working timer to V2 run, attempt, and queue state. Bring over the composer, keyboard, and disclosure transitions while retaining the V2 activity inspector and queue controls.
Keep OV2 web composer and grouping behavior intact; share only the existing command label parser with mobile.
* fix(chat): remove added tool summary status counts
* fix(mobile): keep scroll bounds current after animations
* fix: reconcile main's round-17 features after the rebase
Restores main features dropped by the policy replay: #8569 theme wiring,
settings search rework, #8803 workspace-mutation refresh (v2-adapted),
video + image previews (web and mobile, v2-adapted), #8862 Expo glass,
and the round's docs. Timeline thinking rows (#8984) stay on the v2
work-live system.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(web): port working and thinking timeline rows to orchestration v2
The v2 equivalents of main's #8984 and #8922: a "Working for ..." header
anchors the active run, the trailing live tool row survives between
actions in past tense instead of vanishing, and a shimmering Thinking
row marks reasoning gaps. During workspace preparation the header shows
"Setting up worktree..." (driven by the local dispatch flag or the v2
run's preparing status, so remote viewers see it too), the composer
footer span is gone, and draft promotion waits until the run starts or
startup fails instead of navigating mid-preparation.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix: reconcile main's round-18 features after the rebase
Adopts the round's main features into the v2 architecture: the #9023
media rework (streamed videos, media-file assets, protocol-relative
links), #9098 shared live-activity row folded into the v2 working and
thinking rows, the #9084/#9078 Claude model catalog for v2 consumers,
a native #9005 OpenCode child-session abort in the v2 adapter, #9013's
landed LegendList patch, and per-environment sidebar provider entries.
For #8600 the server-side pieces land, but auto-settle evaluation stays
client-side (reading the new server-owned settings) until the v2
orchestrator grows its own settlement reactor; main's v1-only reactor
and coalescer additions are dropped with the rest of the v1 path.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(server): evaluate automatic thread settlement in the v2 orchestrator
Ports #8600's server-owned settlement to orchestration v2 instead of
keeping client-side evaluation. A ThreadSettlementService sweep runs at
startup, on auto-settle settings changes, and once per minute: it
evaluates inactivity and merged or closed pull requests over v2 thread
shells and dispatches the new guarded thread.auto-settle command, which
rejects threads that changed after the sweep's snapshot or carry any
explicit override, then reuses the orchestrator's settle lifecycle.
With the server deciding, the clients drop their effectiveSettled
evaluation and partition on the persisted settledOverride like main.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix: reconcile main's round-19 features after the rebase
Adopts main's round-19 features into the v2 stack: payload-budgeted
orchestration replay (#8992), sidebar row subscription leases (#9052),
tool group virtualization and scroll anchoring (#9106), repeated-command
and browser-group presentation, inline assistant citations (#9146),
per-cwd provider skills discovery (#8778), Claude composer skill
dispatch (#9128), grok health probe and model negotiation (#9154), and
the failed-tool thinking fallback (#9165).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): right-align the stash shoulder tab again
Round 17 adopted main's #8850 ComposerBanner.Attachment (mx-auto plus the
standalone drawer-inset width) without main's matching mounts, so the
stash tab's ml-auto lost to the attachment's auto right margin and the
tab centered over the composer. Column now spans its attachments like
main does, the stash tab zeroes the right margin, and the stash menu
keeps the full dock width.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(web): realign the composer and timeline with main
The branch had drifted from main's composer and work-log design in ways
unrelated to orchestration v2: a pre-revert "Working for" shoulder tab
on the composer (main reverted #8693 and re-landed #8734 without it),
an inline stash variant plus in-flow stash menu, expanded tool rows that
hid their icons, an unmounted woke-thread banner, a composer scroll
observer main never had, and a right-panel toggle that lost its
showRightPanelControl gate so it rendered twice with the panel open.
ChatComposer and its satellites now start from main's files with only
the v2 delta re-applied (dispatch modes, queued-message editing, runtime
request ids, response capability). Background tasks surface as a
ChatView banner in main's backgroundLiveness shape instead of a
composer tab. SimpleWorkEntryRow takes main's PlainWorkEntryRow body
with the V2ItemInspector kept behind the expander.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix: reconcile main's round-20 features after the rebase
Renumbers the v2 migrations 044-052 to 045-053 behind main's
044_ClearAutomaticProjectModelDefaults, and adopts main's sticky
new-thread selection (#9164), local-only worktree bases in the v2 launch
path (#8751), the PR summary read for settlement (#9176), Claude per-cwd
skills (#9210), the provider editor redesign with the branch's dedicated
environment fields re-grafted (#8508), and the client half of
continue-threads-across-restart (#9167). The server-side continuation
markers stay unported: they live in the v1 session directory, and v2
recovery terminalizes running runs on restart, so the capability is
withheld until the v2 runtime carries them.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(server): reduce v2 recovery and runtime resource usage
Index event sequence lookups, recover only threads with pending work, and page projection verification and rebuilds. Bound provider event logging and omit turn histories when resuming Codex threads.
Allow delegated thread identifiers through relay routes. Add focused regression coverage and document the performance constraints.
* fix: reconcile main updates with orchestration v2
Retain main's composer, work-log, settings, mobile and performance changes through c8f77e0d441 while preserving v2 runs, queued messages, provider handoffs and durable history.
Port native compaction and logout, asynchronous Codex questions, provider usage accounting, automatic settlement and PR refresh into the v2 services. Bound live event retention during replay and delivery, measure thread replay before decoding, and read checkpoint metadata without loading transcripts or patches.
Keep main migrations through 047 and move the v2 migrations to 048–058. Preserve the existing branch history and the pre-rebase backup.
Model: GPT-6. Harness: Codex.
* fix(orchestration): stabilize Codex turn mapping and settlement
- Preserve Codex turn identity while suppressing duplicate diff notifications
- Optimize settlement projections and isolate thread visit hand…
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…247) * feat(web): rename and delete a Squadron from the scope control The sidebar Squadron scope dropdown gains a section for the selected Squadron with Rename and Delete actions. Rename opens a dialog prefilled with the current name that trims and disables submit on empty or unchanged input. Delete opens a confirmation naming the Squadron, stating that thread labels lose their Squadron home and that live members or Crews block deletion, and shows the server's 409 message verbatim. Both actions call the owning environment, then force-refresh that environment's Squadron directory, requested thread homes, and the Fleet read. After delete, the ambient scope and any draft carrier that pointed at the deleted Squadron reset to All Squadrons. The client commands are a local stub against the agreed PATCH/DELETE routes until client-runtime ships renameSquadron/deleteSquadron. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(web): delete confirmation says Squadron history goes with it The server removes the Squadron's ledger and placement history inside the delete transaction, so the confirmation dialog now states that alongside the existing consequences. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * feat(server): squadrons can be renamed and deleted over J5 HTTP Squadrons could only be listed and created. Add PATCH and DELETE routes on /api/j5/squadrons/:id backed by new ledger and management operations. Rename trims the name and rejects a blank one. Delete is one transaction that first counts members, active Crews, machine participants, and the ON DELETE RESTRICT history tables (comm events, placement events, command receipts) and refuses with a typed SquadronDeleteBlockedError naming the blockers, returned as HTTP 409; CASCADE tables go with the row. Contracts and client-runtime fetchers renameSquadron/deleteSquadron are appended next to the existing Squadron schemas. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * feat(client-runtime): rename and delete Squadron commands Expose renameSquadron and deleteSquadron on the J5 environment atoms next to createSquadron so the web scope control can run them. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * refactor(web): Squadron rename and delete use client-runtime commands client-runtime now ships renameSquadron and deleteSquadron in the J5 environment atoms, so the temporary HTTP stub in the web state module is gone and the web client points at those commands. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * refactor(web): Squadron rename and delete call the environment atoms directly Matches createSquadron: squadronClient runs j5Environment.renameSquadron and j5Environment.deleteSquadron itself, so the alias object and the stub's leftover imports leave the web state module untouched. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * feat(server): squadron delete purges history and refuses only live state A delete that only worked on never-used Squadrons did not meet the ask. Only unarchived agent members and running Crews now block deletion, named with counts in the 409 message. Otherwise one transaction removes every squadron-keyed row in dependency order, including the ON DELETE RESTRICT history tables (comm events, command receipts, placement events), before the squadron row, so threads that were homed there read as unknown-home. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * test(server): squadron delete refusal test no longer depends on list order Two Squadrons created in the same millisecond tie on created_at and fall back to random id order, so the assertion sorts the names first. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * test(server): squadron delete purge list is checked against the live schema The purge now iterates one exported list of every table with a foreign key onto j5_a2a_squadron, and a test reads PRAGMA foreign_key_list for the migrated schema and asserts the list matches, so a future referencing table fails the test instead of surfacing as a constraint error. The archived-member delete test also seeds placement history and asserts the thread reads as unknown home afterward. Blockers stay unarchived agent members and running Crews; a send-only machine credential is purged with its Squadron because it means nothing without it. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * fix(web): rename Squadron dialog keeps its footer inside the popup The dialog popup lays out header, panel, and footer as a flex column. The rename form sat between them as a plain block, so the scroll panel grew past the popup border and the Cancel and Rename buttons rendered on a strip below it. The form is now a flex column like the upstream wrapper in the publish dialog, so the footer sits inside the popup again. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * chore: integrate upstream orchestration V2 into J5 (#178) * fix: port main fixes stranded by the v2 rewrite - Port thread pinning (#5312) into the orchestration-v2 command pipeline: thread.pin/unpin commands, thread.pinned/unpinned events, pinnedAt on the v2 thread state and projected shells, promotion semantics (pin clears settle/snooze, settle clears pin) matching the v1 decider, and client pin/unpin operations in the v2 dispatch style. - Port the regenerated-title context anchoring (#5365) into ThreadTitleRegenerationService: pin the first user message ahead of the retained tail when the digest is truncated. - Re-apply the right-panel controls positioning from #5260 to the v2 ChatView title bar controls. - Repair merge artifacts: committed conflict markers in BranchToolbar, duplicate capability keys, duplicate CommandPalette import, v1 turn naming in DiffPanel's focus-refresh effect, onSend signature merge. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * test(server): align migration expectations with renumbered ids Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(orchestrator): Surface waiting background work (#4378) * fix(web): align git action progress button layout - Match progress button spacing and single-line height to static git actions * fix: repair conflict-marker artifacts from rebase auto-resolutions rerere replayed stale resolutions during the rebase and committed nested conflict markers in several files. Restore the branch-intended v2 shapes and re-graft main's compatible additions (pending-card opacity comments, theme-editor keybinding test, mobile scroll re-arm effects from #5566). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(server): renumber v2 migrations after main's 037_ProjectionTurnsKeysetIndex Main owns migration numbering: 037_ProjectionTurnsKeysetIndex landed on main (#5493), so the v2 migrations shift from 037-045 to 038-046. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: port main fixes stranded by the v2 rewrite (round 2) Native subagent observability (#5219), wired per its spec's v2 merge plan: - getWorkflowScript RPC re-homed onto the v2 WS surface (contracts, rpc group, ws handler, auth scope, client atom). - AgentsPanel fed by the spec's mapper swap: projectedSubagentsToRuntime maps orchestration-v2 subagent entities into the panel model; deriveAgentPanelModel's v2Projection leg is now live and the v1 fold never runs. Agents surface wired into ChatView + RightPanelTabs. Other ports and reconciliations: - Shell reconnect-loop fix (#5561) ported into the v2 shell sync (same-session resubscribes resume from the in-memory cursor), with the cursor-resume regression test adapted to v2 fixtures. - Mobile end-follow latch (#5566) ported onto the v2 ThreadFeed. - Claude ede_diagnostic interrupt classification (#5557) ported into ClaudeAdapterV2 (aborted_tools/aborted_streaming => interrupted; CLI telemetry never becomes the failure banner). #5559 needs no v2 port (unknown system subtypes are already ignored). - Plan sidebar removed from the v2 ChatView/ChatComposer per main's plans-fold-into-chat rework (#5558); rightPanelStore stays at main's surface set. - SettingsPanels rebuilt as main's refactored version plus the branch's composer-context setting; sidebar snooze respects the time format (#4438 follow-through). - v1-only leftovers deleted: zombie v1 adapters/ingestion/tests the v2 rewrite removes, the v1-bound transfer-budget CI harness (#5350, needs a v2 rebuild), and main's v1 client pagination machinery (#5493 client side; the 037 keyset migration is kept — server-side v2 windowing is a follow-up). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * chore(web): prune plan-sidebar leftovers after the inline-plans rework Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): port the refined live-follow gesture gating to the v2 timeline The rebase kept the LegendList 3.3.3 upgrade and patch from #5449 and the mobile end-follow latch from #5566, but the v2 MessagesTimeline/ChatView still carried the branch's blunt any-gesture-breaks-follow listeners. Port main's #5566 web mechanics onto the v2 follow architecture: - resolveTimelineIsAtEnd measures the 40px follow re-arm band from real geometry (contentLength/scroll/scrollLength minus the composer inset), keeping the isNearEnd fallback for older state shapes. - Follow now breaks only on gestures that can actually leave the live edge: upward wheel with overflowing content, touch drags that exited the end band, scrollbar drags vs content clicks, and keyboard navigation (PageUp/Home/ArrowUp) — previously keyboard scrolling never broke follow and the next stream chunk yanked the view back down. - Listener attach retries across frames so a thread switch cannot mount the list without its opt-out listeners. Deliberately not ported: #5449's shouldRestorePosition disclosure anchoring and follow-gated maintainScrollAtEnd — the v2 timeline keeps maintainScrollAtEnd={false} with its own follow scrolls and anchor system; flipping that core is a separate change. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): let LegendList own end-follow and disclosure anchoring (#5449) Complete the #5449 architecture on the v2 timeline, following the LegendList author's direction to lean on the list's native mechanisms instead of app-side scroll layers: - maintainScrollAtEnd is enabled and owned by LegendList, gated off only while the user reads history (liveFollowEnabled), while a sent turn anchors near the top (anchoredEndSpace), or during the two-frame settle of a fold toggle. - maintainVisibleContentPosition compensates size changes natively ({data, size, shouldRestorePosition}); fold toggles anchor compensation to the toggled row via a disclosure anchor key, so the trigger stays under the pointer instead of the viewport chasing the end. - ChatView's hand-rolled streaming follow (double-rAF scrollToEnd on every data change) is gone; the app now only owns streaming adjustments during anchored-end-space mode, mirroring main. - timelineLiveFollowEnabled state mirrors the follow refs so the render-visible gate switches native follow off when a gesture breaks follow and back on when the viewport returns to the end band. Timeline tests updated to assert the native-ownership invariants. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): show Git action success inline in panel - Keep success feedback visible in the Git action control for 10 seconds - Move the running elapsed timer into the panel menu slot * fix: repair rerere-damaged files and reconcile main's round-3 features with v2 Post-rebase reconciliation sweep: - Sidebar: main's folded Sidebar.tsx/Sidebar.logic.ts adapted to v2 shells (latestRun/runtime naming, waiting status instead of monitoring), with subagent-thread filtering and main's pinned-reorder helpers re-exported - Pinned drag reorder (#5581) ported into v2: thread.pin orderKey + thread.pin.reorder command, thread.pin-reordered event, Orchestrator fold, ProjectionStore/Maintenance, client-runtime commands and shell mapping - Project favicon (#4849-era) and defaultThreadEnvMode flowed through v2 contracts (OrchestrationProjectShell, application event payloads) - ChatView: main's #5592 header props, pull-request right-panel surfaces, liveAgentCount badge (#5745) wired into the v2 panel layout - enableAssistantStreaming -> enableLegacyTokenStreaming rename applied to v2 RunExecutionService and replay testkit - Removed v1 zombies resurrected by the rebase (provider service/reaper/ ingestion + v1 layer tests, server.test.ts, integration harness) - routeTree: main's tree + branch's /settings/scheduled-tasks route - Misc marker-sweep syntax repairs (rpc.ts, entities.ts, localApi.test.ts, rightPanelStore.test.ts, GitManager.test.ts, mobile model menu helpers) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * chore(server): renumber v2 migrations 038-046 to 041-049 after main's 038-040 Main released ProjectionThreadsPinOrderKey (038), ProjectionProjectsDefaultThreadEnvMode (039) and ProjectionProjectFaviconPath (040), so the branch-private v2 stack shifts up by three. Registry ids were already 41-49; this renames the files and identifiers to match and updates the ledger expectations and through-id boundaries in the migration tests (released boundary 37 -> 40). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(server): port round-3 main fixes into the v2 orchestrator - a6c9b41f90 (agents open pasted images): ClaudeAdapterV2 now grants the attachments dir alongside cwd via additionalDirectories and appends '[Attached ... is saved at: path]' lines to the turn text so tools can dereference pasted images (pixels alone are not tool-readable). - 5bb8c03664 (settle leaves monitors running): thread.settle now joins archive/delete in the provider-session detach set, so PR monitors, dev servers and subagent fleets stop when the user parks the thread. The settle guard already rejects active runs, and serialized dispatch closes the re-engage race the v1 fix handled with onlyIfSettled. - e70cdb478d (Claude resume handshakes) and 2c7267ad43 (reaper vs live background subagents) are already covered structurally in v2: results are turn-scoped with explicit zero-turn handshake drops, and idle release is pinned while background work is pending. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * test(web): restore main's right-panel migration expectations after the panel-visibility merge The keep-both merge nested main's plan-surface migration test inside a branch popover test and dropped the threadPanelVisibilityByThreadKey key from the migration results. Restore main's test body and include the branch's (empty) visibility map in the expected persisted shape. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * test(server): expect attachment saved-at lines in ClaudeAdapterV2 turn text Follow-up to the #5757 port: start and steer turns now append the '[Attached ... is saved at: path]' line, so the adapter fixtures assert it. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): restore the branch's slim chat header The round-3 reconciliation took main's ChatHeader wholesale and wired its full prop set, resurrecting the scripts/open-in/git-actions cluster the branch had deliberately relocated into the thread panel. Restore the 79-line slim header (project favicon + name + thread title) and its minimal ChatView call. #5592's header actions stay a documented v2 follow-up, as decided in round 2. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(web): add pull request actions to thread details - Add ready, merge, and conflict-resolution actions to the PR row - Share pull request action and handoff logic with the detail panel - Fix thread details scrolling and row alignment * fix(mobile): port main's composer stabilization into the v2 thread screens Round-4 reconciliation of #5986/#5988 with the v2 cutover files: - PendingUserInputCard adopts main's collapsible overlay redesign with the v2 RuntimeRequestId/responseCapability plumbing (dead provider processes still read-only the card) - ThreadFeed adopts the thread-feed-live-follow transition model, the user-scroll settle window, momentum handoff, and env-scoped feed keys while keeping the v2 nearListEnd layout gating - ThreadDetailScreen hides (not unmounts) the composer while a user-input request owns the slot; multi-select answers flow as arrays end to end (threadActivity toggle/build helpers + tests, ThreadUserInputQuestion) - ThreadComposer keeps the v2 canStopThread stop gate under main's onEditorFocusChange rename; standalone stop reuses the shared renderer - Restored the branch's steer stop/send tests alongside main's composer test suite Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): reconcile main's round-5 features after the rebase - PullRequestDetailPanel takes main's #6039 rework wholesale (reactions, update-branch, auto-merge, in-place editing); the thread-details action hooks stay in usePullRequestActions with label maps extended for the new action variants - CommandPalette #6330 provider subtitles adapted to the v2 shell (session -> runtime for provider instance and display name) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(web): prioritize pull request row actions - Add conflict, draft, failing-check, and merge action ranking - Show check progress and pull request details in row tooltips * fix(web): restore main's collapse chrome and tab-status keying on the PR panel The round-5 ChatView reconciliation kept the round-3-era PullRequestDetailPanel call, so the panel mounted without chromeVariant="collapse" — the #6039 scroll-condensing chrome never engaged and the description scrolled under a full-height chrome. Restore main's call exactly: collapse chrome, composerDraftTarget for same-thread hand-offs, the isThreadOwnPullRequest context check, and tab statuses keyed by the active surface id via updatePullRequestTabStatus instead of a key rebuilt from the status payload. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(orchestration): bound thread history and resume payloads - Add paginated thread history with bounded snapshots and replay limits - Trim oversized wire payloads and support progressive mobile history loading * feat(contracts): track thread title regeneration - Add optional title regeneration request and start time to thread shells - Cover cache serialization on mobile and client runtime * fix: reconcile main's round-6 features after the rebase - ChatView: #5880 auto-settle-on-merge setting flows into effectiveSettled, #5644 browser favicon project registration effect, activeProjectRef memo, desktopByTabId on both RightPanelTabs mounts - server: provide ServerSecretStore to the McpSessionRegistry's ServerEnvironment layer (#6325 reads publish opt-in per descriptor) - mobile: 3-way merged main's deltas into the v2 thread screens (NewTaskDraftScreen keeps the branch title seed + main's environmentId, threadListV2 keeps both new test suites, queries imports deduped) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): keep the titlebar layout controls fixed across right-panel toggles Restores main's one-inset rule (#5226) that a rebase resolution had overridden with a conditional right-2 offset, which made the controls jump sideways whenever the right panel opened. Also restores the live-agent count badge on the right-panel toggle (#5745) that the round-6 replay dropped, and applies the same fixed-position rule to the pull requests page: the toggle now stays mounted at one absolute inset in both states, with a footprint spacer in the list header so the refresh button never slides underneath it. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): align titlebar clusters to one shared pixel inset The right controls carry mr-px (main's border compensation for anchoring inside the panel frame), which left the sidebar trigger one pixel closer to its edge and the sheet-mode tab bar one pixel tighter than the closed state. Mirror the pixel on the trigger and the sheet layout-controls slot so all three read the same inset. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): size the titlebar layout-control icons like the sidebar trigger The trigger's icon falls through to the Button default (size-4) while the right cluster hard-coded size-3.5, so the two ends of the titlebar read a pixel apart on every edge. All five layout-control icons now use size-4, matching the trigger and the pull requests page's refresh icon. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * perf(server): keep shell snapshots bounded and active-only - Omit transcript bodies from shell rows - Query archived threads separately and stream compact resume metadata * fix: reconcile main's round-8 features after the rebase Re-applies the deltas that mid-stack blob reverts discarded, and merges main's work into the v2-owned surfaces: - keybindings: main's STATIC_KEYBINDING_COMMANDS rename plus both new commands (rightPanel.toggleMaximized alongside threadPanel.toggle) - OpenInPicker: main's remote-open/SSH routing and favorite-editor shortcut layered onto the branch's panel/toolbar variants; the extracted shouldShowOpenInPicker now takes remoteOpenMode - ChatMarkdown: main's bare-filename resolver (#6297) ported into the branch's module-level component factory, plus #4133 title-attribute stripping on links and images - ComposerPrimaryActions: main's #4781 model (stop stays reachable, send joins it when Enter-to-send is unavailable) carrying the branch's steering send button - ComposerPendingUserInputPanel: main's collapsible redesign with the v2 RuntimeRequestId and responseCapability gate - ChatComposer: main's oversized-prompt submission guard wrapping the branch's dispatch-mode send - preview shell: main's container-aware width clamp ported into the branch's usePreviewPanelInlineSize hook - MessagesTimeline/Sidebar: main's day-aware timestamps, code-font tool bodies and provider accent badges on the v2 runtime shell - index.css: main's @variant dark migration (#6381) replaces the branch's standalone .dark block - contracts: main's send-turn image mime allowlist re-homed to chatAttachment.ts, where v2 keeps the other send-turn limits Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(server): reject replaying a command receipt across threads in v2 Ports v1's #5246 guard into the v2 dispatcher: a stored receipt only proves that this exact command already ran for the thread it was recorded against, so returning it for a command aimed at a different thread reports success for work that never happened there. The check is extracted as canReplayCommandReceipt so the rule is unit-testable, and reuse now fails with OrchestratorCommandIdConflictError like the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(mobile): surface prominent activity status and metadata - Keep prominent activity rows visible with lifecycle status and provider metadata - Move feed sizing logic into tested helpers and preserve native measurement for activity groups * fix: reconcile main's round-9 features after the rebase Re-applies the deltas that mid-stack blob reverts discarded, and merges main's round-9 work into the v2-owned surfaces: - settings: main's Integrations page (#7082) coexists with the branch's Scheduled Tasks page in the path union, section labels, icons, and search catalog - contracts: main's preview appearance/zoom/viewport settings imports restored beside the branch's modelSelection home for ModelSelection - mobile: main's built-in themes (#6619) re-applied to the v2 thread screens and work log (useThemeColor over hand-rolled color-scheme ternaries) - MessagesTimeline: main's #7157 cleanup adopted (toolCallExpandedBody class name unexported, implementation-detail test dropped) - ChangedFilesTree: main's styled tooltip (#7209) carrying the v2 runId - pullRequestDetail tests: branch's row-action coverage renamed onto main's buildAddSelectionToAgentHandoff (#6597) - lint: migrated the six branch-owned native title tooltips that main's new no-native-title-tooltip rule (#7209) flags to styled Tooltips (GitActionsControl, QueuedRunsControl, TimelineSystemDivider, MessagesTimeline intent badge and MCP tool logo) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(server): honor withheld agent browser access in the v2 runtime Ports #7083 into the v2 session path, which replaced the v1 ProviderService where main's gate lives. Instead of withholding the whole t3-code MCP credential — on this branch it also carries the thread orchestration and worktree toolkits — the credential is minted without the "preview" capability when enableAgentBrowserAccess is off, so every preview tool call rejects while orchestration stays available. ProviderSessionManager reads the setting at prepare time (deny on an unreadable settings file, matching main), rotates a reused credential whose capability set no longer reflects the setting, and the session config now carries browserToolsAvailable so the Codex adapter keeps its developer instructions truthful via main's parameterized instruction builders instead of the removed constants. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): restore the titlebar sizing and timeline fade lost to main's style simplification Main's #6381 deleted the shared .workspace-topbar and scroll-fade rules from index.css after inlining them at main's own call sites, but this branch's slim chat chrome still references both classes. The round-8 rebase took the deletion without migrating the branch call sites, so the header collapsed to zero height — the breadcrumb sat on the window edge, timeline rows scrolled unfaded through it, and the thread-details popover anchored to the collapsed header. Restores both as composable utilities in #6381's own style: a workspace-topbar utility for the titlebar rows, and the branch's chat-timeline-scroll-fade mask (soft ramp plus a full-height scrollbar column). Also drops the duplicated media override and its dead settings-page-scroll-fade selector. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * refactor(web): finish aligning the branch with main's style simplification A follow-up sweep against #6381 found the branch still carrying the pre-simplification forms it replaced, which my earlier fix had papered over with a compat utility instead of finishing the migration: - ChatView now uses main's inlined titlebar sizing and the data-workspace-titlebar-controls hook on both control clusters. The class-based markup was silently missing the themed-toggle bridge (html[data-theme-id] [data-workspace-titlebar-controls] …), so custom themes lost their titlebar accent in the thread view. - The scroll-to-end pill becomes main's Button size="xs" variant="glass" instead of a hand-rolled button recreating it. - MessagesTimeline uses main's consolidated topbar-scroll-fade utility; the byte-identical chat-timeline-scroll-fade copy and the workspace-topbar compat utility are gone. - The composer-glass dark rules move into nested @variant dark like main's (the raw .dark duplicates could drift from the nested copies they shadowed), including the branch-only queue strip. - The pre-#6381 dialog-glass/dialog-backdrop/dropdown-glass class rules and their .dark variants are deleted: the #6381 utilities plus call-site shadow utilities own every declaration, and the stale dropdown rule still had the saturate-less backdrop-filter. The dead model-picker-surface dark rule goes with them. index.css now has zero raw .dark selectors outside the variant definitions, matching the doctrine in .macroscope/check-run-agents/ui-consistency.md. Verified against the emitted production CSS: dark variants compile to :is(.dark,.dark *) with their @supports color-mix fallbacks intact. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(orchestration): show provider retries in the work log - Complete retry items when provider activity resumes - Keep retry progress visible across web and mobile clients * fix: reconcile main's round-10 features after the rebase Ten days of main (194 commits, 640 files) re-integrated with the v2 re-architecture. The headline mechanics: - migrations: main added 041-043 (AuthSessionClientConnection, ProjectionThreadLinkedPullRequest, ProjectionThreadsUnsettledAt), so the v2 block renumbers 041-049 → 044-052 with the migration tests shifted to match - contracts: OrchestrationClientOrigin (#7774) and the origin metadata field live in applicationEvent.ts and re-export through the legacy path; ProviderApprovalOption + acceptAlways + mcp-elicitation (#8058) land in providerPolicy.ts; OrchestrationDispatchCommandError (#8824) added; the send-turn image-mime home stays chatAttachment.ts - threadSettled: main's settle-once-on-merge semantics (#7454) and un-settle re-anchor (#8231) hand-merged onto the v2 duck-typed shells (latestRun ?? latestTurn reads); web callers pass the new ChangeRequestSettleSource shape - timeline anchoring: main's #7897 (follow-up sends no longer push to the top) ported by scanning user rows only; the branch test now encodes the new semantics, as does mobile's #7969 settled-pinned shelf behavior - vcs: branch's deleteLocalBranch coexists with main's pruneWorktrees and the #7674 submodule checkout tests - ws: v2 RPC surface keeps its dispatch path; main's attachment upload RPCs (#8048) and client-connection analytics recording are wired; providerUploadFeedback (#7949) fails explicitly pending a v2 route - approvals: main's option-driven approval buttons (#8058) render through the v2 canRespond gate on web and mobile - ChatView/ChatComposer/MessagesTimeline/Sidebar/session-logic/ threadActivity keep the branch's v2 architecture; main's v1-coupled deltas to those files are recorded for follow-up rather than force-fitted Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): stop mis-marking recovered and text-reported tool failures in the v2 work log Ports main's #7999/#7893 failure policy onto the v2 turn-item work log: output text that reports a failure (command not found, ENOENT, nonzero exit markers) now flags a row even when the provider item completed "successfully", while the rendered row judges only its displayed result — a command that merely greps for failure strings stays calm. Success now also requires the failure check to pass, so recovered failures no longer get the blue check. The server half of #7893 needs no port: CodexAdapterV2 already projects item.status directly, so a failed item never masquerades as completed. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(orchestration-v2): project linked pull requests on threads (#8160) Main's thread↔PR linking never reached the v2 runtime: the client types were optional stubs and the v2 server dropped the field, so linking a PR on a v2 environment silently no-opped and #7454's settle-once logic could never match the linked identity. The link now flows end to end: thread.metadata.update carries an optional linkedPullRequest (object to link, null to unlink), the orchestrator folds it into thread state, and both shell builders project it — no migration needed since v2 shells persist as payload JSON. The client command sends the field and the shell mapper surfaces it, so the existing web/mobile badge and settle plumbing light up on v2 threads. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(orchestration-v2): carry approval options and app names to the client (#8058) Round 10 ported main's option-driven approval buttons, but v2 runtime requests had no way to deliver the data — every approval rendered the default button set, MCP app names never showed, and worse, the Codex app-server's mcpServer/elicitation/request went entirely unhandled on the v2 adapter, so ChatGPT-app access requests could never be answered. The v2 approval_request turn item now carries optional appName and options, the client derivation passes them into ThreadPendingApproval, and CodexAdapterV2 handles mcpServer/elicitation/request end to end: unsupported shapes decline immediately (mirroring the v1 runtime), and supported ones surface a mcp-elicitation approval built from the shared describeMcpElicitation/toMcpElicitationResponse helpers, so the persist tiers (session / always) advertise exactly the choices the elicitation can express. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(orchestration-v2): route Codex thread feedback uploads through v2 (#7949) The round-10 rebase stubbed providerUploadFeedback to an explicit error because its v1 ProviderService route died with the v2 rewrite. The route now goes through the v2 runtime: session runtimes may expose an optional uploadFeedback capability, the Codex adapter implements it against the app-server's feedback/upload request, and the WS handler resolves the thread's live provider session through ProviderSessionManagerV2 — failing with a plain-language reason when no session has run, the session is gone, or the driver has no feedback channel. This also un-blocks the dormant mobile feedback UI. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(analytics): credit v2 threads and turns to the starting client (#7774) Main records which client surface started each thread and turn; the v2 dispatch path replaced the v1 handler that did the recording, so v2 environments only reported connections. The v2 RPC layer now records client.thread.started on thread launches (plus client.turn.requested when the launch carries an initial message) and client.turn.requested on message dispatches, using the connection's announced origin. Recording is best-effort — attribution can never fail the user's command. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(grok): fail hung prompts on xAI rate-limit completions (#8358, partial) Ports the rate-limit half of main's #8358 into the branch's reworked XAi extension: a prompt_complete carrying stopReason rate_limit now fails the hung prompt with the -32003 usage-limit error instead of settling it as a normal end_turn, so the turn surfaces "usage limit reached" rather than silently ending. The prompt-completion deferreds carry the error channel end to end. The exit_plan_mode approval gate from #8358 is NOT ported here: it needs a v2 plan-flow design in AcpAdapterV2 (the v1 GrokAdapter it lived in is gone) and is tracked separately. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(orchestration-v2): show live context usage in the meter (#8144) The v2 context meter could only show token counts after a compaction had already happened — v2 had no live usage plumbing at all, so main's compaction-threshold UX was invisible on v2 threads. Provider turns now carry an optional tokenUsage report: the Codex adapter maps the app-server's thread/tokenUsage/updated notification (total breakdown + model context window) onto the active provider turn, ChatView picks the newest report out of the projection, and the meter prefers it over the compaction fallback — so usage and remaining-context percentages update while the turn runs. Claude's v2 adapter does not report usage yet; its meter falls back to compaction items as before. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): keep following the stream after returning to the live edge (#6519) Ports main's anchor-release semantics onto the branch's timeline anchor state: the scroll-to-end pill and a manual scroll back to the live edge both drop the send-time anchored end space before re-enabling follow, and the pill's scroll runs a frame later so the list measures without the anchor space and lands on the true end. Without this the timeline could settle into following-end with the anchor still installed — following nothing. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(grok): capture exit_plan_mode into the v2 proposed-plan card (#8358) Completes the deferred half of #8358: Grok's plan-approval gate now works on the v2 runtime. The XAi extension regains main's exit-plan helpers (request schemas, plan-markdown extraction, the abandoned-with-feedback response, and the plan.md session-path sniffing), and the v2 ACP adapter grows a captureProposedPlan primitive that emits a completed proposed-plan artifact for the active turn — one plan id per turn, so plan.md rewrites and the exit gate update a single card. The Grok flavor wires both ends: tool calls that write plan.md under a Grok session dir surface the plan while plan mode is still active, and x.ai/exit_plan_mode (plus the underscore alias) captures the final plan — request content first, then the sniffed plan.md contents, then the empty-state placeholder — and abandons the native gate so the turn does not hang, mirroring the Claude ExitPlanMode pattern. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): repaint the composer glass and strip the thread-panel popover chrome Two post-rebase regressions from the round-10 index.css merge: - The merge seam ate the closing brace of .chat-composer-glass, silently nesting the entire composer-glass section (shell, host, context strip, shoulder tab, banner cap) inside it as descendant rules that never matched — the composer surface stopped painting and thread content showed straight through the input. The brace is restored and the compensating over-close removed; every composer selector emits at top level again. - The thread-details popover grew dropdown-glass card chrome around the panel: round 9 deleted the legacy components-layer .dropdown-glass rule in favor of the @utility, which the popover's border-0/bg-transparent suppressors no longer outrank (the utility emits later in the layer). The suppressors are now important variants, matching the !overflow override already there. Verified against the emitted production CSS: shell::before is top-level, no descendant-of-glass selectors remain, and the important suppressors (including [backdrop-filter:none]!) are emitted. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): adopt main's attached-composer surface contract so the glass survives shoulder tabs The composer went frameless exactly when the stash or tasks badge was showing: main's #7150 css hides the classic shell chrome whenever the shell :has() a shoulder tab or top drawer and repaints the glass on [data-chat-composer-main-surface] instead — an element the branch's composer body never rendered, since it predates the drawer system while the badges and css came through the rebase in main's new form. The branch composer's frame div now carries the main-surface attribute (with main's relative z-10 stacking) so attached mode paints background, outline, and backdrop on it and the tab connects to the surface, and ChatView applies chat-composer-glass-shell-attached while banner items render in the drawer slot, matching main's externalComposerDrawerAttached wiring. Without a tab or banner the attribute is inert and the classic shell chrome paints as before. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): converge ChatComposer on main's drawer-era body Round 10 restored the branch's pre-#7150 composer body while its satellites (shoulder badges, banner drawers, glass css) arrived in main's new form, and every seam between the two was a visible bug: opaque/frameless composer, overflowing stash tab, detached stash menu. Rebuilt ChatComposer via a reverse three-way merge (main's body as the base, branch delta re-applied): dispatchMode send boundary, live-capable approval gates, latestRun reads, and the v2 context-window meter stay; everything else now matches main, including ComposerPrimaryActions and the sendDisabledReason send gating. Attachment uploads stay off until the v2 claim path lands. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): collapse settled tool runs behind main's summary toggles The v2 timeline rendered every tool call as its own raw row; main's tool-group collapsing (generated "Ran N commands and changed M files" summaries, the live "Running <program>" pill for the active tool run, and the "+N previous log entries" clamp for mixed groups) never made it into the v2 row model. Ported the work-live/work-toggle row kinds and group summarization into the v2 derive, keyed on v2 item types (command_execution, file_change, file_search, dynamic_tool, subagent) and runId lifecycle instead of v1 activities. Expanded groups keep the branch's richer per-entry detail rows (V2ItemInspector) — only the collapsed presentation converges on main. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): surface v2 todo-list plans as task progress v2 already projected todo_list plan artifacts (deriveActivePlanState existed with zero consumers), but nothing rendered them: todo_list turn items showed as a bare "Updated tasks" work row, the composer Tasks drawer never appeared, and the working row had no current-step label. todo_list items now become inline turn-plan chips (mini step segments, current step, N/M count, expandable step list) that fold with their settled turn, ChatView derives the composer Tasks drawer progress and steps from the running run's plan artifact, and the working row shows "Working for Xs · <current step>" like main. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): show the command on collapsed tool rows, not its stdout Collapsed command rows rendered "Command" plus the raw result JSON as the preview; the command itself is the useful collapsed line, so it now renders as the row text (whitespace-collapsed, truncated) with stdout and the full payload behind the expander. Tool-like headings drop the bold foreground for the muted secondary-label the summary rows use, and the "+N previous tool calls" toggle loses its bold black label for the same muted treatment. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): collapsed tool rows preview inputs for every tool type Extends the command-row fix to the whole preview: file-change rows were still leaking raw diff lines into the collapsed line. workEntryPreview now resolves input-first — command, then touched-file paths, then detail (which is input for the remaining types: search patterns, reasoning text, error messages) — so outputs only appear behind the expander. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: reconcile main's round-11 features after the rebase Round-11 rebase onto main (25 commits). Reverse-merged main's new work into the v2 cutover files: #8395 muted ordinary-tool-failure treatment (v2-adapted workEntrySignalsSevereFailure keyed on error items), #5931 sidebar project-filter combobox + #4c51 keyboard pin/settle with their ChatView support graph, the auto-settle-mode migration through threadSettled/threadListV2, #8235 file/unknown attachment schemas moved into chatAttachment.ts with nullable attachment paths, #8481 client analytics through the v2 ws layer, #8480 OpenCode server owner wired into the driver beside the v2 orchestration adapter, and the mobile semantic-theme migration applied to branch-only components. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(opencode): route child-session approvals through the v2 adapter Ports the v2-applicable half of #8480 (the v1 adapter rewrite was not carried; this branch's OpenCode path is OpenCodeAdapterV2). Permission and question asks from child sessions — task subagents and their descendants — were dropped because the adapter only looked up root thread sessions. Related sessions now map back to the owning root state (registered from task parts and session.created/updated parent chains), and an ask that arrives before the relation is known resolves it inline via session.get with a short forked backoff, then surfaces the approval on the root turn. Replies already route by native request id. Interrupts now tolerate the abort racing turn settlement instead of failing the stop. Covered by a new opencode_child_approval replay fixture where the child asks for bash permission before the task part reveals the relation. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: reconcile main's round-12 features after the rebase Small rebase onto main (4 commits). The composer stash-shortcut label and the mobile start-task menu refactor merged onto the v2 composer bodies, and main's new auto-settle list tests are ported to the v2 thread-list test file (latestRun/RunId shapes). The auto-settle machinery itself already matched main from the round-11 reconcile. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: restore main's automatic thread settling after the revert Round-13 rebase onto main (2 commits). Main reverted the auto-settle opt-in (#8596 undoing #8321), so the branch drops the ported autoSettleMode machinery and returns to sidebarAutoSettleOnMerge with settling-by-default, keeping only the v2 shell-shape delta in threadSettled. The unpin confirmation (#7313) merges into the v2 thread-actions hook, and the mobile list tests re-sync to main's reverted semantics in v2 shapes. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): restore the full-screen file-drop target over the chat column Main's #6636 workspace file drop (drag anywhere over the chat column to attach, with the dashed overlay) split across ChatView and the composer. The round-10 rebase restored ChatView from the pre-#6636 backup wholesale and only the composer half was ever re-applied, so the drop target, overlay, and drag-state plumbing vanished while addDroppedFiles sat unused on the composer handle. Re-applies main's ChatView half verbatim; the shared workspaceFileDrop module was already identical to main. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(server): claim uploaded attachments at v2 dispatch Closes the last gap from the rebase follow-up list (#8048/#8161 for v2): pending uploads staged via the attachment upload URL flow were never claimed by the v2 orchestrator, so the composer kept supportsAttachmentUploads off and fell back to inline dataUrls with no progress UI and no big-file support. message.dispatch and thread.launch now claim pending refs at intake — verify the staged file, copy it under a thread-scoped id (the pending copy stays as the retry source), rewrite the refs, and release the claimed copies if the dispatch fails. A launch carrying uploads requires its thread id up front. The web composer reads the attachmentUploads capability again like main, which lights up the upload progress overlay, retry-on-failure, and PDF/ZIP attachments. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(web): show attachments on queued messages and edit them in the composer Queued rows now render image thumbnails, and the pencil action loads the queued message into the composer instead of an inline input: text and stored attachments are editable (attachments removable, new images addable), sending saves the queued run in place, and the user's own draft is stashed untouched for the duration. queued-run.edit gains an optional full-replacement attachments list end to end. Built by Claude Fable 5 on Claude Code. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): drag-to-reorder queued messages and retire stale pending rows Replaces the queue rows' up/down arrow buttons with a drag handle (arrow keys still work on the focused handle). Also prunes optimistic queued messages once the projection holds them: keying the prune on turn items alone left a phantom clock row behind whenever a queued run was removed or steered before it ever started. Built by Claude Fable 5 on Claude Code. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: reconcile main's round-14 features after the rebase Round-14 rebase onto main (7 commits, mostly the Expo SDK 57 upgrade plus the mobile glass restore and codex app-server buffering fix). The mobile composer merged main's restored glass chrome onto the v2 body, and the lockfile is regenerated from main's SDK-57 lock with the branch's extra dependencies installed on top. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): dedupe the composer glass styles and align the chat column width The rebase left a stale early copy of the composer glass-host, context-strip, and shape() fallback rules that the identical later block always overrode. The composer shell and queue/context strips also kept main's 48rem width while the timeline moved to the 46rem content lane; they now share --chat-content-max-width so one owner defines the chat column. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(mobile): replace remaining dark: utilities with adaptive semantic tokens The v2 thread surfaces still styled borders, fills, and status text with dark:/light: pairs, which do not follow registered custom themes and now fail the no-mobile-uniwind-theme-escape-hatches lint. Convert them to adaptive tokens, adding the missing amber/sky badge and neutral hairline/fill entries to the theme generator. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(lint): allowlist the queue and relationships interop boundaries ThreadQueueControl and ThreadRelationshipsBanner read theme variables only to tint SymbolView icons and color native modal chrome, the same reviewed interop pattern as the existing thread-feed entries. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(server): inject HostProcessPlatform into the Grok plan extractor The plan.md path check read process.platform and process.env directly; thread the host platform reference and the adapter's provider environment through GrokAdapterV2Options instead. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * chore: retrigger ci The pull_request workflows never fired for 6c3b84bbfc; only the pull_request_target ones ran. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * chore: refresh macroscope ui-consistency check Its findings were fixed in 9abca06b28 and the review threads are resolved; the check only re-evaluates on push. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: reconcile main's round-16 features after the rebase Round-16 rebase onto main (5 commits, headlined by #8236 file attachments in the client). Main's upload-aware send path — capability probes, upload await/retry gating, uploaded-ref turn attachments with dataUrl fallback, draft release on success — is woven into the v2 dispatch flow, timeline user rows render file attachments as download links with the ChatView download handler, and the provider settings editor cleanup keeps the branch's environment-field rows. The codex feedback client flow stays unported, replay testkit configs gained main's environmentThemesDir, and the rpc/settings/docs unions carry both sides. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(web): keep failed tool items in the collapsed group summaries An ordinary exit-code failure knocked its whole tool group out of the "Ran N commands" summary (and the live pill mid-run) into the raw "+N previous tool calls" clamp: v2's derived tone marked any status=failed item as "error", which the grouping treats as a non-tool row. That inverts v1's semantics and #8395's muted-failure rule — the failed lifecycle status already carries the X marker and the summary's includes-a-failure hint, so the tone override goes. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(server): keep Claude session approvals ephemeral R03: Rescope Claude permission suggestions to the current session and add a whole-tool session rule when the SDK provides no suggestion. * fix(orchestration): reanchor unsettled threads Carry unsettledAt through v2 thread state and shells, stamp explicit and activity-driven reactivation, and clear the stamp when settling. Finding: R11 Implemented by GPT-5.6 Sol in Codex. * fix(server): observe pre-aborted Claude approvals R15: Race Claude approval decisions with cancellation while checking already-aborted signals and cleaning up the listener on every exit. * fix(server): include service launcher in bundle build Build the service launcher after the clean CLI pack so publish and background-service installation receive the required artifact. Audit: R01 * fix(orchestration): preserve legacy thread metadata Import pin order, snooze state, unsettle time, and linked pull requests. Repair prior imports only where the stored v2 property is absent, so later v2 changes remain authoritative. Finding: R08 Implemented by GPT-5.6 Sol in Codex. * fix(web): honor disabled legacy plan mode Force the effective composer interaction mode to build when the legacy plan setting is off, including threads and drafts with a saved plan mode. Audit: R13 * fix(server): preserve Claude subagent models R16: Apply authoritative assistant snapshot models to Claude subagents and buffer snapshots that arrive before task_started. * fix(orchestration): honor migrated thread visibility in search Use v2 ownership and lifecycle metadata when a legacy transcript belongs to a migrated thread. Legacy transcript rows remain searchable until lazy hydration finishes. Finding: R09 Implemented by GPT-5.6 Sol in Codex. * fix(orchestration): recreate missing worktrees before turns Prune stale git worktree registration and recreate the saved branch at the saved path before provider startup. Recovery remains best effort so normal provider errors still report when repair is impossible. Finding: R10 Implemented by GPT-5.6 Sol in Codex. * fix(clients): restore Codex feedback submission Intercept /feedback in web and mobile, show the upload result and feedback ID in the thread, and block duplicate submissions while an upload is active. Audit: R12 * fix(server): preserve generic provider attachments R02: Append persisted paths for every uploaded file on provider sends and steering while reserving native image payloads for supported images. * fix(web): load workspace markdown images through assets Classify markdown image sources and request environment-scoped asset URLs for workspace files while leaving ordinary web images direct. Audit: R17 * fix(web): preserve Windows markdown paths Normalize drive-path links and image sources before sanitization so file chips and signed workspace images receive usable paths. Audit: R18 * fix(server): keep current provider context usage R07: Project Codex last-turn and Claude assistant context usage, and retain the latest usage when terminal provider-turn updates omit it. * fix(web): restore markdown file chip actions Keep ordinary file-chip clicks in the in-app preview while restoring modifier-click editor opening, configured editor labels, remote gating, and reveal-in-file-manager actions. Audit: R19 * fix(web): scope markdown actions to their environment Use the owning thread or pull request environment for editor, shell, and remote-open actions instead of whichever environment is active. Add a multi-environment regression test for the action hooks. Audit: R20 Implemented by GPT-5.6 Sol with Codex. * fix(protocol): reject incompatible orchestration peers Advertise and validate an explicit orchestration protocol before clients open RPC sessions. Announce the same protocol on WebSocket upgrades so hosts reject older clients before request decoding while preserving existing auth and relay parameters. Audit: D03 Implemented by GPT-5.6 Sol with Codex. * docs: explain legacy thread migration Document which thread metadata and transcript data migrate, which runtime history does not, and how the fresh provider continuation uses the latest 32,000 characters. Add a safe read-only recovery procedure without claiming an export API. Audit: D02 Written by GPT-5.6 Sol with Codex. * docs: state portable handoff limits Explain the eligible timeline items, whitespace-normalized 240-character prefixes, omitted tail risk, and practical preparation for provider or fork handoffs. Distinguish this rule from the legacy import's 32,000-character transcript suffix. Audit: D04 Written by GPT-5.6 Sol with Codex. * chore(repo): remove tracked audit scratch files Remove obsolete implementation plans and the probe write marker so temporary work artifacts no longer ship with the repository. Audit: H01 Implemented by GPT-5.6 Sol with Codex. * fix(server): guard OpenCode prompt admission races R14: Hold idle completion through prompt admission, reconcile status only for the current admission generation, and invalidate admission before abort. * fix(server): restore Claude structured questions Project AskUserQuestion as a structured user-input runtime request and return keyed answers to the Claude SDK instead of routing the tool through generic approval. Finding: R04 Model: GPT-5.6 Sol via Codex * fix(server): project Claude plans and todos Translate TodoWrite and ExitPlanMode tool input into canonical todo-list and proposed-plan artifacts so every client can render Claude planning state. Finding: R05 Model: GPT-5.6 Sol via Codex * perf(orchestration): bound history reads in SQL Load at most one turn-item page per thread in a fork lineage before decoding, keyed by the stable history cursor. Restrict message, plan, and handoff reads to that page plus live actionable state so cold opens and older-page requests no longer decode complete historical tables. Finding: P01 Implemented by GPT-5.6 Sol in Codex. * perf(orchestration): bound complete thread snapshots Budget the serialized bounded projection after retaining live control state. Cap historical control arrays and large plan or handoff details only on the bounded route; the full thread-detail route remains available for complete text. Finding: P02 Implemented by GPT-5.6 Sol in Codex. * fix(server): restore Claude resume compaction Pass the automatic compaction window to Claude and route resume-return dialogs through structured user input so users can compact, continue, or permanently dismiss the prompt. Finding: R06 Model: GPT-5.6 Sol via Codex * fix(server): allow protocol negotiation in CORS Permit the canonical orchestration protocol header in browser API preflights so cross-origin web and desktop clients can negotiate compatibility while retaining authorization and DPoP headers. Finding: D03 Model: GPT-5.6 Sol via Codex * fix(server): preserve provider usage in persisted turns Merge terminal provider updates with stored context usage before replacing the SQLite payload. Keep newer usage reports authoritative and verify the persisted projection after reload. Finding: R07 follow-up Model: GPT-5.6 Sol via Codex * fix(server): preserve Claude planning lifecycle Keep typed plan and todo records distinct from generic tool events, activate captured plans, and supersede older planning state within the owning thread. Ignore nested todo snapshots for the parent and retain identity across duplicate SDK messages. Finding: R05 follow-up Model: GPT-5.6 Sol via Codex * fix(server): normalize Claude question answers R04 follow-up Convert client multi-select answer arrays to the comma-separated string shape required by the pinned Claude SDK while preserving single-select strings. Implemented by GPT-5.6 Sol via Codex. * fix(server): correlate OpenCode prompt admission Stale cached user and status events could admit and complete a newly submitted OpenCode prompt. Generate the native message ID before submission and only advance admission when that exact message is observed. Finding: R14 Implemented by GPT-5.6 Sol with Codex. * fix(clients): anchor feedback in conversation order R12 follow-up Insert persistent feedback blocks by their timestamp within the canonical timeline while preserving projected row order. Keep real optimistic sends appended and suppress duplicate local messages already committed by the server. Implemented by GPT-5.6 Sol via Codex. * fix(web): retain markdown workspace ownership R20 follow-up Give inspector reasoning markdown its projected source thread and retain the explicit environment fallback for proposed plans without a thread reference. Workspace links and images now resolve through their owning environment after removal of the active-environment fallback. Implemented by GPT-5.6 Sol via Codex. * fix(orchestration): page history through its true end Read the inclusive cursor, a full history page, and a look-behind row so older history does not terminate after one page. Finding: P01 pagination termination Model: GPT-5.6 Sol via Codex * fix(server): cancel pending OpenCode prompts safely Cancel pending SDK requests before aborting the native session. Preserve per-admission cancellation state and treat stopped initial prompts as interruption instead of provider failure. Finding: R14 prompt cancellation Model: GPT-5.6 Sol via Codex * fix(orchestration): retain nested fork history when paging Keep the original cursor owner through ancestor traversal and preserve the history budget across empty intermediate forks. Verify exact paged history against the complete nested projection. Finding: P01 nested lineage Model: GPT-5.6 Sol via Codex * fix(server): recover OpenCode status reconciliation Retain pending admission after transient status failures and use one generation-owned retry worker. Ignore stale timers and duplicate evidence so older prompts cannot finish newer steering. Finding: R14 status reconciliation Model: GPT-5.6 Sol via Codex * fix(orchestration): select visible history before limiting SQL Keep hidden local and inherited rows from consuming history pages. Preserve stop-request dependencies, source-run cutoffs, and imported history while loading related metadata from the selected cohort and using indexed watermark lookups. Finding: P01 bounded history visibility Model: GPT-5.6 Sol via Codex * fix(web): port composer activity and grouping to orchestration v2 * fix(web): align queue headers and prevent stash overlap * fix(web): share the outline for joined composer tabs * fix(web): keep stash separate from the composer activity column * refactor(web): use shared banner rows for queued messages * fix(web): keep queued message editing inside the queue panel * fix(web): keep queued messages in place while editing * fix(web): match composer actions to draft and modifier state * fix(web): keep composer shortcut tooltip stable on Mod * feat(web): summarize T3 orchestration actions * feat(mobile): port chat summaries and transitions to orchestration v2 Adapt grouped tool summaries and the floating working timer to V2 run, attempt, and queue state. Bring over the composer, keyboard, and disclosure transitions while retaining the V2 activity inspector and queue controls. Keep OV2 web composer and grouping behavior intact; share only the existing command label parser with mobile. * fix(chat): remove added tool summary status counts * fix(mobile): keep scroll bounds current after animations * fix: reconcile main's round-17 features after the rebase Restores main features dropped by the policy replay: #8569 theme wiring, settings search rework, #8803 workspace-mutation refresh (v2-adapted), video + image previews (web and mobile, v2-adapted), #8862 Expo glass, and the round's docs. Timeline thinking rows (#8984) stay on the v2 work-live system. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(web): port working and thinking timeline rows to orchestration v2 The v2 equivalents of main's #8984 and #8922: a "Working for ..." header anchors the active run, the trailing live tool row survives between actions in past tense instead of vanishing, and a shimmering Thinking row marks reasoning gaps. During workspace preparation the header shows "Setting up worktree..." (driven by the local dispatch flag or the v2 run's preparing status, so remote viewers see it too), the composer footer span is gone, and draft promotion waits until the run starts or startup fails instead of navigating mid-preparation. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: reconcile main's round-18 features after the rebase Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dro…
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopts the round's main features into the v2 architecture: the #9023 media rework (streamed videos, media-file assets, protocol-relative links), #9098 shared live-activity row folded into the v2 working and thinking rows, the #9084/#9078 Claude model catalog for v2 consumers, a native #9005 OpenCode child-session abort in the v2 adapter, #9013's landed LegendList patch, and per-environment sidebar provider entries. For #8600 the server-side pieces land, but auto-settle evaluation stays client-side (reading the new server-owned settings) until the v2 orchestrator grows its own settlement reactor; main's v1-only reactor and coalescer additions are dropped with the rest of the v1 path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
What changed
Agent-referenced images and videos outside the workspace could not be previewed, selecting MP4s in Files tried to read them as text, and web attachment previews downloaded entire videos into blobs.
Preview supported media from Markdown links, reference links, recognized inline-code paths, image/video embeds, and viewed-image work logs. Absolute paths refer to the environment host; relative paths resolve against the thread workspace or worktree.
Show paused video previews with an initial frame, native controls, and expansion on web and mobile. Web chat embeds and file viewers share MediaVideoPlayer; the other web previews share first-frame preparation. Mobile reuses the bounded native-thumbnail cache and extraction queue. Inline playback begins only after Play, while opened viewers load without autoplay.
Use the same filmstrip artwork for supported video formats in web and mobile file chips and file lists. Extension detection follows the shared video MIME map.
Play videos in both file viewers without text reads or text preloads. Web video attachments pass signed URLs directly to the player so the existing Range/206 route handles buffering. Workspace refreshes wait for active playback to pause; ordinary signed-URL renewal preserves paused position.
Authorize local media with expiring signed URLs bound to the canonical path and device/inode identity. Validate a regular media file, then stream from the same guarded descriptor. Refresh descriptor metadata before setting response lengths and ranges. Do not copy previews into attachment storage. Removed, moved, or atomically replaced files invalidate old URLs. Mutable host videos omit cache validators and use
private, no-store, while ordinary byte ranges still return206.Render the shared web viewer through a body portal so clipped Markdown containers cannot contain it. Failed direct HTTP(S) and protocol-relative previews retain Open original navigation; generated host-media URLs are not exposed as original links. Protocol-relative media links open the viewer on both clients, with native clients explicitly using HTTPS.
Retry, workspace refresh, and mobile Save/share request fresh capabilities. Signed-URL renewal does not reset a loaded player's position.
Keep existing workspace HTML/PDF/SVG restrictions and explicit-download behavior. Enable streaming through the Electron protocol.
Preserve original source references in shared media actions across inline previews, expanded previews, and file viewers. Web images support save/copy; video downloads use browser controls. Mobile reuses native sharing.
Markdown source → rendered UI
These captures use a synthetic thread and files on the environment host, including files outside its workspace.
Image and video embeds
Raw Markdown source → rendered UI below
Both screenshots show paused previews. The recordings show playback followed by expansion into a paused viewer.
Recordings: web 28s and Android 22s
Before: empty web video preview
With preload disabled, no frame appeared until Play:
Expanded, paused video viewers
Links, reference links, and inline-code paths
Raw Markdown source → rendered UI below
Before: video chips used a generic file icon
Opening an image reference and reporting a removed file
Remote previews that are actually web pages
The following Markdown file intentionally links to HTML responses with media-looking paths. The shared web viewer covers the whole window and offers Open original when decoding fails. It never navigates automatically.
Raw Markdown source → rendered UI below
File viewers: paused frames and streaming playback
Before: reported web file-viewer failure
This is the reported failure for an MP4 opened in the web file viewer. The after captures use separate synthetic fixtures.
Web opens a 90-second, 45 MB test-pattern video. Android opens the five-second
media/workspace-video.mp4; its reused artwork says "Outside-workspace video", but this file is inside the workspace.Recordings: web 31s and Android 25s
The Chromium pass showed a paused initial frame at 0.1 seconds with only 3.01 seconds buffered from the 90-second video. Metadata/frame loading used
Range: bytes=0-and a laterRange: bytes=458752-, both answered with206 Partial Content. Playback and a seek to the middle worked through the same native player. The app does not fetch a playback blob.Preview loading now deliberately reads video data before Play. Web preloads metadata only near the viewport for inline videos; mobile gates native thumbnail extraction on visible chat rows. Preload is a browser hint, not a strict byte limit. No server-generated poster files or imported video copies are needed.
Review regression evidence
Authored dimensions, unavailable media, and protocol-relative links
The deliberately unavailable URLs return HTML to exercise decode failures. Retry keeps the fallback usable, Open original retains the authored URL, and image/video dialogs share the same error card. The sized video remains paused at its initial frame.
Raw Markdown source → rendered UI below
Image and video error dialogs
Android file viewer, playback, and explicit sharing
The latest Android pass opens a video in Files, expands the paused preview, plays it to the end, and invokes Save or share video. Only that explicit share operation downloads a client-side copy. No share target was selected. Returning from the share sheet preserves the final frame without a loading spinner.
Play Android file viewer, playback, and sharing recording · 29s · Paused expanded preview · Ended state after sharing
Workspace links open the file viewer
Workspace image and video chips now open the normal file viewer on web and mobile. Out-of-workspace media links and the expand controls on inline embeds keep their media previews. The demo now labels deliberately absent files as expected failures.
Raw Markdown source
Watch the web link-to-file-viewer recording
Watch the Android link-to-file-viewer recording
Verified the 90-second video plays in both clients. The web file viewer uses the signed HTTP streaming URL, not a blob; a range request returned
206withContent-Range: bytes 1048576-1048703/45066088. The routing follow-up passed 158 existing focused tests, web/mobile typechecks, and scoped lint and formatting. No new test dependencies or DOM harnesses were added.Source paths and media actions
Previews retain the authored path or URL separately from their playback URL. Web hover tooltips show that reference, and right-click menus copy it. Workspace media can copy its relative path or open the file viewer. Images can be saved or copied to the clipboard. Mobile uses long-press and overflow menus with native Save or share.
Raw Markdown source for the previews below
Web hover tooltip showing the original path
Before: previews did not expose source actions
Expanded previews and file viewers
Watch web source actions, clipboard paste, expansion, and file-viewer navigation · 45s · Watch Android expanded actions and file-viewer navigation · 18s · Android native video share sheet
Video downloads use the browser
Removed the web context menu's fetch-based Save video action. The native player keeps its Download control; source-copy and file-viewer actions remain in the context menu. No application-side video fetch is needed for downloading. An undecodable video offers Open in browser, while remote failures retain Open original. Existing local blob attachments use a native download link.
Raw Markdown source for the external video
Failed video recovery
The intentionally invalid MP4 exercises decode failure in the file viewer. The browser link opens the original streamed bytes without a fetch-and-blob download.
Verified actual clipboard paste for the full host path, authored external URL, and copied image. The saved image matches the original SHA-256. Android shared the original filename through the native share sheet, and expanded previews returned to the file viewer. Source actions never copy generated playback capabilities. Empty file-viewer padding no longer opens the image menu. Focused tests, scoped typechecks, lint, and formatting passed; no UI/DOM harnesses were added.
Verification
git diff --checkpassed.bytes=0-31request returned206with exactly 32 bytes; unverifiableIf-Rangecorrectly returned200.Follow-ups from the rendering audit
Built with Codex in T3 Code.
Note
High Risk
Changes signed asset authorization, symlink-safe file open, and video HTTP semantics (security-sensitive), plus broad client routing from markdown/workspace media to new preview flows.
Overview
Adds end-to-end host media preview so images and videos referenced in chat, workspace files, and markdown can play without reading MP4s as text or buffering whole files into blobs.
Server introduces a
media-fileasset with signed URLs tied to canonical path plus device/inode identity. Responses stream from a guarded descriptor with byte-range (206) support,private, no-storefor videos, and invalidation when files move or are replaced atomically.Mobile gains inline and full-screen
MediaVideoPlayer, shareduseMediaActions(copy path/URL, open file viewer, native share), workspace video preview in Files, and richer markdown handling (protocol-relative URLs, UNCfile:paths, path-shaped inline code, video file icons). Asset URLs can be refreshed on retry/share without resetting playback position; list viewability gates thumbnail work.Desktop widens CSP
media-srctohttp:/https:and enablesstream: trueon the custom scheme for streaming playback.Web adds
useAssetUrlRefreshfor re-minting capabilities after file changes (aligned with mobile).Reviewed by Cursor Bugbot for commit 0977bfc. Bugbot is set up for automated code reviews on this repo. Configure here.
Note
Add
media-fileasset type with signed streaming URLs and video preview across web, mobile, and desktopmedia-fileasset resource in packages/contracts/src/assets.ts; the server issues identity-bound tokens (device/inode) and serves guarded byte-range streams via openMediaFile in assetFileResponse.loadVideoPreviewUrlanddownloadVideoPreview.http:/https:inmedia-srcand enablesstream: truefor the custom scheme.media-filebind to device/inode identity; if a file is replaced or truncated in place, resolveAsset rejects the token and the client must calluseRefreshAssetUrl/useAssetUrlRefreshto re-sign. The oldworkspace-fileresource tag for images is replaced bymedia-fileacross presentation.ts and test expectations, which may break out-of-tree consumers expecting_tag: "workspace-file".Macroscope summarized 0977bfc.