tiny-harness

Appearance

Evidence: the delegation example with a real model (T13) ​

R5 and R6, added in PR review. Run on 2026-10-13 with gpt-6.1-sol, both agents on 127.0.0.1, embedded Temporal, TEMPORAL_API_KEY unset, OPENAI_API_KEY from the environment (not recorded). Task ids are shortened.

Result: pass. The front desk had no tools and was told nothing about the ledger. Its model read the ledger's card, delegated the question by itself, and answered with facts that only the ledger's tool holds.

Procedure ​

As in testing-plan.md § Verification environment (T13). Each server was started from a script file and stopped by process group:

sh
uv run tiny-harness --config examples/delegation/ledger.toml serve      # :8091
uv run tiny-harness --config examples/delegation/front-desk.toml serve  # :8090

The question was sent to the front desk as one JSON-RPC SendMessage with X-Participant-Id: alice: "What is the status of order 48213?"

1. The ledger serves the card that [card] configures (R5.2) ​

GET http://127.0.0.1:8091/.well-known/agent-card.json (excerpt):

json
{
  "name": "Order ledger",
  "description": "Answers questions about the status of a Nimbus Kitchen order: shipping, delivery and refunds. Give it the order number.",
  "version": "0.1.0"
}

2. The front desk lists the ledger at start ​

Front-desk log, at startup:

json
{"level": "INFO", "logger": "tiny_harness.runtime", "msg": "agents roster: ledger"}

No truncation warning: the ledger's card is within the caps.

3. The front desk delegates, unprompted (R6.2) ​

The first status reply to the user: "I'm checking the order ledger for the latest status of order 48213." The front-desk task's sub_tasks:

json
[{"task_id": "cd675ce4….1", "agent": {"kind": "agent", "id": "ledger"}, "state": "COMPLETED"}]

The ledger's task (read with GetTask as the front desk's identity, tiny-harness):

  • goal: "Look up Nimbus Kitchen order 48213 and report its current status, carrier, shipping and delivery dates, and any refund information available."
  • The ledger called its ledger/orders.get_order tool (tool.invoked pre/post in its log), then answered: "Order 48213 (Nimbus 900 blender) is shipped via Harbour Couriers, with a shipping date of October 10, 2026 and expected delivery on October 14, 2026. The ledger has no actual delivery date or refund information recorded."

4. The front desk answers with the ledger's result ​

The front-desk task's final status message:

Order 48213 (Nimbus 900 blender) has shipped.

  • Carrier: Harbour Couriers
  • Shipped: October 10, 2026
  • Expected delivery: October 14, 2026

The ledger has no confirmed delivery date or refund information recorded.

The carrier and the dates exist only in examples/delegation/ledger/mcp_ledger.py, which the front desk cannot reach. The front desk's plan recorded both steps as done ("Check order", then "Report status").

Observations ​

  • One attempt, no retries.
  • Not part of this change: both agents name themselves tiny-harness as a participant, so the ledger's task lists tiny-harness as both reporter and assignee. This is the existing participant id, which [card] does not change.