Quality Gates Parallel (Native Multi-Agent)

Orchestrator integration for launching 4 quality subagents in parallel using Claude Code 2.1+ native Task tool with teammate coordination.

Quick Start

# Launch quality checks after implementation
/quality-gates-parallel src/auth.ts --complexity 7

# Read aggregated results post-analysis
/quality-gates-parallel --read-results <run_id>

Native Multi-Agent Architecture (Claude Code 2.1.16+)

Based on: claude-sneakpeek native-multiagent-gates

Features Available

• TaskCreate: Create tasks for subagents
• TaskUpdate: Update task status
• TaskList: List all tasks
• TaskGet: Get task details
• Parallel execution: Multiple agents work independently
• Result aggregation: Collect findings from all agents

Workflow

Phase 1: Launch Parallel Quality Checks

After implementation (orchestrator step 6b), launch 4 subagents:

// Pseudo-code for orchestrator integration
1. Classify task complexity (1-10)
2. If complexity >= 5:
3.    Create 4 tasks using TaskCreate:
4.      - Security auditor (sec-context-depth)
5.      - Code reviewer (code-reviewer)
6.      - Code cleanup (deslop)
7.      - Prose cleanup (stop-slop)
8.    Tasks execute in parallel (non-blocking)
9.    Continue orchestrator workflow
10. Else: Skip quality checks (low complexity)

Phase 2: Read Results (Pre-Validation)

Before validation (orchestrator step 7), poll for results:

# Run results reader
.claude/scripts/read-quality-results.sh <run_id>

Phase 3: Orchestrator Decision-Making

Orchestrator reads aggregated results and decides:

// Pseudo-code for decision logic
results = readQualityResults(run_id)

if (results.total_findings == 0) {
    // No issues - proceed to validation
    proceedToValidation()
} else if (results.critical_findings > 0) {
    // Critical issues - block and fix
    blockMerge()
    requireFixes()
} else {
    // Minor issues - advisory only
    proceedWithWarnings()
}

Quality Agents (4 Parallel)

1. Security Auditor (sec-context-depth)

Agent: security-auditor or glm-reviewer Purpose: 27 security anti-patterns (OWASP/CWE) Findings: P0 (Critical), P1 (High), P2 (Medium) Command: /sec-context-depth <file>

Coverage:

• 86% XSS failure rate detection
• 72% Java AI code vulnerability detection
• SQL injection, command injection, XSS
• JWT none algorithm, weak hashing, ECB mode

2. Code Reviewer (code-reviewer)

Agent: code-reviewer or codex-cli Purpose: Official Claude Code plugin with 4 parallel agents Features: Confidence scoring (≥80 threshold) Command: /code-review <file>

Architecture:

• Agent #1: CLAUDE.md compliance
• Agent #2: CLAUDE.md compliance (redundancy)
• Agent #3: Bug detection (changes only)
• Agent #4: Git blame/history analysis

3. Code Cleanup (deslop)

Agent: refactorer or gemini-cli Purpose: Remove AI-generated code slop Command: /deslop

Removes:

• Extra comments inconsistent with codebase
• Extra defensive checks/try/catch blocks
• Casts to any for type issues
• Inline imports (move to top)

4. Prose Cleanup (stop-slop)

Agent: docs-writer or minimax Purpose: Remove AI writing patterns from prose Command: /stop-slop <file>

Removes:

• Filler phrases ("Certainly!", "It is important to note")
• Structural clichés (binary contrasts, dramatic fragmentation)
• Stylistic habits (tripling, metronomic endings)

Integration with Orchestrator

Step 6b.5: Quality Parallel (NEW)

Location: After implementation (6b), before validation (7)

Trigger: complexity >= 5 OR security-related code

Execution:

# Non-blocking parallel launch
.claude/scripts/quality-coordinator.sh <target_file> <complexity>

Output: JSON with 4 task definitions

Step 7: Validation with Quality Results

Before validation: Read aggregated results

# Poll for completed checks
.claude/scripts/read-quality-results.sh <run_id>

Output: Aggregated JSON with all findings

Decision Logic:

• 0 findings: Proceed to validation
• Critical findings: Block and require fixes
• Minor findings: Advisory warnings

Results Storage

.claude/quality-results/
├── aggregated_<run_id>.json        # Aggregated results
├── sec-context_<run_id>.json       # Security findings
├── code-review_<run_id>.json       # Code review findings
├── deslop_<run_id>.json            # Code cleanup findings
├── stop-slop_<run_id>.json         # Prose cleanup findings
├── *_<run_id>.done                 # Completion markers
└── coordinator.log                # Execution log

Usage Examples

Manual Execution

# Launch quality checks
./.claude/scripts/quality-coordinator.sh src/auth.ts 7

# Read results
./.claude/scripts/read-quality-results.sh 20250128_221437_12345

Orchestrator Integration

# In orchestrator step 6b (after implementation)
quality_check_result=$(./.claude/scripts/quality-coordinator.sh "$file" "$complexity")

# Parse result and create tasks
if [[ "$complexity" -ge 5 ]]; then
    # Create 4 tasks using TaskCreate
    # Tasks execute in parallel
    # Store run_id for later retrieval
fi

# In orchestrator step 7 (before validation)
quality_results=$(./.claude/scripts/read-quality-results.sh "$run_id")

# Parse results and make decision
critical_count=$(echo "$quality_results" | jq '.summary.critical_findings // 0')
if [[ "$critical_count" -gt 0 ]]; then
    # Block and require fixes
    echo "CRITICAL: $critical_count security issues found"
else
    # Proceed to validation
    echo "No critical issues, proceeding to validation"
fi

Scripts

• quality-coordinator.sh: Launch 4 quality tasks in parallel
• read-quality-results.sh: Poll and aggregate results

Hooks

• quality-parallel-async.sh: Async hook for Edit/Write operations
  • Uses async: true in settings.json
  • Non-blocking background execution
  • Results stored for later reading

Version History

• 1.0.0 (2026-01-28): Initial native multi-agent integration
  • Based on Claude Code 2.1.16+ Task tool
  • 4 parallel quality agents
  • Result aggregation and polling

References

• Native Multi-Agent Gates: claude-sneakpeek documentation
• Claude Code 2.1.16+ features: Swarms, TeammateTool, teammate coordination
• Quality consolidation: docs/analysis/QUALITY_PARALLEL_CONSOLIDATION_v2.80.3.md
• Async hooks correction: docs/analysis/ASYNC_HOOKS_CORRECTION_v2.80.2.md