fix(streaming): guard against IndexError in Responses API stream accumulator

[MaxwellCalkin]

Summary

Fixes IndexError: list index out of range that occurs intermittently inside ResponseStreamState.accumulate_event() and handle_event() when output_index references a snapshot output item that has not yet been added.

Fixes #2852

Problem

When streaming Responses API events with tools (e.g., web_search, code_interpreter) or when resuming a stream via starting_after, the stream accumulator can crash with:

File ".../openai/lib/streaming/responses/_responses.py", line 344, in accumulate_event
    IndexError: list index out of range

This happens because several event handlers in both accumulate_event() and handle_event() access snapshot.output[event.output_index] without verifying the index is within bounds. The response.created event initializes the snapshot with an empty (or partially populated) output list, and subsequent events like response.content_part.added, response.output_text.delta, or response.function_call_arguments.delta may reference indices not yet populated by response.output_item.added.

Fix

Adds event.output_index < len(snapshot.output) bounds checks before every snapshot.output[event.output_index] access in both methods:

accumulate_event() — when the index is out of bounds, the snapshot update is silently skipped. The final response.completed event contains the full response data regardless, so no information is lost.

handle_event() — when the index is out of bounds, the raw (un-enriched) event is emitted to the caller instead of crashing. This preserves the event stream and lets consumers handle the event directly.

Affected event types:

• response.content_part.added
• response.output_text.delta
• response.output_text.done
• response.function_call_arguments.delta

Test plan

• Verify the fix handles out-of-bounds output indices gracefully
• Confirm normal streaming (in-order events) is unaffected — the bounds check is a no-op when indices are valid
• Confirm get_final_response() still returns the complete parsed response

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: cf39cab8d9

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 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve parsed done events when output index is missing

When response.output_text.done arrives with an out-of-bounds output_index, this branch now emits the raw event object, which does not have the parsed field that responses.stream() consumers expect for response.output_text.done events. In the exact tool/resume scenarios this patch targets, user code that accesses event.parsed after checking event.type == "response.output_text.done" will still fail (now with AttributeError instead of IndexError), even though parsing can still be performed from event.text without needing snapshot state.

Useful? React with 👍 / 👎.