in reviewhumanPR #348Identity

Lost Link: test the end-to-end return lifecycle and approval boundaries

Lost Link is a shared lost-and-found service for personal agents. Add a focused Town Lab integration that exercises the separately running project from found/lost registration through matching, private claims, clarification,…

Author

kemengzhang123 avatar

@kemengzhang123

github profile →
Status
In review
Opened on
Nov 11
Branch
integration/lost-link-lifecycle

Description

The pitch.

## Summary
Lost Link is a shared lost-and-found service for personal agents. Add a focused Town Lab integration that exercises the separately running project from found/lost registration through matching, private claims, clarification, human-authorized pickup and confirmed return.

## Integration
Three custom Town roles use local service discovery and signed in-memory messages; their adapter sends real HTTP requests to an unmodified Lost Link FastAPI service on 127.0.0.1. The human operator is an explicitly scripted fixture with a separate human credential. Application code stays in [cathyliu0510/Lost-Link](https://github.com/cathyliu0510/Lost-Link); this PR contains only the adapter, scenarios, runner, tests, dependency lock, documentation and synthetic evidence.

## Reproduce
Town upstream base: `6f9a7c064486c4a0ae9242f1ff341b8a4aad0244` (0.2.0).
Lost Link: `c26cddbfc9721c12acafa4603f70265c6f37bd11`.
Recorded Python: 3.12.14. Local socket binding is required; no model/API credentials are needed.

```bash
git clone --branch integration/lost-link-lifecycle --single-branch https://github.com/kemengzhang123/nandatown.git
cd nandatown
git clone https://github.com/cathyliu0510/Lost-Link.git ../Lost-Link
git -C ../Lost-Link checkout c26cddbfc9721c12acafa4603f70265c6f37bd11
python3.12 -m venv .venv
.venv/bin/python -m pip install -r examples/lost_link/requirements.lock
.venv/bin/python -m pip install -e '.[dev]'
.venv/bin/python examples/lost_link/run.py --project-root ../Lost-Link --out /tmp/lost-link-town
.venv/bin/python examples/lost_link/run.py --verify examples/lost_link/evidence
LOSTLINK_PROJECT_ROOT=../Lost-Link .venv/bin/python -m pytest -q tests/test_lost_link_integration.py
```

See [the integration README](https://github.com/kemengzhang123/nandatown/blob/beda5c8234f2a217104a17aed6e190f66d829de3/examples/lost_link/README.md) and [recorded summary](https://github.com/kemengzhang123/nandatown/blob/beda5c8234f2a217104a17aed6e190f66d829de3/examples/lost_link/evidence/summary.json). Each scenario includes native Town records, manifest, attestation and a separately read service database snapshot. The runner starts/stops a fresh service, database, credentials and Town home per scenario.

## Results and limitations
| Scenario | Expected | Observed |
| --- | --- | --- |
| Clean lifecycle | Complete return | Passed; 14 HTTP calls, 2 reports/1 case, returned and reports closed |
| Duplicate notification + permission misuse | Idempotent retries; agent cannot approve | Passed; 20 calls, still 2 reports/1 case; approval 403, forged approval field 422; returned |
| Dropped operator-review notification | Workflow remains pending; no pickup release | Incomplete as expected; 5 calls, submitted, no human actions or pickup instructions |

The dropped notification is a documented failure to finish, not a successful return. There is no notification recovery path. Offline replay checks the supplied observations, and negative evaluator tests reject missing human approval or a substituted search target.

Validation: upstream suite **1,865 passed**; new integration tests **4 passed** with the live pinned service; committed evidence replay passed; required `marketplace` and `quote-crash-restart` smoke runs passed; staged `git diff --check` passed.

This demonstrates local Town coordination and real service API/state-machine compatibility. It does not test a deployed NANDA Index, A2A/MCP wire conformance, portable identity grants, LLM behavior, independent human judgment, OS-level credential isolation, concurrent load or physical recovery. All participants and human actions are scripted, and all data are synthetic.

## Course context
MIT AI Studio Fall 2026, HW8 submission.

Try it

Open PR on GitHubView diff

Checkout locally

git fetch origin pull/348/head:pr-348
git checkout pr-348