Summary

Hermes' zai provider successfully calls Z.AI's GLM models (e.g. glm-5.1 via the coding plan endpoint https://open.bigmodel.cn/api/coding/paas/v4), but reasoning_content is always empty — even on questions that clearly trigger chain-of-thought when the same model is hit directly.

Root cause: Hermes signals "reasoning capable" for open.bigmodel.cn in _supports_reasoning_extra_body (run_agent.py:7501), which causes the chat- completions transport to inject:

extra_body["reasoning"] = {"enabled": True, "effort": "medium"}

But Z.AI's API doesn't recognize the OpenRouter-style reasoning field. It silently ignores it. Z.AI's documented thinking parameter is:

extra_body["thinking"] = {"type": "enabled"}

(same shape as the existing is_kimi branch in chat_completions.py:232-240).

Without thinking set, GLM-5.1 / GLM-4.6 / GLM-4.5 produce regular completions with no reasoning_content deltas in streaming responses, and no message.reasoning_content field in non-streaming.

Reproduction

Direct cURL against the same endpoint, same model, same payload:

curl -sS https://open.bigmodel.cn/api/coding/paas/v4/chat/completions \
  -H "Authorization: Bearer \$GLM_API_KEY" \
  -d '{
    "model": "glm-5.1",
    "messages": [{"role":"user","content":"Xiao Ming has 3 apples at \$2 each and 2 lbs of bananas at \$5/lb. Walk through the math step by step."}],
    "thinking": {"type": "enabled"},
    "stream": true
  }'

→ 542 streamed `reasoning_content` chunks + 118 `content` chunks.

Same payload through Hermes' agent loop with the same model and endpoint: → 0 `reasoning_content` chunks, only `content`.

The only difference between the working cURL and Hermes' wire payload is the `thinking` field. Verified by dumping Hermes' actual `stream_kwargs` right before the SDK call and replaying it via the OpenAI Python SDK 2.32.0 — adding `extra_body={"thinking": {"type": "enabled"}}` to the same payload immediately produces hundreds of `reasoning_content` deltas.

Proposed fix

Mirror the `is_kimi` branch for Z.AI in `chat_completions.py`:

1. In `run_agent.py` (next to `_is_kimi`), compute: ```python _is_zai = base_url_host_matches(self._base_url_lower, "open.bigmodel.cn") ``` Pass `is_zai=_is_zai` to `build_kwargs`.

2. In `chat_completions.py` `build_kwargs`, after the `is_kimi` block: ```python if params.get("is_zai", False): _zai_thinking_enabled = True if reasoning_config and isinstance(reasoning_config, dict): if reasoning_config.get("enabled") is False: _zai_thinking_enabled = False extra_body["thinking"] = { "type": "enabled" if _zai_thinking_enabled else "disabled", } ```

3. Optionally, exclude `open.bigmodel.cn` from `_supports_reasoning_extra_body` so the no-op `reasoning` field is no longer sent (cosmetic — Z.AI ignores it anyway, but it pollutes the wire payload).

This makes `reasoning_config.enabled = False` correctly disable thinking on Z.AI, mirroring the Kimi behavior. Honors the same on/off semantics. Effort levels could be added later — Z.AI's docs don't currently document an effort scalar for the `thinking` parameter.

Environment

• Hermes commit: `f67a61dc` (recent main)
• Provider: `zai`
• Endpoint: `https://open.bigmodel.cn/api/coding/paas/v4\` (coding plan)
• Models tested: `glm-5.1`
• OpenAI SDK: `2.32.0`

References

• Z.AI thinking docs
• GLM-5.1 model card
• Coding plan GLM-5.1 setup