fix

<user_request> $ARGUMENTS </user_request>

<value_banner> Display this ONCE at the start of the command:

/dev:fix — Production-Grade Bug Fix Beyond /dev:debug, this command adds:

• TDD state machine: RED → VERIFY RED → GREEN → REGRESSION CHECK (always, not opt-in)
• Dual multimodel consensus review (Phase A: root cause; Phase B: patch quality)
• 3-strategy fault localization with context budget enforcement
• Optional deployment monitoring (Sentry, CloudWatch, log files)
• Downgrade offer at VALIDATE: skip Phase B + monitoring if fix is clean

For quick debugging or one-off patches, use /dev:debug instead. </value_banner>

<critical_override> THIS COMMAND OVERRIDES THE CLAUDE.md TASK ROUTING TABLE FOR AGENT SELECTION.

WHY: The CLAUDE.md routing table maps "Debugging" to code-analysis:detective, but this command needs dev:debugger for root cause analysis and dev:developer for applying fixes. code-analysis:detective is READ-ONLY and cannot write code or tests.

AGENT RULES FOR THIS COMMAND:

• Stack detection → dev:stack-detector agent (subagent_type: "dev:stack-detector")
• Error analysis and root cause investigation → dev:debugger agent (subagent_type: "dev:debugger")
• Writing tests and applying patches → dev:developer agent (subagent_type: "dev:developer")
• Phase A vote, internal Claude (root-cause hypothesis) → dev:debugger with EVALUATE ONLY instruction, run_in_background: true
• Phase B vote, internal Claude (patch quality) → dev:reviewer (subagent_type: "dev:reviewer"), foreground, handed the TARGET / FOCUS / OUTPUT / MODELS contract lines. The reviewer judges the patch it is handed; the debugger wrote the diagnosis and does not grade its own work. It writes its own report in its own format; this command maps its **Verdict**: line to a vote (PASS → APPROVE, FAIL → REJECT, CONDITIONAL → ABSTAIN) and never hands it the vote schema
• Phase B tally when MODELS is not "none" → dev:synthesizer (subagent_type: "dev:synthesizer"), foreground, given the vote files and never the code. At MODELS "none" there is one mapped vote and it is the tally — no synthesizer
• Multimodel vote (external models) → claudish MCP tools (team/create_session), run_in_background: true
• Post-deploy monitoring → inline Bash
• Validation → inline Bash (run tests directly)

DO NOT use code-analysis:detective (READ-ONLY — cannot apply fixes or write tests). DO NOT use dev:researcher (researches topics, does NOT debug code). DO NOT use dev:architect (plans architecture, does NOT debug or fix code). </critical_override>

<critical_constraints> <todowrite_requirement> You MUST use Tasks to track all 8 phases plus the two review gates.

  Before starting, create this todo list:
  0. Initialize (session setup, flag parsing, skill loading)
  1. REPRODUCE (confirm bug, detect stack, capture error signature)
  2. LOCALIZE (3-strategy fault localization with context budget)
  3. PLAN (root cause analysis + Phase A multimodel review)
  4. PATCH (TDD: RED → VERIFY RED → GREEN, apply minimal fix)
  5. VALIDATE (full test suite + quality checks + downgrade offer)
  6. REVIEW-B (multimodel patch quality vote — after full validation)
  7. MONITOR (optional 3-tier deployment monitoring)
  8. DOCUMENT (fix-report.md + git commit)