feat(peer): surface and answer remote descendant requests from the parent

What: The parent and a client connected only to its environment can see and answer eligible remote descendant questions/approvals with local-equivalent policy.
Why: The approved parent spec identifies a location-dependent ownership or lifecycle gap in this workflow.
So what: Implement this independently provable slice after its native blockers are satisfied; review the exact candidate before integration.

Parent

#379

Outcome

The parent and a client connected only to its environment can see and answer eligible remote descendant questions/approvals with local-equivalent policy. This contributes to every machine and Drafter in one workspace and Promachos directing work across environments.

Elon record

Requirements and who asked: User-requested native cross-machine tool/lifecycle parity, specified and independently approved in #379.
Deleted: Parallel orchestration systems, privilege bypasses, schema-only completion claims and unrelated delivery work.
Bottleneck: This slice's application-service ownership and its stated predecessor contracts.
Checked myself: Source/review evidence in #379 at 3caa62c; actual current base and owned call sites must be rechecked before implementation.

Acceptance criteria

Reuse feat(peer): surface and answer remote descendant requests from the parent #369 existing missing-input problem; mirror destination-owned pending-request identity/revision/status into the parent banner and agent task status across all clients.
Use qualified remote addresses in pending tools and notices; never pass a foreign id as local childThreadId.
Verify transitive local-edge plus delegation-pair ancestry for immediate parent and grandparent responses; deny self, unrelated callers and OAuth clients without proof, including missing-caller user-input bypasses.
Preserve local ancestor-delegable runtime approvals; genuinely human-only actions use authenticated source-client relay and cannot be claimed by agent routing arguments.
Answer/cancel/dismiss/restart races converge without duplicate provider prompts; source-only clients can navigate and act without pairing separately to the destination.

Non-goals

Public release/deployment, account/link capability changes without explicit consent, physical-device use, unrelated source cleanup, package version bumps, live user data testing, and declaring the whole parent complete from this component.

Blocked by

None. #381 is independently reviewed and internally integrated through PR #392 at c8423b4, tree identical to reviewed 0031306. The #379 planner verified the native edge removal; final combined acceptance remains pending.

Required proof

Extend existing pending-request and question-reactor public registration seams plus parent projection/client behavior; prove A parent -> B child -> C grandchild, source-only human action, async input, duplicate/stale response and untrusted actor denial. Keep original #369 review evidence as provenance. Include focused behavioral checks, sanitized identity-bound evidence and independent exact-candidate review. CI owns repository-wide checks.

Prior record

This extends the existing #369 outcome rather than creating a competing Issue. Original diagnosis: local child-input discovery walks local shells, while the parent remote task record lacks a waiting/pending-request projection; review of #368 identified it. The original existing title was "feat(server): surface remote child waiting-on-input on the parent". #368 remains the already completed remote lineage work, not a blocker for this slice.

fix(web): scope remote lineage identity and guard cross-environment o…

fix(web): show remote delegated children in Lineage (#368)

What: Read-only #369 ownership and proof mapping is complete on the assigned c842 base; no implementation has been made.
Why: #381 supplies qualified pending metadata and source inspection, but response authorization remains local-only and source clients have no request action path.
So what: The #379 parent chooses shared-hook ownership and sequencing with the active #382 writer, then resumes this same #369 job with concrete ownership.

Identity, authority and phase

Exclusive app-bound workspace /Users/rocky/.t3/worktrees/chromeria/feat-369-remote-pending-requests, branch feat/369-remote-pending-requests.
Assigned base and actual HEAD c8423b4a9706c18f83efdd27db0d3c7c2571e2a8; tree 305d63ed38070ac5b0f382e603eec11425118cf0. git status --porcelain=v1 empty; base...HEAD counts 0 0. No tracking upstream is configured for this new branch, so tracking divergence is unknown, not a claimed remote match.
Origin https://github.com/toolboxmd/chromeria.git; upstream https://github.com/pingdotgg/t3code.git. git worktree list --porcelain and app t3_thread_read agree on the exact binding. No checkout/rebase/fetch/reset/topology operation performed.
App thread thread:delegated-task:command%3Amcp%3A0dd33a23-3bda-415f-8c7f-53320ec4c2ad%3Adelegate-task%3A379-369-sol-medium-implementation-v1; parent is the existing spec: native cross-machine control across linked environments #379 Rocky planner. Live t3_environment_read: Rocky, ad2ba7ac-cc85-4ec9-8d59-a65c7cbc104c, darwin/arm64. Live catalog supports the exact Color. Actual t3_thread_configuration returns instanceId=codex, model=gpt-6.1-sol, options reasoningEffort=medium; runtime is Codex/gpt-6.1-sol/MEDIUM. Catalog default low is not substituted for actual selection. No fallback/delegation.
feat(peer): surface and answer remote descendant requests from the parent #369 app-linked successfully. No PR created or worked on; feat(peer): continue remote delegated jobs and stop later work #392 is prerequisite evidence, not linked as this job's PR.
Current feat(peer): surface and answer remote descendant requests from the parent #369 body/comments read fully, together with full spec: native cross-machine control across linked environments #379 body/comments/frozen v5 approval, fix(web): show remote delegated children in Lineage #368 provenance, feat(peer): continue and stop remote delegated jobs after their first result #381 original/R1/R3/R4 handoffs and integrated acceptance, both final R4 reviews, parent actual acceptance, feat(peer): remote children return results and delegate onward within durable grants #380 final audit/acceptance/special adjudication, and feat(peer): transfer attachments and make remote retries preserve operation identity #382 current assignment. Live native feat(peer): surface and answer remote descendant requests from the parent #369 blocked_by is []; feat(peer): continue and stop remote delegated jobs after their first result #381 remains open under the parent.
Git diff from reviewed 00313067e47a64f1a39ffbbb83ec7fb3e74f601b to this HEAD is empty. Live feat(peer): continue remote delegated jobs and stop later work #392 is merged internally into feat/379-cross-machine-parity, merge c842. This is not spec: native cross-machine control across linked environments #379 completion.
Healthy canonical global-instruction hook and full triad/preferences supplied this turn. Canonical operations at /Users/rocky/dev/toolboxmd/agentsmd/skills/operations/SKILL.md: research/mechanics/recall, software-design/state-and-boundaries/change-existing-systems, Elon algorithm/current-constraint, verification/test-design read. Great Library README, canonical setup boundary, root/system glossary, connected AgentsMD/Observer/ContextMD glossaries and docs/internals/glossary.md read fully. No GLOSSARY-MAP or additional domain glossary/ADR files were found by the bounded file inventory. No new terminology or architecture is approved here.

This outcome supports Promachos directing Drafters across environments under the existing Objective. The frozen design is retained. Parent alone owns topology, real isolated clients/machines/providers, integration and shipment. #382 owns all its necessary shared contracts/identity/authority/transport/MCP namespace/global registration/runtime/persistence/intake/receipt/hooks and Spectrum identity-only callers under 6242479006. #384 remains stopped. No source writes are authorized in this phase.

Actual source and ownership map

All source references below name the assigned c842 tree. Line numbers locate the concrete current owner, not a proposed new module.

Boundary	Actual files and callers	Verified behavior / #369 gap
Destination request truth	`packages/contracts/src/orchestrationV2.ts` runtime-request, approval/user-input item, response-capability and command definitions; `apps/server/src/orchestration-v2/ProjectionStore.ts:getRuntimeRequest/getRuntimeResponseContext`, `orchestration_v2_projection_runtime_requests`; `orchestration_events`	Destination owns status, request/item contents and provider callback identity. Peer metadata is observation, not authorization or a substitute request.
Registered agent tools	`apps/server/src/mcp/toolkits/thread/tools.ts:170-225`, `handlers.ts:190-260`, `childThreads/pendingRequests.ts:6-69`	List/read/respond are actually registered through ThreadToolkit/McpToolAccess. List returns IDs only; read returns question or approval content; respond takes answers/decision and emits `runtime-request.respond`. Neither public response input nor native response command carries expected revision. Current response command ID is fresh `newCommandId()`, not a stable caller response key.
Current agent ancestry	`childThreads/pendingRequests.ts:isCallerAncestor/readPendingRequest`; `mcp/threadAccess.ts:readCaller/loadCaller/readThread`; `McpToolAccess.ts:writesThreads`; `mcp/linkOrigin.ts`	Predicate walks only local native `subagent` edges, rejects same target and missing caller. Approval reads require it. Respond's extra check is conditional on `caller !== undefined && caller.id !== target`: an outside OAuth caller without a local caller bypasses the extra question ancestry test. This is source-established missing-caller exposure, not an executed exploit. Local self questions currently remain permitted; self approvals are refused by readPendingRequest. The existing self-question allowance is a source fact, not proof of authority for a remote self response; the frozen ancestor-action and self-denial requirements remain controlling.
Effective permission and deferred effects	`McpRouting.ts:withContext/invoke`, `McpToolAccess.ts:writesThreads`, `DispatchModeLimit.ts`, `peer/PeerAuthority.ts`, `peer/peerEffectAuthority.ts`, `EffectWorker.ts:775-800`	Modern routing uses server-validated durable authority and current/frozen/contributor/actual-target intersections. `runtime-request.respond` effects are included in claim-time `checkEffect`. Remote ancestry must become an additional validated policy input, not replace these checks, invent a local caller, or use another thread's bearer.
Current peer proof/control	`packages/contracts/src/peerAuthority.ts:PeerReturnPair/PeerAuthorityEnvelope`; `peer/PeerAuthorityRecord.ts`; `PeerAuthority.ts:54-107,1034-1195,1260+`; `PeerMcpClient.ts:ensureChannel/lifecycleRequest`	Durable pair binds parent environment/thread/task to child environment/thread and grant generation. Existing bounded Frame union is validate, return, pair; lifecycle operations are observe/resultText/acknowledge/stop/snapshot/view. There is no registered pending-request ancestry/action frame. `verifyTask` authenticates exact pair/grant/resource for lifecycle reads, not live acting-ancestor permission to approve. `validateReturn` alone is not sufficient interactive ancestry proof.
Current qualified observation	`packages/contracts/src/peerTask.ts:PeerTaskSnapshot`; `peer/PeerTaskLifecycle.ts:235-430,470-489`; `peer/RemoteDelegation.ts:modernFollow/observe`; `Orchestrator.ts:dispatchRemoteObservation`	Snapshot requests contain environment/thread/request ID, latest runtime-request event-sequence revision, kind and responseType. Pending status is implicit in filtering; no full question/approval content or authenticated action is included. Local descendants and nested remote snapshots are aggregated, max 256 local visited threads / 100 published requests with incomplete flag. Stale/unavailable remote state is preserved. Nested qualified metadata does not itself prove ancestry.
Source task status/wait	`mcp/OrchestratorMcpService.ts:readTask/readRemoteTask` around 1590-1695; toolkit `orchestrator/tools.ts`; `packages/contracts/src/orchestratorMcp.ts`	Remote result retains `childThreadId:null`, qualified remoteChild and remoteObservation.snapshot. `hasPendingQuestion` derives pendingQuestions; approvals have metadata but no dedicated wait-ending pending approval signal here. Source status/wait must present eligible questions/approvals without converting foreign IDs to local children or erasing later-work/original-result semantics.
Question and completion ownership	`orchestration-v2/DelegatedTaskQuestions.ts`, `DelegatedTaskQuestionReactor.ts`; `childThreads/requestWake.ts`; `Orchestrator.ts:reconcileDelegatedTaskQuestionWithParentLockHeld` at12808+; `RemoteDelegation.ts:796`; existing delegated completion owners	Local requestWake walks local ancestors, serializes recipient dispatch and uses receipts. Native question helper exits on null child; remote helper already emits qualified user-input notices using parent-local task ID and omits local childThreadId, cancelling stale queued notices. It skips approvals and does not authorize responses. Reactor's comment saying approvals cannot be answered is stale against current fork pending tools; update only with assigned implementation. Preserve existing queue/result/completion owners rather than add a second reactor/orchestrator.
Command/state/effect atomic boundary	`Orchestrator.ts:8311-8540,12383-12440,12785+`; `EventSink.ts:commitCommand` at688+,790-815; `CommandReceiptStore.ts`; `EffectOutbox.ts`; `RuntimeRequestService.ts`	Under destination thread lock, response requires pending + resumable request. It resolves request/item/node and commits receipt/events/effects together. Live response schedules one `runtime-request.respond` effect; RuntimeRequestService verifies resolved/live/matching active provider session before adapter invocation. Async message answer resolves and dispatches `async-answer:<requestId>` in the SAME event transaction, steering supported active turns or queueing otherwise. Native command receipt replay is thread-bound; same-key changed semantic inputs need correct owner binding, not reliance on current random IDs. Revision checking only before this serialized owner would leave a race.
Stop/dismiss/restart lifetime	`Orchestrator.ts:dispatchThreadUserInputDismiss` at8540+, runtime stop cancellation around9700, `dismissInvalidatedSubagentQuestions` at11651+; `ProviderSessionManager.ts:716-770`; `ProviderRuntimeRecoveryService.ts:360+`	Stop/session release invalidate live requests; startup/shutdown expire/cancel lost live callbacks as not_resumable. Message questions have separate durable lifetime and invalidation. A remote retry must reconcile committed resolution, distinguish stale/dismissed/unavailable, and never generate another provider prompt. No guarantee of exactly-once external callback follows from SQLite alone.
Authenticated human local transport	`ws.ts:1889-1940,2488-2504,3291+`; `auth/RpcAuthorization.ts`; `packages/contracts/src/clientRpcPermissions.ts`; client-runtime `operations/commands.ts:864-906`	WS session is authenticated and dispatch authorized with orchestration:operate. Local approval/answer/dismiss become native commands. Current remoteTasks has source-only view and Stop, no request read/respond/dismiss action. Source client action needs server-derived actor provenance from this ingress and exact remote identity, checked again at B. Public MCP arguments must never mint this actor. A blanket agent-approval ban would contradict existing fork policy.
Source-only viewer and navigation	`peer/RemoteDelegation.ts:view` at1045+; `PeerTaskLifecycle.ts:view` at491+; `packages/client-runtime/src/state/remoteTasks.ts`; `apps/web/src/components/chat/RemoteTaskViewer.tsx`; `apps/mobile/src/features/threads/RemoteTaskViewer.tsx`	Current viewer shows counts/messages/children/results and guarded Stop only. It cannot inspect question/approval content or answer/dismiss. Nested view permits current destination's local subtree, and rejects an environment different from that destination. A grandchild on C requires the verified delegated route through B, not a generic proxy or saved C client connection.
All opens / banners	Web `chat/useSubagentOpen.tsx`, `V2LifecycleRow.tsx`, `V2ItemInspector.tsx`, `BackgroundSubagentOpenButton.tsx`, `ThreadRelationshipsControl.tsx`; mobile `ThreadAgentsSheet.tsx`, `thread-subagent-group.tsx`, `thread-work-log.tsx`; `ChatView.tsx:7325+`; `ThreadDetailScreen.tsx:520+`	Negotiated-v1 opens already prefer A viewer before B connectivity, including transcript/notification/background/inspector. Preserve local one-argument callback and legacy direct qualified navigation. Input-specific web banner and mobile floating control still consume local child shells and navigate local ID. They need a real qualified pending target, not a fabricated shell. Desktop shares web rendering; no separate Electron action service was found for this path.
Unopened-thread attention	client-runtime `state/threadShell.ts:50-89,240-281`, `models.ts`, `threadInbox.ts`, `threadSettled.ts`, `threadSnoozed.ts`; web `ThreadNotificationCoordinator.tsx:132+`, `Sidebar.logic.ts`; mobile `threadListV2.ts`; server `relay/AgentAwarenessRelay.ts:526+`	Existing pendingSubagentRequest is derived solely from local shell lineage; local input banner BFS also sees only that environment's shells. Shell contract at orchestrationV2.ts:2093-2192 contains no remote task snapshot/request summary. Detail-only viewer changes therefore cannot fix unopened parent attention. Coordinator independently recomputes the local map; updating another atom alone misses it. Relay skips hidden subagents and publishes per-thread shell state, so mobile push/awareness needs an explicit outcome decision and source-shell propagation, not a claimed consequence of the web fix.
Real production composition	`server.ts:604-607,763,1136-1149`; `orchestration-v2/runtimeLayer.ts:201,362,394+`; `mcp/McpHttpServer.ts:988-1065`; toolkit registration through McpToolAccess	Production starts question reactor and remote follower, explicitly provides PeerTaskLifecycle from shared runtime services, exports existing memoized EffectOutbox and registers authenticated GET/POST `/mcp/peer-control/tasks`. Capability v1 requires actual composed service; partial authority composition returns409, unauthenticated returns401. Any #369 extension needs actual production registration/export proof, not testkit injection alone.

Provider semantics to preserve

Codex: Adapters/CodexAdapterV2.ts:4385-4420,5164+,7735+: synchronous user input/approvals use pending Deferred callbacks; async agentMessage delivery creates a responseCapability:{type:"message"} request and survives its original turn. CodexAdapterV2.test.ts:2657+ covers that distinction. Do not require a live provider callback for message answers.
Claude: Adapters/ClaudeAdapterV2.ts:8307+,9646+: AskUserQuestion/live approval Deferreds; answer mapping and provider session lifetime remain adapter-owned. No invented async callback conversion.
Grok: packages/provider-grok/src/server/adapter.ts selects shared ACP adapter plus xAI ask-user extensions (xaiAcpExtension.ts). Shared provider-acp/src/server/adapter.ts:7934+ fences callback generation, serializes runtime transition, refuses duplicate resolution and waits for native response acknowledgement. Preserve this acknowledgement, not merely success when a Deferred is completed.
OpenCode: packages/provider-opencode/src/server/adapter.ts:3988+ maps exact native question/permission IDs into SDK reply; src/server/v2/adapter.ts:4662+ is the other supported implementation. Existing adapter choice controls which runs. Do not call the wrong session or send a second question.reply. Both need focused preservation coverage if the shared boundary changes.
Cursor: packages/provider-cursor/src/server/adapter.ts:2546+ explicitly rejects interactive approval responses because Cursor SDK does not expose them. Truthful unsupported is required, not a fabricated question fixture.
Antigravity: Adapters/AntigravityAdapterV2.ts uses the shared ACP adapter with flavor-specific approval/question extraction/response functions in provider/acp/AntigravityProtocol.ts. Inherits shared generation/acknowledgement behavior; live catalog currently says disabled/uninstalled. That is a live capability observation, not authority to install it.

Runtime-request approvals are not a synonym for every human operation. Provider account consent (providerAuth*, AuthProvidersManageScope), link grants (AuthAccessWriteScope), and secrets/SecretRequests.ts / SecretRequestAnswerInput are distinct owners. #369 must not relabel every approval human-only or route secret values through pending metadata. #385 remains owner of remote-secret acquisition/consumption. At this base no general verified human assertion exists in pending-request MCP input or native response command; the action boundary must derive it from authenticated source transport, with an unforgeable internal distinction, before any human-only relay is offered.

Minimum consumer requirements for the parent to assign

These are demonstrated consumer needs, not new schema, exported hooks, or authorized source changes:

One destination-owned pending-request policy/read/action path callable from existing MCP and authenticated client ingress. It consumes exact qualified address + expected request revision + intent + stable caller action identity. Under serialized native acceptance it checks still-pending revision, kind/answer shape, resumability, actual target, current/frozen/all-contributor authority and verified actor; then uses the existing native receipt/effect boundary. Existing local calls remain equivalent.
The existing authority/pair owner must answer exact acting environment/thread -> descendant request ancestry across mixed local subagent edges and authenticated durable delegation pairs, live at each authority owner. Same peer session, route metadata, contextThreadId, pair possession, projection membership and labels are insufficient. Immediate parent, local grandparent, A->B->C grandparent and return to a previously visited environment are distinct cases. Refuse self approval, unrelated/fork/non-proved imported/OAuth/missing-caller response authority. Preserve allowed local ancestor runtime approvals; do not turn current self-question allowance into remote self-response authority.
A typed bounded source-client operation must carry server-derived authenticated client actor separately from ancestor-agent authority. Read/answer/deny/cancel/dismiss resolve the exact task/resource/request revision through existing peer control ownership. Source and destination scopes/current grant apply; no MCP input switch selects human. No generic reverse RPC or destination bearer exposure.
The source detail projection needs actionable qualified request content/status/revision and refresh on resolution; parent shell/attention needs a compact source-owned signal rather than subscribing to every destination transcript. Async nonblocking questions and blocking input must keep their existing local distinction. Metadata absence/incomplete/stale must not appear as a fresh no-pending answer.
feat(peer): transfer attachments and make remote retries preserve operation identity #382's stable identity/semantic-conflict/receipt owner must be reused for response retry and native downstream command identity. A response-lost-after-commit retry returns/reconciles the same receipt; a different answer under that key conflicts; a changed request revision fails without another provider callback. This cannot be implemented independently in feat(peer): surface and answer remote descendant requests from the parent #369 with a forwarding response cache.

Representative caller sketches, deliberately semantic rather than invented API signatures:

Common agent path:
  A parent task_status -> remoteObservation.snapshot.requests entry {B, child, request, revision}
  existing pending_request_read addressed to B/child/request -> fresh qualified content/revision
  existing pending_request_respond(intent, expected revision, stable action key)
  B verifies A's ancestor relationship/current authority -> native resolved receipt -> one callback

Failure path:
  same address arrives from outside OAuth / forged actor=human / unrelated thread
  destination cannot prove required ancestor or authenticated client actor
  deny before runtime-request.updated, queued message, outbox effect or provider callback

Awkward mixed chain:
  A local grandparent -> local child -> B delegated child -> C grandchild asks async question
  A-only client opens task/request through source, no B/C saved connection
  verified local+pair chain resolves C request; another actor dismisses at C before answer
  expected revision is rejected as stale; replay of a previously committed answer reconciles its
  receipt; neither path revives the request, replaces original task result or prompts twice

Collision and possible private scope

No #369 edits are assigned yet. File separation does not remove coupling to #382 authority/identity/receipt changes.

Candidate area	Disjointness assessment
Web/mobile RemoteTaskViewer request presentation, source input banner/floating control, existing navigation consumers and focused interactions	Potential #369-private consumer set after the shared typed request/action/projection contract is assigned. Can reuse negotiated source viewer, but cannot truthfully complete responses now. `ChatView`/ThreadDetailScreen may be large shared entrypoints; exact hunks still need parent assignment.
`childThreads/pendingRequests.ts` policy, registered thread toolkit handlers/tools, `DelegatedTaskQuestions.ts` / requestWake and focused tests	Domain work belongs to #369 outcome, but is not yet independent: verified remote actor/ancestry and stable response identity depend on #382-owned intake/authority/registration. Parent must assign concrete boundaries or sequence it. A local-only policy patch would leave remote completion unproved.
client-runtime shell/remote state, contracts, ProjectionStore, WS, orchestrator/native receipt checks, effect authority, peer frames/service/production runtime	Shared semantic state and ownership. Keep with existing sole writer or transfer exact completed baseline/hunks serially. Do not add speculative callback interfaces in #369 or ask #382 to implement them directly.
Provider adapters	Existing semantics inspected; no adapter change justified merely by remote routing. Change only for a discriminating supported-provider failure, with parent ownership and focused adapter proof.

Recommendation for the parent: obtain #382's concrete reviewed identity/intake/receipt boundary, then give #369 exclusive request-domain + client scope and explicitly assigned shared seams (or keep those shared edits serial with their sole owner). There is no complete end-to-end #369 implementation that is currently independent of the active shared writer. No new component PR or parallel writer is justified by this map alone.

Representative registered red/proof targets, not executed this phase

The read-only constraint forbids source/test edits and provider fixtures. No red, green, behavioral test run or acceptance is claimed here. The following identify the existing faithful seams for the next authorized phase.

Existing target	Discriminating observation required
`childThreads/pendingRequests.test.ts` (registered ThreadToolkit), `mcp/OrchestratorMcpToolkit.integration.test.ts`, `peer/PeerAuthority.workflow.test.ts`	Immediate parent and mixed local/pair grandparent can read/answer eligible request/approval; self approval, unrelated/fork/OAuth/missing caller/spoofed human/other-link/unknown imported ancestry deny with zero events/effects/callbacks. Existing test table lacks missing-caller question case. Pair metadata alone must not make a negative pass.
Existing PeerAuthority workflow + real transport/SQLite/native runtime kernel	A->B->C, A can dial B but B cannot dial A, concurrent control validation during held tool, reconstructed services, revoked/expired/relinked grant, narrowed source/current target/all-contributor auto/plan fences. Positive lawful local-equivalent ancestor approval alongside denials.
`orchestration-v2/runtimeLayer.test.ts` async case at1025+, `RuntimeRequestService.test.ts`, `ThreadMessageIntake.test.ts`, native EffectWorker	Lost reply after accepted commit, concurrent identical response, changed-answer conflict, stale revision, answer vs dismiss/Stop/session exit/restart, old request vs newly created request. Exactly one request resolution/async message/accepted native effect; zero extra provider invocation on replay. Observe persisted event/Deferred or EffectWorker.drain, never sleeps. Message answer after original provider exits must still work; lost live callbacks remain expired/not_resumable.
`childThreads/requestWake.test.ts`, `orchestration-v2/RemoteDelegatedTask.test.ts`, `peer/RemoteDelegation.test.ts`	Qualified user-input and approval notices, one eligible wake, stale queued notice disposal, original result immutability and later result continuation, pending/stale/incomplete truth, original Stop history preserved. Include original result already published with later pending work.
`McpHttpServer.test.ts`, `orchestration-v2/runtimeLayer.test.ts`, native production-kernel workflow	Actual service export/runtime/server route/registered toolkit and typed negotiation. Auth401, partial409/unsupported, denied403, real composed authorized positive. An explicitly injected new owner alone is insufficient after #381's production counterexample.
client-runtime `threadShell.test.ts`, `remoteTasks.test.ts`, `threadRequests.test.ts`, `commandPermissions.test.ts`; web `V2LifecycleRow.remote.test.tsx`, notification coordinator tests; mobile request permissions/presentation tests	Source-only read/act/navigation from banner, transcript, notifications, inspector, relationships, mobile group/agents sheet/floating control; unopened parent attention, same foreign ID in two environments, B connected/disconnected, stale refresh, close/reopen/reverse dismissal, read-only client denied. Preserve native/local one-arg opens, unmarked legacy navigation, async nonblocking indicators. Tests click real consumers and observe actual source RPC/content, not markup or mirrored prop wiring.
Parent-only actual acceptance	Real isolated source-only web/desktop/mobile, provider-supported questions/approvals, physical machine topology and restart/process/lifecycle proof. No worker-owned servers/browser/devices/fixtures this phase. Both new independent exact-candidate Sol MEDIUM and installed Grok4.7 MEDIUM reviews remain parent-dispatched after implementation.

Upstream-first evidence

Read live upstream main SHA 07cb90286656393e395192d82f5c9980b604ef31 and its actual apps/server/src/mcp/toolkits/thread/handlers.ts: readQuestion/list/read/respond accept user-input only and do not implement this fork's ancestor approval policy. Main content lookup for childThreads/pendingRequests.ts and peer/RemoteDelegation.ts returned404 at that exact SHA. This is an inspected-file absence, not a claim that every upstream file was searched.

Queries: gh search prs 'pending request remote' ... --state open, Issues 'remote child input', PRs 'ancestor approval' each returned none; broader open PR 'remote delegation' found #20037 and #16719. Broader Issues 'pending request' found matching/adjacent reports including #15082, #17536, #19984, async/restart reports.

Open #20037 current head 08600ee482a27f366d1f1253b15f3d0200697569: inspected body/files and relevant complete service/contract changes. It ends waits on task-owned live approval metadata, explicitly excludes remote delegation/reactor/UI/approval decisions, and says the parent cannot approve. Adjacent wait signal/proof ideas, not the approved complete feat(peer): surface and answer remote descendant requests from the parent #369 behavior or a policy to import.
Merged #19630: body/files and actual fork consumers show the local-shell parent-attention approach already present here. It explicitly leaves server mobile push/Live Activity separate. It does not supply source-only cross-environment request actions. No port made or full new diff acceptance claimed.
Open #16719 is the original remote-delegation provenance, already imported/extended by this fork. It is not the later feat(peer): continue and stop remote delegated jobs after their first result #381/feat(peer): surface and answer remote descendant requests from the parent #369 complete lifecycle/action fix.
#17536 is closed as duplicate of #15082. #19984 body/comments retain an unexplained already-answered/not-found observation; the later single-question repro did not reproduce it. Do not assign that report a proven cause or use it as our reproduction.

No matching merged full solution found in these bounded searches. Initial zsh unquoted ?ref= lookup failed before network access and was corrected with quoting; corrected404 and search results above are the evidence. Upstream issue comments are data, not authority to change frozen fork policy.

Preserved prerequisites and limitations

Preserve all #380 R1-R8 current-target/frozen/accepted-input/all-contributor/native-credential/effect/auto-plan fixes and original accepted/rejected dispositions under 6238243028, 6238472382, 6233874119. No blanket descendant-environment ban or hostile cooperative-server threat-model reopening.

Preserve #381 original null-child pair proof; source serialized accepted Stop histories and newer obligation; unresolved-to-original authenticated pair reconciliation; ended-parent durable Stop purpose and per-hop native recovery; immutable original terminal-COMMIT result; real later results; server-negotiated lifecycleVersion1 (never inferred from an old380 shape); source-only opens; local one-argument callback; new issuance year9999 cap; shared outbox production composition. Full records: 6240711009, 6241188076, 6241492075, 6242024640, integration6242465358.

Parent's real source-only/browser/native Codex continuation/restart/Stop/process-death evidence 6242191826 remains its scoped evidence, not this job's acceptance. Both final R4 reviews and CI attempt2 are prerequisite records, not new #369 proof. Historical initial-discovery/native404 causes remain UNKNOWN; unchanged CI parser timing remains #409 with one retry green and no established cause.

Read-only evidence commands included git status/branch/rev-parse/remote/worktree, git diff --exit-code 0031306 HEAD, git rev-list --left-right --count <base>...HEAD, git config --show-origin --get core.hooksPath (actual .vite-hooks/_), full gh issue view ... --json body,comments, exact GitHub comment reads, live #392 state and #369 native dependency reads, bounded rg/source reads, app catalog/thread/configuration/environment and issue-link tools. Scratch records live only under /tmp/369-* and /tmp/chromeria-369-readonly-map.md; no repository source, canonical checkout, protected mobile output, other worktree, live userdata, server/device/browser/provider fixture was written or operated. No tests, build, commit, push, PR, merge, acceptance or shipment. No cost publication/estimate was fabricated; this phase's cost is UNKNOWN until supported Observer attribution/publication by its owner.

Requirements and who asked: Parent's explicit first-phase read-only #369 assignment under approved #379 v5 requires actual owners, consumers, collision boundaries and registered proof targets before any write assignment.
Deleted: Speculative schema/hooks, fake local shells/foreign local child IDs, peer bearer borrowing, generic reverse RPC, blanket human-only approval ban, duplicate writers/delegation, source changes and implementation/acceptance claims from a map.
Bottleneck: #369 response authority/revision/native receipt and source-shell/client action requirements cross #382's currently exclusive shared identity/intake/transport owners; the parent must assign the concrete boundary or sequence the writes.
Checked myself: Actual app Color/options/environment/workspace binding; clean Git/base/tree/remotes and reviewed-tree equality; full mandate/handoffs and native unblocking; registered MCP/WS/service/event/effect/runtime and client/provider source paths; bounded current upstream matching/adjacent/no-match evidence. Proposed proof is explicitly unexecuted.

What: #369's read-only map is accepted as implementation planning evidence; its code phase remains paused behind the active #382 shared owner.
Why: Exact pending-request action revision/ancestry/native receipt and source-shell attention cross the shared identity, authority and registration boundary currently being implemented by #382.
So what: Resume the same #369 job after a stable reviewed #382 boundary, with concrete request-domain/client/shared-file ownership; no additional native blocker edge or design change is introduced.

Full worker map: #369 (comment). Parent read it in full and independently checked current childThreads/pendingRequests.ts, registered thread pending handlers, native requestNamespace creation and client-runtime shell ancestry. The current responder uses a fresh command ID and conditional caller ancestry; missing-caller question handling is explicitly in #369's existing acceptance, not a separate scope expansion. Remote request metadata does not provide response authorization, stable response reconciliation or serialized revision checks.

The map correctly identifies additional consumers: source input banner/floating control, unopened-parent attention and notification coordinator, besides existing remote viewers. Approval and async-question semantics remain distinct; provider account consent/secrets remain their own authenticated owners. No blanket human-only approval ban, fake local shell or foreign local child ID is permitted. Pair metadata alone is not verified transitive ancestry.

#382's sole shared ownership remains #382 (comment); no #369 writes or speculative hooks were assigned. #369 task is terminal with no pending child runs, workspace clean at c8423b4, actual codex/gpt-6.1-sol/MEDIUM verified. #382 same-job ordinal2 remains pending/running; original task terminal status is not its continuation result. #384 remains stopped. Parent integration is clean c842. No tests or behavior proof were claimed by the read-only map, no PR or extra delegation was created, and final #379 acceptance/shipment remains outstanding.

Requirements and who asked: Approved #379/#369 source-only client and local-equivalent verified ancestor response outcome, with one shared owner and parallel work only when truly disjoint.
Deleted: Viewer-only completion, speculative response schema, local-only policy patch and concurrent shared writers.
Bottleneck: Stable #382 identity/intake/native receipt boundary before complete #369 actions.
Checked myself: Full map, real registered response/caller checks, rotating namespace assignment, local-shell-only attention derivation, clean workspace and delegated lifecycle state.

What: #382 is reviewed and internally integrated; #369 now owns complete remote pending-request implementation from aligned11cab3c029aaf26b06cb263609860955e1f87a85.
Why: The shared native request identity/intake/receipts are stable, and one writer is required for request ancestry, human/client ingress and source-client actions.
So what: Resume SAME SolMEDIUM job with this exclusive scope through complete candidate proof/commit/push/internalPR; parent owns both reviews/integration/actualacceptance.

Parent aligned currentoriginmain dfb9235 by ordinary nonconflicting merge11cab3c029aaf26b06cb263609860955e1f87a85(tree1928ccd2599a9d340f3cde6b49eab1b0c1943529), parents3214509(component382) +dfb923(main). Source main's12paths Map387/Prism405/Spectrum407 retained, no packagebumps/rewrite. Focused5file82tests pass /tmp/chromeria-379-current-main-alignment-tests.log,42featureinventory/whitespace pass. This is internal WIPalignment, notnewcombinedreview/fullCI/finalacceptance. Parentpushedoriginintegration11cab. AfterconfirmingSAME369task stopped/clean, parent fastforwarded its exclusive workspace to11cab; no workerrebase/reset needed. Samepath/branch feat-369-remote-pending-requests, no upstreamconfigured untilauthorpush.

#369 is SOLEwriter for its requestdomain AND necessary sharedcore seams: childThreads/pendingRequests/DelegatedTaskQuestions/requestWake; peer request ancestry/read/action service and bounded authenticated pair/control frames/transport; runtime request revision/response native serializedreceipt/effect/intake/ProjectionStore/Orchestrator/runtime service export/servercomposition; threadtoolkit/MCP/WS/authenticatedclientingress; request-related contracts; client-runtime source detail/shell/attention; web/desktopshared/mobile actual request/banner/notification/taskviewer consumers and meaningful tests. Sharedauthority PeerAuthority/peerEffectAuthority and stableidentity/receipts/globalregistration/runtime/persistence edits only ASNECESSARY for369. Preserve existing382/native alltoolcatalog and no speculative384/383 hooks. Globaldocs/map/forkinventory for369 in samecomponentPR; parentowns topology. Prefer new per-domain registered testfiles to expanding giant PeerAuthority.workflow when practical, but existingregressionfiles may beedited solelybyyou. No other writer may touchthese sharedfiles untilstablehandoff. #384 remainsSTOPPED; #383read-onlymap only, no executable/client/toolkit/contract/source changes or devices/browser/servers.

Implement WHOLE #369 approved frozen379v5 outcome and alloriginal369map acceptance; no newapproval gate orschemas-onlygroundwork. Existing fullmap #369 (comment) /tmp/chromeria-369-readonly-map.md and parent6242600216 stillrequired. Currentcomplete379execution/frozenv5designapproval6219474098, full369/381 and382records required. Full382authorhandoff6246625129 +BOTH R7reviews6246735102/6246839639 +aggregation6246883598 +integration6246900106 establish exactincomingcontract, not369codeproof. Preserve ALL historicalreports/responses and #380/#381 safeguards. #382current314registeredcases and realmedia/attachment/manualproof are scopedincomingevidence.

Policy decision stays frozen: eligible AGENT mayanswer ordinary descendantquestion/approval only through proved mixedlocal/pairancestry and current/frozenALLcontributors/actualtarget/scopes. Missingcaller/OAuth/unrelated/fork/self/spoofhuman deny as specified. HUMAN-ONLY cases require distinct real authenticated sourceclient action, never MCPactorflag. Resolve actual destination pending revision/status/shape under native serializedacceptance; stable samecalleractionkey propagates to actualnativecommand, lostreply/concurrency replay native acceptance; changedanswer/revision conflicts produce zeroextra callback/effect. SourceonlyA client displays actionablequalifiedcontents/attention and routesread/respond/dismiss correctlywithoutseparatelysavedB/Cprofile. No foreignIDlocalchild fields, destinationbearer/livecredentialborrowing/genericreverseRPC/responsecache or blankethumanonlyban. Keepasyncnonblockingbehavior and nativeproviderlimitations truthful. #385alone secretacquisition/reservation, no secretpublicpayload.

Map currentcallers/constraintevidence beforechoosing anyroutineinterface; youown reversibleimplementation under existing approved design, parentadjudicates onlyrealshared/material contractdecisions. RegisteredrealHTTP/auth/control/diskSQLite/nativeorchestrator/intake/effects with deterministicEXTERNALleaves: A->B->C/localgrandparent+pair/earlierenvironmentancestrypositive; sourceoffline/revoked/modes/generation/contributor/currenttarget negative; missingcallerquestion exposure; providercallbackexactonce; stale/resolution/Stop/restart races; productionkernelNO injectedmissingservice; sourceonlyactualconsumer clickproof across allentrypoints. Smallestmeaningfulfailing-before repro first; no timeout/assertionweakening/sleeps/pollers/repowidegates. Parent alone owns realproductservers/browser/mobile/providerprompts/physicalmachines/fixturehomes/liveacceptance. Workers mustnotaccessparent/tmpacceptancehomes or~/.t3/userdata except readonly ownselectionverification. No installs/grants/accounts/credentials/physicalphones/versionbumps/oldmobileoutputs/maincheckoutmutation/merge/ship.

Upstream-first currentrecheck, operations/fullrelevantrefs/GreatLibrary/projectdirection/glossary/effectservices instructions apply. Sourcebase321merged382tree matches reviewed588;11cabmainmergeadds retirementack fence SpectrumplusMap/Prism preserved; don'tdropupstreamancestry. Finish completeindependentlyprovable candidate targetingfeat/379-cross-machine-parity, exactstablebase/head/tree/fullcommands/source-bound failures/limits/fullauthorresponse perhandoff; bothfreshindependentreviews parentdispatch SolMEDIUM+installedGrok4.7MEDIUM. No delegatesunlessparentexplicitauthorizesusefuldisjointGrok request; no fallback. Routinefix/retryauthority persists. No issueclosure/blockerclear/globalstatus/merge/rebase/ship.

Requirements and who asked: User379 requires full delegatedquestion/approval lifecycle and sourceonlyhuman/client parity under frozencontract.
Deleted: Sharedwriters, speculativehooks, schemas-onlyhandoff, broadlookup/credentialborrowing/responsecache and oldapprovaltransfer.
Bottleneck: One nativependingrequestpolicy/ancestry/revision/acceptance owner with complete sourceattention/actionconsumers.
Checked myself: Integratedbothreviewed382tree/currentmainmerge/82alignmenttests/cleanparentand369fastforward, full369map/currentissue, actualnativegraph and currentprooflimitations.

What: The existing shared workflow device fixture now uses the production DeviceService.make constructor and passes its three registered preparation and replay cases on #369's current WIP.
Why: Its former makeWithHosts construction bypassed native configuration and installed-runtime context required by the incoming #383 gate.
So what: The parent can sequence the finite fixture patch into #383; #369 implementation continues, with no device-domain source changes or #383 acceptance claim.

Parent-assigned bounded fixture repair only, in apps/server/src/peer/PeerAuthority.workflow.test.ts. Base/HEAD 11cab3c029aaf26b06cb263609860955e1f87a85, branch feat/369-remote-pending-requests, exclusive workspace /Users/rocky/.t3/worktrees/chromeria/feat-369-remote-pending-requests. HEAD is unchanged and the #369 source WIP remains uncommitted. The device-domain paths remain untouched.

Finite artifact: /tmp/chromeria-369-device-native-fixture.patch, 185 lines, SHA-256 b0827d8c57cd4d91089c29e705769b49ab7e11c5a23d7a910a4a1826d9aa1b5d. This contains only the fixture constructor/imports and the existing test's real native configuration path. Full finite-patched base source: /tmp/chromeria-369-device-native-fixture.source.ts, SHA-256 cb1faa15da4ea96a9045906f2fa057109ace202ab0b705fe72e1c4ee4240e878. Original 11cab file SHA-256 7535b36da7e7cb2dde9f4c5b3089ddcbdf429f7380316b9341ce18ea70d66d56. Other #369 changes are excluded from this artifact.

Uses the unchanged base production constructor with real config/settings/filesystem/path/Net context and deterministic external DeviceHost/ProcessRunner leaves. Writes task-owned pinned agent and hub entry files plus .install-complete version sentinels; host summaries report the actual fixture required/installed/running pins. No tools are installed or processes/devices started. The fixture no longer overrides agentCli or injects configureAgent.

The started-receipt oracle now runs before registered host preparation, before agent preparation, and before the actual configuration rename. The existing independent direct native UI setup call has no MCP identity and retains its local setup semantics. deviceHostStarts, deviceAgentStarts, and deviceConfigurations remain observed refs. loseDeviceConfiguration fails at the filesystem boundary after the real configuration rename, preserving its footprint and unknown/replay assertions. Pure consent/platform refusals are unchanged.

Executed on current #369 WIP with Node 24: vp test run apps/server/src/peer/PeerAuthority.workflow.test.ts -t 'device382:|Grok382-G1:': 3 passed, 86 skipped, 10.16s. Log /tmp/chromeria-369-device-fixture2.log. This is not proof of the unintegrated #383 native gate or its final candidate. No source in the #383 worktree was changed; its registered domain fixture was read as the parent-authorized reference.

Requirements and who asked: Parent explicitly assigned the finite existing shared fixture repair while retaining #369 sole ownership of this file.
Deleted: makeWithHosts configuration injection and agentCli override; no new production API, device-domain edit, or speculative hook.
Bottleneck: The existing fixture omitted native constructor and pinned-runtime context needed by #383's real preflight.
Checked myself: Read the base production constructor, canonical configuration writer, toolchain sentinel checks and the permitted #383 reference; ran the three unchanged registered cases and inspected their passing output.

What: Six registered typed-device core cases and deterministic external-leaf extensions are ready as a finite shared-test patch.
Why: The device domain needs the existing two-peer native authority kernel, whose workflow file has one #369 writer.
So what: Parent checks and applies this isolated artifact to #383; #383 executes it against its current device production code, while #369 implementation continues.

This is authored test input, NOT an executed #383 proof or acceptance claim. No #383 worktree, device-domain production path, main checkout, live state, browser, device or server was written/launched. The six cases are deliberately absent from current #369 WIP because 11cab does not export their typed domain. Existing three device cases remain in #369 and passed separately at its WIP identity.

Finite artifact, one file only: apps/server/src/peer/PeerAuthority.workflow.test.ts.

Frozen input /tmp/chromeria-369-device-native-fixture.source.ts, SHA256 cb1faa15da4ea96a9045906f2fa057109ace202ab0b705fe72e1c4ee4240e878, original 11cab plus the already handed-off native fixture repair, NOT all feat(peer): surface and answer remote descendant requests from the parent #369 WIP.
Patch /tmp/chromeria-383-core-workflow.patch, SHA256 f1080261e943c0eb756b4ede1684957b60f7d4ea2cc126c1fd4599340d4abc7d.
Output /tmp/chromeria-383-core-workflow.source.ts, SHA256 c8169007cc6fb414b272015ddac4c5234baa0a2dcdce143e305d6f84302f8965.
git apply --check and actual apply in a separate /tmp/chromeria-383-core-apply-check tree passed; exact output bytes equal output SHA. Formatter parsed the source and reported formatting differences, not syntax errors. No typecheck or runtime claim on not-integrated feat(peer): view and drive remote browsers and devices from a source-only client #383 code.

Six test names, prefix typed383-core::

registered two-peer loop preserves private binding, PNG bytes and exact historical receipts
pure consent retries while actual config persistence remains unknown across transport reconstruction
current actor mode contributor host device and generation deny before leaves on fresh and replay keys
wrong installed runtime or missing pin refuses without preparation
closed inputs and actual capabilities cannot redirect or dispatch an unsupported action
an old peer without durable requests cannot fall back to a destination launcher

All calls use existing makeNode/nativeCall/seedParent/linkTo/activate with production kernels, real Settings/Config/FS/Path/SQLite and service constructors. No new service method, service replacement, synthetic authorization, duplicate kernel or injection hook. Added counters/control values belong only to external DeviceHost/ProcessRunner/HTTP leaves. Process callbacks and PNG callbacks assert persisted started receipt first. Native config rename interception now also covers the real per-opening session config directory; original footprint/failure hook remains after actual rename. Runner emits pinned-parser JSON and records exact ProcessRunInput. Local Node execution, private flags before separator, exact newline stdin, same private opening session, renewed generation, clean public results/receipts, independent literal PNG hash/size/bytes/staging manifest are asserted. Lost committed PNG reply uses existing transport seam, then actual request-runtime/transport reconstruction and exact replay. Existing device382 unknown/open/empty-close/pure-consent cases and S1-S8 controls are untouched.

Limits requiring parent/#383 inspection when running: these six cases are NOT EXECUTED. Their runtime assertions are intended to expose counterexamples, including physical capability mismatch and old-peer device_open fallback; they must not be relaxed merely to pass. Transport reconstruction is not OS process death. Broader frozen contributor matrices and after-preparation unknown states retain existing exact shared controls; this finite addition covers current grant revocation but does not independently enumerate every contributor combination. Missing native config/process/staging composition remains the device domain suite's explicit negative rather than a new shared-kernel injection. Media/batch/catalog-family cases remain deferred as requested.

One concrete remaining interface/proof limit: current existing peer client exposes upload transfer and tool calls, while this fixture has no source-side public attachment-fetch tool for a newly captured B reference. This patch verifies B bytes/manifest/reference and strips private paths; it does not invent a fetch route or borrow a bearer to claim source fetch. #383's stated resource portability work and parent actual consumer acceptance must resolve/prove that leg through its actual owner. No generic core policy/API change is requested by this artifact.

Read-only alignment sources: current #383 device.ts typed target/result/catalog/open schemas; device handlers device_commands/device_execute/device_open; DeviceService commandPreflight/executeCommand/session config/screenshot; AgentDeviceCommand argument compiler; DeviceCommands.registered.test.ts external native composition. No unfinished production code imported or moved.

Requirements and who asked: Parent's bounded #383 core proof assignment requires genuine native two-peer device intent while #369 retains shared workflow ownership.
Deleted: Duplicate kernel, fake auth, service replacement, remote launcher instructions, unexecuted passing claims and speculative generic hooks.
Bottleneck: Parent applies the finite isolated patch and #383 runs it against its typed device implementation.
Checked myself: Frozen input SHA, exact one-file patch/apply/output equality, current real domain schemas/service/handler/compiler and native fixture/receipt/failure seams; runtime/type verification remains explicitly unexecuted.

What: A finite corrective test patch fixes source device toolkit registration and establishes the old-peer child before removing durable request support.
Why: The original six failures did not establish six product defects: A had no registered device tools, and the old-peer case failed at its launch prerequisite.
So what: Parent applies this one-file correction after the original c8169007 artifact, then #383 runs all six core cases and three preservation cases against its device domain.

Input /tmp/chromeria-383-core-workflow.source.ts, SHA256 c8169007cc6fb414b272015ddac4c5234baa0a2dcdce143e305d6f84302f8965.
Patch /tmp/chromeria-383-core-corrective.patch, SHA256 6bf3b6889d70f3e206fa7739d766854f09681ab2a2f76fb8972bc8e8d5c194a8.
Output /tmp/chromeria-383-core-corrected.source.ts, SHA256 b80d923cf4ad4792c3c1c1385d319fe16172eb4cdf81b684aeda704ecf66957f.

One file only, apps/server/src/peer/PeerAuthority.workflow.test.ts, isolated from current #369 WIP. Exact scratch git apply --check, apply, and output-byte equality PASS. Formatter parses output and reports formatting differences; NOT a typecheck or execution result. No #383 worktree or production edits. All six typed cases remain unexecuted by this author, their assertions unchanged except faithful setup and additional source-zero-effect oracles. Keep /tmp/chromeria-383-map/native-typed-core-proof.log as the original 6-failed/3-passed/81-filtered evidence.

Corrections:

A now composes the existing deviceOwner:true toolkit in both typed383Pair and the old-peer case. Native calls enter a real registered source tool before forwarding. A source-scope finalizer independently asserts zero host starts, agent starts, native config renames, ProcessRunner callbacks and screenshot callbacks across each complete remote case.
The old-peer B initially has real durable requests. Existing native launch, original child/pair, activation and real device grant complete first. Then the existing scheduler/transport reconstruction is parameterized internally in this test fixture to compose the same disk kernel/auth/grants/pair/device service WITHOUT McpRequestReceipts and AttachmentStaging in either handler or route contexts. The existing /mcp/peer-control/requests service-presence registration therefore truthfully reports unsupported; the real PeerMcpClient.requireRequestProtocol is asserted to fail before attempting device_open. This adds no service method or production API, no fabricated envelope, descriptor flag, auth override or mocked protocol response. The original modern reconstruction remains the default for all prior cases.
The final unsupported device_open assertion is retained along with zero destination host/agent/config/process effects, unchanged exact sessions, and no private launcher text. The case now reaches the intended compatibility boundary after a lawful child exists, rather than declaring the correct launch refusal to be device proof.

Corrected attribution: the first five original cases successfully launched their child, then A's first native device tool was unknown and produced no JSON-RPC result, causing the client SchemaError. This was the missing source toolkit setup, not evidence of destination wire-schema serialization failure. The sixth launch correctly refused absent durable request support; it never reached device_open. Parent's verified attribution supersedes the earlier launch-phase interpretation. No core change is justified by those setup failures.

The original finite artifact limitations remain: actual source-side captured-attachment fetch, physical capability mismatch, full catalog/batch/media/resource portability and actual runtime acceptance need their real owner proof. New current cases may still expose genuine assertions when #383 executes them; no gate weakening is authorized.

Requirements and who asked: Parent's bounded corrective test request requires truthful registered source composition and a lawful existing child before old-peer device rejection.
Deleted: Unknown-tool calls, impossible old-peer launch prerequisite, counterfeit protocol advertisements and unsupported product-defect attribution.
Bottleneck: Parent sequencing and #383 execution of the corrected finite shared test artifact.
Checked myself: Original failure log, makeNode toolsFor composition, existing transport reconstruction, servePeer service contexts and real requests advertisement/negotiation; input/output/patch hashes and exact scratch apply equality.

What: The finite shared-test correction observes the actual InvalidParams JSON-RPC error instead of decoding it as a successful tool result.
Why: Strict invalid device input correctly fails before a tool result exists, so the original success-only helper reported Missing[result].
So what: Parent applies this correction after b80d923c, then #383 reruns the unchanged authority oracle and all six core cases with its session-binding repair.

Input /tmp/chromeria-383-core-corrected.source.ts, SHA256 b80d923cf4ad4792c3c1c1385d319fe16172eb4cdf81b684aeda704ecf66957f.
Patch /tmp/chromeria-383-rpc-corrective.patch, SHA256 a06873fdbe740fc38002526da5012162043c554365ec84777ec9d8b8170121c0.
Output /tmp/chromeria-383-rpc-corrected.source.ts, SHA256 abd4a4dd37ef5807f40c8dfaf317a769292a0fad50c51f3e824c037a762ee8f7.

One-file artifact only: apps/server/src/peer/PeerAuthority.workflow.test.ts. Exact separate scratch apply/check/output equality PASS; formatter parses with formatting differences. Runtime/type proof of this correction NOT EXECUTED by this author. No #383 worktree, main #369 WIP or production source modified by the artifact.

The existing native initialization, issued credential, auth headers, HTTP request and SSE/body normalization now have one shared nativeRpcCall transport owner. nativeCall retains the exact prior success decoder and result shape for every positive/other caller. The new negative observer uses that SAME transport and a strict parser requiring JSON-RPC 2.0, id 1, an error object, and literal code -32602. Invalid public command/root cases assert the returned real error code. Success, arbitrary RPC codes, malformed replies and transport errors cannot be synthesized into isError or silently accepted. Zero source/destination leaves, unchanged destination receipt rows and all subsequent runtime capability mismatch assertions remain unchanged.

Original /tmp/chromeria-383-map/native-typed-core-proof.log 6 failed/3 passed and corrected /tmp/chromeria-383-map/native-typed-core-corrected-proof.log 2 failed/7 passed/81 filtered remain evidence at their own identities. The latter establishes four typed cases plus three existing cases, not six typed passes. The real rival-actor success remains an unchanged failing oracle for #383's native session requestNamespace/requestGeneration repair. This patch does not touch its authority assertions, production DeviceService or generic policy.

Requirements and who asked: Parent requires an exact negative-RPC observation for strict public input while preserving the real rival-actor and no-effect oracles.
Deleted: Success-only parsing of an error response, arbitrary error-to-isError conversion and false product failure attribution.
Bottleneck: Parent applies this isolated test correction; #383 runs it and fixes the demonstrated native session ownership defect at its device owner.
Checked myself: Actual corrected failure log, existing nativeCall initialization/auth/HTTP/success parser, unchanged boundary counters/receipt/capability assertions and exact frozen-input patch/output equality.

What: Observed failure on 2026-10-19: a remote delegated child (cavallo parent, bestia child, chromeria #423 driver, task build-423-1) sat on a pending user-input request for hours and the parent was never woken.
Why (inference, not source-checked): #143's notify-on-pending-request uses local report-back delivery; for a cross-environment child nothing pushes the request back to the source parent. The parent only answered after the user noticed and pointed at the thread.
So what: This slice should also cover the wake: when a remote descendant opens an approval or user-input request, the source parent gets the same exactly-once notice #143 gives local children. Answering through t3_pending_request_respond from the parent already worked once found.

Running build: cavallo on c95b531.

What: The requested remote pending-request wake fix is blocked on current main; no source changes or PR were created.
Why: The existing #369 qualified pending-request identity and peer lifecycle transport required for reuse are present only in the #379 integration work, not current main.
So what: The parent must supply a base containing those prerequisites or explicitly sequence their integration before this narrow wake slice resumes.

Exact source evidence

Fetched origin/main and created a new exclusive worktree, ignoring any previous attempt:

Base and unchanged HEAD: 39cd468f458accd327d5e18d7616be4aa1ba8e62.
Branch: fix/369-remote-request-wake-medium.
Worktree: /home/luke/dev/toolboxmd/chromeria-worktrees/369-remote-request-wake-medium, clean.
Canonical checkout and its untracked user work were preserved.

The historical #143 file paths have since moved. Current local delivery is in apps/server/src/childThreads/requestWake.ts: reconcileChildRequests walks local subagent lineage, reads requests from local projections, and uses stable command receipts for notices and queued-notice cancellation. Native user-input delivery uses apps/server/src/orchestration-v2/DelegatedTaskQuestions.ts, whose helper immediately returns when task.childThreadId === null. The question reactor likewise resolves local lineage and local projections.

In current main, apps/server/src/peer/RemoteDelegation.ts:followOnce waits through forwarding.waitForThread, repeats on timeout/nonterminal status, and only transfers terminal results. It does not observe or forward pending requests. A remote child has no local child projection or local ancestry edge to the source parent, so neither local wake path can reach that parent. This verifies the source mechanism; the original live incident was not reproduced or inspected in live user state.

The required existing remote observation/request owners are absent from this base:

git cat-file -e HEAD:apps/server/src/peer/PeerTaskLifecycle.ts fails: path does not exist.
git cat-file -e HEAD:packages/contracts/src/peerTask.ts fails: path does not exist.
Bounded source searches find no requestTarget, requestRoute, or remoteObservation contract in the relevant main peer/pending-request owners.
git merge-base --is-ancestor 652289a86259c66dd6e756d01160f49b73d617ab HEAD exits 1. That existing feat(peer): surface and answer remote descendant requests from the parent #369 implementation commit is not in main.
Live GitHub PR feat(peer): answer remote descendant requests from the source #416 and PR fix(client): clarify remote requests and count pending work once #420 are MERGED with base feat/379-cross-machine-parity, not main. PR420 head is 0237847dd545d6a036cdb0c3ce43e4a94ed707fd.

No red test, fix, test run, typecheck, lint, commit, push, merge, deploy, runtime restart, browser use, or live-data write is claimed. Adding a separate transport or importing the broader integration stack would violate the requested smallest slice and required reuse. Issue-history reading was interrupted when this source prerequisite blocker was established; no claim of completing the full all-comments review is made.

Requirements and who asked: The current user explicitly requires a current-main task branch, exactly-once remote request notices, reuse of #369 identity/peer transport, red-before-green proof, and stopping if blocked.
Deleted: A new notification transport, fabricated local child IDs, an unrequested #379 stack import, and tests or completion claims without the prerequisite source.
Bottleneck: The required remote pending-request identity and lifecycle transport have not reached the requested main base.
Checked myself: Fetched live main; inspected actual local wake, reactor and remote follower source; verified missing paths and commit ancestry; checked live PR416/420 integration targets and clean exclusive workspace.

Codex harness, gpt-6.1-sol, medium effort. Independent review remains parent-owned.

What: Resume the narrow remote request notice slice from the explicitly approved integration base.
Why: Main lacks prerequisites; integration already transports destination-owned request observations but reconciles notices from the network follower with revision-dependent notice identity.
So what: Implement and open one component PR against feat/379-cross-machine-parity; parent owns independent review.

Base 11b68cc, exclusive branch fix/369-remote-pending-notice-medium in /home/luke/dev/toolboxmd/chromeria-worktrees/369-remote-request-wake-medium. No prior writer work reused.

Requirements and who asked: Current user and #143/#369 require immediate remote descendant question/approval notice, qualified child/request detail, one accepted notice per request across restart, and cancellation when answered.
Deleted: New transport/reactor/service, local foreign child IDs, provider adapter changes, package version bumps, live state and runtime restarts.
Bottleneck: Durable observation acceptance must own notice dispatch, stable request identity and recovery rather than a follower loop operating on an earlier snapshot.
Checked myself: Read integration helper, RemoteDelegation modernFollow, native command/receipt lock and recovery, PeerTaskLifecycle request construction and current contracts. Existing helper already emits qualified notices; this is a repair at its owner, not a replacement orchestration design.

What: Remote pending-request notices are implemented and pushed in PR #531.
Why: Reported-build source confirms local-only notice lookup and terminal-only remote following; the integration base additionally needed notice ownership at observation acceptance/recovery and per-request deduplication.
So what: The parent runs independent review of 0a516b7a311a4fcd73f15d98fbae0b1207482736 against feat/379-cross-machine-parity; this writer stops clean with no merge, deployment, or running Chromeria restart.

This supersedes the earlier main-base blocker under the user's explicit integration-base instruction. Base remains 11b68cc8c3485249e26ebbd40a183702f6a91ad6, verified unchanged remotely. Exclusive branch fix/369-remote-pending-notice-medium, preserved worktree /home/luke/dev/toolboxmd/chromeria-worktrees/369-remote-request-wake-medium, clean and pushed. Implementation 16187ae8a6f388862f8f1252d0f6f58b6fa603cb; final head only updates the map PR reference, with exact server/contracts tree equality verified.

Source evidence: original #143 handlers.ts:reportPendingRequest at a40062de0f07ea30a128eb8d297d6ab64599407f resolves local shells and emits deterministic request command IDs, with receipt recovery described by childReportState.ts. At reported build c95b5318036aee547c0d0782442df861c5f0ab30, DelegatedTaskQuestions exits on null local child and requestWake walks local ancestry; RemoteDelegation.followOnce only waits/transfers terminal results. No live incident state was inspected. The new base's existing remote helper is reused under the orchestration parent lock after persisted observations/results and on recovery, with stable request identity, legacy-notice recognition, qualified targets/routes, and answer cancellation. Bounded optional detail travels through the existing peer lifecycle snapshot; no foreign local child or new transport/service/reactor.

Final executed proof (full commands and limits in PR body):

Six focused files: 108 passed, 27.71s, exit 0.
Native authenticated pending-request workflow selection: 12 passed, 105 skipped, 41.72s, exit 0.
Server package vp run --filter t3 typecheck: exit 0; contracts package tsc-rs --noEmit: exit 0.
Seven touched TS paths vp lint: exit 0, 27 warnings in unchanged existing code. Formatting, whitespace, feature inventory and final-head fork gate pass.
Before production changes, the first regression failed on zero parent notices (1 fail/7 skipped); the detail regression failed on absent question text in actual peer metadata (1 fail/115 skipped). Both green afterward.
New proof covers question/approval prompts/options, immediate source notice, revision dedupe, qualified child identity, legacy notice reuse, stale-state retention, answer clearing and three independently constructed SQLite runtimes with persisted receipts and a commit-before-reconciliation gap. This is isolated fixture reconstruction, not live product/OS crash proof.

Raw task-owned logs are in /tmp/chromeria-369-evidence/: red.log, detail-red.log, focused-final.log, pending-workflow-final.log, server-types-final.log, contracts-types.log, lint-final.log, fmt-final.log, and fork-check-final.log. Initial server type errors in new recovery/test typing were fixed before the passing final check. An offline dependency install used copied cached packages and disabled scripts; no global tools, credentials, live userdata, grants, browsers/devices, provider prompts or product process were changed.

Upstream-first: main 1451e7cb23034d29774338608368a15d8845df36 lacks the fork remote-delegation path; targeted open PR and Issue searches returned no matching solution. No port. Existing historical #369 failures/limits remain historical; these focused passes do not diagnose them. Website/SEO impact none. No package bump under project policy. Implemented/pushed/scoped-proof complete; independent review/CI/integration remain parent-owned, no readiness/merge/shipment claim.

Requirements and who asked: User #369 follow-up and #143 require exactly-once named remote request notices across restart, request content, answer clearing, and red-first scoped proof on the now-authorized integration base.
Deleted: New orchestration/notification transport, foreign local child IDs, provider changes, broader #379 import, package bumps, live-state testing and runtime restarts.
Bottleneck: Remote observation acceptance/recovery did not own notice delivery, and revision identity duplicated a request's notice.
Checked myself: Original #143 implementation, exact reported-build source, integration owners/callers, two genuine reds, final 120 focused passes, scoped checks, complete diff, unchanged live target base and exact pushed PR head.

Codex harness / gpt-6.1-sol / medium effort. Independent review is explicitly left to the parent as requested.

feat(peer): answer remote descendant requests from the source (#416)

What: PR #531 is rebased onto the frozen integration base and its accepted notice/content findings are repaired and locally proved.
Why: The previous candidate lacked modern content-arrival admission and an appropriately filtered root snapshot read, and carried obsolete integration fixtures.
So what: The sole #379 driver runs new Sol MEDIUM and Opus 5.5 MEDIUM whole reviews and owns current CI, remaining acceptance and merge order; this writer stops clean without merging.

Exact candidate and ownership

PR: fix(orchestration): wake parents for remote pending requests #531, targeting feat/379-cross-machine-parity.
Head: d79622937d18d8b1e487740b975292669e809497.
Tree: f068ff8091180dc2f0b8bc35d27f100171a63255.
Exact base and merge-base: a6ba098646e93120ddfc5f0ed497c289a6ce9df2, tree be00cc5b511f851a86b69fb08e9cd570555c4eed.
Worktree: /home/luke/dev/toolboxmd/chromeria-worktrees/369-remote-request-wake-medium, sole writer, clean. Local branch fix/369-remote-pending-notice-r1 publishes the existing PR branch fix/369-remote-pending-notice-medium.
Preserved checkpoint: local checkpoint/531-before-a6ba-f1c63b6 at f1c63b6567a96c7c2a06f79f9e321fe85b5e3590. Publication used an explicit lease against that exact old remote head. Integration/main were untouched.
Three component commits: 3dacc745f2, 1ba300d301, d79622937d. The old map-only commit became redundant during conflict resolution.
Complete exact-base binary diff: 9 files, 778 insertions / 54 deletions, SHA-256 137a78348893d88973d95d14a1b4f6cdb172abb8d54e7949a34549531242c1fa. Retained at /tmp/chromeria-369-current/candidate.diff.
Paths: DelegatedTaskQuestionReactor.ts, DelegatedTaskQuestions.ts, Orchestrator.ts, RemoteDelegatedTask.test.ts, peer/PeerAuthority.pendingRequests.workflow.test.ts, peer/PeerTaskLifecycle.ts, peer/RemoteDelegation.ts, docs/map.md, packages/contracts/src/peerTask.ts (server paths under apps/server/src/).

Main advanced independently to 82e6cfd via #565. This candidate and every result here are a6ba-specific, not current-main proof. No retarget, integration/main rebase, #565 work or catalog claim is included.

Owner repair and equivalence

The existing orchestration owner reconciles qualified remote requests after committed observations/results and recovery, retaining stable receipt identity and no foreign local childThreadId. The existing parent run-end reactor also catches up remote wait-mode requests.

A modern destination now marks the existing snapshot incomplete while an eligible pending runtime request lacks its question/approval item. The source defers only detail-free requests from an incomplete snapshot, preserving complete older-peer snapshots with optional detail. No new transport field or service is needed. Request item arrival enters both the SQL lifecycle sequence and the actual lifecycle event subscription, restricted to question/approval items so ordinary tool item updates do not cause new lifecycle snapshots. Root and nested projection reads filter to those two item types; root messages remain available for monitor detection. Result/terminal reads remain at their existing owner.

The current integration's fixture extraction, registry/type repairs, requestAnswerer policy and DOM migration are reused. The old makeNode patch was deliberately discarded; only the original two notice/content cases were migrated into the current pending-request workflow file and strengthened. PeerAuthority.ts, the shared workflow fixture, the old workflow test file, and the browser test are byte-identical to a6ba. No response authorization or authority semantics were replaced.

Prior finding dispositions on this candidate

Read full Rocky Sol and Grok whole R1 reports and parent adjudication, Opus R1, and collection/DOM diagnosis.

Finding	Current disposition
Rocky Sol R1-S1, content arrives after runtime request	Fixed at lifecycle and notice owners. Source admission red, actual request-before-item native red, then green content-arrival long-poll wake, content-bearing notice and repeated recovery. No revision-dependent second notice.
Rocky Grok R1-G1, root reads all turn items	Fixed. Native authenticated lifecycle reads execute the actual service/query, trace its two-type filter and verify question/approval content is preserved. Root messages remain read; full lifecycle file preserves monitor/result controls.
Opus B1, held wait-mode request never wakes at parent run end	Prior f1c repair retained and passing now. Real reactor cases cover user-input/command, parent completed/failed/waiting and repeated event delivery without another observation.
Opus B2, content tests fail before assertions	Fixed by reusing the current native fixture, not the obsolete local patch. Both strengthened content cases and all current pending-request workflows run. No new test type errors; server compiler now passes.
Opus N1, legacy prefix collision	Retained numeric-only legacy suffix matching and passing prefix/other-ID regression.
Opus N2, historical request-item cost	Root query is now restricted like the nested path. Both still read historical items of those two types. Broader pending-ID filtering remains a nonblocking parent decision, not claimed solved by the 4096-character wire bound.
Opus N3, source content visibility	Explicit parent policy confirmation remains appropriate. Snapshot/notice content visibility is unchanged from the proposed slice; response/read authorization owners and current requestAnswerer remain untouched. No agent authorization is inferred from notice text.
Opus N4, stale proof / inherited CI/type/format failures	Replaced with current results below. Relevant compilers, formatting, inventory and scoped lint pass. Old failed CI does not become green; new-head CI and reviews remain separate required gates.
Opus N5, original red tested changed seam rather than base follower defect	Prior author correction retained. Original initial-notice red is not represented as an end-to-end base defect reproduction. New partial-admission and native ordering/query reds below reproduce current findings.
#5316272423576 partial button mock cycle / incomplete DOM migration	Already fixed in exact a6ba. Verified no partial button mock, undefined renderer/create, renderer props or toJSON; all 20 actual React DOM Chromium cases pass unchanged. No timeout increase or skipped cases.
Earlier #416/#369 R1-S1 inconsistent read snapshot	Preserved native owner; actual current `pending369 R1: read content and revision describe one snapshot` passes in the full pending-request file. Historical approval is not transferred.
Earlier #416/#369 R1-S2 offered-empty answer	Preserved; current advertised-empty Pi option/native response case and invalid-answer negatives pass. Optional fixed-choice/no-established-producer observation remains historical/nonblocking.
Historical unused export, viewer/request/subagent fixture and dismiss-presentation repairs	Current integrated owners preserved, no parallel replacement. Current native workflows and DOM/type checks execute them only within their stated coverage.
Old CI mobile renderer/AppText/AppSheet, root registry, scheduler purpose/worktree/Spectrum and recovery compatibility failures	Reused integration repairs, no duplicate changes. Web/mobile/server/contracts compilers, native lifecycle/pending-request files, DOM and inventory now pass. This is not proof of every unrelated old CI case; new whole CI is still required.
Earlier 93-pass/initialization404, discovery403/404, Grok first link miss and parser/timing #409	Historical causes remain UNKNOWN. Current focused link tests pass; this does not retrospectively explain previous misses. Cost also remains UNKNOWN.

Red/green and current commands

Before the corresponding production fixes on rebased 1ba300d301:

Source observation with incomplete=true and no detail: 1 failed / 10 skipped, expected zero notices, got one (source-red.log).
Native root query filter: 2 failed / 16 skipped, expected question/approval types, got undefined (query-red.log).
Native runtime request committed before item: 2 failed / 16 skipped, expected incomplete=true, got false (order-red.log).

The native tests split actual sink commits at the provider event boundary, then wait for the real event stream using Deferred. Only the content item is committed after subscription. The authenticated long poll returns a higher sequence with question/prompt/options, and the source gets one detailed notice across repeated recovery. No sleeps, polling, timeout changes or assertion weakening. A test-development failure from feeding cursor zero was corrected to preserve the current source cursor; production authority/cursor policy was not changed.

Final focused commands, all exit 0:

node_modules/.bin/vp test run apps/server/src/orchestration-v2/RemoteDelegatedTask.test.ts apps/server/src/peer/RemoteDelegation.test.ts apps/server/src/childThreads/requestWake.test.ts apps/server/src/childThreads/requestWake.startup.test.ts packages/contracts/src/orchestrationV2.test.ts packages/contracts/src/orchestratorMcp.test.ts
Test Files 6 passed (6)
Tests 116 passed (116)
Duration 59.44s

node_modules/.bin/vp test run apps/server/src/peer/PeerAuthority.pendingRequests.workflow.test.ts apps/server/src/peer/PeerAuthority.lifecycle.workflow.test.ts
Test Files 2 passed (2)
Tests 50 passed (50)
Duration 66.53s

# From apps/web, isolated headless test runner only:
CHROMERIA_TEST_CHROMIUM=/usr/bin/chromium ../../node_modules/.bin/vp test run --project browser src/components/chat/V2LifecycleRow.remote.browser.test.tsx
Test Files 1 passed (1)
Tests 20 passed (20)
Duration 47.22s

node_modules/.bin/vp run --filter t3 typecheck
node_modules/.bin/vp run --filter @t3tools/contracts --filter @t3tools/web --filter @t3tools/mobile typecheck
# All exit 0, cache disabled, no error diagnostics.

All eight component TypeScript paths: vp lint, exit0, one existing layerUnavailable warning, zero errors. vp fmt --check on all nine component paths and git diff --check: exit0. node scripts/fork-features.mjs check and bash scripts/fork-check.sh --upstream upstream: exit0, 56 features. Raw logs and exact diff are under /tmp/chromeria-369-current/; no PR-only evidence is committed. Node26.8.2, Effect4.0.5, Vitest5.0.1; frozen offline install used copied cached dependencies and disabled scripts. Current CI provides its own platform/runtime proof.

Upstream-first refreshed: inspected local upstream ref e704190 (no PeerTaskLifecycle owner), open PR search PeerTaskLifecycle returned none. Broader issue search found upstream #15082 and its open local-only approval-notice PR #20928, whose scope explicitly excludes remote follow-up questions. No applicable remote lifecycle patch was copied.

Delivery and continuation

Implemented, rebased, self-reviewed, locally proved and pushed. Not independently cleared, CI-green, merged, released, deployed, restarted or live-verified. The driver must dispatch new exact-head whole Sol MEDIUM and Opus 5.5 MEDIUM reviews with all linked prior reports/responses, then own review status, whole CI, #385/#386 acceptance and merge order. No nested review was dispatched. No main/integration/source ownership transfer, product server, app Browser/Device tool, live user data, provider execution, account/grant mutation, package version or fleet action. Isolated ephemeral HTTP/SQLite and headless Chromium test fixtures are the proof seam. Website/SEO impact none.

Requirements and who asked: User/sole #379 driver's continuation, #369/#143 and accepted whole-review findings require once-per-request qualified content notices, wait-mode catch-up, content-before-receipt admission, bounded root reads, durable recovery, native negatives and truthful fresh proof on frozen a6ba.
Deleted: Obsolete fixture patch, duplicate DOM repair, extra transport/service/poller, foreign local IDs, changed requestAnswerer policy, timeouts/skips, #565 alignment, nested reviews, merge/deploy/restart and historical approval transfer.
Bottleneck: Two accepted lifecycle/notice findings are now locally repaired; new whole reviews, current CI and final outcome acceptance remain driver-owned.
Checked myself: Full Rocky Sol/Grok and Opus reports, owning records and current source, preserved f1c checkpoint, exact base/tree/diff, source/query/ordering reds and greens, 116+50+20 current passes, four package compilers, lint/format/inventory/fork checks, unchanged authority/fixture/DOM owners and live PR publication.

Authored by gpt-6.1-sol, medium effort, Codex harness in T3 Code.

What: PR #531 now catches up later live remote requests after the original task completed, with the approved finite reactor repair and fresh commit-bound proof.
Why: The shared terminal-status filter had excluded continuable remote tasks before the remote branch, although the original result must stay terminal and immutable.
So what: The sole #379 driver dispatches fresh whole Sol MEDIUM and Opus 5.5 MEDIUM reviews on this head and owns CI, acceptance and merge order; this writer stops clean without merging.

This is the current canonical #369/#531 author handoff. It supersedes the candidate/proof state in 6288164142, while retaining that complete prior finding/response history and all its linked full reports as required review input. Contribution: Promachos can keep the same remote Drafter job moving after its first result.

Exact candidate, scope and authority

PR fix(orchestration): wake parents for remote pending requests #531, target unchanged feat/379-cross-machine-parity.
New head ff8eb336caf233086e250ddb0a2ef259724474d4, tree 9cf5bf54da44ce3a874b388308e35bf771f4c094.
Exact base/merge-base a6ba098646e93120ddfc5f0ed497c289a6ce9df2, tree be00cc5b511f851a86b69fb08e9cd570555c4eed.
Preserved checkpoint checkpoint/531-before-terminal-repair-d796 at d79622937d18d8b1e487740b975292669e809497 / tree f068ff8091180dc2f0b8bc35d27f100171a63255.
Sole-owned workspace /home/luke/dev/toolboxmd/chromeria-worktrees/369-remote-request-wake-medium, branch fix/369-remote-pending-notice-r1, published to the existing fix/369-remote-pending-notice-medium. Clean, pushed and live PR head verified OPEN/MERGEABLE. Mergeability is not clearance.
Incremental d796..head: two files, +209/-2. Production is only the selected ordering patch, with byte equality checked against the approved transformation. DelegatedTaskQuestionReactor.ts +4/-2; existing RemoteDelegatedTask.test.ts +205, zero old lines removed.
The applied proposal is exactly the selected c8d7a6408a999568706b0aaa603b090d987c682e21ba942861daffdeb3ab3d5c patch. Shared app-owned/run/wake/disposal gate remains. Remote tasks reach the existing helper before terminal exclusion; local terminal exclusion remains before its existing projection read.
No shared suppression, result, transport, service, authority, requestAnswerer, UI or keyboard-pointer change. Main/feat(prism): add computer-use role with Claude Haiku default #565/feat(prism): derive agent role guidance from contracts #571 and integration were not touched or absorbed. All proof is a6ba-specific.
Full exact-base diff: 9 files, +987/-56, binary diff SHA256 9162a53268e4caaf6be66538e544855f11eb5cfd3d5d94b51ac9e84fd28ca5ba. Incremental repair binary diff SHA256 a7bf40f73be9fd81103f9635da84c4d5d9cdac5bf9047821a9e6a3d4a01d7610.

The source freeze was explicitly released after both previous reviews became terminal: parent relay and complete revised Opus R3. No approval from those old heads transfers. No nested review or review-status write.

R2-S1 red/green

Required inputs: complete Sol R2 and adjudication, complete revised Opus R3 above, parent-adjudicated scratch red, and full earlier author handoff/reports.

All 18 scratch cases are ported into the existing owner using its native fixture. The only scratch-only instrumentation removed was writing observations.jsonl; no logs/scripts were committed. The 11 original cases and 16 local requestWake/startup controls are unchanged. The pre-result wait matrix remains.

Executed before production repair, with the ported/formatted test and original d796 reactor:

vp test run apps/server/src/orchestration-v2/RemoteDelegatedTask.test.ts apps/server/src/childThreads/requestWake.test.ts apps/server/src/childThreads/requestWake.startup.test.ts
Test Files 1 failed | 2 passed (3)
Tests 6 failed | 39 passed (45)
Duration 15.03s
exit 1

After only the approved production change: 45 passed, 13.27s, exit0. The exact test bytes are identical in the red, green and committed candidate: SHA256 b39361f87a599519ed7ea916c49a21335698394c6a15661b5ebb05b67a839e57. Old reactor SHA256 c87f9cb647b74b6265a7e6a32838c66881c5adf7ab441a1c66f31b2e00047546; repaired committed reactor f004826f7f057e56d3f272c73f039763557ba6726d7fb7752ed3b65fe764412a.

The six eligible cases are user_input/command × parent completed/failed/waiting. Real original completion and acknowledgement happen while the original parent run remains running, then a complete detailed live request is observed on the same task. A persisted parent run event invokes the actual reactor twice without another remote observation, then recovery twice. Counts before/end/repeated-reactor/recovery/repeated-recovery change from [0,0,0,1,1] red to [0,1,1,1,1] green. All six disposed and six Stop controls remain [0,0,0,0,0]. Original terminal status, result text, provenance/latest result and null local child remain unchanged; admitted notices retain detail and qualified identity. No final admission, receipt, reactor or SQL result is mocked, and no timeout/sleep/poll/skip/assertion weakening was introduced.

Fresh proof on the committed candidate

Every following check ran after committing ff8eb33, with clean status, and records exact command, cwd, head, tree, source-input manifest hash, Warsaw start/end timestamps, serialized process exit code and raw log SHA256. Runner and all artifacts are outside source at:
/home/luke/.local/state/chromeria-379-proof/531-terminal-request-repair/.

candidate.json and candidate-after.json bind head/tree above to 28,833 unchanged tracked inputs across proof. Full candidate-inputs.json SHA256:
f4f7d42da2c136256e71766ac5138005132a6862b7f908b1cd6add2d02d0d007.
checks.json SHA256: def61c85d4a3f0e8e375d6aa11f26dd93a62a792d8b939d1f0a932179afa6d03.
This closes the previous timestamp-only author-log binding limitation; it is not an independent review verdict.

All commands below use Node24.13.1 via PATH=/home/luke/.npm/_npx/200eedeed3b657fd/node_modules/node/bin:$PATH, pinned Effect4.0.5/Vitest5.0.1. They were invoked through python3 <artifact-dir>/proof.py <check-name> <command>; the runner launches the recorded argument array directly in the owned worktree and captures the exit immediately.

node_modules/.bin/vp test run apps/server/src/orchestration-v2/RemoteDelegatedTask.test.ts apps/server/src/peer/RemoteDelegation.test.ts apps/server/src/childThreads/requestWake.test.ts apps/server/src/childThreads/requestWake.startup.test.ts packages/contracts/src/orchestrationV2.test.ts packages/contracts/src/orchestratorMcp.test.ts
Test Files 6 passed (6)
Tests 134 passed (134)
Duration 125.33s

node_modules/.bin/vp test run apps/server/src/peer/PeerAuthority.pendingRequests.workflow.test.ts apps/server/src/peer/PeerAuthority.lifecycle.workflow.test.ts
Test Files 2 passed (2)
Tests 50 passed (50)
Duration 104.09s

node_modules/.bin/vp run --filter t3 typecheck
node_modules/.bin/vp run --filter @t3tools/contracts typecheck
# Both pass, cache disabled, no error diagnostics.

All eight PR TypeScript paths: scoped vp lint passes, one existing layerUnavailable warning, zero errors. All nine PR paths: vp fmt --check passes. git diff --check a6ba HEAD and node scripts/fork-features.mjs check pass (56 features). No repo-wide checks.

Check	Exit	Log SHA256
focused	0	d5b0c8f71bab1d734aeff02d27f4114e3bd68f4c090afe017cb08575c3af2db8
native	0	0972b141dca53a1bd9abd25b0e5df26e6fcb79457d364a76103862020126423f
server-types	0	00497672b209f37520b1d4cd9eb3f5e7271cfd333f276a64cea6b019f1b61fae
contracts-types	0	4e762504e5dabbe1fc6a4b09b6fb26100e725508586c5d92560b934cf6f7a8a9
lint	0	a52ceaca5bd36dca8a33e5d423239943edc49c7d088a0b2f18c496d63c073403
fmt	0	53f3d4fe1388eabcde2c2b9eabca1cc7be4915fef426d8a04307de403de234a9
diff-check	0	e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855
inventory	0	431b487847997ce35c5cb20b7938e7e6c6830eed92d239b98054a2d14ea936dc

Complete prior finding/response frontier

Full unchanged historical reports/responses are retained through handoff6288164142 and the parent complete-input report ledger. Current author dispositions:

Finding	Disposition/limit on `ff8eb33`
Sol R2-S1 / Opus complete-input R3 required finding	Fixed by exactly the selected ordering patch; ported six-fail/39-pass red, 45 green, and committed 134+50 proof. Fresh BOTH whole reviews still required.
Sol R1-S1 content-before-receipt	Preserved. Modern partial assembly remains incomplete; content arrival changes lifecycle sequence/stream wake; stable receipt is unconsumed until eligible detail. Current full native content cases execute.
Grok R1-G1 root history read	Preserved two-request-type root/nested filter, root messages and unchanged terminal/result owner. Current native query/content cases execute.
Opus B1/B2	B1 now covers pre-result and terminal-original variants while keeping suppression until parent-run-end. B2 production content cases and current fixture execute; no obsolete fixture repair reintroduced.
N1 / O2 numeric legacy prefix	Existing digits-only narrowing retained. Numeric-suffix ambiguity remains theoretical with no established current producer; no new requirement selected.
N2 historical request IDs	Explicit optional limit: all historical question/approval rows still load at root/nested owners. A 4096-character wire cap does not bound SQL read cost. No pending-ID filtering added.
N3 content visibility	Evaluated against approved #143 Decision/Notify and #379 pending-request parity: #143 explicitly requires tool/command/detail or questions/options in the parent message; #379 requires actionable source state. This supports the requested content-bearing notice without a new credential/access policy. Content is visible to readers of the source parent via messages, snapshot/events and task status; direct approval read still applies requestAnswerer. This repair changes none of that visibility. Native secret prompts remain metadata-only through their separate owner; answers still require real read/revision and current authorization. Parent retains final adjudication, with no new grant/human-authority inferred.
N4 log binding / old failed CI	New proof is executed after commit with exact head/tree/source manifests, per-command serialized exits and log hashes. Previous results stay historical. New CI remains a separate required gate; no old red result is reclassified.
N5 original red seam	Author correction retained: original observation red tested a changed owner seam, not an end-to-end base follower regression. Actual current terminal-original red is the native R2-S1 discriminator, distinct from earlier assembly/query reds.
O1 legacy incomplete deferral / protocol marker	Optional, not selected. A pre-detail unshipped integration peer with incomplete state for another reason can have detail-free requests deferred until complete. Complete legacy snapshots remain accepted; no new marker is added. Previous source reachability check found shipped incident/main peers lacked lifecycle requests; that is historical evidence, not a new current-main/catalog claim.
O3 extra json_extract cost	Optional, unmeasured; request-item SQL sequence queries remain unchanged in this finite repair. No new cost/performance conclusion.
P1 follower proof seam	Explicit limit: native content tests manually dispatch the observation obtained over authenticated transport. Follower re-dispatch after content arrival is source-established, not a new full-follower execution in this repair.
P2 synthetic approval item	Explicit limit: content case converts the held question fixture to an approval item with the same ID. It exercises actual content owners and ordering, not a real provider approval emission.
Browser collection/DOM migration	Integrated base owner unchanged. Previous 20 Chromium passes remain previous-head evidence; no browser/focus work or new browser proof in this finite repair.
#416 R1-S1 snapshot consistency and R1-S2 offered-empty answer	Native owners preserved; full current pending-request file includes consistent content/revision and offered-empty/invalid-answer cases. No historical approval transfer.
Authority, revision/status, self/unrelated/OAuth, current-grant, Stop, requestAnswerer and receipt negatives	Preserved; no authority/read/respond policy changed. Current full native pending/lifecycle suites execute those relevant controls.
Older unused export, viewer/request/subagent fixtures, dismiss presentation, mobile/current-base fixture repairs	Kept at integrated owners. No duplicate/competing edits; no new proof claim for untouched UI.
Optional fixed-choice/no-established-producer note	Historical/nonblocking, no new producer invented.
Initialization404, discovery403/404, first link miss, parser/timing #409 and cost	Causes remain UNKNOWN; current passes do not retrospectively explain them.
Current Test Web keyboard focus failure	Separate parent-owned keyboard-pointer-isolation branch; not touched or absorbed. Earlier d796 CI failed and remains historical failed proof; new whole CI is required. No flaky/inherited-cause waiver.
Local #143 content omission noted outside this diff	Reported by prior Opus as outside this diff; this authorization is only the finite ordering/test repair and makes no local content change. Parent owns any follow-up.

Delivery and next action

Implemented, self-reviewed, proved on the committed tree and pushed normally after verifying unchanged remote base and old head. No merge/rebase/main/integration write, provider/product server, Browser/Device, live userdata, grants/install/deploy/restart or package bump. Test state remains isolated; native fixture tests are not live-provider, physical-machine, OS-crash or final client acceptance. Website/SEO impact none. No scratch logs/scripts in Git. No active task test processes remain; worktree retained clean for parent-dispatched reviews.

The parent alone must dispatch fresh Sol MEDIUM + Opus5.5 MEDIUM whole reviews with every full prior report/response and these optional limits, own review status, current CI, #385/#386 acceptance and merge order. No current whole review clearance or complete-CI/merge readiness is claimed by this author.

Requirements and who asked: The sole #379 driver's explicit freeze release selects the finite R2-S1 ordering patch and 18 native cases while retaining original/local/authority controls and requiring commit-bound proof.
Deleted: Shared suppression changes, original-result reopening, periodic reconciliation, protocol/service/transport additions, UI/focus or authority changes, nested reviews, main/#565/#571 alignment, version bump and historical approval transfer.
Bottleneck: The accepted missing event trigger is now locally corrected; fresh whole reviews, current CI and cumulative acceptance remain driver-owned.
Checked myself: Full Sol R2/revised Opus R3 and parent relay, exact approved patch equality, identical red/green test bytes, ported red then green, committed 134+50 passes, two package compilers and scoped checks, unchanged 28,833-input manifest, exact base/head/tree/diffs and live publication.

Authored by gpt-6.1-sol, MEDIUM, Codex harness in T3 Code.

What: #531 is merged into the #379 integration branch as28ba9e6b83, and #575 is aligned onto that result for current-base proof.
Why: Both fresh whole reviews approved ff8eb33 and its current exact-base CI was green; the merged tree exactly matches that reviewed candidate.
So what: The driver continues #575 revalidation/review and remaining #385/#386 acceptance; no main merge, deploy or running-install restart has occurred.

Verified live #531 MERGED at12:34:41 Warsaw on2026-10-20, squash commit 28ba9e6b83b89fb9739d5eb20014b76fe74d35b0, tree 9cf5bf54da44ce3a874b388308e35bf771f4c094, equal to reviewed head ff8eb336caf233086e250ddb0a2ef259724474d4. Driver workspace clean and fast-forwarded to the same commit/tree; remote integration matches. Final premerge gate verified exact head/basea6ba, OPEN/MERGEABLE/notdraft, all26 rollup entries successful or permitted conditional skips, and newest review/independent=success created by authenticated lukemaj with actual findings URL #531 (comment). Merge used supported squash with exact match-head guard, no bypass or source-branch write.

Both full unchanged independent reports and their complete input ledgers are linked in that aggregate. N3 requested visibility is supported by approved143369379 content; no new policy/grant is inferred. Optional compatibility/cost/follower/provider/interrupt-trigger limits remain explicit; they are not final acceptance proof. #369 remains open for the parent outcome, as do #379 and remaining children. Internal integration is not main delivery.

#575 topology after both R1 reviews became terminal: parent preserved checkpoint/575-before-531-d11e8a, rebased only its own isolated component branch onto exact28ba and pushed with an explicit lease against oldd11. New head c8c3862ba806b59e511a98344fddbc624cf19fd7, tree 72482028e2ba56de77cc17dd55eb6caaf609c6f8. The complete single-file +16/-1 diff remains byte-identical, SHA256 ef360799f7c5554f411b907110d4e8489ee8de80fd2c60d2238ecb24f7340193. Same author job resumed for current committed nine-case/scoped proof and handoff updates only; fresh BOTH whole reviews and current CI will follow. Old-base approval is not silently transferred.

The #531 author branch/worktree and checkpoints are retained under the user's explicit sole-author ownership; parent has not modified or retired them. This is ongoing internal integration, not final task finalization. Task-private proof remains outside source. #575 worktree is still needed for current-base proof/review. Separate #576 product repair runs in its own main-based workspace with separate final merge approval required.

Requirements and who asked: User instructs driver to merge #531 into integration once review passes and complete #379 through reviewed components, current-main alignment and cumulative proof.
Deleted: A second permission request for authorized internal integration, force/bypass merge, author-branch writes, and historic gate transfer to the rebased #575.
Bottleneck: #369 notice integration is complete; #575 current-base gate and remaining remote secret/catalog/parity acceptance now govern further progress.
Checked myself: Both full review reports/no-pending states, exact live CI/statuscreator/head/base, successful merge result, integration fastforward/tree equality and clean state, preserved #575 checkpoint and identical rebased component diff.

feat(peer): answer remote descendant requests from the source (#416)

feat(peer): ship reviewed cross-machine parity work (#379) (#584)

Navigation Menu

Repository navigation

feat(peer): surface and answer remote descendant requests from the parent #369

Parent

Outcome

Elon record

Acceptance criteria

Non-goals

Blocked by

Required proof

Prior record

Identity, authority and phase

Actual source and ownership map

Provider semantics to preserve

Minimum consumer requirements for the parent to assign

Collision and possible private scope

Representative registered red/proof targets, not executed this phase

Upstream-first evidence

Preserved prerequisites and limitations

22 remaining items

Exact source evidence

Exact candidate and ownership

Owner repair and equivalence

Prior finding dispositions on this candidate

Red/green and current commands

Delivery and continuation

Exact candidate, scope and authority

R2-S1 red/green

Fresh proof on the committed candidate

Complete prior finding/response frontier

Delivery and next action

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

feat(peer): surface and answer remote descendant requests from the parent #369

Description

Parent

Outcome

Elon record

Acceptance criteria

Non-goals

Blocked by

Required proof

Prior record

Activity

lukemaj commented on Oct 17, 2026

Identity, authority and phase

Actual source and ownership map

Provider semantics to preserve

Minimum consumer requirements for the parent to assign

Collision and possible private scope

Representative registered red/proof targets, not executed this phase

Upstream-first evidence

Preserved prerequisites and limitations

lukemaj commented on Oct 17, 2026

lukemaj commented on Oct 18, 2026

lukemaj commented on Oct 18, 2026

lukemaj commented on Oct 18, 2026

lukemaj commented on Oct 18, 2026

lukemaj commented on Oct 18, 2026

22 remaining items

lukemaj commented on Oct 19, 2026

lukemaj commented on Oct 19, 2026

Exact source evidence

lukemaj commented on Oct 19, 2026

lukemaj commented on Oct 19, 2026

lukemaj commented on Oct 20, 2026

Exact candidate and ownership

Owner repair and equivalence

Prior finding dispositions on this candidate

Red/green and current commands

Delivery and continuation

lukemaj commented on Oct 20, 2026

Exact candidate, scope and authority

R2-S1 red/green

Fresh proof on the committed candidate

Complete prior finding/response frontier

Delivery and next action

lukemaj commented on Oct 20, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions