Start here: reproduce a lost response, then help choose the first real provider

[ArtAndrew]

Your agent sent a request. The provider committed it. The response disappeared. Should the agent retry?

CommitLatch is an experimental Python SDK that records intent before dispatch, reconciles authoritative receipts and blocks a repeated dispatch when the outcome remains UNKNOWN.

Try the failure

git clone https://github.com/ArtAndrew/commitlatch.git
cd commitlatch
PYTHONPATH=src python3 -m commitlatch http-demo

No API keys or model calls. The loopback HTTP fixture closes the connection after committing to its own database.

Strategy	Effects	POSTs	Outcome
Blind retry	2	2	Duplicate
Native idempotency key	1	2	Success
CommitLatch + matching receipt	1	1	Success
CommitLatch + unavailable lookup	1	1	UNKNOWN; repeat blocked

Native idempotency works when its contract covers the operation. Keep using it. CommitLatch adds a durable intent record and an explicit recovery decision; it does not manufacture evidence a provider cannot supply.

Methodology and limits · LangGraph example · When to use it

Help shape the next adapter

Please reply with:

1. The provider and state-changing operation you need.
2. Public documentation for its idempotency keys or receipt lookup.
3. A synthetic failure scenario you want reproduced.
4. What happened when you ran the demo, including Python/OS if it failed.

Counterexamples to the recovery contract are especially welcome. Use synthetic data only. This alpha has no production provider adapter and makes no universal exactly-once guarantee.