This fixes the issue where conversation histories with extended thinking could have tool_result messages appearing before their corresponding tool_use messages, causing Anthropic API rejections.

What was happening:

Reasoning blocks between tool calls caused the converter to create malformed message sequences
Multiple consecutive assistant messages with tool calls weren't getting paired with their results properly
Anthropic API requires strict tool_use → tool_result ordering

The fix:

Added_fix_tool_message_ordering() method that only activates whenpreserve_thinking_blocks=True
Analyzes the entire conversation to identify tool call/result pairs
Reconstructs messages ensuring proper tool_use → tool_result sequencing
Handles edge cases like consecutive tool calls and misplaced results
Non-breaking - only affects extended thinking scenarios

Repro script available in the original issue.

fix: reorder tool messages to resolve interleaved thinking bug

d81891f

chatgpt-codex-connectorbot reviewed

Sep 24, 2025

View reviewed changes

Copy link

chatgpt-codex-connectorbot left a comment

Choose a reason for hiding this comment

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

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

src/agents/extensions/models/litellm_model.py

Comment on lines 482 to 485

		elifrole=="tool":
		# Skip - these will be handled as part of tool pairs above
		used_indices.add(i)

Copy link

chatgpt-codex-connectorbotSep 24, 2025

Choose a reason for hiding this comment

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

Preserve unmatched tool_result messages

In_fix_tool_message_ordering() therole == "tool" branch simply marks the index as used and skips emitting the message. When a tool message has no matchingtool_call_id or the corresponding tool call is absent from the trimmed history, this logic drops the tool result entirely. Previously those messages were forwarded to the provider; now they silently disappear, which removes tool outputs from the conversation and can change agent behaviour. Consider emitting unmatched tool messages (or raising) rather than discarding them.

Useful? React with 👍 / 👎.