Workflow self anneal

Workflow Self-Anneal

Purpose

Perform a controlled, auditable improvement pass on instructions and process docs to reduce repeated failures while preserving human control.

This skill optimizes process reliability, not model internals.

Hard safety constraints

1. No unbounded loops or autonomous "improve forever" behavior.
2. Max one anneal pass per invocation; max three proposed doc changes.
3. No deploy/publish success claims without verification evidence.
4. Do not modify security/permission policy automatically; propose only.
5. If evidence insufficient, return NO_CHANGE with missing-evidence list.

Trigger conditions

• Same failure pattern ≥2 times in 7 days.
• Session crash/restart/tooling instability.
• Instructions contradicted by observed actions.
• Pre-merge gate requires hardening evidence.
• Repeated carryover/blocked items across iterations.
• PR/merge latency breaches targets or reopened issues trend up.
• Commits exist without linked GitHub issues (process drift).
• Consultant check registry returns ≥1 FAIL on domain:governance.
• Fleet check fleet-002 (cost-budget) FAIL for ≥2 consecutive cycles.

Consultant Integration

Self-anneal may be invoked automatically from consultant-feedback.js when gov-002 (baton-artifact-presence) or gov-003 (event-emission) FAIL — these are workflow process failures, not implementation bugs. Check IDs are managed in scripts/global/consultant-checks.js (3 domains: governance/tools/fleet).

Input collection

Collect: execution artifacts (logs, errors, checks), instructions/workflow docs, recent commits/PR notes. If unavailable, declare missing and stop.

Analysis protocol

1. Detect mismatch: expected vs observed behavior.
2. Classify root cause: ambiguity|missing guardrail|stale instruction|tool fragility|human override.
3. Assess recurrence risk: low|medium|high by frequency and blast radius.
4. Detect Agile drift signals: carryover, blocked age, review/merge latency, reopen rate.
   • Ticket linkage: git log --oneline | wc -l vs --grep='#'. If <80%, flag ticket-linkage-drift.
   • PR coverage: commits vs gh pr list --state all. If 0 PRs, flag pr-coverage-drift.
   • Event gaps: count events in .dashboard/events.jsonl per role transition. If <1 event per baton handoff, flag event-emission-drift.
5. Propose minimal fix: smallest docs/workflow delta preventing recurrence.
6. Define verification gate: objective checks confirming the fix.

Output format (required)

Return exactly this structure:

SELF_ANNEAL_REPORT
context: <session-start|post-failure|pre-merge|post-release>
scope: <workflow|stability|qa|git>
observation: [...]
expected_behavior: [...]
mismatch: [...]
root_cause: <ambiguity|missing guardrail|stale instruction|tool fragility|human override>
risk: <low|medium|high>
drift_signals:
- metric: <carryover|blocked-age|review-latency|merge-latency|reopen-rate|ticket-linkage|pr-coverage>
   trend: <improving|stable|degrading>  evidence: <artifact>
proposed_changes:
1) file: <path>  change_type: <add|edit|remove>  rationale: <why>
verification_plan:
- check: <objective>  pass_condition: <measurable>
decision: <apply|defer|NO_CHANGE>
missing_evidence: <none or list>

Change targeting guidance

Prioritize: project runbooks/checklists, workflow docs, system-stability docs, learnings docs.

Stop conditions

Return NO_CHANGE if: change cannot be validated objectively, evidence is missing, change expands permissions, or same fix was already applied.

For iterative anneal loops (refinement passes), call shouldStop(...) from scripts/global/anneal-stop.js after every iteration (Epic #1568 AC-5, #1574). Stop when: deterministic gates green, OR iterations >= 3, OR |rubric_mean_delta| <= 0.5. Continuation past a stop requires ANNEAL_OVERRIDE_CONTINUE: rationale comment from the operator and is logged as event:kill-switch-bypass.

Quality bar

Good anneal output is: specific (concrete artifacts), minimal (smallest viable delta), testable (pass/fail checks), traceable (rationale + risk), trend-aware (drift reducing).

Visual verification rule (publish/UX scope)

For UX-rendered changes: require post-change visual inspection artifact. Do not allow ticket closure from DOM/CI checks alone. If visual evidence is missing, decision must be defer.