Real-Time Self-Healing Multi-Agent Cockpit for Code Generation & Context Drift Prevention
CodexMap real-time telemetry cockpit: live AST dependency mapping, real-time context drift scoring ($S_1, S_2, A, T, D$), and automated closed-loop self-healing over WebSockets.
Watch High-Definition 80s Video Demo (MP4)
When autonomous AI agents or multi-agent loops generate software codebases, they frequently suffer from context drift:
- As files accumulate, local edits inadvertently introduce broken dependencies, missing parameters, and incompatible interface signatures across the call graph.
- Syntactic safety slips (unhandled exceptions, loose equality, dangerous
evalcalls, or TypeScript: anytypes) bypass traditional compiler checks. - Existing tools lack real-time visibility, causing AI generation runs to cascade into unrecoverable compilation failures—an effect known as Architectural Collapse.
CodexMap solves this by establishing a real-time developer cockpit and closed-loop self-healing framework. It maps code generation as a live interactive graph, continuously evaluates 5-dimensional drift scores in background agent threads, and automatically dispatches scoped LLM re-anchoring routines to repair drifted files without human intervention—all while strictly preventing infinite repair loops.
-
4-Agent Actor Pipeline: Dedicated concurrent sidecars for file-mapping (
Cartographer), drift-scoring (Sentinel), LLM code repair (Healer), and WebSocket streaming (Broadcaster). -
5-Component Scoring Matrix: Real-time evaluation combining semantic vector distance (
$S_1$ ), vectorless lexical BM25 matching ($S_2$ ), architectural tree coherence ($A$ ), static AST type safety heuristics ($T$ ), and PageRank graph centrality penalty ($D$ ). -
Closed-Loop Self-Healing: Automatically catches drifted files (
$S_{\text{final}} < 0.40$ ) and triggers scoped LLM re-invocations to rewrite and re-anchor code targets. -
Infinite Loop Prevention (Re-Anchor Registry): Uses cryptographic SHA-256 state-signature tracking to cap repair retries and stop oscillating sister-node edit loops (
$0%$ runaway repair risk). -
Model-Agnostic Engine Adapters: Standardized plug-and-play adapter interface supporting Google Gemini, OpenAI Codex, and deterministic simulation engines (
--engine fake). - Zero-Jitter Visual Cockpit: Cytoscape.js interactive knowledge graph with decoupled layout computation, animated score deduction bars, and live API token/cost financial auditing.
sequenceDiagram
autonumber
actor Dev as Developer / Generator
participant C as Cartographer Agent
participant DB as SQLite DB (WAL Mode)
participant S as Sentinel Agent
participant H as Healer Agent
participant LLM as Gemini / Codex Engine
participant B as Broadcaster Agent
participant UI as Browser Cockpit (WS Client)
Dev->>C: Write / Modify File (Disk Watcher)
C->>DB: Extract AST, Map Nodes/Edges (grade='pending')
DB-->>S: Trigger Sentinel via notify()
S->>S: Run Embeddings (S1) & BM25 (S2)
S->>S: Run Static Rules (T) & PageRank Penalty (D)
S->>DB: Write Score (S1, S2, A, T, D, S_final)
DB-->>B: Broadcast node_grade & cost_update via events_queue
B->>UI: Stream live updates over WebSockets
Note over S,H: Loop check: If S_final < 0.40 & AUTO_HEAL = true
S->>DB: Queue node in heal_queue
DB-->>H: Trigger Healer via notify()
H->>H: Check Re-Anchor Registry (State Hash Loop Check)
H->>LLM: Dispatch repair prompt with AST context
LLM-->>H: Return corrected code JSON
H->>Dev: Overwrite local file on disk
H->>DB: Mark heal_queue status = 'done'
Every node (file or function) in the graph is evaluated against the target prompt using the composite formula:
| Component | Metric Name | Purpose | Calculation Method |
|---|---|---|---|
| Semantic Similarity | Evaluates conceptual prompt alignment | Cosine similarity on 1536-dim code/prompt embeddings | |
| Vectorless Lexical RAG | Ensures exact keyword & class presence | BM25 term frequency matching across tokenized nodes | |
| Architectural Consistency | Penalizes unstable/broken dependencies | Coherence score based on parent context & child error rates | |
| Type Safety & Heuristics | Detects syntax & error-handling hazards | Static pattern inspection (eval, ==, : any, raw db. calls) |
|
| Centrality Penalty | Penalizes drift on central orchestrators | PageRank-weighted deduction: |
# Run with Google Gemini engine
GEMINI_API_KEY="your-api-key" npx codexmap run "Build a secure REST API with JWT auth" --engine gemini
# Run with local simulation/fake engine (offline & zero cost)
npx codexmap run "Build a banking app with accounts" --engine fake# Clone the repository
git clone https://github.com/your-username/codexmap.git
cd codexmap
# Install dependencies
npm install
# Run doctor diagnostic
node bin/codexmap.js doctor
# Start local daemon
node bin/codexmap.js run "Build an express service" --engine fake --watch ./output| Flag | Default | Description |
|---|---|---|
--engine |
codex |
Engine adapter (gemini, codex, or fake) |
--watch |
./output |
Target folder to monitor and map |
--port |
3333 |
HTTP Cockpit dashboard port |
--ws-port |
4242 |
WebSocket real-time telemetry streaming port |
--auto-heal |
false |
Automatically heal nodes with |
--no-open |
false |
Prevent opening browser on startup |
--model |
gemini-1.5-flash |
LLM model name for generative engine |
- Multi-Agent Runtime: Core orchestration pipeline with Cartographer, Sentinel, Healer, and Broadcaster.
- Native Gemini Adapter: Dependency-free Google Gemini API engine adapter supporting structured JSON responses.
- Re-Anchor Loop Prevention: Cryptographic state-hash registry preventing infinite agent repair loops.
- Interactive Cockpit UI: Modern Cytoscape graph canvas, formula breakdown cards, deduction bars, and live API cost budgeting.
- SQLite WAL Datastore: Thread-safe concurrent logging and event queues with busy-timeout configurations.
Before publishing a new release:
npm run release:preflightThis runs syntax checks, unit tests, engine contract validations, and package tarball verification.
npm publish --provenance --access publiccodexmap/
├── bin/
│ └── codexmap.js # CLI entry point (run, doctor, engines, ui)
├── agents/
│ ├── cartographer.js # AST dependency mapping agent
│ ├── sentinel.js # 5-component scoring & drift detection agent
│ ├── healer.js # Autonomous LLM code re-anchoring agent
│ └── broadcaster.js # WebSocket real-time event streaming agent
├── engines/
│ ├── contract.js # Engine adapter interface contract
│ ├── gemini.js # Native Google Gemini API engine adapter
│ ├── gemini_worker.js # Child-process worker for Gemini HTTPS calls
│ ├── codex.js # OpenAI Codex engine adapter
│ └── fake.js # Mock/simulation engine for deterministic testing
├── lib/
│ ├── db.js # SQLite database (WAL mode, event queues)
│ ├── cost.js # Token expenditure & API cost tracker
│ ├── runtime.js # Daemon lifecycle coordinator
│ └── atomic.js # Atomic file operations & directory helpers
├── ui/
│ ├── index.html # Cockpit dashboard application entry
│ ├── graph.js # Cytoscape.js interactive knowledge graph
│ ├── panel.js # Detailed score breakdown & formula card
│ └── ws.js # WebSocket client event broker
└── assets/
├── codexmap-demo.gif # Embedded animated demo preview
└── demo.mp4 # Full high-definition video walkthrough
Distributed under the MIT License.
