fix: add loop detection for doom loop models (GLM-4.6, Grok Code) - fix #3444 #3445
+33
−0
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
ElecTwix
changed the title
Comment on lines
+917
to
+918
| "glm-4.6": `⚠️ Loop detected! You have been calling "%TOOL%" with the same parameters repeatedly. This is causing an infinite loop. Please STOP calling this tool and try a different approach or modify the parameters.`, | ||
| "grok-code": `⚠️ Loop detected! You're stuck repeating "%TOOL%" with identical inputs. Break this pattern and try a different strategy or change the parameters.`, |
There was a problem hiding this comment.
should it be model specific? Almost seems like it should be model agnostic
There was a problem hiding this comment.
Made it model-specific since only GLM-4.6 and Grok Code are known to have this issue. Other models work fine and we don't want to add overhead for them. Could make it agnostic later if more models show this behavior.
Comment on lines
+1066
to
+1071
| const key = `${value.toolName}|${JSON.stringify(value.input)}` | ||
| const count = (loopCounts.get(key) || 0) + 1 | ||
| loopCounts.set(key, count) | ||
|
|
||
| if (count >= LOOP_DETECTION_THRESHOLD) { | ||
| await Session.updatePart({ |
There was a problem hiding this comment.
Is this how everyone else does loop detection too?
There was a problem hiding this comment.
I don't know if this is how everyone else does it, but this approach should work. Tool+input tracking with thresholds using simple Map-based counting with O(1) performance keeps it lightweight and effective.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
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.
fix #3444