Malformed thinking block in toolcall (glm-4.7-free) #6039

Activity

1. kevingatera

   added

   bugSomething isn't workingSomething isn't working

   on Dec 23, 2025
2. github-actions

   assigned

   fwang

   on Dec 23, 2025
3. github-actions

   added

   zenRelates to ZenRelates to Zen

   on Dec 23, 2025
4. github-actions commented on Dec 23, 2025

   

   github-actionsBot

   on Dec 23, 2025 – with GitHub Actions

   Contributor

   More actions

   This issue might be a duplicate of existing issues. Please check:

   • SSE Stream Bug: Out-of-Order thinking_delta via LiteLLM → AWS Bedrock #3596: SSE Stream Bug: Out-of-Order thinking_delta via LiteLLM → AWS Bedrock - similar malformed thinking block issues in streaming
   • Expected thinking or redacted_thinking, but found tool_use #3077: Expected thinking or redacted_thinking, but found tool_use - related thinking block ordering problems
   • Loop in certain models #3743: Loop in certain models - GLM models occasionally showing tool call malfunction issues
   • GLM-4.6 model gets stuck in infinite loop, repeating same actions #3444: GLM-4.6 model gets stuck in infinite loop - similar behavior with GLM models getting stuck on tool calls

   Feel free to ignore if your specific case with glm-4.7-free differs from these issues.
5. ramarivera commented on Dec 23, 2025

   

   ramarivera

   on Dec 23, 2025

   More actions

   +1ing this one, the "it seems to have stopped completely" thing happens both on Zen GLM 4.7 and Z.AI coding plan's
6. rekram1-node commented on Dec 23, 2025

   

   rekram1-node

   on Dec 23, 2025

   Contributor

   More actions

   glm-4.7-free was incorrectly not configured as an interleaved reasoning model.

   This should be fixed if you do opencode models --refresh

   Your cache will automatically refresh but this just forces it to so that you can have the fix sooner
7. rekram1-node

   closed this as completedon Dec 23, 2025
8. rekram1-node commented on Dec 23, 2025

   

   rekram1-node

   on Dec 23, 2025

   Contributor

   More actions

   fix
9. ramarivera commented on Dec 23, 2025

   

   ramarivera

   on Dec 23, 2025

   More actions

   oh this is really sad for me :(
   I was working on a PR to to detect and extract the tool call from inside the thinking block cuz I assumed it was a GLM issue that required normalization in packages/opencode/src/provider/transform.ts

   Guess I should have asked first 😓
10. ramarivera commented on Dec 24, 2025

    

    ramarivera

    on Dec 24, 2025

    Last edited by ramarivera

    More actions

    so, even with this fix, I still get tool calls in thinking blocks which causes the opencode workflow to stop e.g.

    Thinking: <tool_call>bash<arg_key>command</arg_key><arg_value>cat packages/portal/.generation-cache.json | jq '."volume", "artifact", "generatedAt", "tokens" | select(.volume == "openspec") | .artifact | .generatedAt | .tokens' 2>/dev/null</arg_value><arg_key>description</arg_key><arg_value>Check cache for openspec artifacts</arg_value></tool_call>
    

    EDIT: more examples

    Thinking: <tool_call>use_skill<arg_key>skill</arg_key><arg_value>handoff</arg_value></tool_call>
    

    👍React with 👍2Reacted by Danila Poyarkov and zizineco
11. crystech commented on Dec 25, 2025

    

    crystech

    on Dec 25, 2025

    More actions

    so, even with this fix, I still get tool calls in thinking blocks which causes the opencode workflow to stop e.g.

    Thinking: <tool_call>bash<arg_key>command</arg_key><arg_value>cat packages/portal/.generation-cache.json | jq '."volume", "artifact", "generatedAt", "tokens" | select(.volume == "openspec") | .artifact | .generatedAt | .tokens' 2>/dev/null</arg_value><arg_key>description</arg_key><arg_value>Check cache for openspec artifacts</arg_value></tool_call>
    

    EDIT: more examples

    Thinking: <tool_call>use_skill<arg_key>skill</arg_key><arg_value>handoff</arg_value></tool_call>
    

    +1 here. same issues on GLM 4.7 local llm
12. rekram1-node commented on Dec 25, 2025

    

    rekram1-node

    on Dec 25, 2025

    Contributor

    More actions

    if u are still running into it it is entirely dependent on ir provider, if opencode is properly sending thinking chunks (the interleaved setting for model is correct) then it is on the provider of the model to properly parse the tool calls out and send them as such
13. lkonga

    added a commit that references this issue on Jan 2, 2026

    fix: configure GLM-4.7 as interleaved reasoning model

    ...

    c02eea8