openai-compatible provider: chat.params reasoning effort is dropped for glm-* ids but serialized for gpt-5* ids

openai-compatible provider: chat.params reasoning effort is dropped for glm-* ids but serialized for gpt-5* ids

OpenCode version: 1.18.31

Reproduction

Use a custom OpenAI-compatible provider pointed at a local OpenAI-compatible gateway. Configure a glm-* model with reasoning enabled and declare selectable low, high, and max variants, each with its matching reasoningEffort. A chat.params hook writes output.options.reasoningEffort; send a request via an agent whose message carries variant high, then inspect an allowlisted request-body capture.

Expected: the configured reasoning level serializes as reasoning_effort.

Actual: for glm-* model ids such as glm-5.3-flash, the outgoing body contains no reasoning_effort key at all. With the same options value, gpt-5* control requests serialize reasoning_effort. GPT-5-family handling also has a hardcoded medium default transform, yielding an asymmetric model-family behavior: one configured effort is silently absent while another family gets an implicit effort and serializes effort.

The report concerns serialization behavior only; the provider is a generic local OpenAI-compatible gateway. No credentials or endpoint details are needed to reproduce it.

Please make options reasoning effort propagate consistently for generic OpenAI-compatible models, rather than dropping it for glm-* model ids while relying on GPT-5-specific heuristics and defaults.