fix(responses): generalize tool call parsing for reasoning models

[sapountzis]

Description

Fixes a bug in parse_responses_tools method that caused failures when using reasoning models where completion.output[0] may not exist or may not have the expected name attribute.

Problem

The previous implementation made unsafe assumptions about the response structure:

1. Conditional complexity: Used using_inbuilt_tools parameter to handle different response types with branching logic
2. Unsafe indexing: Directly accessed completion.output[0] without checking if it exists or has the expected properties
3. Reasoning model incompatibility: Reasoning models can return responses where the first output element isn't the expected tool call

Solution

• Unified approach: Removed the using_inbuilt_tools parameter and conditional logic
• Safe iteration: Now iterates through all messages in completion.output to find the matching tool call
• Robust error handling: Properly handles cases where no matching tool call is found

Changes

• Removed using_inbuilt_tools parameter from parse_responses_tools method signature
• Replaced unsafe completion.output[0] access with defensive iteration
• Simplified the logic by removing the conditional branch for inbuilt tools
• Updated error handling to be more explicit about missing tool calls

---

Important

Fixes parse_responses_tools to safely handle reasoning models by iterating over completion.output and removing unsafe assumptions.

• Behavior:
  • Fixes bug in parse_responses_tools to handle reasoning models where completion.output[0] may not exist or lack name attribute.
  • Iterates through completion.output to find matching tool call, removing unsafe direct access.
  • Handles cases where no matching tool call is found with explicit error.
• Code Changes:
  • Removes using_inbuilt_tools parameter from parse_responses_tools.
  • Simplifies logic by removing conditional branches for inbuilt tools.
  • Updates error handling for missing tool calls.

^{This description was created by for 9eca15c. You can customize this summary. It will automatically update as commits are pushed.}

[ellipsis-dev]

Important

Looks good to me! 👍

Reviewed everything up to 9eca15c in 52 seconds. Click for details.

• Reviewed 53 lines of code in 1 files
• Skipped 0 files when reviewing.
• Skipped posting 3 draft comments. View those below.
• Modify your settings and rules to customize what types of comments Ellipsis leaves. And don't forget to react with 👍 or 👎 to teach Ellipsis.

1. instructor/processing/function_calls.py:556

• Draft comment:
  Removed the unused 'using_inbuilt_tools' parameter from the function signature. This simplification cleans up the interface and removes redundant branching logic.
• Reason this comment was not posted:
  Confidence changes required: 0% <= threshold 85% None

2. instructor/processing/function_calls.py:561

• Draft comment:
  The loop now iterates safely over 'completion.output' to find the matching tool call. Consider using a for-else construct to make the intent even clearer.
• Reason this comment was not posted:
  Confidence changes required: 33% <= threshold 85% None

3. instructor/processing/function_calls.py:567

• Draft comment:
  Enhanced error handling is now in place if no matching tool call is found. It might be beneficial to ensure that 'completion.output' is always a list to avoid unexpected errors.
• Reason this comment was not posted:
  Confidence changes required: 33% <= threshold 85% None

Workflow ID: wflow_xoFCSICcoQAyXZAS

^{You can customize by changing your verbosity settings, reacting with 👍 or 👎, replying to comments, or adding code review rules.}

[jxnl]

@claude please review this

[claude]

Claude encountered an error —— View job

Failed with exit code 128

---

I'll analyze this and get back to you.

[sapountzis]

@jxnl are the tests and claude expected to succeed? My changes are not related to the errors

[jxnl]

I just approved workflows to run. Since they cost credits, I try not to make every pull request run everything.