Mistake Memory Guardrails

Mistake Memory Guardrails

Global guardrails for repository edit tasks.

Scope

Apply this skill to any repository edit task, including code, docs, tests, config, infra, and planning files.

Execution Order

1. Read AGENT_MISTAKES.md. If missing, create it using the required schema.
2. Read target files and map applicable scope_tags.
3. Run guardrail checks before proposing or applying changes.
4. Run a post-edit self-check against known mistake patterns before final response.
5. Record every detected mistake occurrence.

Guardrail Checks

1. Requirement fidelity and scope discipline:

• Change only what the user requested.
• Reject unrelated refactors and speculative additions.

2. Cross-file consistency for shared facts/configs/interfaces:

• Keep shared names, paths, status values, and contracts aligned across touched files.
• Flag stale duplicated facts and mismatched defaults.

3. Assertion quality:

• Replace vague claims with measurable criteria.
• Require explicit thresholds, formulas, or pass/fail conditions when asserting quality or correctness.

4. Reproducibility metadata when claims depend on run conditions:

• For benchmark/performance/behavior claims, capture model/tool version, dataset revision, seed, environment, and run conditions when relevant.
• Separate deterministic correctness checks from environment-sensitive comparisons.

5. Safety and regression checks:

• Verify edits do not silently break existing behavior.
• Call out potential regressions and missing validation coverage.

Mistake Recording Rules

• Record every detected mistake occurrence.
• Deduplicate by normalized pattern + scope_tags + prevention_rule.
  • Normalization: lowercase, trim whitespace, collapse repeated spaces.
• If a deduplicated entry exists:
  • Increment occurrence_count.
  • Update last_seen.
  • Append new evidence (file:line and/or commit).
• If no matching entry exists:
  • Append a new entry with all required fields from AGENT_MISTAKES.md.

Enforcement Behavior

• If a known rule is violated, warn and revise the proposal before final response.
• Fix violations before finalizing whenever possible.
• If unresolved, explicitly report the risk and why it could not be fully resolved.

Output Contract

When reporting results, provide:

1. What was checked.
2. What was changed and why.
3. Remaining risks or unresolved ambiguities.
4. Whether AGENT_MISTAKES.md was created or updated.