POOLSIDE_STANDALONE_MODEL not working

[creepscope]

pool --version = 1.0.16
trying to add a OpenAI-Compatible endpoint which is hosting laguna-s-2-1.
Terminal Command I run:

POOLSIDE_STANDALONE_BASE_URL="https://openai.###.###/v1" \
POOLSIDE_API_KEY="sk-############" \
POOLSIDE_STANDALONE_MODEL="poolside/laguna-s-2-1" \
POOLSIDE_STANDALONE_CONTEXT_LENGTH=200000 \
pool

It just opens with the BASE_URL above listed as the agent server but then model is listed as * and I get the error:

• Worked for 1s
• Error during ACP method session/prompt
  ⎿ 400 Bad Request: /chat/completions: Invalid model name passed in model=*. Call `/v1/models` to view available models for your key.
• Run /logs to save debug logs, or /feedback to report a bug on GitHub.

I have also tried passing the model with "pool -m laguna-s-2-1" but it just says model not found in the pool client.

[agentcooper]

Does your OpenAI API compatible endpoint have /v1/models? What does it reply with?

Also you can try not passing POOLSIDE_STANDALONE_MODEL when starting pool and then doing /models to open the model selector.

[creepscope]

the models endpoint returns an asterisk:

curl https://openai.#####.####/v1/models -H "Authorization: Bearer sk-#########"
{"data":[{"id":"*","object":"model","created":12345678,"owned_by":"openai"}],"object":"list"}%   

I have been under the assumption the models listing endpoint doesn't need to work as long as we can pass the exact model ID in the request.

"So if you already know the model ID (e.g. claude-opus-5), just hardcode or configure it and call messages.create(model="claude-opus-5", ...) directly — no need for /v1/models to be reachable, working, or called first.

[creepscope]

We turned on the /v1/models endpoint and now everything it lists appears in the picker (both Desktop and pool cli), so I can no longer really test the "single preconfigured model" issue.

[IgorGanapolsky]

Reproduced the same class of failure on pool 1.0.15 (macOS, darwin/arm64).

The public README currently documents POOLSIDE_STANDALONE_MODEL as sufficient for any OpenAI-compatible endpoint. In practice, if GET /v1/models returns a wildcard id (*) — which some gateways do — pool still opens with model * and the first session/prompt 400s (Invalid model name passed in model=*).

That matches the 1.0.16 report here: enabling a real /v1/models list made the picker work, which is “env override is not applied unless the id appears in the models list.”

Suggested contract:

1. If POOLSIDE_STANDALONE_MODEL is set, send that id even when /v1/models is *, empty, or unreachable.
2. Until then, document the current rule: POOLSIDE_STANDALONE_MODEL is ignored unless the id appears in GET /v1/models.

pool source is not in this repo, so this stays a contract request rather than a code PR. Happy to re-test a documented override from 1.0.15. Not affiliated.