fix(responses): only dedup user messages when function_call_outputs are present

[derekhiggins]

Fixes #6018

Summary

• Restrict user message deduplication to only apply when function_call_output items are present in the input
• Prevents dropping the new user message when using previous_response_id with repeated content

Test plan

• Unit test covering the dedup-with-function-call-outputs path
• Integration test with Anthropic provider and previous_response_id

---

[devin-ai-integration]

Devin Review found 1 potential issue.

View 2 additional findings in Devin Review.

[devin-ai-integration]

🟡 Duplicate-user-message check uses mutable tool_call_results dict that may already be emptied by preceding function-call processing

The new condition and tool_call_results on line 383 checks whether the tool_call_results dict is non-empty to determine if function_call_outputs are present. However, tool_call_results is mutated during the same loop — entries are deleted at utils.py:332 (del tool_call_results[input_item.call_id]) each time a matching OpenAIResponseOutputMessageFunctionToolCall is processed. If all function_call_outputs in the input have matching function_calls in the same input, and those function_calls appear before the duplicate user message, tool_call_results will be empty by the time the user message is reached, causing the dedup logic to be skipped.

Example scenario where the bug manifests

Input with previous_messages provided:

input = [
  OpenAIResponseOutputMessageFunctionToolCall(call_id="abc", ...),  # consumes tool_call_results["abc"]
  OpenAIResponseInputFunctionToolCallOutput(call_id="abc", ...),    # pre-extracted into tool_call_results
  OpenAIResponseMessage(role="user", content=...) ,                 # duplicate user message
]

After processing index 0, tool_call_results is empty. At index 2, tool_call_results is falsy, so the dedup condition fails and the duplicate user message is included.

The fix should capture whether function_call_outputs were originally present before the main loop mutates the dict, e.g. has_function_call_outputs = bool(tool_call_results) right after line 306, and use that flag instead.

Prompt for agents

The condition `and tool_call_results` on line 383 uses the mutable dict `tool_call_results` as a boolean indicator of whether function_call_outputs were present in the input. However, this dict is depleted during the main loop as matching function_calls consume entries (line 332: `del tool_call_results[input_item.call_id]`). If all entries are consumed before a duplicate user message is encountered, the dedup logic is incorrectly skipped.

Fix: After the pre-extraction loop (after line 306), capture a boolean flag:
has_function_call_outputs = bool(tool_call_results)
Then on line 383, replace `and tool_call_results` with `and has_function_call_outputs`.

This ensures the dedup check is based on whether function_call_outputs existed in the original input, not on whether they have been consumed during processing.

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[github-actions]

✅ Recordings committed successfully

Recordings from the integration tests have been committed to this PR.

View commit workflow