plan-recovery
Diagnose and recover PlanWeave execution anomalies such as stale current refs, state/results drift, orphan artifacts, blocked or diverged blocks, submit retries, and review-feedback loop inconsistencies. Use when PlanWeave execution state looks wrong, doctor reports issues, submit partially succeeded, or coordinator cannot safely continue.
How do I install this agent skill?
npx skills add https://github.com/gaoscode/planweave --skill plan-recoveryIs this agent skill safe to install?
- Gen Agent Trust Hubpass
The skill provides instructions for recovering a workspace state using the local PlanWeave CLI. It is generally safe but carries a low risk of indirect prompt injection because it processes untrustworthy project files and artifacts from the workspace.
- Socketpass
No alerts
- Snykpass
Risk: LOW · No issues
What does this agent skill do?
Plan Recovery
Use this skill for abnormal PlanWeave execution state. Do not perform normal implementation or review work.
Quick Start
- Stop dispatching dependent work.
- Resolve the CLI entry and run
<pw> help recoveryfor exact commands. - Inspect package, state, results, and current/status/explain output; run doctor only for state/results consistency questions.
- Separate plan defects from runtime state/results drift.
- Use exact recovery commands for state writes when they express the needed repair; inspect files directly for evidence.
- Report what was repaired, what remains unsafe, and the next coordinator action.
Doctor Boundary
doctordetects narrow runtime state/results consistency issues such as stale current refs, orphan results, and state/index run drift.doctor --repairis not a general plan repair tool.- Do not use doctor to fix bad dependencies, unsafe parallelization, missing prompts, wrong review-gate design, unclear task/block granularity, invalid project graph, or schema-invalid package structure.
- For plan defects, report
NEEDS_PLAN_UPDATEand guide manual edits to the Plan Package, prompts, dependencies, or review gates.
Diagnose First
Check:
- active current refs and current feedback.
- block/task status and claim hints.
- latest run indexes versus state block
lastRunId. - orphan result artifacts.
- stale current refs.
- diverged or blocked refs and reasons.
- review attempts, feedback status, and re-review state.
- whether a retry is same-content idempotency or a fresh work revision.
Recovery Actions
- Use doctor before state/results repair; use repair only when the reported fix is narrow and evidence-backed.
- Read package/state/results directly when CLI output is incomplete, contradictory, or too summarized for diagnosis.
- If CLI cannot safely claim/submit, have the coordinator read package/state/results and manually assign exact refs until write-back is safe.
- Use blocked/unblocked commands for external prerequisites or temporary stops.
- Use diverged/resolve-divergence when implementation reality no longer matches the Plan Package.
- For submit partial success, look for persisted report/result artifacts before creating new runs or attempts.
- For feedback resolved but still current, re-check current/status before submitting anything.
- For parallel mismatch, compare sequential and parallel claimability before declaring no work.
- If parallel mismatch comes from package design, adjust dependencies, parallel safety, or review gate placement manually instead of running repair.
Manual Fallback
Direct file inspection is part of recovery:
- read
project-graph.jsonwhen present, each canvaspackage/manifest.json, source prompts,state.json,results/, andplanweave schema project/manifest/state/layoutwhen structure is suspect. - verify metadata belongs to the same task/block/run before trusting indexes.
- use CLI/runtime commands for mechanical workspace operations, validation, recovery transactions, and runtime state/results changes.
- edit Plan Package semantic files directly inside CLI-returned workspace paths when the repair is a plan update:
project-graph.jsondependencies, canvasmanifest.jsontasks/blocks/edges/acceptance/prompt paths, and source prompt Markdown. - use narrow CLI edit commands when they exactly express the semantic change; otherwise update the source package files and prompts directly.
- edit
state.jsonorresults/directly only when no recovery command expresses the fix, the exact invariant is known, and the user accepts the risk. - after direct plan edits, run canvas-scoped validation for edited canvases and project validation when
project-graph.json, canvas edges, orcrossTaskEdgeschanged. - preserve evidence paths and explain any direct state/results write.
Stop Conditions
Stop and ask for user or plan update when:
- package dependencies or prompts are wrong.
- state cannot be repaired without guessing.
- multiple current refs conflict with actual work ownership.
- result metadata points to the wrong block or task.
- review/feedback state is contradictory after doctor.
- a recovery command would hide a real product or contract defect.
Recovery Report
Report:
- verdict:
RECOVERED,NEEDS_PLAN_UPDATE, orBLOCKED. - symptoms and evidence.
- commands run.
- files inspected.
- repairs made or intentionally skipped.
- next safe command for
plan-coordinator.
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/gaoscode/planweave/plan-recovery">View plan-recovery on skillZs</a>