run -m with an unqualified model id fails after session start with "Unexpected server error" #46991

What happens

opencode run -m <model-id> with an id that carries no provider/ prefix creates a session and then fails on the first turn with:

UnknownError: Unexpected server error

Nothing is written to the server log, --print-logs adds nothing, and the exit status is the same one a genuine provider error produces. The same invocation with -m <provider>/<model-id> works.

Why this is hard to diagnose

The failure happens after session creation, so it does not read as an argument problem. In a headless run driving several agents, the visible symptom is that the top-level process works and every child dies with an opaque error, which sends you looking at the provider, the model, or concurrency rather than at the flag.

There is a second path into it that is easy to hit: a config file can carry a model id per role. If the top-level -m is qualified but a role's id is bare, the run starts, the first agent answers, and every subsequent agent dies. Nothing in the output distinguishes that from a provider outage.

Reproduce

1. Configure any OpenAI-compatible provider in opencode.jsonc, for example a provider named acme with one model my-model.
2. opencode run -m my-model --format json 'say ok' → session is created, then UnknownError: Unexpected server error.
3. opencode run -m acme/my-model --format json 'say ok' → works.

Version: 1.18.27, headless, Linux container.

Suggested fix

Resolve -m before the session is created and fail with a message naming the problem, e.g. model id "my-model" is not qualified; use "<provider>/my-model" and list the configured providers. Resolving an unqualified id against a single configured provider would also be reasonable, but the explicit error is the part that matters — the current behaviour spends a whole run to say nothing.