UNPKG

claude-flow-novice

Version:

Claude Flow Novice - Advanced orchestration platform for multi-agent AI workflows with CFN Loop architecture Includes Local RuVector Accelerator and all CFN skills for complete functionality.

201 lines (165 loc) 7.54 kB
--- description: Run ACE curator to merge reflection deltas into adaptive context with deduplication tags: [context, ace, curation, merge, deduplication] --- # Context Curation Command Spawn a `context-curator` agent to merge reflection deltas into the `adaptive_context` table using deterministic rules, semantic deduplication, and version control. **Usage:** ```bash /context-curate [--reflection-id=<id>] [--auto-merge] [--similarity-threshold=0.85] ``` **What This Does:** 1. Spawns `context-curator` agent (specialized for ACE curation) 2. Fetches pending reflections from `context_reflections` table 3. For each extracted lesson: - **Check for duplicates** using semantic similarity (embeddings) - **Merge similar bullets** if similarity > threshold - **Add new bullets** if novel insight - **Increment counters** for reinforced lessons - **Version bullets** for significant edits 4. Updates `adaptive_context` table with deterministic merge logic 5. Records curation decisions in `context_merge_log` (audit trail) 6. Updates bullet confidence scores based on reinforcement **Arguments:** - `--reflection-id=<id>`: Curate specific reflection (default: all pending) - `--auto-merge`: Skip human review, auto-merge with high confidence (≥0.8) - `--similarity-threshold=<0.0-1.0>`: Semantic similarity for merging (default: 0.85) - `--require-validation`: Flag merged bullets for human/PO validation - `--dry-run`: Preview merge actions without applying - `--category=<type>`: Filter by bullet category - `--min-confidence=<0.0-1.0>`: Only merge bullets above confidence threshold (default: 0.5) - `--output=<path>`: Save curation report to file **Merge Logic (Deterministic):** 1. **New Bullet (No Similar Match):** ```sql INSERT INTO adaptive_context (bullet_id, category, content, confidence_score, ...) VALUES ('STRAT-042', 'strategy', 'Use Redis pub/sub...', 0.85, ...); INSERT INTO context_merge_log (merge_type, bullet_id, new_content, curator_reasoning) VALUES ('new_bullet', 'STRAT-042', 'Use Redis pub/sub...', 'Novel strategy, no similar bullets found'); ``` 2. **Similar Bullet Found (Similarity > Threshold):** ```sql -- Increment helpful counter UPDATE adaptive_context SET helpful_count = helpful_count + 1, confidence_score = MIN(1.0, confidence_score + 0.05), usage_count = usage_count + 1, last_used_at = CURRENT_TIMESTAMP WHERE bullet_id = 'STRAT-001'; INSERT INTO context_merge_log (merge_type, bullet_id, similarity_score, curator_reasoning) VALUES ('increment_helpful', 'STRAT-001', 0.92, 'Reinforced by new reflection'); ``` 3. **Merge Semantically Overlapping Bullets:** ```sql -- Create new version merging STRAT-001 + STRAT-042 INSERT INTO adaptive_context (bullet_id, category, content, version, parent_bullet_id, helpful_count, ...) VALUES ('STRAT-001-v2', 'strategy', 'Merged content...', 2, 'STRAT-001', 15, ...); -- Archive old bullets UPDATE adaptive_context SET is_active = 0, archived_at = CURRENT_TIMESTAMP WHERE bullet_id IN ('STRAT-001', 'STRAT-042'); INSERT INTO context_merge_log (merge_type, merged_from_bullet_ids, new_content) VALUES ('merge_similar', '["STRAT-001", "STRAT-042"]', 'Merged content...'); ``` 4. **Archive Harmful Bullets:** ```sql -- Auto-archived by trigger when harmful_count >= 5 AND helpful_count < 2 UPDATE adaptive_context SET is_active = 0, archived_at = CURRENT_TIMESTAMP WHERE bullet_id = 'ANTI-013'; INSERT INTO context_merge_log (merge_type, bullet_id, curator_reasoning) VALUES ('archive', 'ANTI-013', 'Harmful pattern confirmed by multiple tasks'); ``` **Semantic Similarity (Optional):** If embeddings available in `embedding_vector` column: ```javascript // Cosine similarity between bullet embeddings const similarity = cosineSimilarity(embedding1, embedding2); if (similarity > threshold) { // Merge or increment counter } else { // Add as new bullet } ``` **Curation Prompt Template:** ``` You are a Context Curator for the ACE (Adaptive Context Extension) system. **Pending Reflection:** {reflection_data} **Existing Bullets (Semantic Neighbors):** {similar_bullets} **Your Mission:** For each extracted lesson in the reflection: 1. **Check for semantic similarity**: - Compare with existing bullets - Use similarity threshold: {threshold} 2. **Apply merge logic**: - **Similarity > {threshold}**: Increment helpful_count, boost confidence - **Similarity 0.6-{threshold}**: Consider merging bullets (create new version) - **Similarity < 0.6**: Add as new bullet 3. **Validate quality**: - Is the bullet actionable? - Is it specific enough? - Does it have clear conditions/context? - Would it help future agents? 4. **Assign priority** (1-10): - High (8-10): Critical strategies, common patterns, safety rules - Medium (5-7): Optimization tips, domain insights - Low (1-4): Edge cases, situational patterns 5. **Update counters**: - Identify helpful_existing_bullets → increment their helpful_count - Identify harmful_existing_bullets → increment their harmful_count **Output Format (JSON):** { "curation_actions": [ { "action": "new_bullet|increment_helpful|increment_harmful|merge_similar|archive", "bullet_id": "STRAT-042", "target_bullet_id": "STRAT-001", // For increment/merge actions "similarity_score": 0.92, "reasoning": "Why this action is appropriate", "merged_content": "New merged content (if merge action)" } ], "validation_required": ["STRAT-042"], // Bullets needing human review "summary": "Curated 5 lessons: 2 new, 3 reinforced existing" } ``` **Post-Curation:** 1. Update `context_reflections.curator_status = 'merged'` 2. Record all actions in `context_merge_log` 3. Emit metrics to `metrics` table 4. Print summary: ``` ✅ Curation Complete: reflection-abc123 📊 Merge Actions: - STRAT-042: New bullet added (confidence: 0.85) - STRAT-001: Helpful count +1 (now 16 helpful, 0 harmful) - PATTERN-017: Helpful count +1 (now 10 helpful, 0 harmful) 📈 Context Stats: - Total active bullets: 127 - Avg confidence: 0.78 - Bullets needing validation: 3 💡 Next: Run /context-inject --category=strategy to add top bullets to CLAUDE.md ``` **Periodic Maintenance:** Run curation with `--maintenance` flag to: - Deduplicate similar bullets (similarity > 0.90) - Archive bullets with harmful_count ≥ 5 - Prune bullets unused for >90 days (if helpful_count < 3) - Merge semantically overlapping bullets - Rebalance priority scores **Integration with CFN Loop:** - **Loop 3 completion**: Auto-curate agent reflections - **Loop 4 approval**: PO reviews high-priority bullets - **Phase completion**: Periodic maintenance curation - **Epic completion**: Comprehensive deduplication pass **ACL & Security:** - Curation actions require ACL Level 4 (Project) - Merge log audit trail at ACL Level 5 (System) - Sensitive bullets flagged for validation - Human review required for confidence < 0.6 **See Also:** - `/context-reflect` - Extract lessons from execution - `/context-query` - Search bullets by category/tags - `/context-inject` - Add bullets to CLAUDE.md - `/context-stats` - View bullet statistics and health