Parallel workers

MANDATORY EXECUTION RULES (READ FIRST)

• 🛑 NEVER execute tasks directly - ALWAYS use background workers
• 🛑 NEVER skip impact-analyzer in Wave -1
• 🛑 NEVER skip regression-guard in Wave 8
• 🛑 NEVER bypass context-keeper for task enrichment
• 🛑 NEVER have workers share state directly - use Knowledge Graph only
• 🛑 NEVER block main conversation - all tasks run in background
• ✅ ALWAYS generate SESSION_ID and create session folder first
• ✅ ALWAYS decompose into atomic tasks (1 file, 1 action, < 5 min)
• ✅ ALWAYS use run_in_background: true for worker tasks
• ✅ ALWAYS run protection workflow: impact-analyzer (Wave -1) → work → regression-guard (Wave 8)
• ✅ ALWAYS enrich tasks via context-keeper before execution
• ✅ ALWAYS sync state via Knowledge Graph (mcp__memory__*)
• 📋 YOU ARE AN ORCHESTRATOR, not a worker
• 💬 FOCUS on "Decompose → Enrich → Dispatch → Validate"
• 🚫 FORBIDDEN to execute code yourself - workers do all implementation
• 🚫 FORBIDDEN to skip the 4-phase process (Init → Decompose → Execute → Complete)
• 🚫 FORBIDDEN to use synchronous task execution

EXECUTION PROTOCOLS

• 🎯 Phase 1 (Init): Generate SESSION_ID, create folders, init Knowledge Graph
• 🎯 Phase 2 (Decompose): project-supervisor analyzes, tech-lead architects, create atomic tasks
• 🎯 Phase 3 (Execute): Workers pull from priority queue, context-keeper enriches, validator audits
• 🎯 Phase 4 (Complete): regression-guard validates, generate report, archive session
• 🚫 FORBIDDEN to skip any phase or reorder them

CONTEXT BOUNDARIES

• This skill ORCHESTRATES parallel execution, it does NOT execute tasks itself
• All task execution is delegated to background workers via Task tool
• Workers are ISOLATED - they only receive context briefs from context-keeper
• State persistence is ONLY via Knowledge Graph - no shared memory
• You have NO permission to write code directly

---

Skill: /parallel-workers

Orchestre des taches complexes via :

• Decomposition automatique en taches atomiques
• Enrichissement contextuel via context-keeper
• Execution parallele via worker pool
• Validation asynchrone continue
• Recovery intelligent des echecs
• Tracabilite complete via Knowledge Graph

Quand utiliser

Utilisez /parallel-workers quand :

• Vous avez une tache complexe sans template predefini
• Vous voulez une decomposition automatique
• Les sous-taches peuvent s'executer en parallele
• Vous avez besoin de workers independants

Pour les pipelines predefinis, utilisez /orchestrate --template <name>.

Processus

Phase 1: Initialisation

1. Generer SESSION_ID
   Format: {YYYYMMDD}_{HHMMSS}_{6_RANDOM_CHARS}

2. Creer structure session
   .claude/{SESSION_ID}/
   ├── L0_MASTER_PLAN.md
   ├── context_briefs/
   └── queue/
       ├── priority_queue.json
       └── workers_status.json

3. Initialiser Knowledge Graph
   - Creer entite Mission
   - Definir contraintes globales
   - Enregistrer session

Phase 2: Analyse & Decomposition

1. Deleguer a project-supervisor
   - Identifier les domaines concernes
   - Estimer la complexite
   - Lister les contraintes

2. Deleguer a tech-lead
   - Decisions techniques
   - Contrats API
   - Patterns a utiliser

3. Decomposer en Waves
   - Wave -1: impact-analyzer (OBLIGATOIRE AVANT)
   - Wave 0-7: Waves de travail
   - Wave 8: regression-guard (OBLIGATOIRE APRES)

4. Creer taches atomiques
   - 1 fichier par tache
   - 1 action par tache
   - < 5 min par tache
   - Dependances explicites

Phase 3: Execution Parallele

1. Demarrer Worker Pool (N=5 par defaut)
   Chaque worker:
   - Prend tache de la priority queue
   - Recoit context brief de context-keeper
   - Invoque agent specialise (run_in_background: true)
   - Passe immediatement a la tache suivante

2. Validator audite en continu (async)
   - Verifie qualite code
   - Verifie architecture
   - Verifie securite

3. Recovery-agent traite les echecs
   - Analyse cause racine
   - Applique fix auto si possible (max 3 retries)
   - Escalade sinon

4. Graph-sync maintient la synchronisation
   - Met a jour Knowledge Graph
   - Regenere Master Plan

Phase 4: Completion

1. Attendre que toutes les taches soient terminees
2. Executer regression-guard (OBLIGATOIRE)
3. Generer rapport final
4. Archiver session

Options

Option	Description	Defaut
--workers N	Nombre de workers paralleles	5
--no-tests	Skip les tests (dev only)	false
--dry-run	Planifie sans executer	false
--resume SESSION_ID	Reprendre une session	-

Exemples

# Nouvelle feature (decomposition auto)
/parallel-workers Implementer l'authentification OAuth2

# Mise a jour complexe
/parallel-workers Mettre a jour toute la stack vers les dernieres versions

# Mode dry-run (planification seule)
/parallel-workers --dry-run Ajouter un systeme de notifications

Difference avec /orchestrate

Aspect	/parallel-workers	/orchestrate
Mode	Standard (auto)	Injectable (template)
Execution	Workers paralleles	step-orchestrator sequentiel
Input	Description textuelle	--template, --workflow, --file
Decomposition	Automatique	Pipeline YAML

Agents utilises

• project-supervisor : Analyse initiale
• tech-lead : Architecture
• impact-analyzer : Protection AVANT
• context-keeper : Enrichissement contexte
• validator : Validation continue
• recovery-agent : Gestion echecs
• graph-sync : Synchronisation KG
• regression-guard : Protection APRES

---

SUCCESS METRICS

✅ SESSION_ID generated and session folder created ✅ Tasks decomposed into atomic units (1 file, 1 action, < 5 min each) ✅ Wave -1 executed: impact-analyzer ran BEFORE any modifications ✅ Wave 8 executed: regression-guard ran AFTER all modifications ✅ All workers used run_in_background: true ✅ context-keeper enriched every task with context brief ✅ Knowledge Graph updated throughout (via graph-sync) ✅ No direct execution - all work delegated to workers ✅ Parallelism achieved: 4-5 workers active simultaneously

FAILURE MODES

❌ Executing tasks directly instead of dispatching to workers ❌ Skipping Wave -1 (impact-analyzer) - protection is MANDATORY ❌ Skipping Wave 8 (regression-guard) - validation is MANDATORY ❌ Not using run_in_background: true - blocks the conversation ❌ Workers sharing state directly instead of via Knowledge Graph ❌ Tasks not atomic (multiple files, multiple actions, > 5 min) ❌ Skipping context-keeper enrichment ❌ Not generating SESSION_ID and session folder ❌ CRITICAL: Going "free style" without following the 4-phase process

CRITICAL REMINDERS

<critical> 1. YOU ORCHESTRATE, YOU DO NOT EXECUTE 2. Protection workflow is NON-NEGOTIABLE: impact-analyzer (Wave -1) → regression-guard (Wave 8) 3. ALL tasks MUST use `run_in_background: true` 4. Workers are ISOLATED - context comes ONLY from context-keeper briefs 5. State synchronization is ONLY via Knowledge Graph - no shortcuts 6. Atomic tasks: 1 file, 1 action, < 5 minutes - NO EXCEPTIONS </critical>

---

Voir aussi

• /orchestrate : Mode workflow injectable
• /workflow : Wrapper simplifie pour templates
• ~/.claude/docs/distributed-orchestration.md : Architecture detaillee