fix: resolve reported runtime, catalog and delivery failures - #686

What & why

This batch addresses eleven reported failures across session navigation, task authoring, Loop recovery, native worktree delivery, SQLite contention, catalog traffic, desktop startup/update recovery, and ACP process lifecycle. The fixes use the existing owners for state, authorization, transactions, and terminal transitions. They do not disable validation, widen active-session exemptions, change the store writer's per-operation attempt budget, or hide backend failures.

CodeRabbit and Greptile remediation

Published remediation head: 3938b308bb71306d15f69e2787fe2bb765060ea9. The inventory includes every initial inline finding (4 Greptile + 14 CodeRabbit), the CodeRabbit review-body nit, its general/inconclusive checks, the subsequent Greptile inventory-read, automatic-recovery and real-receipt contention findings, the CodeRabbit outside-diff initial-boot finding, and the source-backed ACP timestamp disposition. The changes below are implemented. CodeRabbit approved this head after a fresh review and Greptile reviewed it at 5/5. All 26 inline threads are resolved, and the latest CodeRabbit review-body import nit is fixed. Current-head CI is complete: 33 successful checks, 17 expected release/nightly skips, zero failures and zero pending checks. The unchanged literal sixty-minute Electron regression passed; its current-head artifact is detailed below. Previous green statuses belong to ab9bed777d18c8870d2d1e3217e3d4aea1a0e0f1 and do not verify this remediation head.

General review dispositions and excluded-artifact evidence

Linked Issues: the inconclusive review excludes Markdown/MDX/skill files. Both .agents/skills/cy-create-tasks/references/task-template.md and extensions/spec-cycle/skills/cy-create-tasks/references/task-template.md explicitly require a double-quoted YAML title, escaping embedded quotes/backslashes while preserving the decoded first H1. The existing importer suite covers colon, quotes, backslashes/hash and quoted YAML indicators [logs] {core} &alias *ref !tag | >; the malformed unquoted-colon negative control remains intact. The original public importer walk and native publication receipts are recorded in the tracked QA report.

Three concrete documentation defects were corrected: managed delivery always requires reviewed --include and --expected-scope; the official HTTP action documents delivery_id and expected_head with the trusted native caller boundary; generated command output now uses the owning Cobra example with all required flags. Example:

compozy worktree deliver feature/alpha --delivery-id reviewed-alpha --expected-head <reviewed-head> --include internal/example.go --expected-scope <reviewed-scope> --message "fix: reviewed change" --base main --title "fix: reviewed change" --body "Reviewed scoped delivery" -o json

Catalog sixty-minute provenance: the original 07:39–08:39 UTC window belongs to Web index e6bc4e764b56a49213be2cc80a298d4fd0ec684f4a09d45a902b945a1fef25bf and Go binary 3b964b0f62628df3e75c45f98cf8d5e7f84e0871b2d7ffa41afa35b4a812c2e9: 1,964 responses, all 200, rolling peak 44/52, history 374→554. Subsequent repaired-flow walks used the separately recorded 71ea3236… Web index. This review remediation changes catalog causal code, so neither historical receipt is claimed as a sixty-minute pass for the new head. New blocked-versus-healthy, facets and debounce invariants are owned by the named canonical CI suites. The existing packaged Electron shell suite now automatically runs a literal sixty-minute regression with 600 stopped sessions, a real active Loop and continuous public session activity, asserting every rolling sixty-second endpoint budget (26 lists, 13 facets, 13 summaries, 52 combined). Native visibility/focus, monotonic duration, current head, runtime digest, actual client headers and rate samples are retained as a CI artifact on success or failure. This current-head acceptance passed in packaged Linux Electron CI: 3,600,906.654 ms, 600 verified stopped seed sessions and 281 additional public create/stop cycles. Across 2,244 renderer requests from one real client, independently recomputed rolling maxima were 20 lists, 10 facets, 10 summaries and 40 combined, below the unchanged 26/13/13/52 ceilings. All 281 native samples were visible/focused with the real Loop running; violations and cleanup errors were empty. The retained artifact records checkout ca0bd7fe98e838b05760bdda28b85d37c541adda, whose PR parent is the published source head 3938b308bb71306d15f69e2787fe2bb765060ea9, and runtime SHA-256 1d98033d1cb669d9e4a4a61bfcbc32c37ba871ac43ecd0a9def87628beabac18. The historical macOS soak remains qualified by its separately recorded assets; it is not relabeled as this new Linux run.

Docstring Coverage: the aggregate bot metric is below its 80% threshold but identifies no function-level documentation defect. The explicit repository coding policy (.agents/skills/eng/eng-code-guidelines/references/coding-style.md) defaults to no WHAT comments and permits a short non-obvious WHY comment, rather than multi-paragraph docstrings. Safety boundaries received targeted concise explanations; bot thresholds/configuration were not changed and an 80% result is not claimed. The aggregate suggestion is disposed as a policy conflict/inconclusive metric. CodeRabbit independently inspected the sources and withdrew both the #663 generation-template premise and the aggregate docstring suggestion; no duplicated prose assertion or weaker importer validation was needed.

Validation for this remediation

No local gate/test/typecheck/build/cross-build was run. Owning generators regenerated OpenAPI, Web types and CLI docs; the actual pre-commit formatter/lint tasks passed without suppressions. CI now invokes the existing real-Git TestWorktreeLifecycleIntegration and TestWorktreeManagedDeliveryIntegration on Go shard 2, following the existing targeted-integration pattern. CI on the first remediation head exposed a widened query-key array type, excessive React hook complexity, and a unit fixture referencing an integration-only helper. The second remediation commit uses the canonical readonly query-key tuple, moves page composition into the existing palette session projection owner, and checks the exact ACP agent response through the available unit helper. Assertions and gate thresholds remain intact. The subsequent Greptile 4/5 review identified an unavailable-directory inventory being treated as absent; the third remediation commit distinguishes these states and adds the real-filesystem regression above. Remote Go lint also required an explicitly checked context-owner assertion and owning formatter output, which were corrected without suppression. One remote shard hit a gotestsum download HTTP 403 before tests; this was infrastructure evidence rather than an assertion failure. The recovery follow-ups provide lifecycle-owned retries for unavailable inventory and wrapped transient SQLite errors, including initial boot, while protecting concurrent active receipts. The latest canonical contention fixture locks the actual receipt database and observes its production writer; no auxiliary database error substitutes for that boundary. The existing desktop lane also runs the new literal sixty-minute catalog regression and retains its artifact on success or failure. The existing palette regression now waits for its exact original remote-search assertion after the intentional debounce, retaining filters, rows, chips, cursor assertions and its default deadline. All 25 inline threads were resolved with owning CI evidence on a49e92f5fca192dfeeb6ff427e07a784de2f31f4, which Greptile reviewed at 5/5 and CodeRabbit approved. That head passed frontend, Desktop checks, Go lint, daemon/worktree unit suites and actual Git integrations. Its new Electron run verified all 600 stopped sessions and retained a setup-failure receipt, then correctly refused the fixture's undeclared UDS start transport. The fixture now declares its actual CLI/UDS binding; duration, budgets and assertions remain unchanged. CodeRabbit approval and Greptile 5/5 on 15d73b3 are confirmed, with zero unresolved inline threads. Its CI passed 31 contexts and all 29 other desktop E2E tests, but the new soak correctly failed after 61.518 seconds of measurement: 600 verified stopped sessions, a running real Loop and one renderer client produced rolling maxima 20/15/10, combined 45, exceeding only the fixed facets ceiling of 13. The artifact preserved the failed measurement and no cleanup errors (run). The terminal badge was polling the shared facets query independently of live stream reconciliation. The production hook now uses the shared live clock for healthy reads, retains visible disconnected polling and retries failed reads at the existing thirty-second error interval even with a live stream, addressing the subsequent Greptile quiet-workspace recovery finding. Its canonical hook regression uses actual TanStack Query and an adapter I/O mock to verify request counts and fresh badge values across live wakes and disconnected fallback. The Electron budgets, assertions and literal duration remain unchanged. Greptile reviewed d8bed70 at 5/5 and resolved the new quiet-workspace thread. CodeRabbit reported one review-body import nit, now corrected through the session public barrel in 3938b30. CodeRabbit approved 3938b30 after its fresh review and Greptile reviewed the exact same head at 5/5; all 26 inline threads are resolved and the review-body import nit is fixed. Current-head CI passed all 33 successful contexts with 17 expected skips, no failures and no pending checks. Frontend logs confirm 6,786 Web tests, including all 17 canonical OS attention tests. All eight Go race shards, targeted real-Git integration, runtime E2E and four Web E2E shards passed. Packaged desktop logs confirm all 30 E2E tests passed, including the literal sixty-minute window with the independently verified 20/10/10/40 rolling peaks described above. Local, remote branch and PR head all match 3938b308bb71306d15f69e2787fe2bb765060ea9, and the owning worktree is clean. Current-head CodeRabbit approval, Greptile 5/5 and zero unresolved threads were reconfirmed after CI completion. Earlier-head evidence and the historical soak remain qualified. No merge or release is requested.

Issue-by-issue changes

#657 and #659 — stopped-session deletion and scoped navigation

The existing stopped system-session deletion path already had the correct runtime/store behavior on current main. Its real Manager/HTTP/SQLite regression now establishes that behavior. The navigation failure had a different cause: workspace session detail reads omitted the operator's aggregate profile scope, so an otherwise valid session could produce a profile mismatch and 404. Detail reads explicitly request the aggregate profile view while the server retains authoritative workspace and agent identity checks.

Presence cancellation uses the shared error mapper instead of turning an aborted request into a 500. Genuine service failures remain failures. Global workspace scope also enables its canonical all-workspaces catalog reads; an unresolved project still waits for a concrete workspace.

#663 — valid task titles and terminal importer failures

Both shipped and local task-authoring templates emit escaped quoted YAML title scalars and keep the decoded title aligned with the Markdown H1. Authoring/preflight instructions use the existing strict importer. Existing importer/provider tests cover colons, quotes, backslashes, hash characters, and malformed negative controls.

Real default-policy QA exposed a second defect: tool_invalid_input entered automatic failed-only succession, quarantined the sole producer, and left the Loop running without runnable work. The coordinator now closes an unhandled permanent invalid-input failure before another generation. Authored correction routes, allow-fail, and transient/quality repair preserve their existing precedence. Canonical lifecycle and real-daemon tests cover both the default policy and halt, including released forbid-concurrency admission. Parallel failure selection checks all unhandled outputs so an earlier transient error cannot hide a later permanent invalid-input failure; existing target-unavailable precedence is preserved.

#665 — updater operation owner and deadlines

The updater rejected applying-phase cancellation before checking whether the recorded holder still existed, and the CLI described every rejection as a live installer. Reads, acquisition, and commit-fence reconciliation now settle dead/expired operations through the journal owner. A dead applying holder can be canceled; live execution and irreversible handoff remain protected. update --check describes checking without applying because journal reconciliation can update stale operation state.

#666 — Linux desktop discovery and URI registration

Linux scheme registration names the actual executable, includes %u, registers the MIME scheme, and records the current version. CLI discovery accepts the authenticated running shell or an executable installation, including the newest valid AppImage. Deleting a desktop shortcut no longer makes an installed/running app disappear, and direct launch does not require that shortcut.

Public commands were exercised on isolated macOS state. A disposable Linux environment with real KDE/KIO, xdg tools, desktop-file validation, and the service cache verified URI dispatch and executable launch after shortcut deletion. The executable is the process I/O boundary in that test; a packaged Electron/AppImage installation on a physical KDE desktop is a separate platform claim.

#675 — adopt the exact recovered child only with current proof

A settled failed parent output previously could not refresh its now-successful child, and failed-only succession discarded the child identity. The parent now reconciles before selecting failure/succession. Public replay also caught the real settled-failure payload's omitted optional diagnostic target; adoption now accepts that omission while rejecting a contradictory target and still requiring the authoritative original child ID and canonical completed task receipt. Reuse requires the exact workspace, parent, Loop name, original canonical node-task receipt, successful upstream producers, terminal done/no-op child, and unchanged rendered child inputs.

Input equality alone cannot prove that mutable files are unchanged. Optional inputs.reviewed_worktree is a JSON string strictly decoded into a typed proof of the original worktree ID, exact selected paths, and scoped exit-plan fingerprint. It uses the existing DSL string input type. The registered worktree owner independently rechecks HEAD, branch, selected file bytes/modes, and selected index state. Absent, malformed, foreign, or changed proof takes the normal fresh-review path. Recovery never replaces the original fingerprint with a newly computed digest to authorize reuse. The boot factory passes this verifier only after establishing the runner's required store capabilities, preserving the existing disabled-capability nil runner contract.

#676 — durable native delivery from a managed session

compozy worktree deliver accepts a reviewed intent from validated native session identity. A real bound-session submission exposed the existing SQLite action constraint's missing deliver value; migration 00125 widens it losslessly and preserves all operation fields, foreign keys, and the active-operation uniqueness index. The daemon acquires an exact-checkout lifecycle fence, rejects unrelated active sessions, stops the submitting caller, and owns commit/push/draft-PR effects after the caller exits. A versioned, atomically persisted and fsynced journal records the original identity, selected scope, HEAD, base, expected tree, destination, and phase.

Restart recovery repairs operation receipts and reconciles the same intent. Actual recovery on the final native binary durably failed both the interrupted receipt and its recovery receipt when the original selected inputs had changed; the journal retained the original authorization and HEAD stayed unchanged, with no commit, push, or PR effects. Canceled intents do not replay effects. Push names one recorded destination and one explicit SHA-to-branch refspec. Forge lookup/create requires exact open head branch, base, SHA, and draft evidence before reuse; an ambiguous provider response is reconciled before another create. Ordinary non-managed forge operations retain compatibility with old extensions. Before accepting a new commit, the owner verifies the captured immutable SHA's exact tree, original parent, and reviewed message. An externally replaced HEAD with the same tree cannot authorize delivery or a prepared no-op; the recorded, verified SHA is the only push source.

The first post-migration public handoff admitted a durable operation and stopped its caller, then durably failed on two genuine pre-commit lint diagnostics with no commit, push, or PR. The API-client test keeps its original client-ID assertion with safe optional access, and the shell reuses its existing sessions-modal owner for dialog composition, keeping the scoped shell body within the existing component-size rule without moving catalog hooks or suppressing diagnostics. Existing suites exercise actual asynchronous Submit/Recover execution with real linked Git, a local bare remote, and an HTTP forge boundary across eight phases, including persisted completion/cancellation and ambiguous response recovery. Separate identity and session-fence suites cover spoofing, workspace isolation, and admission/stop behavior.

#677 — selective reviewed commits preserve unrelated work

Exit plans and commit/commit-push accept repeatable exact include paths and an expected-scope fingerprint. Validation rejects traversal, Git metadata, and symlink paths. The fingerprint binds HEAD, branch, selected content/modes, and complete per-path staged blob identity, and is checked again under repository serialization before staging.

Git stages/commits only the selected files. An isolated alternate index computes the expected selected tree, and the actual resulting commit must match before publication. This catches hooks that expand a commit. Per-path full Git blob OIDs also prevent a large earlier binary diff from truncating the proof of a later selected file; already-staged tracked deletions remain selectable. Unrelated staged and unstaged files, including private task artifacts, remain outside the delivered tree. The commit owner also binds the resulting immutable SHA to exactly one parent: the HEAD reviewed before final fingerprint validation. A same-tree external HEAD replacement is rejected before completion or publication. The existing whole-worktree action remains available.

#678 — writer admission before consuming a pooled connection

A real two-connection reproduction showed queued writers consuming pool slots before BEGIN IMMEDIATE, starving the active writer's commit-fence read. Per-database writer admission now happens before connection acquisition. Final native delivery recovery exposed a second failure: local writer waiting incorrectly spent the finite SQLite BUSY attempt budget, leaving an exit receipt running when terminal persistence could not acquire admission. Local admission is now acquired once with context cancellation before the unchanged SQLite retry loop. Waiting reserves no pooled connection and spends no SQLite attempt; actual external BUSY still uses the original bounded budget. Existing canonical tests cover cancellation, registry cleanup, pool capacity, and real receipt persistence after a long local writer releases.

Named transaction diagnostics distinguish waiting/held writers and unknown external writers. Hot session/heartbeat/tool/resource/task paths use the existing write owner. Safe tool phase evidence distinguishes event-write contention, registry resolution, and skill resource loading without logging resource content, credentials, or raw tokens. Historical unknown lock holders are not retroactively attributed to this reproduced cause.

#679 — bounded catalog reconciliation and row reads

Catalog events previously invalidated and canceled in-flight reads repeatedly, amplifying one lifecycle trigger into several full walks. Reconciliation now coalesces wake windows, finishes in-flight reads, retains pending dirty work, defers hidden/unfocused reads, and backs off errors. Client requests carry a bounded opaque instance identifier in daemon request logs.

Durable attention filtering occurs in SQL before the page cut, and catalog page reads no longer sweep history or acquire a writer. Appended recent/activity/attention/navigator/creation indexes support the keyset orders. An opt-in skip_total count-free page preserves existing default exact-total clients.

The Web's palette and session groups use count-free bounded pages with explicit Previous/Next navigation. Each automatic wake rereads the currently visible page, including its held cursor, rather than every historical page. Search and badge filters run on the server before paging, so older matching sessions remain reachable. A separate scoped /api/sessions/facets metadata aggregate supplies exact chips/group totals without rich history hydration. Exact pending terminal-approval metadata owns the terminal badge, including workspace/profile isolation and Global cache readiness. The dock reads one row in its original creation order. Palette navigation preserves its four-band attention order before paging; Unicode-aware title/agent search matches the visible selection contract, including contextual Greek final sigma and the dotted Turkish capital I. The new search mode uses the existing Unicode library consistently in the SQL scalar function, live overlay, and query normalization; legacy search retains its behavior. Loop session selection uses bounded pages, server search and direct selected-ID hydration. Collapsed groups do not fetch hidden row histories, and keyboard navigation reuses current visible group pages. Palette entity projections, session-page rows, and list controls are composed separately from their existing state/query owners so the changed React functions remain cohesive within the project's complexity rules. Existing cursor fingerprints remain compatible, while new search/order modes fence their own cursors. Legacy counted clients keep their contract; exact aggregates and arbitrary text searches are not advertised as constant-time operations.

The literal sixty-minute native Electron window completed from 2026-09-30T07:39:10.868302Z through 08:39:10.868302Z on frozen Go binary 3b964b0f62628df3e75c45f98cf8d5e7f84e0871b2d7ffa41afa35b4a812c2e9 and Web index e6bc4e764b56a49213be2cc80a298d4fd0ec684f4a09d45a902b945a1fef25bf. The real catalog grew from 374 to 554 records through 180 ordinary public create/stop cycles while the authored Codex Loop remained running. All 1,964 observed session-prefix HTTP responses were 200. Rolling-minute peaks were 22 list reads against 26, 11 facet reads against 13, and 11 summary reads against 13: combined 44 against 52. All 119 native samples showed a visible 100-row page; one natural unfocused sample is retained, so continuous focus is not claimed. The external monitor corrected a pre-navigation client ID by rereading the complete original interval, then included preserved rotated logs after log rotation; neither correction restarted the native runtime, clock, activity, or limits. An earlier eleven-minute interval was interrupted and remains explicitly not a pass.

This is historical proof of the automatic catalog invariant on the recorded original assets; the review-remediation changes above require their own current-head evidence. Later CI remediation restores the separate background notification/title contract, resolves the dock query on explicit clicks, and repairs a forced-new window command revision race. Before this review round, the automatic catalog query options, reconciliation, paging, facets, and summary code under measurement remained unchanged. Rebuilt final assets separately passed the repaired interactions and catalog smoke. The actual final Electron index 71ea3236e585a54471d994843a07ea687a1797a34badc28b6af7086f13d18b57 showed hidden-title growth and clearing, cold/newest/minimized dock selection, the eight-terminal cap dialog, cursor navigation, contextual Greek search, and clearing back to the complete population. The hidden-title walk used one owned instance and the other walks used a subsequent same-byte restart after host fullscreen focus limitations; synthetic visibility was never used. The manual terminal click did not reproduce the 66 ms race; the canonical queue regression and unchanged E2E own that boundary. A sixty-minute soak of these later bundle bytes is not claimed.

#682 — bounded ACP control startup and verified process cleanup

ACP control requests have a bounded deadline. Only a private session/new stage timeout can receive one fresh-process retry, while the caller remains alive and the previous process has verifiably exited. Initialization/authentication/configuration failures and caller cancellation do not retry. Prompt and clarification lifetimes retain their existing owners.

Stop cleanup bounds graceful exit plus process-tree drain and escalates through existing platform process utilities. Real subprocess tests cover silent startup, first-attempt stall followed by success, both attempts stalled, and removal of leader/child before retry or terminal stop. Windows cross-build evidence is separate from native Windows execution.

Impact, compatibility, and generated artifacts

• Existing SQLite, config, workspace files, session history, and persisted layouts are preserved. Migrations 00123 and 00124 add catalog indexes without changing or deleting user rows. Migration 00125 copies every existing exit-operation row into the widened constraint before replacing the table and recreating its uniqueness index.
• New public fields/routes/flags are additive. Counted catalog defaults and whole-worktree commit behavior remain available. Optional forge identity evidence is required only for the new managed delivery contract.
• HTTP/UDS/OpenAPI, Go/TypeScript SDK contracts, Web generated types, and generated CLI documentation are changed through their owning generators.
• Worktree, child, profile, workspace, and session boundaries are validated by their existing authoritative owners. No cross-workspace Git scan, raw token publication, blanket active-session exemption, or forge identity inference is introduced.
• The single cross-surface audit is in docs/_memory/change-impact.md; official Compozy skill references, site docs, task-authoring references, and affected QA scenarios co-ship.

Historical validation before review remediation

Focused owning suites, real runtime walks, local checks, and current-head CI are separate evidence. All eleven selected runtime journeys passed with the platform, provider, and exact-build qualifications documented in docs/qa/reports/2026-09-29-reported-issues.md. At the final delivery-gate checkpoint, exact-head CI passed with 32 successful contexts and 17 expected release/nightly skips, with no failures or pending gate contexts. The strict QA evidence audit also passed with no blockers or warnings. The owned lab was torn down at 2026-09-30T09:46:32Z, with clean=true, cleanup_clean=true, no surviving owned processes, and no listeners on its three owned ports. An interrupted local test run is not a pass.

Owning focused checks already include real SQLite contention/paging/index plans, linked Git selective commits and hook rejection, managed Submit/Recover execution, real ACP subprocess lifecycle, real daemon task-import terminality, provider HTTP identity reconciliation, scoped Web query/navigation behavior, and public macOS/KDE command walks. Earlier generated contract checks and the complete affected Go lint passed with zero issues. The final public-wire, migration 00125, and contextual writer-admission corrections passed the remote Go, generated-contract preflight, and runtime E2E lanes on source head b4410915c1ccc6a3eefb14b91ce063ce0b9b398d; the owning generator completed for migration 00125. Final-head CI passed on ab9bed777d18c8870d2d1e3217e3d4aea1a0e0f1 (workflow): preflight, frontend, React Doctor, Desktop, all eight Go race shards, Go build/boundaries, runtime E2E, all four Web E2E shards, Desktop E2E, and the aggregate checks are green. Frontend logs record 6,780 passing Web tests, including all 53 canonical window-manager tests. All 49 contexts present at the final delivery-gate checkpoint were inspected: 32 successful and 17 expected release/nightly skips, with zero failures and zero pending contexts. Marking the unchanged head ready subsequently starts independent CodeRabbit/Greptile reviews and a React Doctor rerun; their live state is separate from that completed gate run. At the user's request, the final local gate was interrupted and remaining delivery gates run in CI; the interrupted local test run is not claimed as a pass. The final public child-recovery replay passed on the fixed binary: an unchanged candidate let the paused parent finish in generation 1 with its original recovered child; changing only selected working bytes preserved the original HEAD/index/proof but required generation 2 and a new child. Literal Web deletion passed for both the default and non-default profiles using the real Delete confirmation and independent correctly scoped 404 reads. Final native boot recovery passed with durable failed receipts, preserved original authorization, and no Git effects. This PR itself passed the real native managed-delivery journey: the bound agent submitted once through the hosted terminal tool, the daemon stopped that caller, the official commit tasks passed, and the durable operation/journal completed. The initial native publication receipt binds local HEAD, the explicit remote branch, and GitHub to 268933c20e844f42355feb06af5b06ae4f4dff42; GitHub independently confirmed one open draft PR and the working tree was clean. The completed sixty-minute catalog window is recorded above with its actual build provenance. Final repaired-flow asset walks are complete. Final-head CI is green, the post-cleanup strict QA evidence audit passed, and owned-lab teardown completed. The isolated lab retains its manifest, original failed audit, corrected final verification metadata, audit result, ownership revalidation, teardown result, and listener proof. No unrelated process was stopped and durable artifacts were preserved.

CI remediation

The first published head exposed genuine React complexity diagnostics, two Go test-style diagnostics, stale product naming, and an exact shipped-example documentation contract. These were repaired in their owning files without configuration changes or suppressed diagnostics. The documentation example now includes the shipped reviewed-worktree input.

The runtime E2E attention journey also failed identically on the accepted main commit a15f541f93e1da6d5d1678b22e2312101a1cda4f (baseline CI). Its binding noop legitimately emits the completion wake added by that main change, but the existing fixture returned the first wake for the child before inspecting its reason. The same canonical integration helper now awaits the child’s needs_attention wake specifically. Every behavioral assertion, five-second wake timeout, and overall journey deadline is retained; no production notification is removed and no retry/deadline is enlarged. The corrected journey runs remotely in this PR’s CI.

CI also exposed two production regressions from the catalog work. Background notification polling had been disabled despite the existing title-count contract; the notification owner now polls at five seconds with thirty-second error backoff in background as well as foreground. Catalog page/summary reads keep their hidden pause. A dock click could treat a coalesced, cached empty page as current after another client created a session. The explicit action now joins/refetches its existing scoped, count-free one-row query before selecting create versus open, reports query failures, and permits a subsequent click to retry. Existing latest-created selection remains the owner, and automatic catalog wake behavior is unchanged. The dock unit fixture awaits the asynchronous action with all prior assertions intact; both failing canonical title and real dock E2E cases are retained.

The final Web shard exposed an existing queued window-command race: terminal adoption began window.navigate, then New terminal was clicked 66 ms later while adoption was in flight. The new-window request captured revision 1 at enqueue, reached the serialized dispatcher after revision 2 had been accepted, and received window_manager_revision_conflict. It never issued a ninth terminal-create request, so changing the cap dialog or its test would conceal the cause. Deliberate forced-new/stack-target opens now use the existing serializer’s current execution revision. Ordinary semantic identity lookups keep their captured revision fence and conflict recovery, preserving the canonical concurrent-open protection against duplicate windows. A controlled pending-adoption regression was added to the same runtime suite; all existing assertions and E2E-009 deadlines remain intact. The following CI head exposed a TypeScript-only fixture mistake: the test attempted to assign through a readonly snapshot map. The final fixture constructs an immutable copied snapshot instead; production code, assertions, and deadlines are unchanged.

Agent contributions and delivery policy

Codex coordinated GPT-6.1 Sol workers for ACP lifecycle, scoped session navigation, catalog pagination, SQLite contention, Loop recovery, managed delivery, selective Git, desktop/update handling, and task authoring. Owning tests and retained real SQLite, Git, daemon, provider, and UI receipts support the findings above; all selected runtime journeys are complete and their evidence boundaries are explicit.

• The user explicitly requested that delivery gates run in CI to avoid consuming local machine resources. The owned local gate and all its descendants were stopped; unrelated worktree processes were preserved. No final local gate pass is claimed.
• New and changed observable invariants are covered in existing canonical suites. The recovered-child wire, migration 00125, and contextual writer-admission regressions passed remote Go/runtime CI. The new forced-window queue regression passed the final remote frontend lane, and all applicable checks passed on the final published head.
• All applicable CI checks on the pre-review head and the post-cleanup QA evidence audit passed. Owned-lab teardown is complete. The newly published review-remediation head requires fresh CI and bot evidence, tracked above. No merge or release is requested.

Issues

Closes #657
Closes #659
Closes #663
Closes #665
Closes #666
Closes #675
Closes #676
Closes #677
Closes #678
Closes #679
Closes #682

Summary by CodeRabbit

• New Features

  • Linux desktop installations can register as handlers for CompozyOS links.
  • Session browsing adds searchable, paginated catalogs, workspace counts, creation-time sorting, and attention-based navigation.
  • Worktree actions support commits scoped to selected files and managed draft deliveries.
  • Loop inputs can select sessions from a searchable catalog.

• Improvements

  • Pull-request checks match branches, commits, and draft status more precisely.
  • Agent startup and shutdown handle stalled requests with bounded timeouts, cleanup, and eligible retries.
  • Opening desktop links better handles running installations and Linux AppImages.
  • Session views refresh more efficiently, with clearer loading and error states.
  • Unicode search and reviewed-file checks improve discovery and scoped-commit safety; managed deliveries recover more reliably after interruptions.

• Bug Fixes

  • Session links and lookups work more reliably across profiles and workspaces.