Fix Bug Pipeline

This skill triggers the Fix Validation pipeline to systematically investigate, fix, and validate a bug.

When to Use This Skill

Load this skill when the user request matches ANY of these patterns:

DO NOT use this skill for:

Why This Skill Exists

Without this skill, conversational prompts like "please investigate and fix this bug" may cause Claude to skip pipeline stages and fix directly. This skill ensures deterministic execution of all Fix Validation pipeline stages.

Usage

/fix-bug <path> [description]

Arguments:

• $1 (required): Path to code with the bug
• $2 and beyond (optional): Description of the issue - recommended for better analysis

Examples:

/fix-bug src/auth/login.ts "Users report login fails for new accounts"
/fix-bug tests/fixtures/fix-validator/simple-fix/ "Cannot read property displayName of undefined"
/fix-bug src/api/routes.ts

Pipeline Stages

When invoked, follow the Fix Validation pipeline exactly:

IssueAnalyzer (bulwark-issue-analyzer)          // Sonnet - root cause analysis
|> FixWriter (bulwark-implementer)              // Opus - implement fix
|> (if !tests_cover_scenario                    // Conditional: only if tests don't already exist
    then TestWriter |> TestAudit                // Opus writes, then audit for T1-T4
    else Skip)
|> FixValidator (bulwark-fix-validator)          // Sonnet - validate against debug report
|> CodeReviewer (general-purpose)               // Sonnet - review fix
|> (if !approved
    then IssueAnalyzer                          // Loop back
    else Done)
|> LOOP(max=3)                                  // Max 3 iterations

Execution Instructions

Stage 1: IssueAnalyzer

MUST spawn bulwark-issue-analyzer agent via Task tool:

Task(
  subagent_type="bulwark-issue-analyzer",
  model="sonnet",
  prompt="GOAL: Analyze the bug and produce a debug report..."
)

Input: Path from $1, description from $2 onward

Output: Debug report at logs/debug-reports/{issue-id}-{timestamp}.yaml

Do NOT skip this stage. The debug report is required for subsequent stages.

Stage 2: FixWriter

MUST spawn bulwark-implementer agent via Task tool:

Task(
  subagent_type="bulwark-implementer",
  prompt="GOAL: Fix the identified issue based on the debug report.
  CONSTRAINTS: Only fix the identified issue. Write tests for the fix. Max 3 quality gate retries.
  CONTEXT:
    mode: fix
    debug_report_path: logs/debug-reports/{issue-id}-{timestamp}.yaml
    root_cause: {from Stage 1}
    affected_files: {from Stage 1}
    fix_approach: {from Stage 1}
  OUTPUT: Implementation report at logs/implementer-{id}-{timestamp}.yaml"
)

Input: Debug report from Stage 1

Output: Implementation report at logs/implementer-{id}-{timestamp}.yaml

SA6 Note: The implementer returns pipeline suggestions with MANDATORY language in its summary. Evaluate each suggestion per SA6.

Do NOT implement the fix yourself. The implementer agent handles quality gates and structured output.

Stage 3: TestWriter (Conditional)

Condition: Check debug report's validation_plan.recommendation.new_tests_needed

If tests needed:

1. Write tests that verify the fix
2. Cover the specific bug scenario
3. Follow T1-T4 rules (no mocking system under test)

If tests exist: Skip to Stage 3b or Stage 4

Stage 3b: TestAudit (Conditional)

Condition: Run if any test files were created or modified in Stage 2 (FixWriter) OR Stage 3 (TestWriter). This ensures implementer-written tests receive T1-T4 audit even when TestWriter is skipped.

Action: Run mock-detection on new/modified tests to verify T1-T4 compliance

If T1 violation: Return to TestWriter (or FixWriter if TestWriter was skipped), request rewrite

If T2-T4 violations: Log warning, proceed

Stage 4: FixValidator

MUST spawn bulwark-fix-validator agent via Task tool:

Task(
  subagent_type="bulwark-fix-validator",
  model="sonnet",
  prompt="GOAL: Validate the fix against the debug report...

  CONTEXT:
  Debug Report: logs/debug-reports/{issue-id}-{timestamp}.yaml
  Fix Applied: {description of changes}
  Tests Added: {if any}
  ..."
)

Input: Debug report path, fix details, test details

Output: Validation report at logs/validations/fix-validation-{issue-id}-{timestamp}.yaml

Stage 5: CodeReviewer

MUST spawn general-purpose agent via Task tool:

Task(
  subagent_type="general-purpose",
  model="sonnet",
  prompt="GOAL: Review the fix for correctness, completeness, and safety.
  CONSTRAINTS: Do NOT modify any files. Review only.
  CONTEXT:
    debug_report: logs/debug-reports/{issue-id}-{timestamp}.yaml
    fix_applied: {description of changes from Stage 2}
    tests_added: {from Stage 3, if any}
    validation_results: {from Stage 4}
  OUTPUT: Approval decision (approved: true/false) with concerns and recommendations."
)

Approval Criteria:

• Fix addresses root cause from debug report
• Tests verify the specific bug scenario
• No new issues introduced
• Validation confidence is acceptable (high or medium with justification)

Loop Handling

If rejected and iterations < 3:

• Return to Stage 1 with feedback
• Include previous validation results

If rejected and iterations >= 3:

• Escalate to user
• Summarize all attempts

Progress Reporting

After each stage, report progress to user:

Stage 1 (IssueAnalyzer): Complete
  - Debug report: logs/debug-reports/AUTH-001-20260120.yaml
  - Root cause: {summary}
  - Complexity: {low|medium|high}

Stage 2 (FixWriter): Complete
  - Files modified: {list}
  - Fix: {brief description}

Stage 3 (TestWriter): {Complete|Skipped}
  - Tests added: {count or "existing tests sufficient"}

Stage 4 (FixValidator): Complete
  - Confidence: {HIGH|MEDIUM|LOW}
  - Recommendation: {proceed|revise}

Stage 5 (CodeReviewer): Complete
  - Decision: {Approved|Rejected}

Error Handling

Related Resources

Note: Paths are relative to your skills/agents directory (either skills/ or .claude/skills/).