eforge-build/eforge/eforge-recover

/eforge:recover

Inspect the recovery analysis for a failed PRD and act on the verdict — retry from scratch, continue and repair from preserved compiled artifacts, archive the original, or stop for manual replanning.

Workflow

Call eforge_queue_list to discover failed PRDs, read the recovery sidecar to surface the verdict and rationale, confirm the action with the user, and then call the matching tool. Use eforge_apply_recovery for retry and abandon. Use eforge_continue_repair for the continue-repair action when preserved artifacts are eligible. Manual use of this skill never auto-applies — always confirm. The only automation exception is daemon-owned recovery auto-resume when recovery.autoResume.enabled: true: it is disabled by default, may only auto-queue high-confidence continue-repair that passes compiled-resume artifact eligibility, consumes the configured bounded attempt budget, and emits visible audit/stop reasons before or after it decides. Manual confirmation controls remain available regardless of that policy.

Steps

Step 1: Identify the Failed PRD

If the user supplied a <prdId> argument, use it directly and skip to Step 2.

Otherwise, call eforge_queue_list (no parameters) and filter the response for items where status === 'failed'. Present the list to the user and ask which one to recover. If no failed PRDs are found, tell the user:

No failed PRDs found. Use /eforge:status to check the current build state.

Stop here if no failed PRDs exist.

Step 2: Read the Recovery Sidecar

Call eforge_read_recovery_sidecar with { prdId }.

• If the tool returns a 404 or the response contains a recoveryError field, offer to run the recovery analysis:

No recovery analysis found for {prdId}. Would you like me to run the analysis now? (yes / no)

If the user agrees, call eforge_recover with { setName, prdId } — source setName from the top-level sidecar setName field in a prior sidecar if available, otherwise ask the user to supply it. Then loop back to Step 2.

If the user declines, stop here.

• If the sidecar is present, continue to Step 3.

Step 3: Render the Verdict

Display the recovery report to the user:

PRD: {prdId} Verdict: {verdict} (retry / continue-repair / abandon / manual) Confidence: {confidence} (low / medium / high)

Rationale {rationale}

Completed work {completedWork — bullet list}

Remaining work {remainingWork — bullet list}

Risks {risks — bullet list}

Inspect sidecar-provided continue-and-repair fields. If continueRepairEligibility.eligible === true or recoveryOptions contains a recommended continue-and-repair option such as { kind: "continue-repair", action: "continue-repair", recommended: true }, present exactly one primary action:

This recovery sidecar recommends Continue and repair build from preserved compiled artifacts. Continue with eforge_continue_repair? (yes / no)

If continueRepairEligibility.eligible === false, show the bounded reason and do not recommend continue-and-repair. The mutating eforge_continue_repair tool validates eligibility server-side if the user explicitly asks to try after reviewing the warning.

If recoveryOptions contains kind: "compile-scope-context", treat it as non-mutating guidance: summarize its action, eligibility, source/failure kind, reason, and any bounded decomposition evidence summary, but do not map it to eforge_apply_recovery or eforge_continue_repair.

When a compile-scope-context option includes decompositionEvidence, render it as read-only context-managed planning evidence. Summarize only bounded fields such as failed unit ID, parent unit, depth, triggered limits, blockers, unresolved criteria counts, and split attempts. State that decomposition-exhausted is decomposition exhaustion, not a provider context-window rejection. Do not paste raw source, prompts, transcripts, or agent output, and do not infer an automated recovery action from the evidence.

Do not author replacement PRD content from the recovery sidecar. If decomposition evidence suggests a smaller follow-up scope, tell the user to inspect the bounded evidence and deliberately write or enqueue any replacement PRD outside this recovery action. The engine does not auto-author or auto-enqueue successor PRDs. If the verdict is manual, render the full report and use the manual guidance below.

Step 4: Confirm the Action

Ask the user to confirm the verdict-specific action:

• retry: "Retry PRD {prdId} from scratch? (yes / no)"
• continue-repair: "Continue and repair build {prdId} from preserved compiled artifacts with eforge_continue_repair? (yes / no)"
• abandon: "Archive the failed PRD {prdId} (this cannot be undone)? (yes / no)"
• manual: Render the full markdown report and stop. Tell the user:

This verdict requires manual review / manual replanning. Review the report above, decide the bounded repair plan outside of eforge, and enqueue any replacement work deliberately. No automated apply-recovery action is available for the manual verdict.

Stop here for manual. Do not call eforge_apply_recovery; if it were called it would return { verdict: 'manual', noAction: true } and no mutation would occur.

Step 5: Apply the Recovery

On confirmation, call the appropriate tool:

• retry: Call eforge_apply_recovery with { prdId }. Report: "PRD {prdId} has been queued to retry from scratch."
• continue-repair: Call eforge_continue_repair with { prdId }, adding setName when the sidecar top-level setName differs from the PRD id and profile when the user requests a specific agent runtime profile. Report: "Queued continue-and-repair build for PRD {prdId}. It will wait for scheduler dispatch under the current queue controls."
• abandon: Call eforge_apply_recovery with { prdId }. Report: "PRD {prdId} has been archived."

A dispatched continue-and-repair build preserves normal queue controls: parallelism, pause state, dependency gating, and profile routing. It returns queued metadata such as the PRD id, set name, branches, moved descendants, and optional profile; it does not return a session id or PID.

When to Choose Continue and Repair vs Retry from Scratch

Error Handling

<!-- parity-skip-start -->

<!-- parity-skip-end -->

Related Skills