feat: Microsoft Agent Framework compatibility and bypass mode

[CaddyGlow]

Summary

• Add OpenAI Responses format support to mock handler with proper SSE events
• Add format-aware mock adapter with target format detection
• Add bypass mode to plugin factory with MockAdapter fallback
• Add inject_detection_payload config toggle for generic API usage
• Add openai_thinking_xml streaming configuration support

Replaces #44 (auto-closed when #43 merged). PR 2 of 3 from original #41.

Test plan

• All pre-commit checks pass
• 1101 unit/plugin tests pass

[Copilot]

Pull request overview

Adds Microsoft Agent Framework–compatible behavior to the Codex proxy by making mock/bypass mode format-aware (OpenAI Chat vs OpenAI Responses vs Anthropic), adding a bypass-mode adapter fallback in provider factories, and propagating the openai_thinking_xml streaming setting through format adapters.

Changes:

• Implement OpenAI Responses-format mock streaming (SSE event sequence) and endpoint/format-chain based target format detection in the mock adapter.
• Add provider-factory bypass mode that returns MockAdapter (with clear warning/error handling) plus coverage for bypass behavior.
• Add inject_detection_payload Codex config toggle and propagate openai_thinking_xml via format-adapter ContextVar, with unit/integration tests for MSAF-style workflows.

Reviewed changes

Copilot reviewed 19 out of 20 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
uv.lock	Updates locked dependencies (notably mcp bump) to support new functionality.
tests/unit/services/test_mock_adapter.py	Extends mock-adapter tests for target format resolution and prompt extraction call path.
tests/unit/services/mocking/test_mock_handler.py	Adds prompt extraction tests and validates OpenAI Responses SSE event sequences.
tests/unit/llms/test_llms_streaming_settings.py	Verifies openai_thinking_xml propagation and ContextVar task isolation.
tests/unit/core/test_provider_factory_bypass.py	Adds unit coverage for provider factory bypass mode behavior.
tests/plugins/codex/unit/test_adapter.py	Adds coverage for disabling detection payload injection and preserving user reasoning.
tests/plugins/codex/integration/test_msaf_real_library.py	Adds multi-step MSAF-style workflow tests through the proxy (httpx).
tests/plugins/codex/integration/test_msaf_compat.py	Adds MSAF compatibility integration tests (no CLI injection, no thinking XML).
tests/plugins/codex/integration/test_codex_basic.py	Adds integration validation for bypass-mode Responses streaming events.
ccproxy/services/mocking/mock_handler.py	Implements format-aware mock generation + OpenAI Responses streaming SSE events + prompt extraction.
ccproxy/services/factories.py	Wires OpenAI Responses mock adapter and propagates openai_thinking_xml into adapters/registry.
ccproxy/services/adapters/mock_adapter.py	Adds format inference (format_chain / endpoint) and passes prompt/format into mock handler.
ccproxy/services/adapters/format_adapter.py	Adds configure_streaming() and sets openai_thinking_xml via ContextVar during conversions.
ccproxy/plugins/codex/config.py	Introduces inject_detection_payload configuration option.
ccproxy/plugins/codex/adapter.py	Skips detection/template injection when configured; normalizes input and strips unsupported params.
ccproxy/llms/formatters/openai_to_openai/streams.py	Respects ContextVar-controlled openai_thinking_xml when emitting thinking segments.
ccproxy/llms/formatters/openai_to_openai/responses.py	Respects ContextVar-controlled openai_thinking_xml when merging/serializing thinking.
ccproxy/llms/formatters/context.py	Adds ContextVar helpers for openai_thinking_xml.
ccproxy/core/plugins/factories.py	Adds bypass-mode factory path returning MockAdapter and logging a lifecycle warning.
.ccproxy.codex.msaf.toml.example	Adds example config for MSAF usage (incl. disabling detection injection and thinking XML).

Comments suppressed due to low confidence (1)

ccproxy/services/adapters/mock_adapter.py:116

• handle_request() always creates a new RequestContext with a hard-coded request_id ("mock-request"), even when request.state.context already exists. This makes request/stream IDs collide across requests in bypass mode and breaks log correlation (the upstream middleware/decorators generate per-request UUIDs). Consider reusing request.state.context when present, otherwise generate a unique request_id (e.g., from X-Request-ID header or uuid4()).

        # Get endpoint from context or request URL
        endpoint = request.url.path
        if hasattr(request.state, "context"):
            ctx = request.state.context
            endpoint = ctx.metadata.get("endpoint", request.url.path)

        target_format = self._resolve_target_format(request, endpoint)
        model = "unknown"
        try:
            body_json = json.loads(body) if body else {}
            model = body_json.get("model", "unknown")
        except json.JSONDecodeError:
            pass
        except UnicodeDecodeError:
            pass
        except Exception as e:
            logger.debug("stream_flag_extraction_error", error=str(e))

        # Create request context
        ctx = RequestContext(
            request_id="mock-request",
            start_time=time.perf_counter(),
            logger=structlog.get_logger(__name__),
        )

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

In the exception handler when parsing the request body for model, the log event name is stream_flag_extraction_error, which is misleading (this block isn't extracting the stream flag). Renaming the event (e.g., model_extraction_error or body_parse_error) would make logs easier to interpret.

Suggested change

	logger.debug("stream_flag_extraction_error", error=str(e))
	logger.debug("model_extraction_error", error=str(e))