Devin FoleyandPaperclip
837add0caa
## Thinking Path > - Paperclip keeps agent tasks moving through owned execution and durable waiting paths. > - A consumed review path gets one repair wake before board escalation. > - That wake records an instruction to save a new monitor or another real review path. > - The canonical wake builder omitted that instruction from the agent prompt. > - This pull request delivers it through the existing disposition-repair section. > - The repair limit and execution authority stay unchanged. ## Linked Issues or Issue Description **What happened?** An agent opened a pull request and promised to watch CI and external review. It ended without scheduling a monitor. Its bounded review repair also ended without a maintained path, so the task became blocked for a board decision. The saved repair context contained the instruction, but the canonical wake payload and rendered prompt did not. **Expected behavior** The repair prompt tells the agent to save a one-shot monitor before promising another check. Both fresh and resumed sessions receive that instruction and the one-attempt limit. **Steps to reproduce** 1. Create a review-path recovery with `decideIssueReviewPathRecovery` for a task whose review path was consumed. 2. Pass its recorded context through `buildPaperclipWakePayload` and wake serialization. 3. Render a fresh or resumed prompt. Before this fix, it has the wake reason but no repair instruction. **Paperclip version or commit** Base: `9c442cf48e6`. Current head: `f38d65891c8d9f2b38eb76a8578e870debc2b1e6`. **Deployment mode** Observed on managed sandbox execution with the legacy Claude ACP adapter. The corrected canonical payload serves the shared adapter prompt path. Refs #15764 introduced the bounded exhausted-repair disposition. This change preserves that limit. Open monitor-retry work such as #17171 addresses an already scheduled check; this defect occurs before a check exists. Searches found no duplicate for this missing prompt field. ## What Changed - Build the review-path repair disposition in its owning recovery module, using the server instruction and fixed one-attempt bound. - Project it into the existing canonical `dispositionRepair` field. No new adapter field or execution protocol is added. - Ignore arbitrary replacement instructions and attempt-limit values. A pending repair keeps its instruction across a connection-pool retry that changes the wake reason. Ordinary wakes without the typed repair marker ignore raw instruction text. Existing legacy disposition repairs retain their values. - Add a regression across the real recovery decision, wake builder, serializer and fresh/resumed prompt selectors. Update execution guidance and the affected recovery maps. ## Verification - Verified red before the source fix: two regression cases failed because the instruction was absent; the sibling repair case passed. The same cases pass after the fix. - Passed 194 focused and safety tests across wake preparation, review recovery, adapter prompt handling, and database-backed liveness escalation. The connection-pool retry case also failed before its correction and passes after it. - Replayed the recorded incident wake through the fixed boundary. The old canonical prompt lacked its saved instruction. The corrected serialized prompt includes it with a 1/1 bound. This replay did not execute a provider turn or change the source task. Private evidence stays outside the repository. - Passed `pnpm -r typecheck` after the pool-retry correction. Passed `pnpm build` on final head `f38d65891c8d9f2b38eb76a8578e870debc2b1e6`. Map checks and `git diff --check` pass. - [Full CI](https://github.com/paperclipai/paperclip/actions/runs/40005890857) passed on that head: 54 checks passed and two optional checks were skipped. This includes the full test shards, typecheck, build, and browser suites. The local `pnpm test:run` attempt was stopped after the CI test shards passed. No complete local full-suite pass is claimed. - [Apex review](https://github.com/paperclipai/paperclip/pull/17353#issuecomment-6372160694) is 5/5 on the final head, with no open findings. The first review found the pool-retry case. Its regression and correction are included. The branch has no merge conflicts. - Test setup first required the plugin SDK build, then an empty credential-history database boundary. Those setup errors are not the claimed red. The temporary recorded-wake probe was removed before final validation. - Hosted follow-up acceptance is not run. The source task remains blocked and needs a retry after rollout. This PR does not claim that prompt delivery itself schedules a monitor or that every model will follow the instruction. ## Risks - This fixes instruction delivery. The agent must still save a valid monitor or another waiting path. The existing bounded escalation remains if it does not. - The wake stays a normal-model repair. It does not become status-only recovery and can continue useful work. Existing ownership, approval, budget, pause, and monitor admission gates remain in place. - No schema, credential, telemetry, API, model, or retry-budget change. ## Model Used OpenAI Codex, based on GPT-6, assisted with diagnosis, code editing, shell execution and verification. The exact serving model ID and context-window limit are not exposed in this session. ## Checklist - [x] I have included a thinking path that traces from project context to this change - [x] I have specified the model used (with version and capability details) - [x] I have checked ROADMAP.md and confirmed this PR does not duplicate planned core work - [x] I have searched GitHub for duplicate or related PRs and linked them above - [x] I have either (a) linked existing issues with `Fixes: #` / `Closes #` / `Refs #` OR (b) described the issue in-PR following the relevant issue template - [x] I have not referenced internal/instance-local Paperclip issues or links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip` URLs) - [x] My branch name describes the change (e.g. `docs/...`, `fix/...`) and contains no internal Paperclip ticket id or instance-derived details - [x] I have run tests locally and they pass (194 focused and safety cases; full coverage passed in CI) - [x] I have added or updated tests where applicable - [x] I have updated relevant documentation to reflect my changes - [x] I have considered and documented any risks above - [x] All Paperclip CI gates are green - [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups - [x] I will address all Greptile and reviewer comments before requesting merge --------- Co-authored-by: Paperclip <noreply@paperclip.ing>
36 KiB
Recovery — agent access
Product map: Recovery.
Feature ID: recovery.
Source revision: 3bbc2104b65caf9f7763938a01260a5673e12f7a.
Review depth and executed proof are separate. The operation records enumerate the named behavior and access boundaries. Their evidence states describe the reviewed sources; run results and environment limits belong in the PR or verification artifacts. A listed API or source test does not establish a live authorized journey. Follow the map contract.
Source-reviewed operation contracts at 3bbc2104b65caf9f7763938a01260a5673e12f7a; proposed checks remain not-run. Preserve existing native/legacy ownership, retained-source and historical failed-evidence caveats. Automatic classification is server work, not an agent retry API. REST, native binding, operator, provider and WebSocket clients have distinct authority.
diagnosis
Capability ID: recovery/diagnosis.
Review: operation-reviewed.
Read causes, owner/attempt state and retained obligations without claiming successful recovery.
read-action-and-source
API: present — GET /api/issues/{id}/recovery-actions, GET /api/issues/{id}, GET /api/heartbeat-runs/{runId}.
Agent access: available — permitted agent REST/native call_api reads and CLI issue recovery-actions for that operation; actual native rollout/run binding required.
Policy: Same-company task read authorization, private-task/responsible-user scope and run privacy where run detail is used. Related diagnostics omit unauthorized resource details. Recovery action response includes active plus supported retained settled replay-blocked/native export holds, not unrestricted history. Read revalidation may retire stale action through system lifecycle; no reader participation/physical copyback.
Evidence: source-only — recovery routes, source actions; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Compare exact company/task/action/source run/owner/attempts/next-action across views and reload. Private/foreign read denied. Repeated Done/cancelled read retains exact unresolved physical owner hold.
Gaps: Recorded cause or successful GET isn't a running retry or restored files.
read-blocker-wake-and-subtree-diagnostics
API: present — GET /api/issues/{id}/diagnostics/blockers, /diagnostics/wakes, /diagnostics/subtree.
Agent access: available — permitted REST or native call_api; no dedicated CLI issue diagnostics command confirmed.
Policy: Same-company task read authorization, private-task/responsible-user scope and run privacy where run detail is used. Related diagnostics omit unauthorized resource details. Wake internal IDs depend on company-scope read; linked blocker/child/wake detail filtered. Diagnostics bounded/read-only.
Evidence: source-only — recovery routes, diagnostic discovery; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Read source diagnosis and compare visible blockers/wakes/subtree; unauthorized linked nodes omitted and no mutation/new attempt on repeat.
Gaps: CLI recovery-actions is supported; don't invent diagnostics subcommands. Coverage of every diagnostic inference remains separate.
inspect-run-recovery-host
API: present — GET /api/issues/{id}/recovery-actions; GET /api/issues/{id}; GET /api/heartbeat-runs/{runId}. The run UI host adds no separate recovery API.
Agent access: available — permitted agent REST/native call_api reads and CLI issue recovery-actions for that operation; actual native rollout/run binding required.
Policy: Same-company task read authorization, private-task/responsible-user scope and run privacy where run detail is used. Related diagnostics omit unauthorized resource details. Run workspace controls only for failed workspace_validation_failed run linked to source task with live workspace_validation action; other failures use source task.
Evidence: source-only — run host, recovery routes; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Exact eligible run and task agree; deny missing authority and hide stale/other-code controls; reload same persisted action.
Gaps: Component host test doesn't prove actual filesystem/provider repair.
bounded-continuation
Capability ID: recovery/bounded-continuation.
Ordinary task Retry uses POST /api/issues/{id}/retry. This is authenticated
board task-mutation authority, not agent retry authority. The only request
fields are a UUID idempotency key and observed source run (or null). It accepts
no paths, stop attestations or client-selected strategy. The source-scoped
exact-run and maintenance APIs retain their existing contracts. Connected-chat
Retry reuses chat admission and its external identity. It can persist a wait
only for the exact current system recovery generation. The chat owner resumes
that same deferred receipt through normal admission after containment. The
board intent and external principal remain separate authorities; current task,
connection, pause and budget gates still apply. See the
durable contract. Agent Chat also requires its
current conversation owner and feature enablement at admission and locked
dispatch; generic board mutation permission does not replace chat ownership.
Local PostgreSQL and assembled
browser qualification do not establish hosted agent access or recovery.
Review: operation-reviewed.
Supported restoration and automatic attempts retain source/owner/attempt authority.
observe-automatic-attempt
API: present — GET /api/issues/{id}/recovery-actions; GET /api/issues/{id}; GET /api/heartbeat-runs/{runId}. Automatic recovery has no generic agent-enqueue endpoint.
Agent access: available — permitted agent REST/native call_api reads and CLI issue recovery-actions for that operation; actual native rollout/run binding required.
Policy: Server-owned bounded continuation uses exact run/source identity, current owner/normal admission and durable episode limits. Stop, pause, budget, durable wait, owned lifecycle and existing execution remain gates; native exact completion proposal requires source/session/turn+termination proof, not just terminal row.
Evidence: source-only — source actions, legacy bounded decision, native recovery; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Observe attempts/retry timing and actual useful result; concurrent sweeps don't reset budget or duplicate turn. Operator Stop/chat idle must not trigger unwanted recovery.
Gaps: Native/Legacy and live provider takeover need separate proof; preserve existing failed callback evidence.
A review-path repair uses the existing canonical disposition-repair prompt section. The server fixes its instruction and one-attempt bound. This adds no API, monitor authority, credential access, or bypass for task and approval gates. A pending repair retains that section across connection-pool retries. Ordinary wakes without the typed repair marker ignore stale raw instruction text. Prompt controls cover fresh/resumed delivery and existing disposition repairs; hosted execution and durable monitor acceptance remain separate verification.
resolve-permitted-source-action
API: present — POST /api/issues/{id}/recovery-actions/resolve with exact actionId/outcome/sourceIssueStatus/resolutionNote.
Agent access: partial — eligible REST/CLI issue recovery:resolve; native call_api standard/skill_test. false_positive/cancelled and special reconciliation/workspace paths board-only.
Policy: Agent needs issue:mutate, authenticated run, low-trust exclusion and cross-issue allowance; source assignee/action owner or admitted active-checkout management override. Action ownership does not itself grant source-task mutation rights. Board and agent route policies remain distinct. Status changes separately require source owner/current execution participant/management grant or guarded exact-owner handback. blocked needs unresolved first-class blocker. Restored todo cannot bypass task/project/subtree/budget/review/approval/competing-run gates. Dedicated workspace causes reject generic retry.
Evidence: source-only — recovery routes, resolution schema; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Eligible owner resolves fresh action, reload recorded outcome/state and useful continuation. Peer/recovery-owner source completion, stale target, no blocker, pause/budget/review gates denied without side effects.
Gaps: CLI body lacks executionReconciliation/workspaceBaseRef; no generic agent permission for operator source repair.
replay-and-deliver-resolution
API: present — POST /api/issues/{id}/recovery-actions/resolve; GET /api/issues/{id}/recovery-actions. Server wake, reconciliation and chat outboxes own delivery.
Agent access: partial — permitted resolution followed by current eligible agent context; no arbitrary client self-wake.
Policy: Lock exact source/action. Same settled actionId returns retained receipt/no repeated log/wake. Todo wake only when required state/owner changed; reconciled executions use own durable delivery, restricted chat uses exact authorized failed request and real board user/current channel authority. No generic duplicate wake alongside durable owner.
Evidence: source-only — recovery routes, reconciliation outbox; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Repeat/concurrent same action and lost acknowledgment -> one recorded resolution/eligible continuation. Changed/new action/source never overwritten; restricted chat either retains exact provenance or stays unresolved.
Gaps: Receipt success isn't provider execution; actual retry may remain gated/pending.
source-action
Capability ID: recovery/source-action.
Review: operation-reviewed.
Source-scoped repair ownership and retained files survive task status changes.
read-and-follow-owner
API: present — GET /api/issues/{id}/recovery-actions; GET /api/issues/{id}; GET /api/heartbeat-runs/{runId}. Linked repair tasks retain their own access checks.
Agent access: available — permitted agent REST/native call_api reads and CLI issue recovery-actions for that operation; actual native rollout/run binding required.
Policy: Same-company task read authorization, private-task/responsible-user scope and run privacy where run detail is used. Related diagnostics omit unauthorized resource details. Original assignee, recovery owner and return owner remain distinct. Completion/cancellation doesn't remove physical workspace hold; retired notices shouldn't hide retained obligation.
Evidence: source-only — recovery routes, source actions; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Follow source/repair/return owner, reload exact IDs; foreign/mismatched source hidden. Unresolved file hold persists after Done/cancelled.
Gaps: Repair task existence does not authorize source mutation or prove recovery.
record-retained-file-repair
API: present — POST /api/issues/{id}/recovery-actions/resolve with exact actionId, restored outcome, unchanged source status and workspace repair evidence.
Agent access: restricted — board with accessible source and exact file/stop evidence. Current branch does not additionally enforce runtime:manage; agent REST/native denied.
Policy: Original terminal retained source validated independently: exact issue/run/lease receipt, stop/process/coordinator/no unreleased authority. Successful legacy source may record files only without changing its result. Preserve current status/owner/newer active action; no replay/wake. Same repair evidence receipt dedupes. Do not document stronger runtime grant than route enforces.
Evidence: source-only — recovery routes, source/stop validation; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Repair exact historical files, reload unchanged current owner/status/newer action and source result; no wake. Deny agent/foreign/run mismatch/running source/no stop evidence. Duplicate returns saved repair.
Gaps: Existing desired runtime-manager wording is stronger than implementation; product policy decision separate from map edits. Actual files/artifact must be verified before attestation.
add-new-execution-evidence
API: present — POST /api/issues/{id}/recovery-actions/resolve with executionReconciliation on the exact settled automatic replay-blocked action.
Agent access: restricted — board; exact source read, restoration/status and independent execution evidence.
Policy: Can reopen matching settled no-replay disposition only restored/todo with no competing active action and new accepted source proof; plain repeat returns receipt. Active coordinator/newer turn/retained uncertain authority disallows blind replay.
Evidence: source-only — recovery routes, reconciliation proof; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — New verified evidence permits one eligible delivery; duplicate safe. Competing current action or missing/mismatched evidence denied and audit history preserved.
Gaps: Elapsed timeout/client providerStopped flag alone is not proof.
exhaustion
Capability ID: recovery/exhaustion.
Review: operation-reviewed.
Exhausted and overdue recovery advertises only the current eligible operator action.
read-attempt-budget
API: present — GET /api/issues/{id}/recovery-actions; GET /api/issues/{id}; GET /api/heartbeat-runs/{runId}.
Agent access: available — permitted agent REST/native call_api reads and CLI issue recovery-actions for that operation; actual native rollout/run binding required.
Policy: Durable maxAttempts is execution budget; exhaustion retires automatic wake/monitor/timeout paths. Due-but-not-started retry differs from live retry. Same episode identity persists through equivalent ref spellings/new run IDs/classifier upgrades.
Evidence: source-only — source actions, attempt display; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Exhaust episode, reload count/gate/no additional automatic wake; overdue retry remains visible as missed rather than working.
Gaps: Reading attempts isn't qualification of every provider retry lane.
request-explicit-retry
API: present — POST /api/issues/{id}/recovery-actions/resolve with exact actionId, restored outcome and todo source status, or the separately authorized workspace operation below.
Agent access: partial — source-permitted resolution; board-escalated/reconciliation/workspace decisions keep their explicit board gates.
Policy: Exhausted disposition still blocked, exact board-owned deliberate wait/board-escalation action and recorded return owner assigned/invokable; no blockers/pending interaction; normal handback pause/project/budget/approval/run gates. Cannot reset attempts from client.
Evidence: source-only — recovery routes, retry control; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — One click waits real acknowledgment, reload one continuation; repeated click stale-owner/paused/blocker/question/budget cases rejected. No new episode budget from repeated request.
Gaps: An acknowledged retry can remain pending or fail; report actual useful result separately.
workspace
Capability ID: recovery/workspace.
The internal preservation resume path grants no board or agent execution API. It requires the exact source action, company, run, lease and environment, a held physical workspace owner, and its current stop-only cleanup claim. The ordinary resume and destruction fences still apply to protected sources. The assembled-runtime regression checks this separation before any provider call.
Managed missing-source reconciliation uses the server-only Cloud
POST /v1/tenant/sandbox-absence protocol. A host-bound capability authenticates
the stack; board and agent credentials cannot call it. Core requires a fresh
Cloud signature for the exact company, issue, run, lease, allocation and physical
owner. This can establish execution absence, never recovered files or successful
task completion. Normal Retry admission and external-effect fences still apply.
Signature and PostgreSQL regressions are local evidence, not deployed acceptance.
Review: operation-reviewed.
Each workspace repair/export/reconstruction/physical-owner operation has its own proof and authority.
recover-retained-files
API: present — POST /api/issues/{id}/recovery-actions/recover-workspace with actionId/runId/idempotencyKey.
Agent access: restricted — board task-read and issue:mutate; not runtime:manage. Agents denied through REST/native fallback.
Policy: Exact terminal retained source/lease/reference; independent stop and physical ownership checks; no newer writer. Bounded queued repair preserves newer files and has retry state. Matching key/recovered state dedupes; permanent blocked source refuses. Recovery starts no provider turn.
Evidence: source-only — recovery routes, retained recovery; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Allowed board requests exact source, inspect persisted copied files/artifact and cleanup. Repeat key/transient retry bounded. Agent/wrong company/newer writer/uncertain owner/no stop denied.
Gaps: Client never certifies recovered files by posting this request; hosted recovery separately not-run.
retry-accepted-result-export
API: present — POST /api/issues/{id}/recovery-actions/retry-workspace-export with exact actionId/runId/repairNote.
Agent access: restricted — board task read + runtime:manage; native agent call_api preserves denial.
Policy: Eligible native export-repair cause, saved result/coordinator/retained source and independent stop-only proof. Reuses original result/lease and bounded finalization, not provider execution. Pending duplicate audits once. Unsafe-export automatic path/manual admission and active exporter/new owner rejected.
Evidence: source-only — recovery routes, export retry; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Repaired sandbox export commits original accepted result without new provider turn; repeated click one audit/intent. Deny mismatched source/cause, active writer and unprivileged actor.
Gaps: Provider stop/probe/export must be actual evidence, not inferred from 202.
resume-after-owner-stop
API: present — POST /api/issues/{id}/recovery-actions/resume-workspace-finalization with actionId/runId/ownerToken/stopped flag/stopEvidence.
Agent access: restricted — board task read + runtime:confirm_owner_stop; active company owner/admin, trusted local board or current instance admin. Ordinary runtime manager/agent denied.
Policy: Exact native owner/result/task/run; independent platform verification all controller+copyback children stopped. Lease expiry/advisory disconnect not proof. Active lock/local writer/newer run blocks. Exact replay dedupes; reconciler exports/commits saved result and retains approval/monitor waits without fresh provider turn.
Evidence: source-only — recovery routes, owner proof, administrator gate; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — After actual platform stop proof permitted admin records exact owner; inspect original saved result/files/holds and no fresh turn. Deny wrong owner/member/runtime manager/agent/live writer/stale Cloud admin.
Gaps: Automatic managed-cloud signed stop proof is separate; never invent platform attestation for local acceptance.
reconstruct-destroyed-source
API: present — POST /api/issues/{id}/recovery-actions/reconstruct-workspace with action/run/statusVersion/task attachments/lost-work acknowledgement/outcome evidence.
Agent access: restricted — board task read + runtime:manage.
Policy: Server confirms original source destruction, hashes same-task saved attachments, exact current status/source/no contradictory execution. Retained/uncertain source denied. Same request hash dedupes, changed request/state conflicts. Preserve failed old export/history/budget; one fresh normal-admission continuation from recorded sources. Intervening run cancellation proof remains exact.
Evidence: source-only — recovery routes, reconstruction; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Inspect acknowledged lost edits/selected hashes and actual reconstructed files. Retry/concurrent same request one continuation; wrong/private attachment, retained source/status change/false destruction denied.
Gaps: Original unexported edits remain lost; fresh reconstruction isn't successful original export or permission to replay completed external effects.
source-files-only-repair
API: present — POST /api/issues/{id}/recovery-actions/resolve with unchanged source status and exact file-repair evidence.
Agent access: restricted — board with readable source and exact source/action/run proof. This branch does not independently check runtime:manage.
Policy: Record recovered-file evidence only after independent stopped-source validation. Keep the original result and current task status, owner and execution unchanged; do not reopen or wake work. No invented runtime grant can be inferred from the map.
Evidence: source-only — recovery routes, file evidence; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Repair a matching retained-file obligation and reload its evidence. Reject missing/stale stop/source proof and agent identity. Compare unchanged task status, accepted result, owner and run count.
Gaps: This is the actual route boundary. The stronger runtime:manage requirement on dedicated workspace operations must not be imputed to generic file-only resolution.
repair-starting-branch
API: present — POST /api/issues/{id}/recovery-actions/resolve with the matching workspaceBaseRef repair shape.
Agent access: restricted — board with runtime:manage, readable source, exact recovery action and eligible source mutation authority.
Policy: Require the current blocked configuration_incomplete/base_ref_unresolved failure, matching assignee/project/requested ref and effective git-worktree strategy. Replacement branch differs from the failed ref. Reject stale configuration, incompatible reconciliation evidence or mismatched source; normal retry admission remains required.
Evidence: source-only — recovery routes, file evidence; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Repair a matching failed starting ref, reload the saved effective strategy and observe one eligible retry. Reject ordinary non-manager, agent, changed ref/project/owner/configuration and stale action.
Gaps: The simple CLI resolve wrapper does not express this branch-repair shape. Source-file-only repair above has different authority and effects.
reconcile-branch
API: present — POST /api/execution-workspaces/{id}/reconcile-branch modes forward, override or quarantine_restore.
Agent access: restricted — board + workspace privacy + runtime:manage; agent/native fallback denied.
Policy: Source task required. Forward re-inspects ancestry or confirmed missing recorded ref+clean resolved target. Override requires reason and current inspection. Quarantine_restore preserves dirty work on rescue ref, respects physical/runtime contention. No browser ancestry hint substitutes source check.
Evidence: source-only — branch route, branch/filesystem checks, mode schema; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Exercise offered mode on disposable branches, preserve before/after files/rescue ref and audit. Deny unrelated workspace/private/agent/no grant, wrong ancestry/no reason/live writer. Reload binding and actual files.
Gaps: No dedicated CLI reconciliation wrapper found; 202/toast isn't filesystem integrity proof.
unknown-outcome
Capability ID: recovery/unknown-outcome.
Review: operation-reviewed.
Inspect actual side effects before explicit source-specific operator resolution.
inspect-and-reconcile-execution
API: present — GET /api/issues/{id}/recovery-actions; GET /api/heartbeat-runs/{runId}; POST /api/issues/{id}/recovery-actions/resolve with exact run and executionReconciliation evidence.
Agent access: restricted — source read plus board execution-reconciliation decision; agent may inspect own permitted context, not attest provider state.
Policy: Server independently checks terminal source/task/owner or reviewer, stopped leases/process, no coordinator/successor/unreleased authority and workspace evidence. Outcome completed/not_performed/mixed informs continuation without repeating completed effects. Matching source-only file repair requests do not enqueue work.
Evidence: source-only — recovery routes, execution proof; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Inspect external provider first; wrong source/flag without stop evidence/live owner denied. Record truthful outcome -> one eligible continuation or retained hold; duplicate doesn't replay.
Gaps: Unknown stays unknown until corroborated. Evidence text alone can't override independent fences.
resolve-chat-publication
API: present — POST /api/chat-endpoints/{endpointId}/publications/{publicationId}/resolve action mark_delivered/retry_anyway/cancel, optional fileTransfer phase/version.
Agent access: restricted — board endpoint/company + tools:manage_connections (local implicit/instance-admin exceptions) and board user ID.
Policy: Only current delivery_unknown; credential lease/exact publication; stage-specific Teams/Slack/Photon rules and current phase apply. mark delivered/cancel settle; retry_anyway explicitly records duplicate risk and schedules bounded send. Terminal interaction prompt can't retry stale actions.
Evidence: source-only — channel management, publication resolution; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Inspect real provider, choose truthful resolution, reload/audit exact source/result. Agent/no connection grant/wrong endpoint/stale phase/nonunknown state denied. Retry must not silently duplicate.
Gaps: Accepted send request isn't recipient receipt; per-provider live proof required.
replay-failed-publication-or-resolve-action
API: present — POST /api/chat-endpoints/{endpointId}/publications/{publicationId}/replay; POST /api/chat-endpoints/{endpointId}/actions/{actionId}/resolve.
Agent access: restricted — endpoint-management board/company/tools:manage_connections; resolve requires board user ID.
Policy: Replay only failed state, expected attempt generation, live pending interaction prompt; delivery_unknown and stage-owned file transfer reject generic replay. Action resolution dispatches authoritative provider_effect or Slack task-start contract with mark_delivered/retry_anyway/cancel; no arbitrary execution retry.
Evidence: source-only — replay and effects, action kind dispatch, route gates; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Blind uncertain replay denied, failed eligible replay one generation. Resolve actual provider effect/task-start after inspection; races/wrong source/stale action no extra effect.
Gaps: No dedicated agent/CLI decision path; provider-specific stage semantics remain explicit gaps.
resolve-uncertain-email
API: present — POST /api/companies/{companyId}/email/deliveries/{publicationId}/resolve outcome sent/failed plus providerMessageId.
Agent access: restricted — board company + tools:manage_connections (local implicit/instance-admin exceptions).
Policy: Only current uncertain send under endpoint/thread lease. sent requires live provider read proving exact inbox, x-paperclip-publication-id and sent label; failed marks provider-not-sent and requires new explicit send. Repeat settled request conflicts; never auto-resend unknown email.
Evidence: source-only — email manager, provider proof; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Use actual uncertain disposable send; wrong inbox/header/label/message ID rejected, truthful sent/failed persists. Failed resolution emits no new send; competing decisions one outcome.
Gaps: Fixture AgentMail response isn't live provider inspection; public assistant/native agent cannot approve unknown outcome.
user-continuation
Capability ID: recovery/user-continuation.
Review: operation-reviewed.
Saved post-stop instructions wait for real cleanup and normal execution authority.
save-and-observe-wait
API: present — POST /api/issues/{id}/comments; GET /api/issues/{id}; GET /api/issues/{id}/comments; GET /api/issues/{id}/queued-comments; GET /api/heartbeat-runs/{runId}.
Agent access: partial — human author creates own instruction; agent REST/CLI/native report_progress creates attributed agent content, never human identity.
Policy: Ordinary source-task comment/steering company/privacy/responsible-user policy; saved message independent of provider stop. Existing operator Stop with no new instruction respected. No new client API to manufacture saved-human provenance.
Evidence: source-only — recovery routes, saved continuation, waiting messages; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Stop disposable owned run, submit distinct follow-ups during cleanup; reload sees ordered waiting messages without premature turn. Deny foreign/private/wrong identity.
Gaps: Comment acknowledgment isn't agent answer or cleanup completion.
admit-once-after-stop
API: present — GET /api/issues/{id}; GET /api/heartbeat-runs/{runId}. Server reconsideration uses normal admission; there is no unrestricted retry endpoint.
Agent access: partial — eligible admitted agent/native bound wake receives saved input; delivery remains server-owned.
Policy: Exact prior run/lease/workspace stop proof, no competing execution, current owner/responsible user, normal subtree/project pause/budget/review gates. Durable ordered message/wake receipts avoid duplicates; never borrow authority from stale callback/other routine. Unused hold-denied admission needs no-execution/dispatch evidence before automatic reconsideration.
Evidence: source-only — legacy gates, native ownership, queued Stop; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — After confirmed cleanup one useful continuation consumes saved order; repeat sweep/reload no duplicate. Pause/budget/approval/unknown external outcome and reassignment prevent stale admission. Queued Stop leaves another routine cycle intact.
Gaps: Legacy/native/channel runtimes separately qualify; preserve original undispatched-hold and callback regressions.
reconnect
Capability ID: recovery/reconnect.
Review: operation-reviewed.
Browser transport recovery refetches persisted state without implying runner recovery.
subscribe-and-filter
API: present — GET /api/companies/{companyId}/events/ws with WebSocket Upgrade; GET /api/issues/{id} and GET /api/heartbeat-runs/{runId} remain separate REST reads.
Agent access: partial — browser session/trusted Cloud identity or local implicit board; standard unrevoked same-company agent API key via actual WS client. Native call_api cannot Upgrade; scoped keys/native JWT/personal MCP bearer not this path.
Policy: Upgrade membership/company checks; scoped key scopeConfig rejected. Delivery-time key/member and issue/project/run/skill/case privacy rechecks; redacted invalidation payload doesn't confer detail read. Personal MCP Events is a distinct protocol/authority.
Evidence: source-only — WS admission/filter; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Permitted browser/key gets only authorized events; foreign/scoped/revoked key/member/private resource denied/filtered after subscription. REST details still enforce own ACL.
Gaps: Actual WS client isn't automatically present in native run; avoid query tokens in evidence despite source compatibility.
disconnect-poll-and-refresh
API: present — GET /api/issues/{id}; GET /api/heartbeat-runs/{runId}; reconnect through the same WebSocket protocol. No event-replay API is implied.
Agent access: partial — LiveUpdatesProvider browser recovery; agents can poll permitted reads or use actual installed protocol client.
Policy: Bounded reconnect backoff; failed socket fallback polling, healthy open stops polling and invalidates active queries including initial-subscription gap. Events aren't replayed. Hidden/unmounted stops relevant loops; stale socket ignored; reconnect toast suppression distinct from content delivery.
Evidence: source-only — browser recovery; reviewed at 3bbc2104b65caf9f7763938a01260a5673e12f7a. No live provider proof.
Verification: not-run — Disconnect while server changes task/run/decision; reconnect/visibility return/refetch matches reload without duplicate comments/decisions/toasts. Initial-open gap and hidden/unmount loops checked.
Gaps: Healthy socket proves transport only, not provider restart or successful failed task. Mocks not deployed network proof.