Wrench-bot Factory

Agentic factory maintenance, from diagnosis to verified onchain settlement.

Wrench-bot reads sensor logs to identify the root cause of a machine failure, purchases replacement parts through Reap Agentic Payments, and pays technicians in USDC after verifying the repair. Every purchase follows spending limits and approval rules set by the factory manager.

Winner · Best Bridge Between Onchain and Real World

Reap × 65labs Agentic Buildathon · Singapore · 9 October 2026 · Team Reapin' It In

Quick start · Architecture · Integrations · Evaluation guide

---

Project highlights

Claim	Where to verify
Real agentic purchasing. Live Reap catalog search, product → variant → quote → shipping → checkout → hosted approval → order status, against the Reap sandbox.	server/src/reap/client.js, server/src/reap/purchase.js, docs/PAYMENTS.md
Clear payment authority. The model only recommends. Code enforces an auto-approve limit, a monthly budget, a confidence threshold and a trusted-store list (AUTO / ESCALATE / BLOCK), and Reap's hosted page holds the final approval.	server/src/agent/policy.js, docs/AGENT.md
A decision model that earns its place. Diagnosis is a typed choice made by the OpenAI Decisions API. In Brownout, the fault code blames the servo; the model reads the sagging 24 V rail and blames the power supply. The game shows the naive "fault-code rule" next to it.	server/src/agent/decide.js, server/src/sim/diagnostics.js
Closed loop with ground truth. After the swap, the simulation's telemetry decides whether the fix worked. If it didn't, the agent rules that part out and re-diagnoses.	server/src/agent/orchestrator.js, docs/SIMULATION.md
Onchain, for real. The factory treasury owns a Kwal USDC vault on Ink Sepolia, deployed and funded onchain. Technicians are paid with real USDC transfers, released only after a verified repair.	Vault deposit tx · Payout tx · server/src/chain/usdc.js
Every state is designed. Auto-buy, needs-your-decision, blocked store, quote expired, store can't fill, payment failed, fix didn't take, Kwal unavailable → card fallback.	docs/PAYMENTS.md
Full audit trail. A manager dashboard lists every purchase, approval, block, payout and agent decision, kept across shifts.	server/src/agent/ledger.js, GET /api/ledger

→ Full evaluator guide with a criteria-to-evidence matrix and a 5-minute verification path: docs/JUDGING.md

---

The problem

When a production line stops, every minute costs money. Getting it running again is slow and manual: someone reads the alarms, guesses the part, finds a supplier, gets purchasing approval, waits for delivery, books a technician and pays them. And alarms mislead: the part that throws the fault code is often not the part that failed (a sagging power supply makes a healthy servo lose position), so the wrong part gets ordered and the line stays down longer.

Wrench-bot runs that whole loop, diagnose → source → pay → repair → verify → settle, inside spending rules the manager sets, and asks the manager exactly when it should.

What happens when a machine breaks

	1. Diagnose from the logs. The agent reads live sensor signals and the plant log (never the hidden part health) and asks the OpenAI Decisions API which part is the root cause. You see the model's probabilities next to the naive fault-code rule, plus the evidence lines it used. 2. Find the part live. Reap catalog search, filtered by spec and price, with trusted stores ranked first. 3. Get a real quote from the merchant: price, shipping, tax, expiry.
	What the agent sees. Click any machine for its parts, each signal against its warn and fail thresholds, and a live sparkline. This is exactly the telemetry the decision model reads.
	4. Check the rules in code. Within the limits and confident → the agent proceeds. Over the limit, over budget or not sure enough → "Needs your decision", with the reasons and the cost of waiting (−$60/min). Untrusted store → blocked, with no checkout created. 5. Pay through Reap's hosted approval.
	6. Ship, repair, verify. A technician swaps the part; the simulation's telemetry decides whether it worked. In Brownout the servo swap doesn't take, so the agent rules it out, finds the 24 V power supply and fixes it on try 2. 7. Settle onchain. The technician's escrow is released as a real USDC transfer on Ink Sepolia, linked in the log.
	Manager dashboard. Every purchase (store, amount, rail, auto or manager approval, order id), every block, every onchain payout with its transaction link, and every agent decision with its confidence, kept across shifts and exportable as CSV.

_{Left: the plant log; WARN/FAULT lines from the machines, AGENT lines for every action Wrench-bot takes. Right: a real technician payout on the Ink Sepolia explorer.}

Payment authority, in one table

Who	Can do	Enforced where
Decision model (OpenAI Decisions API)	Recommend a part, a listing, a technician, with probabilities	Nowhere near money: its output is an input to the policy
Wrench-bot (code)	Spend alone only if: total ≤ auto-approve limit and ≤ remaining monthly budget and confidence ≥ threshold and the store is trusted	policy.js evaluate(), with reason strings shown to the manager
Manager	Approve or reject anything escalated; change limits, budget, confidence bar and trusted stores live (Manager's Desk)	Reap's hosted approval page (passkey); PUT /api/policy
Onchain treasury	Hold the vault funds (hard cap) and pay technicians after a verified fix	Kwal vault contract; chain/usdc.js; payouts are idempotent per escrow

Integrations

System	Status	Proof
Reap Agentic Payments (sandbox)	Live: catalog search (~1,750 products found), quotes, checkouts, hosted approval, card enrollment (VISA ending 1811)	Observed live quotes: proximity sensor $35.68, 2× servo $93.15, PoE switch $64.06
OpenAI Decisions API (gpt-6-luna)	Live: typed choice / yes-no decisions with per-option probabilities	Brownout: picks the 24 V PSU while the fault code says servo (docs/AGENT.md)
Kwal vault (Payward)	Live onchain: vault 0x1473…572f deployed and funded with 8 USDC, "ready for checkout"	deployment tx · deposit tx. Vault-paid parts fall back to the Reap card while Kwal's quote endpoint returns 400 ParticipantBadRequest
Ink Sepolia USDC	Live: technician payouts from the treasury	payout tx (amounts scaled 1:100 so testnet funds last)
Factory, technicians, deliveries	Simulated	docs/SIMULATION.md

Scenarios (each demonstrates one idea)

Scenario	What you'll see
Sensor burnout	A $15 sensor dies; within the limit → the agent handles the whole loop
Brownout	Misleading fault code; model vs fault-code rule; escalation; wrong part → re-diagnosis → right part
Grinding noise	Predictive maintenance: rising vibration → part ordered before the line stops
Month-end crunch	$38 left in the budget; over-budget orders escalate to the manager
Supplier gap	The only listing is from an untrusted store → blocked until the manager trusts it → retry
Free play	Parts fail one at a time; chaos tools and the Manager's Desk

Architecture

flowchart LR
  G[Game · Three.js + DOM UI] -- REST --> S[Express API]
  S -- SSE events --> G
  S --> SIM[Simulation engine<br/>hidden health → signals → plant log]
  SIM -- machine down / warnings --> O[Orchestrator<br/>incident pipeline]
  O --> D[decide · OpenAI Decisions API]
  O --> P[policy · hard spending limits]
  O --> C[catalog · live Reap search]
  O --> R[Reap Agentic client<br/>quote · checkout · approval]
  O --> K[Kwal vault rail]
  O --> U[Ink Sepolia USDC payouts]
  O --> L[ledger · manager dashboard]

Loading

Details: docs/ARCHITECTURE.md · docs/AGENT.md · docs/PAYMENTS.md · docs/SIMULATION.md · docs/API.md · docs/JUDGING.md

Quick start

npm install
cp .env.example .env      # REAP_API_KEY, OPENAI_API_KEY (+ optional onchain / card settings)
npm run dev               # server :8787 · game http://localhost:5173

No keys? npm run dev:mock runs everything offline with simulated payments.

• Card for Reap checkouts: npm run reap:enroll -w server, enter the card on Reap's hosted page, then set REAP_ENROLLMENT_ID.
• Onchain: TREASURY_PRIVATE_KEY (an Ink Sepolia wallet with test USDC) enables real technician payouts. The Kwal vault is created with the Kwal agent skill.
• Public demo ("Judge mode"): JUDGE_MODE=1 keeps the AI, catalog and quotes live, completes checkouts as clearly labelled demos (visitors can't tap our card's passkey), caps tiny onchain payouts, and resets the factory when idle. Deploys as one Render web service (render.yaml).

What we learned about Reap's sandbox

• Reap-Version: 2025-02-14 is required, and the only way to discover it is the validation error.
• Checkout returnUrl must be public HTTPS; localhost is rejected.
• One quote per merchant; mixed carts are rejected (AGENTIC_REQUEST_REJECTED).
• Every sandbox checkout needs a one-tap approval on the hosted page, even with X-Simulate-Checkout: COMPLETED. Mandates (pre-approved terms) aren't live yet, so the app's own policy decides when the agent may ask, and Reap's page stays the final human check.

Limitations

• The factory, technicians and deliveries are simulated; the catalog, quotes, checkouts, vault and payouts are real sandbox/testnet systems.
• Vault-paid parts fall back to the Reap card while Kwal's quote endpoint fails.
• Technician payouts are testnet USDC, scaled 1:100.

---

_{Built in one evening at the Reap × 65labs Agentic Buildathon by Team Reapin' It In. Docs index: docs/.}