Parallel Workers

Skill para execução paralela de tarefas de implementação usando git worktrees isolados.

Overview

Esta skill permite que múltiplas tarefas de código sejam desenvolvidas simultaneamente por workers independentes, cada um operando em seu próprio git worktree. Inspirada no claude-orchestrator, foi adaptada para ser multiplataforma e integrar com o ecossistema do SDLC Agêntico.

Architecture

┌─────────────────────────────────────────────────────────┐
│                    Main Repository                      │
│                  (feature/epic-33-*)                    │
└────────────┬────────────────────────────────────────────┘
             │
             ├─→ Worker 1: ~/.worktrees/project/task-001/
             │   Branch: feature/task-001
             │   State: WORKING
             │   Agent: code-author
             │
             ├─→ Worker 2: ~/.worktrees/project/task-002/
             │   Branch: feature/task-002
             │   State: PR_OPEN
             │   Agent: test-author
             │
             └─→ Worker 3: ~/.worktrees/project/task-003/
                 Branch: feature/task-003
                 State: NEEDS_INIT
                 Agent: iac-engineer

Worker State Machine

UNKNOWN ──→ NEEDS_INIT ──→ WORKING ──→ PR_OPEN ──→ MERGED
   ↑            ↑             ↓            ↓
   └────────────┴─────────────┴────────────┘
              (error recovery)

States:

• UNKNOWN: Initial or error state
• NEEDS_INIT: Worker created, needs initialization
• WORKING: Task in progress
• PR_OPEN: Pull request created, awaiting review
• MERGED: Task completed and merged

Usage

Spawning Workers

# Spawn a single worker
python3 .claude/skills/parallel-workers/scripts/worker_manager.py spawn \
  --task-id "TASK-001" \
  --description "Implement user authentication" \
  --agent "code-author" \
  --base-branch "feature/epic-33-claude-orchestrator"

# Spawn multiple workers from spec
python3 .claude/skills/parallel-workers/scripts/worker_manager.py spawn-batch \
  --spec-file .agentic_sdlc/projects/current/tasks.yml

Managing Worktrees

# Create worktree
.claude/skills/parallel-workers/scripts/worktree_manager.sh create \
  project-name task-001 feature/epic-33

# List active worktrees
.claude/skills/parallel-workers/scripts/worktree_manager.sh list project-name

# Remove worktree
.claude/skills/parallel-workers/scripts/worktree_manager.sh remove \
  project-name task-001

Monitoring State

# Check worker state
python3 .claude/skills/parallel-workers/scripts/state_tracker.py get worker-001

# List all workers
python3 .claude/skills/parallel-workers/scripts/state_tracker.py list

# Update state
python3 .claude/skills/parallel-workers/scripts/state_tracker.py set \
  worker-001 WORKING

Security Considerations

Secrets Isolation

Workers operate with sanitized environments:

# Sanitized variables (kept)
- PATH
- HOME
- USER
- LANG

# Removed from worker env
- *_KEY
- *_SECRET
- *_TOKEN
- *_PASSWORD

Validation Before Merge

All workers must pass quality gates before merge:

# From security-gate.yml
- no_hardcoded_secrets
- input_validation_present
- sast_scan_passed

Audit Trail

All worker operations logged to Loki:

{
  "skill": "parallel-workers",
  "phase": 5,
  "worker_id": "worker-001",
  "task_id": "TASK-001",
  "state": "WORKING",
  "timestamp": "2026-01-21T10:30:00Z",
  "correlation_id": "abc-123-def"
}

Integration with SDLC Phases

Phase 4 → 5 Transition

# delivery-planner generates tasks
tasks = [
  {"id": "TASK-001", "agent": "code-author", "desc": "..."},
  {"id": "TASK-002", "agent": "test-author", "desc": "..."},
]

# Spawn workers in parallel
for task in tasks:
    spawn_worker(task)

Phase 5 → 6 Transition

# Wait for all workers to complete
all_merged = wait_for_workers(timeout=3600)

if all_merged:
    # Run phase-5-to-6 gate
    gate_result = evaluate_gate("phase-5-to-6.yml")
    if gate_result.passed:
        transition_to_phase(6)

Observability

Loki Queries

# All parallel-workers activity
{skill="parallel-workers"}

# Errors by worker
{skill="parallel-workers", level="error"} | json | line_format "{{.worker_id}}: {{.message}}"

# Task completion timeline
{skill="parallel-workers", state="MERGED"} | json | line_format "{{.task_id}} completed at {{.timestamp}}"

Grafana Dashboard

New panels added to .claude/config/logging/dashboards/sdlc-overview.json:

• Active Workers (gauge)
• Worker State Distribution (pie chart)
• Task Completion Rate (timeseries)
• Worker Errors (logs panel)

Comparison with Sequential Execution

| Metric | Sequential | Parallel (3 workers) | Improvement | |--------|-----------|----------------------|-------------| | Total time (3 tasks) | 90 min | ~35 min | 61% faster | | Resource usage | 1x CPU | 3x CPU | Scalable | | Context switching | None | Minimal | Isolated | | Merge conflicts | N/A | Zero (worktrees) | Safe |

Limitations

• Max workers: Recommended 3-5 (CPU-bound)
• Disk usage: Each worktree ~= repo size
• Network: Parallel git operations may hit rate limits
• Complexity: Level 2+ only (Level 0/1 remain sequential)

Future Enhancements

• [ ] Distributed workers across machines
• [ ] Dynamic task rebalancing
• [ ] Real-time worker chat/coordination
• [ ] GPU-accelerated task scheduling

References

• Source: claude-orchestrator
• Analysis: .agentic_sdlc/corpus/nodes/learnings/LEARN-claude-orchestrator-patterns.yml
• Epic: Issue #33
• Task: Issue #35