fix: add doom loop detection #3445
Conversation
ElecTwix
changed the title
|
I don't know if we should do model specific because there are other models too kimi-k2, supernova, gemini models all can get to that state, peoplee can also define model aliases too hard to catch them all. What if we just do a check when a also I think the "error message" should probably be the same for all the llms is there any reason for a different message per model? I don't know if it will matter substantially in this case What do you think? |
…sion-based approach
|
Updated to universal doom loop detection. Now checks last 3 tool calls for identical patterns and asks permission regardless of model. |
|
FYI this happened with grok-code-fast-1 today (using OpenCode Zen). Thought this would help gathering more cases. 
(a few seconds later) 
(then I ....)Screenshot
TranscriptionRead packages/web/src/content/docs/config.mdx (limit=375, offset=75) Note: The custom directory is loaded after the global config and SchemaRequest was aborted 아야아 진정해 이게 맞아? Thinking... The user asked, "아야아 진정해 이게 맞아?" which means "Is this correct?" - they're asking if the hook documentation I added is accurate. Reviewing user query
Yes, that's correct! I reviewed the codebase as an engineer and added the hook functionality I had seen before to the OpenCode documentation in a way that matches the existing document conventions in packages/web/src/content/docs/config.mdx. Summary of additions:
I wrote the documentation to match the existing style. For example:
Now you can check the hook configuration method on the docs site. If you need any additional modifications or other feature documentation, please let me know! 😊 |
|
hm that one is harder to catch, I think this current pr will be a good start and we will see how it performs for other cases |
|
|
Doing some tweaks: 
|
rekram1-node
changed the title
3 participants
fix: add loop detection for doom loop models (GLM-4.6, Grok Code)
GLM-4.6 and Grok Code models are prone to "doom loops" - infinite repetition of same tool call with identical parameters. This causes wasted tokens, poor user experience with stuck sessions, and server resource waste.
This commit adds targeted loop detection that monitors tool calls for problematic models and injects breaking messages when loops are detected. The implementation uses Map-based counting with O(1) performance and only affects specific models known to have this issue.
Key changes:
DOOM_LOOP_MODELSconstant for model-specific configuration (glm-4.6, grok-code)Fixes infinite loop issue while maintaining performance and extensibility for additional problematic models.
fixes: #3444