fix: #3308 reject chat custom tool calls explicitly

[Aphroq]

Summary

• Gate Chat Completions custom tool-call rejection behind strict_feature_validation so the default behavior stays compatible.
• Raise UserError for unsupported custom tool calls in both non-streaming and streaming Chat Completions strict mode.
• Track ignored streamed custom tool-call indexes so later chunks without type do not become malformed function-call items.
• Add regression coverage for default compatibility, strict-mode rejection, mixed function/custom calls, and streamed custom chunks.

Test plan

• uv run pytest tests/models/test_openai_chatcompletions_converter.py -k custom_tool_call
• uv run pytest tests/models/test_openai_chatcompletions_converter.py -k "custom_tool_call or mixed_custom"
• uv run pytest tests/models/test_openai_chatcompletions_stream.py -k custom_tool_call
• uv run pytest tests/models/test_openai_chatcompletions.py -k custom_tool_call
• uv run pytest tests/models/test_openai_chatcompletions_converter.py tests/models/test_openai_chatcompletions_stream.py tests/models/test_openai_chatcompletions.py
• bash .agents/skills/code-change-verification/scripts/run.sh

Issue number

Closes #3308

Checks

• I've added new tests (if relevant)
• I've added/updated the relevant documentation
• I've run make lint and make format
• I've made sure tests pass

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1b2c3cbed8

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".

[Aphroq]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. 🚀

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".

[seratch]

Thanks for the fix. One compatibility concern: this changes the released Chat Completions behavior from silently ignoring custom tool calls to always raising. That is safer, but it is still a behavior change for users/providers that currently rely on the old behavior.

Can we gate this behind strict_feature_validation instead?

Suggested shape:

• Pass OpenAIChatCompletionsModel._strict_feature_validation into both Converter.message_to_output_items() and ChatCmplStreamHandler.handle_stream().
• Raise only when strict_feature_validation=True.
• Keep the default False behavior compatible with today’s behavior.
• For streaming, please make sure ignored custom tool call indexes stay ignored across later chunks where type may be omitted, so we do not accidentally build a malformed ResponseFunctionToolCall.

Since existing strict feature validation errors in this model use UserError, I think this should probably use UserError too rather than AgentsException.

[Aphroq]

Handled. PR updates:

• Converter.message_to_output_items() now accepts strict_feature_validation, defaulting to False, so default Chat Completions behavior remains compatible and continues ignoring custom tool calls.
• OpenAIChatCompletionsModel passes _strict_feature_validation into both the non-streaming converter and streaming ChatCmplStreamHandler.handle_stream().
• With strict_feature_validation=True, both non-streaming and streaming paths now reject custom tool calls with UserError.
• The default streaming compatibility path tracks ignored custom tool-call indexes. If later chunks omit type, the same index stays ignored and is not accidentally assembled into a ResponseFunctionToolCall.

Extra edge coverage:

• Converter default compatibility: custom tool call returns no output, matching the old behavior.
• Converter strict mode: custom tool call raises UserError.
• Converter strict mixed case: function + custom does not partially convert the function while hiding the custom call.
• Model get_response strict mode: confirms OpenAIChatCompletionsModel._strict_feature_validation reaches the converter.
• Streaming strict mode: custom delta raises UserError.
• Streaming default mode: custom index followed by no-type chunks remains ignored and does not emit function-call events or final output.

[Aphroq]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. You're on a roll.

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".