Sibling of the max_tokens defect found while fixing it. The same resolved ModelInfo carries a Temperature flag which, like Reasoning, is populated by both the catalog and config paths and consumed nowhere, so the OpenAI provider sends temperature whenever the caller sets one. GPT-5 and o-series models reject a non-default temperature with HTTP 400 in exactly the way they reject max_tokens, so the same class of failure is still live for any caller that sets a temperature. Fix it the same way: carry the resolved flag on the chat request and let the provider omit or transform the parameter for a model class that rejects it, rather than adding a model-name lookup table in the provider. Settling evidence for the claim that it is unreachable: grep the flag's consumers in ai-sdk; for the 400, reproduce with a live call against the vendor or cite the vendor's documented behaviour.
Acceptance criteria
a reasoning-class model with a caller-supplied temperature is not sent a parameter the model rejects, a classic model still receives temperature, and a test pins both wire bodies
Close reason
Moved to ai-sdk; tracked and fixed there.
Imported from beads archie-core-o00p · GitHub #1336 · type bug · priority P2 · created 2026-10-02 · closed 2026-10-05
Sibling of the max_tokens defect found while fixing it. The same resolved ModelInfo carries a Temperature flag which, like Reasoning, is populated by both the catalog and config paths and consumed nowhere, so the OpenAI provider sends temperature whenever the caller sets one. GPT-5 and o-series models reject a non-default temperature with HTTP 400 in exactly the way they reject max_tokens, so the same class of failure is still live for any caller that sets a temperature. Fix it the same way: carry the resolved flag on the chat request and let the provider omit or transform the parameter for a model class that rejects it, rather than adding a model-name lookup table in the provider. Settling evidence for the claim that it is unreachable: grep the flag's consumers in ai-sdk; for the 400, reproduce with a live call against the vendor or cite the vendor's documented behaviour.
### Acceptance criteria
a reasoning-class model with a caller-supplied temperature is not sent a parameter the model rejects, a classic model still receives temperature, and a test pins both wire bodies
### Close reason
Moved to ai-sdk; tracked and fixed there.
---
_Imported from beads `archie-core-o00p` · GitHub #1336 · type bug · priority P2 · created 2026-10-02 · closed 2026-10-05_
sam
added the mediumbug labels 2026-10-27 13:40:33 +00:00
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Sibling of the max_tokens defect found while fixing it. The same resolved ModelInfo carries a Temperature flag which, like Reasoning, is populated by both the catalog and config paths and consumed nowhere, so the OpenAI provider sends temperature whenever the caller sets one. GPT-5 and o-series models reject a non-default temperature with HTTP 400 in exactly the way they reject max_tokens, so the same class of failure is still live for any caller that sets a temperature. Fix it the same way: carry the resolved flag on the chat request and let the provider omit or transform the parameter for a model class that rejects it, rather than adding a model-name lookup table in the provider. Settling evidence for the claim that it is unreachable: grep the flag's consumers in ai-sdk; for the 400, reproduce with a live call against the vendor or cite the vendor's documented behaviour.
Acceptance criteria
a reasoning-class model with a caller-supplied temperature is not sent a parameter the model rejects, a classic model still receives temperature, and a test pins both wire bodies
Close reason
Moved to ai-sdk; tracked and fixed there.
Imported from beads
archie-core-o00p· GitHub #1336 · type bug · priority P2 · created 2026-10-02 · closed 2026-10-05