mirror of
https://github.com/heygen-com/hyperframes.git
synced 2026-11-19 11:26:02 +00:00
An ACP agent reports its models when a session opens, which is the only answer that reflects the user's own account — the same reason Studio asks Codex over its app-server rather than filtering a catalog. The existing rules are unchanged from there: models.dev supplies prices, the harness owns what each model accepts, and cheapest still wins by default. ACP has no separate effort knob. The Codex adapter writes the level into the id (gpt-5.6-sol[high]), so those are folded into one model that offers levels instead of six models that differ by one word, and the chosen level is written back the same way via session/set_model. An agent with plain ids simply has no levels, which is already how Claude Code is treated. Asking now needs the whole harness rather than its kind: two ACP adapters are both "custom" and are two different answers, so the model list, the default model and the default effort all take how the agent is started. Verified against @agentclientprotocol/codex-acp: four models with their real level lists, the account's own default marked, set_model accepted.