fix(streaming): restore real-time chunk delivery in tool/chat paths

[veithly]

Summary

• forward output_queue through tool-call chat path so provider deltas are emitted upstream
• stream tool-call responses directly in OpenAI-compatible / Anthropic / Gemini / Ollama providers
• switch Gemini streaming loops to client.aio.models.generate_content_stream(...) + async for to avoid event-loop blocking
• prevent duplicate full-content enqueue when provider already streamed text (streamed_content metadata)
• add focused regression tests for provider/tool streaming and incremental delivery behavior

Verification

• uv run python -m pytest tests/test_tool_streaming_output.py tests/test_chatbot_integration.py::TestChatBotIntegration::test_ask_tool_with_manager tests/test_agent_llm_integration.py::TestAgentLLMIntegration::test_streamed_tool_response_does_not_enqueue_full_content_twice -q
  • 9 passed

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e16f651871

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Preserve streamed tool-call fragments under one call ID

In the new streaming chat_with_tools path, the accumulator key is recomputed as tc_chunk.id or f"call_{tc_chunk.index}" for every delta, so when later deltas omit id (a common streaming pattern after the first tool-call chunk), argument/name fragments are split across two entries instead of being merged. That can produce malformed tool_calls (e.g., one entry with partial args and another with missing name), which then breaks downstream tool execution in ToolCallAgent.act() when output_queue is enabled.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Normalize Gemini streamed finish reasons before returning

The streamed Gemini tool path now returns finish_reason as str(part_response.candidates[0].finish_reason) (for example "STOP") without mapping it to the normalized values used elsewhere ("stop", "length", "tool_calls"). In streaming agent flows (output_queue set), this can prevent ToolCallAgent._should_terminate_on_finish_reason() from recognizing terminal no-tool responses, causing extra think/act iterations instead of clean termination.

Useful? React with 👍 / 👎.

[veithly]

Addressed the automated review items in commit e578bf7:

• OpenAI-compatible streaming tool-call accumulation now keeps a stable per-index call ID even when later deltas omit id, so argument/name fragments merge into the same ToolCall.
• Gemini streaming paths now normalize finish reasons (for example STOP -> stop) and preserve tool_calls as the standardized finish reason when function calls are emitted.

Validation:

• uv run python -m pytest tests/test_tool_streaming_output.py tests/test_chatbot_integration.py::TestChatBotIntegration::test_ask_tool_with_manager tests/test_agent_llm_integration.py::TestAgentLLMIntegration::test_streamed_tool_response_does_not_enqueue_full_content_twice -q
• 11 passed