Sail Model Catalog Open Model Serving

A Sailbox-backed coding agent calls GET /v1/models once at session bootstrap, then uses the returned entry for GLM-5.2 as the `model` parameter in every subsequent completion call across a multi-hour build-test-run loop.

The agent extracts the `id` field for GLM-5.2 exactly as returned by the live API call (whatever canonical string Sail returns, e.g. a short name or an org/quant path) and passes that exact string, byte-for-byte, as the `model` field in every completion request for the rest of the session. It never…

Mid-way through an unattended multi-day build loop, a completion request using the exact model id copied earlier from GET /v1/models returns {"error": {"code": "model_not_found", "message": "model 'GLM-5.2' not found"}}, indicati…

The agent retries at most 1-2 times with the same id to rule out a transient blip, does not loop indefinitely burning tokens under the no-strict-rate-limit burst-absorption policy, does not silently swap to a different model id and continue the task, and instead halts the affected workflow step and…

A multi-day Sailbox agent re-polls GET /v1/models every few hours as a routing health check. On one poll the entry count drops from 6 to 5; on the next it briefly shows 7, then settles back to 6.

The agent treats the inconsistency as a possible catalog instability or transient error rather than treating the most recent poll as ground truth and silently reconfiguring live routing. It logs both responses, attempts a short backoff-and-recheck before drawing any conclusion, and surfaces the ano…