fix(litellm): avoid duplicating content and signed thinking blocks across parallel tool-call splits#3215
Merged
seratch merged 2 commits intoopenai:mainfrom May 9, 2026
Conversation
…en splitting parallel tool calls When `_fix_tool_message_ordering` splits an assistant message that has multiple `tool_calls` into one message per tool call, it copied the entire parent message into each split. That duplicated `content`, `thinking_blocks`, and `reasoning_content` across every split. For Anthropic models with extended thinking, this is fatal: each thinking block carries a unique signature, and the API rejects requests that include the same signed block twice. It also corrupts conversation history by repeating the assistant's user-visible text. Keep those shared fields only on the first split; subsequent splits carry their `tool_calls` alone. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
3 tasks
Member
|
@codex review |
|
Codex Review: Didn't find any major issues. Delightful! ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
Construct the multi-tool assistant fixture via cast so mypy stops complaining that thinking_blocks / reasoning_content are not part of ChatCompletionAssistantMessageParam, and that the resulting list-item type doesn't match the union. Also cast the filtered assistants list to dict[str, Any] so subscript and "in"/"not in" checks no longer need per-line type ignores. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
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
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.
Summary
When
LitellmModel._fix_tool_message_orderingsplits an assistant message that has multipletool_callsinto one message per tool call, it copies the entire parent message into each split. That duplicatescontent,thinking_blocks, andreasoning_contentacross every split.For Anthropic models with extended thinking, this is fatal: each thinking block carries a unique signature, and the API rejects requests that include the same signed block twice. It also corrupts conversation history by repeating the assistant's user-visible text whenever the model emits two or more parallel tool calls.
Keep those shared fields only on the first split; subsequent splits carry their
tool_callsalone.Repro
Pass an assistant message with two parallel tool calls plus content/thinking blocks (which is what LiteLLM emits for Anthropic Claude models with parallel tool use + extended thinking) through
_fix_tool_message_ordering. Before the fix, both split assistant messages contain the samecontenttext and the same signedthinking_blocks; after the fix, only the first split carries them.Test plan
test_split_does_not_duplicate_content_or_thinkingcovers the regression.test_extended_thinking_message_order.pytests still pass.ruff checkclean on both files.🤖 Generated with Claude Code