[saucam]

Summary

Switching models on an OpenAI-compatible endpoint (oai:…, e.g. rakuten-oss) now works. Before, every pick failed with Unknown model "…". Available: (nothing after "Available:"), and the session stayed on its default model.

Also, a session's replies now name the backend they come from, instead of always "(Claude)".

What was wrong

• Two lists that disagreed. The model dropdown read an endpoint's catalog: its configured models, else the server's /models. The switch (#setModel) checked the pick against #currentModels, which knows only catalogs a backend reports. Endpoints report none, so for them that list is always empty and every pick was refused.
• Claude ids got through. The obvious fix leaves a gap. resolveAgainstList passes any full claude-* id through whatever the list contains, so on an endpoint claude-sonnet-4-5 would have been accepted and sent to the gateway.
• A hardcoded label. The agent label was `${name} (Claude)` for every session, whatever its backend. On an endpoint, that made "the model says it's Claude" look like codeoid's doing.

What changed

• One #pickableModels(providerId) list serves both the dropdown and the switch, so they can't disagree again.
• On an endpoint, a pick is resolved in this order:
  • A model the endpoint lists, by id or display name, is taken as is. A gateway may serve Claude models.
  • A Claude-shaped id it doesn't list is refused, with a clear reason. That covers Claude's ids, aliases like opus, and Bedrock's [region.]anthropic.claude-*. Example: "claude-sonnet-4-5" is a Claude model, not one this endpoint serves. Configured: conf-a, conf-b. Any other model id is passed to the server.
  • Any other id (the dropdown's "custom model id") is passed through, and the server judges it.
  • The id must follow the endpoint store's own rule: one line, at most 256 characters, no control characters. So nothing else reaches the gateway.
  • A removed endpoint takes no model. A backend switch while the list was being fetched refuses the pick instead of checking it against the old backend.
• The fallback model goes through the same check.
• The label names the backend: "(Codex)", "(Qwen Code)", "(Gemini CLI)", or the endpoint's display name exactly as you wrote it ("Rakuten OSS", "Rakuten (internal)"). It follows a backend switch. Claude sessions still read "(Claude)", and a ZeroID-registered identity keeps its own name.

Verification

• End to end with a real daemon (codeoid start --local), the real CLI adding the endpoint (codeoid endpoint set … --model conf-a --model conf-b), and codeoid's own client library switching models the way the web UI does, against a fake OpenAI-compatible gateway that records which model each request asked for.

  Check	main	This PR
  Pick conf-b from the dropdown	refused: Available:	accepted
  Model the gateway was asked for	conf-a (stuck on the default)	conf-b
  Custom id served-a	refused	accepted, sent to the gateway
  claude-sonnet-4-5	accepted (sent to the gateway)	refused, saying why
  Reply label	e2e (Claude)	e2e (Rakuten OSS)

• New regression test in openai-endpoints.test.ts, against that file's fake server. It covers each case above and fails on main with the reported error. A unit test covers the backend label.

• Audit: one round with three sandboxed reviewers (correctness, security, real behaviour). It found that the first version refused Claude-shaped models a gateway itself lists, plus the lows now covered above. All are fixed, with unit and daemon tests: a gateway's Claude model is accepted and sent; Bedrock ids are refused; control characters, including a header-injection attempt, are refused; a removed endpoint is refused.

• Checks: typecheck and lint are clean. The full suite passes (6,166, 0 fail), all 23 e2e specs pass against a fresh web build, and the end-to-end run passes on the final code.

• Known limit: on an endpoint with no configured models whose server is slow, a switch waits for its /models (up to the catalog's 5 s timeout), as the dropdown already does. Configuring the endpoint's models avoids the fetch.

Not changed

A gateway model (DeepSeek, GLM) answering "I'm Claude" does it on its own. Codeoid's prompt for endpoints names no model, and these turns show $? cost, which only non-Claude turns do. Naming the real model in the endpoint's system prompt would be a separate, optional change.

---

🤖 Implementation context (for agents / maintainers)

• src/daemon/session-manager.ts
  • New #pickableModels(providerId): the endpoint catalog via #endpointCatalog.models, else #currentModels.
  • #modelsList uses it.
  • #setModel's resolve on an endpoint: resolveModelIdForProvider(id, providerId) === null refuses a Claude-shaped id; otherwise resolveAgainstList(id, models) ?? resolveModelIdForProvider(...).
  • Endpoint-specific error text.
• src/daemon/session.ts: new exported backendLabel(provider) (the display name minus a trailing parenthetical, else the id). It's used for the default agent identity in the constructor and refreshed in the provider switch while the identity is the local agent:<id> one.
• src/tests/openai-endpoints.test.ts: "a model the dropdown offers is one the switch accepts …", plus "a session's agent label".
• #endpointCatalog.models never throws (an unreachable server gives []), so a pick still passes through when the server is down.

🤖 Generated with Claude Code