Writes, rereads, and environment state

Input or phase

Behavior

model_ids omitted/null/[]/whitespace only

Empty after normalization; configuration is cleared only if the default model is also empty.

Writing an empty configuration

extra_json is ultimately cleared, but any nonempty JSON text supplied is validated first.

Only models supplied

The handler can decode it but does not use it as a write command; an empty model_ids value may clear the configuration.

extra_json

The field type is string and its contents must be valid JSON, not necessarily a JSON object.

Empty body/null or unknown fields

The ordinary decoder does not guarantee strict rejection and may produce an empty configuration. Do not use this as a probe request.

The request is a full PUT replacement, without expected_version or a dedicated rollback field. Before writing, the deployment configuration owner must inspect the configuration source and save a complete backup, including endpoint, which is absent from query responses. The current codec allows this exclusive endpoint only for agent_chat; it includes the address, credentials, and timeout. Neither the query nor write DTO exposes it. Do not include secrets in evidence or documentation.

Even a nonempty model update can remove that endpoint, returning agent_chat to the catalog pool route. Ordinary list responses and operation logs cannot restore the complete routing configuration. Repeating PUT with the original model_ids/extra_json cannot restore endpoint either.

After writing, reread the same slot_id and compare the model order, defaults, and extra configuration. Have the configuration owner confirm persisted values and relevant instance snapshots. HTTP 200 or the presence of a log alone does not establish actual runtime behavior.

The configuration owner persists first, then reloads the current process snapshot; the service subsequently writes the log in a separate system transaction. The old comment claiming a shared transaction does not match the current call order. Reload or subsequent log failures can return an error after persistence, so rollback and joint success of logging and configuration are not guaranteed. A failed reload can leave list responses on the old snapshot; unchanged reread results do not prove that persistence did not occur. Ask the configuration owner to check the source rather than retrying automatically. Activation timing on other instances and global concurrency require verification in the actual deployment.

Operation logs do not contain complete before-and-after values. Recovery requires a full backup of the configuration source and its owner; this API’s response alone is insufficient. There is no idempotency or expected-version contract for the request. Check for other administrators’ changes before recovery, and do not overwrite unknown concurrent writes. Clearing real feature slots such as agent_chat affects usage. An invented, unregistered slot_id cannot isolate an experiment.

The user has been asked to update local credentials.

Ten authentication 401 responses and one earlier public-version 200 have been recorded. The earlier self-reported commit differed from the source; the current deployment has not been reverified.

Last updated on