π The original agent harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build conversational AI systems. Features adaptive memory, self-learning intelligence, federation, vector RAG integration, and native Claude Code / Codex / Hermes and many more Integrated
TypeScript agentic-aiagentic-frameworkagentic-workflowagentsai-agentsai-assistantai-skillsautonomous-agentsclaudecodexdsh-pluginharnessmcp-servermulti-agentmulti-agent-systemsnpmskillsswarmswarm-intelligencetypescript
Checkpoint: File edits
rUvad792751a6d8
diff --git a/README.md b/README.mdindex 14b4d73d76..8a9298cf89 100644--- a/README.md+++ b/README.md@@ -2833,6 +2833,120 @@ Real-world scenarios and pre-built workflows for common tasks.---+## π§ Infinite Context & Memory Optimization++Ruflo eliminates Claude Code's context window ceiling with a real-time memory management system that archives, optimizes, and restores conversation context automatically.++<details>+<summary>βΎοΈ <strong>Context Autopilot</strong> β Never lose context to compaction again</summary>++### The Problem++Claude Code has a finite context window (~200K tokens). When full, it **compacts** β summarizing the conversation and discarding details like exact file paths, tool outputs, decision reasoning, and code snippets. This creates a "context cliff" where Claude loses the ability to reference earlier work.++### The Solution: Context Autopilot (ADR-051)++Ruflo intercepts the compaction lifecycle with three hooks that make context loss invisible:++```+Every Prompt Context Full After Compact+ β β β+ βΌ βΌ βΌ+UserPromptSubmit PreCompact SessionStart+ β β β+ Archive turns Archive + BLOCK Restore from archive+ to SQLite auto-compaction via additionalContext+ (incremental) (exit code 2) (importance-ranked)+ β β β+ βΌ βΌ βΌ+ Track tokens Manual /compact Seamless continuation+ Report % used still allowed with full history+```++### How Memory is Optimized++| Layer | What It Does | When |+|-------|-------------|------|+| **Proactive Archiving** | Every user prompt archives new turns to SQLite with SHA-256 dedup | Every prompt |+| **Token Tracking** | Reads actual API `usage` data (input + cache tokens) for accurate % | Every prompt |+| **Compaction Blocking** | PreCompact hook returns exit code 2 to cancel auto-compaction | When context fills |+| **Manual Compact** | `/compact` is allowed β archives first, resets autopilot, then compresses | On user request |+| **Importance Ranking** | Entries scored by `recency Γ frequency Γ richness` for smart retrieval | On restore |+| **Access Tracking** | Restored entries get access_count++ creating a relevance feedback loop | On restore |+| **Auto-Pruning** | Never-accessed entries older than 30 days are automatically removed | On PreCompact |+| **Content Compaction** | Old session entries trimmed to summaries, reducing archive storage | Manual or scheduled |+| **RuVector Sync** | SQLite entries auto-replicated to PostgreSQL when configured | On PreCompact |++### Optimization Thresholds++| Zone | Threshold | Statusline | Action |+|------|-----------|-----------|--------|+| OK | <70% | `π‘οΈ 43% 86.7K β` (green) | Normal operation, track growth trend |+| Warning | 70-85% | `π‘οΈ 72% 144K β` (yellow) | Flag approaching limit, archive aggressively |+| Optimize | 85%+ | `π‘οΈ 88% 176K β³2` (red) | Prune stale entries, keep responses concise |++### Real-Time Statusline++The statusline shows live context metrics read from `autopilot-state.json`:++```+π‘οΈ 45% 89.2K β π§ 86%+β β β β β β+β β β β β ββ Intelligence score (learning.json + patterns + archive)+β β β β βββββ Intelligence indicator+β β β ββββββββββ No prune cycles (β) or prune count (β³N)+β β ββββββββββββββββ Token count (actual API usage)+β ββββββββββββββββββββ Context percentage used+βββββββββββββββββββββββββ Autopilot active (shield icon)+```++### Storage Tiers++| Tier | Backend | Storage | Features |+|------|---------|---------|----------|+| 1 | **SQLite** (default) | `.claude-flow/data/transcript-archive.db` | WAL mode, indexed queries, ACID, importance ranking |+| 2 | **RuVector PostgreSQL** | Configurable remote | TB-scale, pgvector embeddings, GNN search |+| 3 | **AgentDB + HNSW** | In-memory + persist | 150x-12,500x faster semantic search |+| 4 | **JSON** (fallback) | `.claude-flow/data/transcript-archive.json` | Zero dependencies, always works |++### Configuration++```bash+# Context Autopilot (all have sensible defaults)+CLAUDE_FLOW_CONTEXT_AUTOPILOT=true # Enable/disable autopilot (default: true)+CLAUDE_FLOW_CONTEXT_WINDOW=200000 # Context window size in tokens+CLAUDE_FLOW_AUTOPILOT_WARN=0.70 # Warning threshold (70%)+CLAUDE_FLOW_AUTOPILOT_PRUNE=0.85 # Optimization threshold (85%)+CLAUDE_FLOW_COMPACT_RESTORE_BUDGET=4000 # Max chars restored after compaction+CLAUDE_FLOW_RETENTION_DAYS=30 # Auto-prune never-accessed entries+CLAUDE_FLOW_AUTO_OPTIMIZE=true # Importance ranking + pruning + sync+```++### Commands++```bash+# Check archive status and autopilot state+node .claude/helpers/context-persistence-hook.mjs status++# Manual compact (archives first, then allows Claude Code to compress)+# Use /compact in Claude Code β autopilot allows manual, blocks auto++# Query archive directly+sqlite3 .claude-flow/data/transcript-archive.db \+ "SELECT COUNT(*), SUM(LENGTH(content)) FROM transcript_entries;"+```++### Architecture Reference++- **ADR-051**: Infinite Context via Compaction-to-Memory Bridge+- **ADR-052**: Statusline Observability System+- **Implementation**: `.claude/helpers/context-persistence-hook.mjs` (~1560 lines)+- **Settings**: `.claude/settings.json` (PreCompact, SessionStart, UserPromptSubmit hooks)++</details>++---+## π§ Intelligence & LearningSelf-learning hooks, pattern recognition, and intelligent task routing.