Loop

Agent that learns how to use third-party tools over repeated runs.

Hackathon: Syndicate by Maximor — Track 1 (Automated Agent Engineering)

Built with Agent Orchestrator (AO).

# Terminal 1 — CRM tools (pick one)
npm run tools          # mock Node server (fixture CRM data)
# OR real n8n engine + same fixture CRM payloads:
# npm run n8n:up && npm run n8n:import

# Terminal 2 — one planner run (offline if TensorMux env unset)
npm run agent

# Second run should retrieve the first episode (stdout: episodic_retrieve / episodic_inject)
npm run agent

# Deterministic analyzer proof (flagged naive vs clean)
npm run analyzer:prove

# Two-run lesson promotion (candidate → usable) — writes Neo4j when configured
npm run reflection:prove

# Strategy injection: naive vs injected unseen task
npm run strategy:prove

# Demo spine — seen_a / seen_b / unseen trajectory (Step 9)
# Uses n8n or mock on :5678; Neo4j lessons when NEO4J_* set; offline planner by default
npm run replay:demo

# Demo dashboard — stack status + trajectory + lessons (Step 10)
npm run dashboard
# open http://localhost:3847  → /api/stack · /api/trajectory · /api/lessons

# Inspect stores
npm run working:list
npm run episodic:list
npm run memory:lessons

Mock CRM/support webhooks on http://localhost:5678. See tools/README.md for endpoints, curl examples, and the customer_id dependency the agent must learn.

Preconditions: mock tools on :5678 (npm run tools in another terminal).

Runs the committed scenario pack in scripts/scenarios/replay-demo.json through runOnePlanner:

1. seen_a — historically fails without customer_id resolution (naive path) → candidate lesson
2. seen_b — slight wording variant of the same task → promotes usable lesson
3. unseen — differently worded task; benefits via shared-factor lesson injection

Prints per-run rows (run_id, label, success, tool counts, latency, tokens, lessons retrieved/promoted) then a comparison table. Writes data/replay-demo/trajectory.json for the dashboard. Exits non-zero unless at least one improvement signal holds (fewer failed tools on unseen, success rising, or lessons retrieved on unseen only). Stores under ./data/replay-demo/.

Minimal read-only web view (no auth) for screen recordings:

# Terminal 1 — mock tools (needed before replay)
npm run tools

# Terminal 2 — produce trajectory + lessons (once, or again after code changes)
npm run replay:demo

# Terminal 3 — dashboard
npm run dashboard
# alias: npm run dev:dashboard
# open http://localhost:3847

If no replay has been run yet, both panels show an empty state hinting to run npm run replay:demo. The server reads ./data/replay-demo/ only (same isolation as the demo spine); it does not duplicate planner/reflection logic.

Optional: DASHBOARD_PORT=3847 (default).

See tools/README.md for webhook contracts and agent/README.md for the planner, AO entrypoint, Neatlogs, SQLite working/episodic memory, semantic lessons, strategy injection, demo replay, and dashboard.