Goal Drift Monitor

Type: Local Monitor / API / MCP Server / Dashboard

Value: Tracks whether long-running agents are still aligned with the original goal and alerts when behavior starts to diverge.

Current Status

Complete as a local monitor / CLI / MCP / HTTP dashboard slice. It persists local sessions, scores each turn, and exposes alert receipts for local review workflows.

Shipped Local Scope

• Durable local goal sessions and turn history
• Turn-by-turn lexical drift scoring
• Drift threshold alerts
• Webhook delivery receipts for drift alerts
• CLI for creating sessions, adding turns, summaries, history, alerts, and serving the dashboard
• MCP tools for create, score, summarize, history, alert receipts, list, and delete
• Local HTTP API and dashboard
• Regression tests for persistence, CLI, MCP validation, HTTP/dashboard, and core scoring

Commands

npm install
npm test
npm run build

goal-drift session create --goal "Keep the agent focused on refund policy questions" --threshold 0.35
goal-drift turn add --session <id> --content "The agent is discussing vacation planning"
goal-drift summary --session <id>
goal-drift history --session <id>
goal-drift alerts --session <id>
goal-drift serve --port 4321

The dashboard is available at http://127.0.0.1:4321/dashboard after goal-drift serve.

Local API

• GET /health
• POST /api/sessions
• GET /api/sessions
• GET /api/sessions/:id
• DELETE /api/sessions/:id
• POST /api/sessions/:id/turns
• GET /api/sessions/:id/history
• GET /api/sessions/:id/summary
• GET /api/sessions/:id/alerts
• GET /dashboard

Local Store

By default, local sessions are stored at:

.goal-drift-monitor/sessions.json

Override with:

GOAL_DRIFT_MONITOR_STORE_PATH=/path/to/sessions.json

Current Guardrails

• External HTTP and MCP inputs are validated before mutating state.
• Invalid MCP envelopes return JSON-RPC -32600.
• Invalid MCP tool params return JSON-RPC -32602.
• Webhook threshold breaches and delivery receipts are separate: alertFired means drift threshold breached, while alertDeliveryStatus records whether delivery was configured, delivered, or failed.

Current Limits

• This is a local product slice, not a hosted multi-user SaaS service.
• No public npm publication, hosted account system, billing, team workspace, or live checkout is included.
• Scoring defaults to deterministic lexical matching. Semantic scoring exists in the SDK but MCP/CLI/API currently expose lexical scoring for predictable local operation.
• Email alerts are not implemented; webhook receipts are implemented.
• Delete currently removes the session from the local store. Compliance-grade immutable audit retention remains future scope.

Verification

Fresh suite verification on 2026-05-28:

• npm test passed, 19/19 tests.
• npm run build passed.