[BUG] GLM-4.6 not generating thinking tokens when using OpenAI-compatible custom endpoint

[bizprat]

Problem (one or two sentences)

When using GLM-4.6 through an OpenAI-compatible custom endpoint in Roo Code, the model does not generate thinking tokens. The requests appear to be missing the required "thinking": {"type": "enabled"} parameter that GLM-4.6 needs to activate its reasoning mode.

Context (who is affected and when)

This affects users who are using GLM-4.6 with custom OpenAI-compatible endpoints (not Z.ai or OpenRouter providers) in Roo Code latest versions (3.28.12-3.28.15). The issue occurs consistently with every request since GLM-4.6's recent release (September 2025).

Reproduction steps

1. Configure Roo Code to use an OpenAI-compatible custom endpoint

2. Set the model to GLM-4.6 (glm-4.6)

3. Send any coding request to the model

4. Observe the response from GLM-4.6

5. Notice that no thinking tokens or reasoning blocks are included in the output

Expected result

GLM-4.6 should generate thinking tokens showing its step-by-step reasoning process before providing the final answer, similar to how it works with Z.ai native provider or OpenRouter. The thinking blocks should be visible in the Roo Code interface showing the model's internal reasoning.

Actual result

GLM-4.6 responds immediately without any thinking tokens or reasoning blocks. The model provides direct answers without showing the intermediate reasoning steps, even though GLM-4.6 is designed as a thinking model.

Variations tried (optional)

No response

App Version

3.28.15

API Provider (optional)

OpenAI Compatible

Model Used (optional)

GLM 4.6

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote]

I'm exploring a potential fix for the GLM-4.6 thinking tokens issue with OpenAI-compatible endpoints. I'll investigate adding the required 'thinking' parameter to enable reasoning mode and will open a PR to share progress.

[roomote]

Opened PR #8548 with my attempt at addressing this issue: #8548

The implementation adds the required thinking: { type: 'enabled' } parameter for GLM-4.6 models when using OpenAI-compatible endpoints. All CI checks have passed ✅

The fix includes:

• Detection for GLM-4.6 model variants
• Automatic inclusion of the thinking parameter in requests
• Parsing of thinking tokens from both XML tags and reasoning_content fields
• Comprehensive test coverage

Feedback welcome!

[ChicoPinto70]

Hi, Guys. I tried the PR #8548 with ik_llama.cpp running locally the Ubergarm's GLM-4.6 model with Unsloth's chat template (--jinja --chat-template-file) but I still don't have reasoning in the Roo Code. The reasoning works fine in Roo Code with DeepSeek v3.1 running locally and I also get GLM-4.6 reasoning in the ik_llama.cpp built-in webui and in the continue vscode extension. Do I need to do something else in order to make it works with this PR?

I'm using the openai compatible endpoint in Roo Code with "glm-4.6" alias in the ik_llama.cpp

[kavehsfv]

@roomote Hi
when this feature will be released?

[roomote]

@kavehsfv I have seen your comment about the release timeline. PR #8548 is currently awaiting maintainer review. All CI checks have passed, but there is a compatibility issue reported by ChicoPinto70 with ik_llama.cpp that I will investigate to ensure the fix works across different setups.

I will look into the ik_llama.cpp compatibility and update the PR accordingly. Once the issues are resolved and the PR is approved, it should be included in the next release cycle.