Skip to content

fix: disable sampling params for Claude 4.7+ models#781

Open
tyaroshko wants to merge 2 commits into
mainfrom
fix/disable-sampling-params-unsupported-models
Open

fix: disable sampling params for Claude 4.7+ models#781
tyaroshko wants to merge 2 commits into
mainfrom
fix/disable-sampling-params-unsupported-models

Conversation

@tyaroshko

@tyaroshko tyaroshko commented Jun 12, 2026

Copy link
Copy Markdown
Collaborator

Note

Medium Risk
Changes the shared LLM completion param path for all providers; behavior is limited to omitting sampling params and one retry, but workflows that relied on temperature on listed models will run with provider defaults instead.

Overview
Fixes 400 errors when calling newer Claude models (Opus 4.7+, Fable 5, Mythos, etc.) that reject temperature, top_p, and top_k.

BaseLLM now strips those sampling params before the request when the model id matches SAMPLING_UNSUPPORTED_MODEL_MARKERS, across Anthropic, Bedrock, and Custom LLM paths. If a model isn’t in the list yet, _recover_completion_params retries once without sampling when the error looks like “param not permitted/unsupported,” clears the values on the node for later calls, and leaves real validation failures (e.g. out-of-range temperature) unchanged.

Adds unit tests for marker detection, proactive stripping, and reactive recovery.

Reviewed by Cursor Bugbot for commit 92f9c07. Bugbot is set up for automated code reviews on this repo. Configure here.

@tyaroshko tyaroshko requested a review from a team as a code owner June 12, 2026 08:45

@cursor cursor Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Cursor Bugbot has reviewed your changes and found 1 potential issue.

Fix All in Cursor

❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.

Reviewed by Cursor Bugbot for commit 92f9c07. Configure here.

setattr(self, param, None)
elif self.__pydantic_extra__ is not None:
self.__pydantic_extra__.pop(param, None)
return recovered

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Recovery mutates before retry succeeds

Medium Severity

_recover_completion_params clears sampling fields on the LLM instance before the retried _completion/_acompletion runs. If that retry still raises, the call fails but temperature, top_p, and related extras remain cleared on the reusable node.

Fix in Cursor Fix in Web

Reviewed by Cursor Bugbot for commit 92f9c07. Configure here.

@github-actions

Copy link
Copy Markdown

Coverage

Coverage Report •
FileStmtsMissCoverMissing
dynamiq/nodes/llms
   base.py4707085%34, 36, 153, 159, 372, 418–421, 430–436, 463, 509, 552–553, 557, 655–664, 691–692, 725–735, 741–745, 750–756, 767–768, 778, 816, 818, 852, 854, 946, 1009, 1108, 1184, 1187, 1218–1219
TOTAL33327970170% 

Tests Skipped Failures Errors Time
2466 2 💤 0 ❌ 0 🔥 2m 52s ⏱️

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant