Claude Code Plugins

Community-maintained marketplace

Feedback

Intelligent bug fixing workflow combining root cause analysis, multi-model reasoning, Codex auto-fix, and comprehensive testing. Uses RCA agent, Codex iteration, and validation to systematically fix bugs.

Install Skill

1Download skill
2Enable skills in Claude

Open claude.ai/settings/capabilities and find the "Skills" section

3Upload to Claude

Click "Upload skill" and select the downloaded ZIP file

Note: Please verify skill by going through its instructions before using it.

SKILL.md

name smart-bug-fix
description Intelligent bug fixing workflow combining root cause analysis, multi-model reasoning, Codex auto-fix, and comprehensive testing. Uses RCA agent, Codex iteration, and validation to systematically fix bugs.
tags debugging, rca, codex, testing, essential, tier-1
version 1.0.0

Smart Bug Fix

Purpose

Systematically debug and fix bugs using root cause analysis, multi-model reasoning, and automated testing.

Specialist Agent

I am a debugging specialist using systematic problem-solving methodology.

Methodology (Root Cause + Fix + Validate Pattern):

  1. Deep root cause analysis (5 Whys, inverse reasoning)
  2. Multi-model reasoning for fix approaches
  3. Codex auto-fix in isolated sandbox
  4. Comprehensive testing with iteration
  5. Regression validation
  6. Performance impact analysis

Models Used:

  • Claude (RCA): Deep root cause analysis
  • Codex (Fix): Rapid fix implementation
  • Claude (Validation): Comprehensive testing
  • Gemini (Context): Large codebase analysis if needed

Output: Fixed code with test validation and impact analysis

Input Contract

input:
  bug_description: string (required)
  context_path: string (directory or file, required)
  reproduction_steps: string (optional)
  error_logs: string (optional)
  depth: enum[shallow, normal, deep] (default: deep)

Output Contract

output:
  root_cause: object
    identified: string
    contributing_factors: array[string]
    evidence: array[string]
  fix_applied: object
    changes: array[file_change]
    reasoning: string
    alternatives_considered: array[string]
  validation: object
    tests_passed: boolean
    regression_check: boolean
    performance_impact: string
  confidence: number (0-1)

Execution Flow

#!/bin/bash
set -e

BUG_DESC="$1"
CONTEXT_PATH="$2"

echo "=== Smart Bug Fix Workflow ==="

# PHASE 1: Root Cause Analysis
echo "[1/6] Performing deep root cause analysis..."
npx claude-flow agent-rca "$BUG_DESC" \
  --context "$CONTEXT_PATH" \
  --depth deep \
  --output rca-report.md

# PHASE 2: Context Analysis (if large codebase)
LOC=$(find "$CONTEXT_PATH" -name "*.js" -o -name "*.ts" | xargs wc -l | tail -1 | awk '{print $1}')
if [ "$LOC" -gt 10000 ]; then
  echo "[2/6] Large codebase detected - analyzing with Gemini MegaContext..."
  gemini "Analyze patterns related to: $BUG_DESC" \
    --files "$CONTEXT_PATH" \
    --model gemini-2.0-flash \
    --output context-analysis.md
else
  echo "[2/6] Standard codebase - skipping mega-context analysis"
fi

# PHASE 3: Alternative Solutions (multi-model reasoning)
echo "[3/6] Generating fix approaches..."
# Claude approach (from RCA)
CLAUDE_FIX=$(cat rca-report.md | grep "Solution" -A 10)

# Codex alternative approach
codex --reasoning-mode "Alternative approaches to fix: $BUG_DESC" \
  --context rca-report.md \
  --output codex-alternatives.md

# PHASE 4: Implement Fix with Codex Auto
echo "[4/6] Implementing fix with Codex Auto..."
codex --full-auto "Fix bug: $BUG_DESC based on RCA findings" \
  --context rca-report.md \
  --context "$CONTEXT_PATH" \
  --sandbox true \
  --network-disabled \
  --output fix-implementation/

# PHASE 5: Comprehensive Testing with Iteration
echo "[5/6] Testing fix with Codex iteration..."
npx claude-flow functionality-audit fix-implementation/ \
  --model codex-auto \
  --max-iterations 5 \
  --sandbox true \
  --regression-check true \
  --output test-results.json

# Check if tests passed
TESTS_PASSED=$(cat test-results.json | jq '.all_passed')
if [ "$TESTS_PASSED" != "true" ]; then
  echo "⚠️ Tests failed after 5 iterations - escalating to user"
  exit 1
fi

# PHASE 6: Performance Impact Analysis
echo "[6/6] Analyzing performance impact..."
npx claude-flow analysis performance-report \
  --compare-before-after \
  --export performance-impact.json

# Display summary
echo ""
echo "================================================================"
echo "Bug Fix Complete!"
echo "================================================================"
echo ""
echo "Root Cause: $(cat rca-report.md | grep 'Primary Root Cause' -A 2 | tail -1)"
echo "Tests: ✓ All passing"
echo "Regression: ✓ No regressions detected"
echo "Performance Impact: $(cat performance-impact.json | jq '.impact_summary')"
echo ""
echo "Files changed:"
find fix-implementation/ -name "*.js" -o -name "*.ts" | head -10
echo ""

Integration Points

Cascades

  • Part of /bug-triage-workflow cascade
  • Used by /production-incident-response cascade
  • Invoked by /fix-bug command

Commands

  • Uses: /agent-rca, /gemini-megacontext, /codex-auto, /functionality-audit
  • Chains with: /style-audit, /performance-report

Other Skills

  • Input to regression-validator skill
  • Used by incident-response skill
  • Integrates with code-review-assistant

Advanced Features

Automatic RCA Depth Selection

function selectRCADepth(bugDescription, errorLogs) {
  if (errorLogs.includes("intermittent") || errorLogs.includes("race condition")) {
    return "deep"; // Complex issues need deep analysis
  } else if (errorLogs.includes("TypeError") || errorLogs.includes("undefined")) {
    return "normal"; // Common errors need normal analysis
  } else {
    return "shallow"; // Simple issues
  }
}

Multi-Model Fix Approach

fix_strategy:
  1. Claude RCA → Deep understanding
  2. Codex alternatives → Multiple approaches
  3. Codex auto-fix → Rapid implementation
  4. Claude validation → Comprehensive testing

Codex Iteration Loop

Test → FAIL → Codex fix → Test → FAIL → Codex fix → Test → PASS → Apply
↑                                                                    ↓
└────────────────── Max 5 iterations ──────────────────────────────┘

Usage Example

# Fix bug with description
smart-bug-fix "API timeout under load" src/api/

# Fix with reproduction steps
smart-bug-fix "Login fails on Firefox" src/auth/ \
  --reproduction-steps "1. Open Firefox 2. Try login 3. See error"

# Fix with error logs
smart-bug-fix "Database connection fails" src/db/ \
  --error-logs "logs/error.log"

Failure Modes

  • RCA inconclusive: Request more context, run additional diagnostics
  • Codex fix fails tests: Try alternative approach, escalate if max iterations reached
  • Regression detected: Rollback fix, analyze conflicting requirements
  • Performance degradation: Optimize fix, consider alternative approach