Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Unified context optimization and session state management for Claude Code with 200K token budget management, session persistence, and multi-agent handoff protocols.
Core Capabilities:
200K token budget allocation and monitoring
Session state tracking with persistence
Context-aware token optimization
Multi-agent handoff protocols
Progressive disclosure and memory management
Session forking for parallel exploration
When to Use:
Session initialization and cleanup
Long-running workflows (>10 minutes)
Multi-agent orchestration
Context window approaching limits (>150K tokens)
Model switches (Haiku ↔ Sonnet)
Workflow phase transitions
Key Principles (2025):
Avoid Last 20% - Performance degrades in final fifth of context
Aggressive Clearing - /clear every 1-3 messages for SPEC workflows
Concept: Load context progressively based on relevance and need.
Progressive Summarization:
defprogressive_summarization(context: str, target_ratio: float = 0.3):
"""Compress context while preserving key information."""# Step 1: Extract key sentences (50K → 15K)
summary = extract_key_sentences(context, ratio=target_ratio)
# Step 2: Add pointers to original content
summary_with_refs = add_content_pointers(summary, context)
# Step 3: Store original for recovery
store_original_context(context, "session_archive")
return summary_with_refs # 35K tokens saved
Context Tagging:
# Bad (high token cost):"The user configuration from the previous 20 messages..."# Good (efficient reference):"Refer to @CONFIG-001 for user preferences"
Use Case: Maintain context continuity while minimizing token overhead.
Advanced Documentation
For detailed patterns and implementation strategies:
Reference Guide - API reference, troubleshooting, best practices
Best Practices
DO
Execute /clear immediately after SPEC creation
Monitor token usage and plan accordingly
Use context-aware token budget tracking
Create checkpoints before major operations
Apply progressive summarization for long workflows
Enable session persistence for recovery
Use session forking for parallel exploration
Keep memory files < 500 lines each
Disable unused MCP servers to reduce overhead
REQUIREMENTS
[HARD] Maintain bounded context history with regular clearing cycles
WHY: Unbounded context accumulation degrades performance and increases token costs exponentially
IMPACT: Prevents context overflow, maintains consistent response quality, reduces token waste by 60-70%
[HARD] Respond to token budget warnings immediately when usage exceeds 150K tokens
WHY: Operating in the final 20% of context window causes significant performance degradation
IMPACT: Ensures optimal model performance, prevents context overflow failures, maintains workflow continuity
[HARD] Execute state validation checks during session recovery operations
WHY: Invalid state can cause workflow failures and data loss in multi-step processes
IMPACT: Guarantees session integrity, prevents silent failures, enables reliable recovery with >95% success rate
[HARD] Persist session identifiers before any context clearing operations
WHY: Session IDs are the only reliable mechanism for resuming interrupted workflows
IMPACT: Enables seamless workflow resumption, prevents work loss, supports multi-agent coordination
[SOFT] Establish clear session boundaries when working with multiple concurrent sessions
WHY: Session mixing causes context contamination and unpredictable agent behavior
IMPACT: Improves debugging clarity, prevents cross-session interference, maintains clean audit trails
[SOFT] Create checkpoint snapshots before assuming session continuity
WHY: Context can be lost due to network issues, timeouts, or system events
IMPACT: Provides recovery points, reduces rework time, maintains user trust in system reliability
[SOFT] Load codebase components progressively using priority tiers
WHY: Loading entire codebases exhausts token budget and includes irrelevant context
IMPACT: Optimizes token usage by 40-50%, improves response relevance, enables larger project support
[SOFT] Limit handoff packages to critical context only
WHY: Non-critical context increases handoff overhead and reduces available working tokens
IMPACT: Speeds up agent transitions by 30%, preserves token budget for actual work, reduces transfer errors
[HARD] Execute context compression or clearing when usage reaches 85% threshold
WHY: Approaching context limits triggers emergency behaviors and reduces model capabilities
IMPACT: Maintains 55K token emergency reserve, prevents forced interruptions, ensures graceful degradation
Works Well With
moai-cc-memory - Memory management and context persistence
moai-cc-configuration - Session configuration and preferences
moai-core-workflow - Workflow state persistence and recovery
moai-cc-agents - Agent state management across sessions
