Fixing

Unified skill for fixing issues of any complexity with structured diagnosis.

Common Workflow (Authoritative)

Bug → Mode Select → [Check Memory: standard/deep] → Scout (MANDATORY) → Diagnose
  → [quick: confirm the known cause directly | standard/deep: investigate → sequential-thinking]
  → root cause? → yes → Root-Cause Proof → Complexity → Fix ROOT CAUSE → Verify + Prevent (MANDATORY)
  → pass → Report + capture only when recurrence/salience warrants it
  → fail <3 → re-diagnose | fail 3+ → STOP

This flow is authoritative. If prose conflicts, follow the flow.

Profiles

• quick — known cause and one or two files: minimal scout, compact root-cause proof, edit, focused before/after check, and summary. No default plan, report, wiki, memory, commit prompt, or subagent fan-out.
• standard (default) — full evidence chain, root-cause proof, focused regression test, relevant suite, and review.
• deep — standard plus an approved plan when a contract or architecture changes.

HARD GATE

Do NOT propose or implement fixes before completing Steps 1-2 (Scout + Diagnose). Symptom fixes are failure. Find the cause first through structured analysis, NEVER guessing. If 3+ fix attempts fail, STOP and question the architecture — discuss with user. --quick may use the compact proof, but it never bypasses Scout, Diagnose, root-cause evidence, Verify + Prevent, or the three-failed-attempt stop.

Arguments

• --auto — Standard profile with automatic progress through diagnosis and verification. It stops at any required human gate and never self-approves.
• --review — Human-in-the-loop. Pause at each step.
• --quick — Quick profile for a known cause and ≤2 files.
• --parallel — Parallel developer agents per independent issue.
• --tdd — Force regression test BEFORE the fix (writes the .claude/session-state/tdd-mode sentinel). Without --tdd, regression tests are recommended but not gated. Useful for security-sensitive fixes where you want to prove the bug first.
• --advice — Opt-in strategic supervision for one run. At named checkpoints (GUIDE → RESCUE* → REVIEW → RECHECK*, hard cap 5) the workflow asks the athena agent to assess the situation, recommend an operational path inside the locked scope, and — after the work — review it against the evidence, with the power to return it for correction. Standard and deep profiles only. Supervision is evidence, never approval: it clears no gate, counts as no verification, and the three-failed-attempt human STOP below still fires on its own schedule. Without the flag, zero calls happen and no state is written. Contract: .claude/rules-conditional/advice-supervision-rules.md (load it only when the flag is present).

Plan-First Gate

For moderate/complex bugs:

1. Run mk:investigate to confirm root cause. A supplied diagnostic report is evidence to validate and extend, never a substitute for investigation.
2. If fix affects > 2 files → request an approved plan with mk:plan-creator "bug fix: {symptom and affected area}"
3. Wait for Gate 1 approval

Skip: --quick mode (single file, clear cause).

Step 0 — Mode Selection

Without a mode flag, use the standard profile. Use quick only when its boundary is already proven; otherwise use standard. See references/mode-selection.md for escalation.

Step 0.5 — Check Fix Memory (standard/deep only)

Read .meowkit/memory/fixes.json — it is the canonical, schema-validated store of prior fix patterns. See the source-of-truth rule in .claude/rules/memory-read-rules.md.

• Search for similar symptoms, error messages, or affected modules
• If a matching fix pattern exists → use it as starting hypothesis in Step 2
• If a matching success pattern exists → apply the known fix approach directly

This turns repeated bugs into instant fixes. Skip for quick, or if .meowkit/memory/ does not exist.

Step 1 — Scout (MANDATORY — never skip)

Activate mk:scout to map affected codebase BEFORE any diagnosis:

• Affected files, dependencies, related tests, recent changes (git log)
• Quick mode: minimal scout (affected file + direct deps only)
• Standard/Deep: full scout (module boundaries, test coverage, call chains)
• Risk flags: match the task against .claude/rules/risk-checklist.md (the 9 IDs only — do not invent flags) and hold matchedFlags for the evidence index. If any of AUTH / AUTHZ / DATA_MODEL / AUDIT_SEC / EXT_SYSTEM / PUBLIC_CONTRACT / WEAK_PROOF matches, set risk.requiresHumanApproval = true — auto mode cannot finalize silently (see references/review-cycle.md).

Why mandatory: Without codebase context, diagnosis guesses instead of reasons from evidence.

Step 2 — Diagnose (MANDATORY — never skip)

Capture pre-fix state first: exact error messages, failing test output, stack traces.

Then choose the evidence path that matches the profile:

1. Quick — validate the supplied known cause against the error and affected code directly. Do not delegate to mk:investigate; if the cause is no longer clear, escalate to standard before editing.
2. Standard/Deep — use mk:investigate to collect and validate symptoms, traces, and reproduction steps, then mk:sequential-thinking to generate hypotheses from evidence, test each, eliminate, and conclude. A supplied diagnostic report is evidence to validate and extend, never a bypass.

Load references/diagnosis-protocol.md for the 5-phase protocol: Observe → Hypothesize → Test → Trace → Escalate.

Output: confirmed root cause (not symptom) with evidence chain + confidence level.

BLOCK: If confidence < medium → gather more evidence before fixing. Never fix a "maybe."

Step 2.5 — Root-Cause Proof Checkpoint (HARD GATE)

Operationalizes .claude/rules/core-behaviors.md Rule 6 ("Verify, Don't Assume"). The six fields are the named output of the Step 2 diagnosis (references/diagnosis-protocol.md Phase 4). Do NOT start Step 4 (Fix) until all six are populated — empty fields mean the diagnosis is not yet proven.

Standard/Complex/Parallel — all six required:

1. Exact symptom — copy-pasted error/message/behavior, not paraphrased.
2. Reproduction evidence — either exact command(s) or steps that trigger it every time, or, for an intermittent failure, the observed attempts, failure rate, conditions, and supporting trace/log correlation.
3. Expected vs actual — what should happen vs what does.
4. Root cause with file:line — the specific source location, traced backward from the symptom (never the symptom site).
5. Why now / uncertainty record — what changed or condition makes it surface (regression commit, data state, env, version); if unknown, state that explicitly with eliminated hypotheses and the monitoring or next observation needed. Never invent a cause for timing.
6. Blast radius — other callers, modules, or behaviors the same root cause touches.

--quick compact form (still non-empty — one phrase each): exact compiler/lint error · file · direct cause · command-before · command-after · impacted area.

If any field cannot be filled, return to Step 2 and gather more evidence. Intermittent failures may use bounded probabilistic evidence; a documented unknown is valid only for the uncertainty record, never as a substitute for the root cause.

Write evidence (init): standard/deep runs emit workflow-evidence.json with skill: mk:fix, mode, task, planPath (if the fix escalated to a plan), phase, risk, and fixDiagnosis. Quick runs keep evidence in the response only.

Step 3 — Complexity Assessment

Classify before routing. See references/complexity-assessment.md.

Level	Indicators	Workflow
Simple	Single file, clear error	references/workflow-quick.md
Moderate	Multi-file, root cause multi-step	references/workflow-standard.md
Complex	System-wide, architecture impact	references/workflow-deep.md
Parallel	2+ independent issues	Parallel agents per issue

Task orchestration (Moderate+): references/task-orchestration.md.

Step 4 — Fix Implementation

• Fix must address ROOT CAUSE from Step 2 — never symptoms only
• Minimal changes, follow existing patterns
• If fix deviates from diagnosis → re-diagnose first

Step 5 — Verify + Prevent (MANDATORY)

1. Iron-law verify: Re-run exact pre-fix commands. Compare before/after.
2. Regression test: Follow .claude/rules/tdd-rules.md; when required, the test fails WITHOUT the fix and passes WITH it.
3. Defense-in-depth: Load references/prevention-gate.md — consider entry validation, business logic guards, error handling, type safety.
4. BLOCK: Missing a required regression test is incomplete; record the allowed omission rationale for lint/format/config-only changes.

If verify fails: loop to Step 2. After 3 failures → STOP, question architecture.

Update evidence: standard/deep runs write verification.commands and verification.overall to workflow-evidence.json; quick runs report the same focused before/after check inline.

Step 6 — Finalize + Learn (MANDATORY for Standard/Complex/Parallel; opt-in for Simple)

1. Report: confidence, root cause, changes, files, prevention measures.

2. Write to memory only for recurrence or durable salience — read .meowkit/memory/fixes.json first, then add/update the canonical JSON store only when the pattern is likely to help a future run. Quick runs never write memory.

   • .meowkit/memory/fixes.json — under patterns, add or update:

     {
      "id": "<kebab-slug>",
      "type": "failure",
      "category": "bug-class",
      "severity": "low|medium|high|critical",
      "domain": ["<area1>", "<area2>"],
      "applicable_when": "<one line>",
      "context": "<one line>",
      "pattern": "<one line — what to do or avoid>",
      "frequency": 1,
      "lastSeen": "<YYYY-MM-DD>"
     }
     

   • If the same id already exists, increment frequency and update lastSeen. Do not duplicate entries.

   • DO NOT use ##pattern:bug-class prefixes. That is a user-typed keyboard shortcut; the handler (hooks/handlers/immediate-capture-handler.cjs) only fires on UserPromptSubmit — the human typing the prefix at the start of a message. Agent-emitted ##pattern: text is invisible to the handler. Always call Edit directly. See .claude/skills/memory/references/capture-architecture.md.

   • Scrub secrets / tokens / PII before writing. Edit is not secret-scrubbed; you are responsible.

   • Inside /mk:cook full pipeline: Phase 6 / mk:memory session-capture covers this — do NOT double-write here. Standalone /mk:fix runs OWN the write themselves.

   • Skip when /mk:fix --no-capture was passed, the run is quick, or the result is a one-off with no durable lesson.

3. Delegate to project-manager (Moderate/Complex/Parallel ONLY) per .claude/rules/post-phase-delegation.md Rule 1 (background — include "Run in the background" in the prompt). Skip for Simple complexity — Gate 1 bypass path means no plan to track. Also skipped when MEOWKIT_PM_AUTO=off.

4. documenter agent → update ./docs.

5. Ask user about commit.

6. Terminal wiki handoff (standalone runs only; advisory, fail-open): Inside /mk:cook, SKIP — cook's Phase 6 owns the handoff. For a standalone /mk:fix of Standard/Complex/Parallel complexity, first write a short, durable fix summary (NEVER raw workflow-evidence.json) to tasks/reports/fix-<YYMMDD-HHMM>-<slug>.md with only: symptom, root cause, files changed, verification result, recurrence/friction note, and links to the evidence path. Then hand it to the wiki per .claude/skills/wiki/references/terminal-handoff-advisory.md — resolve the slug (env MEOWKIT_WIKI_SLUG → the sole tasks/wikis/<slug>/wiki.json → else skip + print), fail-open, never wiki approve, no wiki reindex:

   npx mewkit wiki handoff propose \
     --skill mk:fix \
     --from tasks/reports/fix-<YYMMDD-HHMM>-<slug>.md \
     --slug <resolved-wiki-slug> \
     --verified-outcome --recurring-friction
   

   Omit --recurring-friction for a one-off bug. Skip entirely for Simple complexity or when --no-capture was passed.

Workflow Evidence Index

Contract: .claude/rules-conditional/workflow-evidence-rules.md. The index records pointers + summaries of this run; it never approves anything (Gate 2 / ship stay human authority) and carries no score. Generated for standalone Standard/Complex/Parallel fixes; quick and simple fixes keep focused evidence in their response.

Storage path: .claude/session-state/evidence/<YYMMDD-HHMM-slug>/workflow-evidence.json (framework-internal state per skill-authoring-rules.md Rule 2). For a fix that escalated to a plan, use tasks/plans/<plan>/reports/evidence/workflow-evidence.json instead.

Write points (standard/deep only): Step 2.5 (init: skill, mode, task, planPath, phase, risk, fixDiagnosis) → Step 5 (verification) → Step 6 finalize (approvals.gate2/ship as required|not_applicable, memory.fixPatternWritten).

Validate before approval: run node .claude/scripts/validate-workflow-evidence.cjs <path> --phase fix before the user-approval prompt (Step 6 item 5). Surface any EVIDENCE_BLOCKED:<reasons> and fill the missing fields — do not present for approval on a blocked index. A high-risk flag (risk.requiresHumanApproval) forces explicit human approval before finalize regardless of mode.

Evidence ≠ memory: the evidence file is one-run proof; .meowkit/memory/fixes.json is the durable pattern store. Keep them separate — standalone /mk:fix owns its evidence write; inside /mk:cook the pipeline (Phase 6) owns the evidence write, so do NOT double-write. Scrub secrets / tokens / PII and store pointers/summaries only — never raw command logs.

Skill Activation

Always: mk:scout (Step 1) + direct diagnosis (Step 2). Standard/Deep: mk:investigate + mk:sequential-thinking for the full evidence chain. A supplied diagnostic report informs investigation; it never bypasses it. Conditional: mk:brainstorming (complex, multiple approaches) | mk:docs-finder (unfamiliar APIs)

Gotchas

• Guessing root causes: "I think it's X" without evidence → use mk:sequential-thinking to generate + test hypotheses from evidence
• Fixing symptoms: test passes but underlying issue remains → always trace backward: symptom → cause → ROOT cause
• Skipping scout: fixing without codebase context → mandatory scout maps what you're touching
• Missing required regression test: bug resurfaces next sprint → follow tdd-rules.md and record any allowed omission rationale
• 3+ failed attempts without stopping: insanity loop → STOP, question architecture, discuss with user

Full list: references/gotchas.md (update when Claude produces wrong fix patterns)

References

• references/mode-selection.md — AskUserQuestion format
• references/diagnosis-protocol.md — 5-phase structured diagnosis
• references/prevention-gate.md — defense-in-depth + verification checklist
• references/complexity-assessment.md — Simple/Moderate/Complex/Parallel
• references/task-orchestration.md — Task patterns for moderate+
• references/workflow-quick.md | workflow-standard.md | workflow-deep.md
• references/review-cycle.md — Autonomous/HITL/Quick review
• references/skill-activation-matrix.md — When to activate each skill
• references/workflow-ci.md | workflow-logs.md | workflow-test.md | workflow-types.md | workflow-ui.md
• references/parallel-exploration.md — Parallel investigation strategy for multi-hypothesis diagnosis