Repository navigation
Conversation
51d378f to
4bee374
Compare
|
Note Reviews pausedIt looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the Use the following commands to manage reviews:
Use the checkboxes below for quick actions:
📝 WalkthroughWalkthroughClaude background-task launches and owed notifications now contribute to pane status. The hook server and relay schedule status updates when notification leases expire. Wake-up state also informs terminal readiness. The changes add guarded remote interrupt reconciliation across relay, SSH, and WSL connections, with tests and captured fixtures covering task lifecycles, cancellation, expiry, and readiness. Priority: ➖ Normal Merge Risk: 🟡 Moderate · up to Hook events can arrive out of order. If a background agent's end event arrives before its launch is recorded, the pane can be marked finished while the task's notification is still pending. That reintroduces the premature-completion problem this PR is meant to fix. A test-coverage concern about tracking-cap eviction also remains open. Address both before merging. Security Architecture ReviewSecurity architecture risk: 🔵 Low · up to The inspected paths preserve pane, connection and turn ownership while preventing premature completion. Remaining uncertainty concerns lost notifications, restart recovery and mixed-version operation; no introduced security vulnerability was established. Retained concerns Security review detailsSecurity Blast Radius
Trust Boundaries and Controls
Resilience and Maintainability Implications
Hardening Proposals
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
Full details: Docstring CoverageExplanation Docstring coverage is 50.59% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 85 functions across 63 files. (1 skipped: 1 unsupported.)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
🧹 Nitpick comments (1)
src/shared/claude-background-task-inventory.ts (1)
85-87: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low valueSkip empty IDs after trimming.
A whitespace-only
idis pushed as"". Nothing is tracked under an empty ID, so the empty entry is harmless today. It does break the documented contract that unlisted entries are counted but not listed.Proposed fix
- if (typeof obj.id === 'string') { - runningNonAgentTaskIds.push(obj.id.trim()) - } + const id = typeof obj.id === 'string' ? obj.id.trim() : '' + if (id.length > 0) { + runningNonAgentTaskIds.push(id) + }
ℹ️ Review info
⚙️ Run configuration
Configuration used: Repository UI
Review profile: CHILL
Plan: Advanced
Run ID: 7db37d50-723b-4190-94f9-176c9518a999
📥 Commits
Reviewing files that changed from the base of the PR and between 44a2ee5 and 4bee37492040b18c60b032f7e0c28c9bda44f11b.
📒 Files selected for processing (28)
src/main/agent-hooks/server-claude-cancel-captures.test.tssrc/main/agent-hooks/server-claude-owed-notification-expiry.test.tssrc/main/agent-hooks/server-relayed-claude-cancel.test.tssrc/main/agent-hooks/server/server-cleanup.tssrc/main/agent-hooks/server/server-lifecycle.tssrc/main/agent-hooks/server/server-status-update.tssrc/main/agent-hooks/server/server-types.tssrc/main/ai-vault/session-scanner-claude-subagents.tssrc/relay/agent-hook-request.tssrc/relay/agent-hook-result-retry-scheduler.tssrc/relay/agent-hook-server-claude-owed-notification-expiry.test.tssrc/shared/__fixtures__/claude-task-notification-hooks.jsonlsrc/shared/__fixtures__/claude-task-notification-hooks.meta.jsonsrc/shared/agent-hook-listener-claude-task-notification.test.tssrc/shared/agent-hook-listener.tssrc/shared/agent-hook-listener/listener-event.tssrc/shared/agent-hook-listener/listener-state.tssrc/shared/agent-hook-listener/main-agent-turn-state.tssrc/shared/agent-hook-listener/providers/claude-events.tssrc/shared/agent-hook-listener/providers/claude-lifecycle-events.tssrc/shared/agent-hook-listener/providers/claude-pane-hold-evidence.tssrc/shared/agent-hook-listener/providers/claude-roster-state.tssrc/shared/agent-hook-listener/providers/claude-task-notification-delivery.tssrc/shared/agent-hook-relay.tssrc/shared/claude-background-task-inventory.tssrc/shared/claude-owed-notification-expiry-timers.tssrc/shared/claude-owed-task-notifications.tssrc/shared/claude-task-notification-text.ts
Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 8 remain after this review.
![pullfrog[bot]](https://avatars.githubusercontent.com/in/1768019?s=60&v=4){kind=small}
There was a problem hiding this comment.
Important
The hold model is sound and the captured-hook tests are strong, but the shell-owe is gated on the literal 'Stop' turn event, so a shell first seen gone on a StopFailure turn is never marked owed and the row settles done before its pending notification — the same premature-done window this PR closes. I verified this with a scratch replay (details inline).
Reviewed changes
- New hold model:
claudeLaunchedBackgroundTasksByPaneKeyrecords the main agent's launches and marks a task "owed" when it ends before its<task-notification>; owed work now counts as live work inclaudePaneHoldEvidence, so the pane cannot readdonein that window (claude-owed-task-notifications.ts,claude-task-notification-delivery.ts,claude-pane-hold-evidence.ts). - 60-second lease:
ClaudeOwedNotificationExpiryTimersis armed on every stored row by the desktop hook server (commitStatusRowMutationoverride) and by the SSH relay (AgentHookResultRetryScheduler.armClaudeOwedNotificationExpiry), restating a settled row when an idle main agent was owed for a full lease (claude-owed-notification-expiry-timers.ts). turnCompletedAtgate: a turn held only by an owed notification no longer stamps a completion, deferring the announcement to the wake-up turn (claude-events.ts).- Shared parsing and lifecycle wiring:
readClaudeTaskNotificationis shared with the AI-vault scanner; SessionStart/clear/SessionEnd/pane-move/reset all drop what is owed;claudeRunningNonAgentTasksemantics widen to unlisted live work with no new wire field. - Tests: 46 new/updated tests replaying captured Claude 2.1.287 hooks plus hand-built lifecycle/lease cases. I ran the three new prod files (46 passed).
ℹ️ A relay or host restart can strand a held remote pane
Only the execution host holds claudeLaunchedBackgroundTasksByPaneKey and its lease. For a remote pane that is the relay, and the desktop arms nothing for remote rows: server-ingest-remote.ts mirrors only the claudeRunningNonAgentTask boolean into its sets, leaving its own launched map empty, so ClaudeOwedNotificationExpiryTimers.deadline() returns undefined and arm() clears. A relay restart during the owed window discards both the owed record and the timer; the desktop's row is persisted, remains working, and with no later row-producing hook nothing settles it — the remote-pane reaper explicitly requires connectionId === null (server-reaping.ts:33-46) and the connectionId-scoped clear has no production caller. The same shape applies to a desktop restart for a local pane whose agent process survived. The PR lists a live SSH host as unverified and covers "Claude exiting / terminal closing", but not a host restart. Worth confirming whether this is in scope or accepted; if accepted, a line in the Not-verified section would help.
Technical details
# host restart durability
## Affected sites
- `src/relay/agent-hook-request.ts:69` — lease armed only on a live stored hook
- `src/relay/agent-hook-server.ts:195-204` — `stop()` clears the listener state; `start()` restores nothing
- `src/main/agent-hooks/server/server-ingest-remote.ts:250-302` — remote ingress never populates the launched map or lead record, so the desktop arms no lease
- `src/main/agent-hooks/server/server-reaping.ts:33-46` — reaper excludes `connectionId !== null` panes
## Required outcome
- A pane held `working` solely by an owed notification must eventually settle even if the owning host restarts mid-window.
## Open questions for the human
- Is relay/host restart during the owed window in scope, or accepted because a later row-producing hook re-derives the row?deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏
|
Review follow-ups for 16d3ac2:
|
There was a problem hiding this comment.
🧹 Nitpick comments (1)
src/shared/agent-hook-listener-claude-task-notification.test.ts (1)
252-254: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick winAssert which task the cap evicts.
recordClaudeBackgroundTaskLaunchcurrently evicts a settled task. But if it evicts livea2instead of announceda1,a257can still be tracked and the finalworkingassertion can still pass. The other cap test checks only map size and state. Assert thata1is absent anda2anda257remain.Suggested test assertions
post({ hook_event_name: 'Stop', background_tasks: [{ id: 'a257', type: 'subagent', status: 'running' }] }) + const tasks = state.claudeLaunchedBackgroundTasksByPaneKey.get(PANE_KEY) + expect(tasks?.has('a1')).toBe(false) + expect(tasks?.has('a2')).toBe(true) + expect(tasks?.has('a257')).toBe(true) + expect(post({ hook_event_name: 'SubagentStop', agent_id: 'a257' })?.payload.state).toBe(
ℹ️ Review info
⚙️ Run configuration
- Configuration used: Repository UI
- Review profile: CHILL
- Plan: Advanced
- Run ID:
e3e3d32d-17e0-47cb-85ab-95d2a866839d
📥 Commits
Reviewing files that changed from the base of the PR and between 4bee37492040b18c60b032f7e0c28c9bda44f11b and 16d3ac2655226a5ebdfda7a1b5c4925bf2e3d8c0.
📒 Files selected for processing (6)
src/main/agent-hooks/server-claude-owed-notification-expiry.test.tssrc/shared/agent-hook-listener-claude-task-notification.test.tssrc/shared/agent-hook-listener.tssrc/shared/agent-hook-listener/providers/claude-task-notification-delivery.tssrc/shared/claude-background-task-inventory.tssrc/shared/claude-owed-task-notifications.ts
💤 Files with no reviewable changes (1)
- src/shared/agent-hook-listener.ts
Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 9 remain after this review.
There was a problem hiding this comment.
✅ No new issues found. The prior review's two findings are both addressed, and the SessionEnd refactor is sound.
Reviewed changes
StopFailurenow owes a vanished shell: the shell-owe gate inclaude-task-notification-delivery.tswidened from the literal'Stop'toStop/StopFailure, matchingclaude-events.ts's turn-boundary predicate. Adds theowes a vanished shell on a failed turn end toocase.- Cap eviction no longer drops a running task: agent entries carry a
settledflag;recordClaudeBackgroundTaskLaunchevicts only a settled entry and, when none exists, leaves the new launch untracked instead of displacing live work.oweClaudeAgentTaskNotificationclearssettledon a re-owed (resumed) agent. Covered by thepast its cap never forgets a task that is still runningcase. - Empty task ids skipped:
claude-background-task-inventory.tsno longer pushes""for a whitespace-onlyid(the CodeRabbit nit). - SessionEnd drops owed state only for the owner: removed the eager map delete from
normalizeHookPayload's SessionEnd branch. A nested Claude's exit no longer wipes the pane's owed map; owner gating is now solelytransitionHookPresence, and the desktop clears viareconcileEndedProcessForPaneKeys→clearPaneState. New owner-exit and nested-exit tests added.
ℹ️ Relay does not clear its launched ledger on an owner exit
The removed eager delete used to also clean the relay's claudeLaunchedBackgroundTasksByPaneKey. The desktop still clears it (owner SessionEnd → reconcileEndedProcessForPaneKeys → clearPaneState), but the relay's applyEvent has no reconcile equivalent, so a remote pane's ledger entry and its unref'd lease timer survive an owner exit until pane teardown, surface retirement, or the next SessionStart. This is bounded and cannot republish a dead row — the stored row is ended/providerSessionOnly, and transitionHookPresence refuses an exit with no live owner, so the restate is dropped before forwarding — so it is not blocking. Worth noting only because the relay's owner-exit path is not exercised by a test today; the new relay expiry test covers the vanished-shell case.
Technical details
# relay launched-ledger retention after owner exit
## Affected sites
- `src/shared/agent-hook-listener.ts` SessionEnd branch — eager `claudeLaunchedBackgroundTasksByPaneKey.delete(paneKey)` removed in 16d3ac2
- `src/relay/agent-hook-server.ts:283-316` — `applyEvent` has no `reconcileEndedProcessForPaneKeys` equivalent; owner exit only sets `providerSessionOnly`
- `src/relay/agent-hook-request.ts:69` — lease armed for every stored row, including the SessionEnd row
- `src/shared/agent-hook-presence-transition.ts:9-31` — the guard that makes a restate for an ended owner a no-op
## Required outcome
- A remote pane held only by an owed notification must not leave stale ledger/timer state after the owning process exits; or the bounded retention must be an accepted tradeoff.
## Open questions for the human
- Is relay-side ledger retention until pane teardown/next session acceptable, or should the owner-exit path clear it on the relay too?deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏
16d3ac2 to
9541c48
Compare
There was a problem hiding this comment.
Caution
Some comments are outside the diff and can’t be posted inline due to GitHub limitations.
🟡 Minor · Merge owed-task entries during pane authority transfers. · listener-state.ts:264-272
src/shared/agent-hook-listener/listener-state.ts:264-272
🎯 Functional Correctness | 🟡 Minor | ⚡ Quick winMerge owed-task entries during pane authority transfers.
When a transfer moves a source task map onto a destination key that already has a task map, the generic helper replaces the destination map. Any destination-only owed task is lost. The pane-hold resolver reads owed-task state from this map, so the pane can settle before that task’s notification arrives if no other work holds it. Merge task entries at this task-specific move site.
Suggested fix
+function moveClaudeLaunchedBackgroundTaskEntries( + map: Map<string, ClaudeLaunchedBackgroundTasks>, + fromPaneKey: string, + toPaneKey: string +): void { + for (const [key, sourceTasks] of Array.from(map.entries())) { + if (key !== fromPaneKey && !key.startsWith(`${fromPaneKey}\0`)) continue + const destinationKey = `${toPaneKey}${key.slice(fromPaneKey.length)}` + const destinationTasks = map.get(destinationKey) + if (destinationTasks) { + for (const [taskId, task] of sourceTasks) { + destinationTasks.set(taskId, task) + } + } else { + map.set(destinationKey, sourceTasks) + } + map.delete(key) + } +} + export function movePaneCacheState( state: HookListenerState, fromPaneKey: string, @@ - movePaneScopedMapEntries(state.claudeLaunchedBackgroundTasksByPaneKey, fromPaneKey, toPaneKey) + moveClaudeLaunchedBackgroundTaskEntries(state.claudeLaunchedBackgroundTasksByPaneKey, fromPaneKey, toPaneKey)
ℹ️ Review info
⚙️ Run configuration
- Configuration used: Repository UI
- Review profile: CHILL
- Plan: Advanced
- Run ID:
f5d1ca48-2a21-4005-a9a9-da484202a4be
📥 Commits
Reviewing files that changed from the base of the PR and between 16d3ac2655226a5ebdfda7a1b5c4925bf2e3d8c0 and 9541c485b910c4308df27e3d79a7b15518f2469d.
📒 Files selected for processing (7)
src/main/agent-hooks/server/server-cleanup.tssrc/main/agent-hooks/server/server-lifecycle.tssrc/relay/agent-hook-request.tssrc/shared/agent-hook-listener.tssrc/shared/agent-hook-listener/listener-state.tssrc/shared/agent-hook-listener/pane-scoped-cache-entries.tssrc/shared/agent-hook-relay.ts
🚧 Files skipped from review as they are similar to previous changes (1)
- src/shared/agent-hook-relay.ts
Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 9 remain after this review.
9541c48 to
217eb64
Compare
nwparker
left a comment
nwparker
left a comment
There was a problem hiding this comment.
Claude owed task notifications remain host-owned; shared expiry timer, bounded task inventory, cancellation/session cleanup and relay parity reviewed. 66 focused tests and node typecheck pass. Missing task wake-ups can delay idle/completion up to 60 seconds.
Required PR Checks verify passed on this exact commit: 217eb643cc1a5f09731d6b9b6f5da636136d9b84.
nwparker
left a comment
nwparker
left a comment
There was a problem hiding this comment.
The captured task hook sequences reproduce the premature sidebar completion, but the original PR does not yet fix the automation completion path on current main. With real Claude Code 2.1.291 ready OSC bytes fed through OrcaRuntimeService.onPtyData, and a production hook row still working while a wake-up is owed, the actual runtime automation observer reports completed within 100 ms.
Additional production-ingress regressions reproduce an indefinite hold after Ctrl+C followed by a late tool hook on local and relay panes, and duplicate child-end delivery reopening already settled notification debt. Repairs and independent before/after regressions are being added before final approval or merge.
217eb64 to
cf051e0
Compare
|
The original status gap is reproduced through production HTTP/relay ingress. The same corrected captured-native-ready test against the real automation observer also reproduces premature completion on both parent main and the original PR; the repaired head is This head adds the narrow owner task phase, host-authoritative cancellation reconciliation, duplicate-notification controls, and indexed runtime status reads. The final integrated named gate passes 118 tests; node/web typechecks, default lint, the actual changed-code-quality gate (65 files, zero findings), reliability manifest and transcript privacy checks pass. The relevant broader suite passed 1,741 tests with 10 platform/optional skips before the final clean main integration. No current-head hidden-app or physical SSH/WSL/native Windows/Linux proof is claimed. The missing-notification fallback intentionally delays completion up to 60 idle seconds plus the existing poll interval; long foreground finishing work is held until its Stop or fresh native rest. Existing task and timer bounds are verified, and the actual indexed wait performs zero full-store reads with 512 irrelevant rows. The PR body now describes the final implementation and coverage limits. Merge remains held for fresh exact-head CI and current bot findings. @coderabbitai review |
There was a problem hiding this comment.
Actionable comments posted: 1
ℹ️ Review info
⚙️ Run configuration
- Configuration used: Repository UI
- Review profile: CHILL
- Plan: Advanced
- Run ID:
04a978b6-18ea-49f1-ba09-491f8a7b2403
📥 Commits
Reviewing files that changed from the base of the PR and between 9541c485b910c4308df27e3d79a7b15518f2469d and cf051e0e4f12595f64b0b176b338c9d8a12a747f.
📒 Files selected for processing (57)
config/reliability-gates.jsoncsrc/main/agent-hooks/server-claude-task-wakeup-lifecycle.test.tssrc/main/agent-hooks/server/server-cancel-verdict-latch.tssrc/main/agent-hooks/server/server-claude-status-rules.tssrc/main/agent-hooks/server/server-cleanup.tssrc/main/agent-hooks/server/server-constants.tssrc/main/agent-hooks/server/server-ingest-remote.tssrc/main/agent-hooks/server/server-lifecycle.tssrc/main/agent-hooks/server/server-persistence.tssrc/main/agent-hooks/server/server-remote-envelope-normalization.tssrc/main/agent-hooks/server/server-row-child-work-fold.tssrc/main/agent-hooks/server/server-status-identity.tssrc/main/agent-hooks/server/server-status-inference.tssrc/main/agent-hooks/server/server-status-update.tssrc/main/agent-hooks/server/server-types.tssrc/main/agent-hooks/wsl-hook-relay-deps.tssrc/main/agent-hooks/wsl-hook-relay-link.tssrc/main/agent-hooks/wsl-hook-relay-manager.test.tssrc/main/agent-hooks/wsl-hook-relay-manager.tssrc/main/orcad/orcad-entry.tssrc/main/runtime/__fixtures__/claude-ready-task-wakeup.meta.jsonsrc/main/runtime/__fixtures__/claude-ready-task-wakeup.txtsrc/main/runtime/claude-task-wakeup-indexed-read.test.tssrc/main/runtime/orca-runtime-resolve-authoritative-terminal-wait-permission.tssrc/main/runtime/orca-runtime-state-fields.tssrc/main/runtime/orchestration/worker-provider-session.tssrc/main/runtime/tui-idle-claude-task-wakeup.test.tssrc/main/runtime/tui-idle-hook-lane.test.tssrc/main/runtime/tui-idle-hook-lane.tssrc/main/ssh/ssh-agent-hook-interrupt-reconciliation.test.tssrc/main/ssh/ssh-agent-hook-interrupt-reconciliation.tssrc/main/ssh/ssh-relay-session.tssrc/main/startup/main-process-runtime-service.tssrc/relay/agent-hook-envelope-build.tssrc/relay/agent-hook-event-admission.tssrc/relay/agent-hook-interrupt-reconciliation.test.tssrc/relay/agent-hook-interrupt-reconciliation.tssrc/relay/agent-hook-request.tssrc/relay/agent-hook-server.tssrc/relay/relay-agent-hook-runtime.tssrc/relay/wsl-agent-hook-relay.tssrc/shared/agent-hook-cancel-verdict-latch.tssrc/shared/agent-hook-interrupt-reconciliation.tssrc/shared/agent-hook-listener.tssrc/shared/agent-hook-listener/listener-event.tssrc/shared/agent-hook-listener/main-agent-turn-state.tssrc/shared/agent-hook-listener/providers/claude-events.tssrc/shared/agent-hook-listener/providers/claude-lifecycle-events.tssrc/shared/agent-hook-listener/providers/claude-pane-hold-evidence.tssrc/shared/agent-hook-listener/providers/claude-roster-state.tssrc/shared/agent-hook-listener/providers/claude-status-build.tssrc/shared/agent-hook-listener/providers/claude-task-notification-delivery.tssrc/shared/agent-hook-relay.tssrc/shared/agent-hook-row-child-work-fold.tssrc/shared/agent-status-types.tssrc/shared/claude-owed-notification-resource-contract.test.tssrc/shared/claude-owed-task-notifications.ts
💤 Files with no reviewable changes (1)
- src/main/agent-hooks/server/server-constants.ts
Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 6 remain after this review.
…s arrive Claude wakes the main agent for each background task that ends, after the task stops running. "Nothing running" was read as done, so automations closed the terminal before the final turn. Fixes stablyai#23942
…cap and nested exits A failed turn's task list now marks a vanished shell owed like a normal turn end does. Past the cap only an already-announced sub-agent is forgotten. A process exit clears what is owed only where the server admits it from the pane's owner.
…istener-state listener-state.ts passed the 300-line limit once main's and this branch's additions met.
6f7e807 to
1679b0c
Compare
There was a problem hiding this comment.
Actionable comments posted: 1
ℹ️ Review info
⚙️ Run configuration
- Configuration used: Repository UI
- Review profile: CHILL
- Plan: Advanced
- Run ID:
00e43d46-69ec-45fa-adea-977b2b2d012e
📥 Commits
Reviewing files that changed from the base of the PR and between cf051e0e4f12595f64b0b176b338c9d8a12a747f and 1679b0c.
📒 Files selected for processing (4)
config/reliability-gates.jsoncsrc/main/agent-hooks/server-claude-task-wakeup-lifecycle.test.tssrc/main/agent-hooks/server/server-status-inference.tssrc/shared/claude-owed-task-notifications.ts
Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 1 remain after this review.
|
The reported local results do not establish exact-head CI or physical SSH/WSL and native Windows/Linux coverage. This review request does not change the merge hold. ✏️ Learnings added
|
nwparker
changed the title
nwparker
left a comment
nwparker
left a comment
There was a problem hiding this comment.
Independently reproduced premature completion using real captured Claude hook events and ready-title bytes through actual desktop/relay HTTP ingress, the runtime PTY path, and the automation completion observer. The original PR still completed too early. The repaired owner-produced notification/finishing-turn phase now holds completion until the finishing turn ends, including duplicate notifications before Stop for both agents and shells.
Reviewed local and execution-host-owned SSH/WSL cancellation, exact session/launch/turn fences, permission restoration, duplicate task ends, resumed task IDs, missing notifications, lost final Stop, snapshot stripping, and indexed readiness reads. A fresh bot finding on blank/padded launch IDs is fixed with canonical normalization; twelve desktop/relay production-ingress cases reproduce phantom debt or unmatched notifications before the fix and pass afterward. The actual runtime regression with 512 unrelated agents proves zero full status snapshot reads, including empty indexed results. Task records are bounded and timers are coalesced and unref'd; no recurring polling or new status cache was added.
Also reproduced a child ending before its parent launch event: both actual desktop and relay completion observers released too early. The existing bounded task map now preserves provisional end/receipt evidence until an own-main launch confirms ownership. End-before-launch, notification-before-launch, and notification-before-end cases hold through the finishing turn and complete after its Stop without phantom debt. Unknown, restored, wrong-session, and unanchored child evidence cannot mint an obligation; nested children remain provisional. Duplicate ends keep their first clock and real resumed tasks rearm correctly.
A fresh compatibility check exposed an identical pre-existing failure on its exact main baseline: a test expected @streamparser/json to be missing although main already retains its pinned development dependency. The test-only correction executes the actual released dense parser and verifies parsed output, match bounds, and malformed-input rejection. Separate synthetic missing-package refusal tests remain intact; no dependency, production parser, or workflow was changed.
Full unit CI also reproduced main's duplicate directory-cache key enumeration. The existing alias index and subtree purge now reuse the same complete key list only when the React updater still owns that exact cache object; otherwise purge enumerates fresh state. Alias detection remains intact, including Windows case-duplicate keys. The original 2,001-visit budget is unchanged and passes; 61 watcher regressions, node/web typechecks, formatting and changed quality pass. All other prior-head CI checks passed, including physical Windows/macOS SSH hosts and Docker SSH watcher journeys. Fresh full CI on this repaired head is required and has passed before this approval.
Integrated current main after two test conflicts: the original cache-enumeration budget remains strict, and released-parser assertions combine the upstream sparse-input check with bounded matches, path/line and malformed-input controls. All eight author/repair commits and 68 core reviewed files remain intact; all 143 upstream reliability gates survive alongside the new gate. The integrated tree passes 182 affected tests, node/web types, manifest validation and changed quality.
The disclosed low-risk limits are the 60-second idle grace for a genuinely missing notification and the 256-record per-pane tracking bound. Live SSH and native Windows/WSL were not exercised locally; relay/routing/owner tests and fresh platform CI cover the changed paths. The exact configured reliability gate passes 148 tests, plus 40 existing lifecycle/retirement/expiry regressions; typechecks, full lint, and the actual changed-code quality gate pass. Fresh full PR Checks CI on this exact head is green and all review threads are resolved. The repaired behavior satisfies the conditional merge request.
…e the row after the activity window Upstream stablyai#24878 added a relay case where a SessionStart for another Claude session supersedes the row and must void a pending interrupt. The fork's session guard drops a competing session that reports while the owner was active in the last 30 seconds and the sender is not the owner's process, so the row never changed and the interrupt still applied. Production behaviour is the guard's intent; the case now moves the clock past the activity window first, the same way the fork's own relay ownership test admits a genuine replacement. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EjcinZAhBx3gqNYt9urQE2
ELI5
Claude can finish a background task before its follow-up message reaches the main agent. Orca now keeps the task cycle active until that message and its finishing turn complete, so an automation cannot close the terminal in the gap.
What Changed
The execution host tracks bounded task-notification debt and publishes an optional Claude task phase. Terminal waits check that phase against the current session, launch and host, using the existing per-pane status index. Retained ready text cannot finish the task cycle; fresh native rest can recover a lost finishing Stop. Duplicate notifications do not restart the cycle. Current child ends that beat their parent launch stay provisional in the same bounded map; a matching main launch confirms debt, while an earlier delivered notification stays settled.
Required CI also exposed two inherited regressions. The released-parser test now checks real parsing and bounded matches; file watcher reconciliation shares its original key list between the path index and purge. The purge reads fresh keys if a newer state reaches its functional updater, and retains all Windows case aliases.
Canceled local and remote turns keep their original idle boundary despite late tool hooks. Remote cancellation is reconciled to the execution owner with an optional, session/launch/turn-fenced command over the current SSH or WSL transport.
Why
Hook status alone did not prevent the actual automation observer from accepting Claude's native ready title in the owed-message gap. The narrow owner phase fixes that case while preserving ordinary Claude readiness. The execution host owns cancellation and expiry; readers do not invent remote process state. A missing notification expires after 60 idle seconds instead of leaving the pane working forever.
Linked Issue
Fixes #23942
Visual Proof
A real scrubbed Claude PTY capture and its metadata are committed as the automation oracle. Production HTTP/relay ingress and the actual runtime completion observer replay identical captured hook and ready-byte sequences before and after. No current-head rendered Orca screenshot, video, or hidden-app validation is claimed.
Testing
Review
One shared host deadline timer, at most 256 tasks per pane, and zero full-store reads with 512 irrelevant agents on the indexed wait path. Optional wire fields are safe for mixed versions; older hosts retain their existing behavior. No new stream opcode, reader cache or remote timeout authority.
Agent skill upstream boundary
Notes
A lost notification intentionally delays completion by up to 60 idle seconds plus the existing poll interval. Experimental gate documents platform and soak gaps. Final exact-head CI and bot findings must pass before merge.
Checklist