## Problem
Codex receives reasoning text from the Responses stream, but FCC does
not emit reasoning-token usage details. Codex reports zero reasoning
tokens even when thinking is present.
## Changes
| Before | After |
| --- | --- |
| Responses usage reported only input, output, and total tokens. |
Responses usage reports input, output, total, and reasoning-token
details when reasoning text exists. |
| Reasoning text had no adapter-owned token estimate. | Reasoning text
gets a Responses-owned best-effort token estimate. |
| Reasoning estimates could exceed reported output tokens. | Reasoning
estimates are capped at reported output tokens. |
| Text-only Responses usage kept the base shape. | Text-only Responses
usage keeps the base shape. |
Codex includes OpenAI tool_search in the tools array; omit it from the Anthropic payload like other passive hosted tools so fcc-codex requests do not fail at conversion.
## Summary
- Split monolithic `core/openai_responses/conversion.py` and `sse.py`
into focused protocol modules (input, output, stream, tools, reasoning,
events, etc.) behind an `OpenAIResponsesAdapter` facade.
- Wire `ClaudeProxyService.create_response()` through the adapter
instead of importing conversion helpers directly, tightening the
API/import boundary.
- Add Codex bridging for Responses `custom_tool_call` items and document
the adapter architecture in `ARCHITECTURE.md`.
## Test plan
- [x] `uv run pytest tests/core/openai_responses/
tests/api/test_openai_responses.py tests/cli/test_adapters.py
tests/contracts/test_import_boundaries.py`
- [x] Full CI via `.\scripts\ci.ps1`
## Summary
- Add `StreamRecoverySession` in
`core/anthropic/stream_recovery_session.py` to centralize early-retry
classification, holdback buffering, retry counting, and flush/discard
behavior shared by Anthropic and OpenAI transports.
- Refactor `AnthropicMessagesTransport` and `OpenAIChatTransport` to use
the shared session instead of duplicating recovery holdback and
early-retry logic.
- Update architecture docs and tests for per-delta overlap trim on
midstream recovery; bump version to 2.2.1.
## Test plan
- [x] `uv run pytest tests/core/anthropic/test_stream_recovery.py`
- [x] `uv run pytest tests/providers/test_anthropic_messages.py
tests/providers/test_streaming_errors.py`
- Only use provider completion_tokens when it is an int; otherwise estimate
- Coerce message_start/message_delta usage fields to safe integers in SSEBuilder
- Add regression tests for null upstream completion_tokens and builder edge cases
Claude Code could crash (e.g. undefined access on usage) when NIM/GLM or
similar sent usage with null token fields in streamed message_delta.
- Make AnthropicToOpenAIConverter stateful: assistant text after tool_use is
deferred until matching tool_result, then replayed as a follow-up assistant
turn.
- After native streamed tool_use, emit top-level SSE error on transport
failure instead of assistant text_delta (avoids bad transcript shape).
- Add NIM preflight, streaming, converter, and product smoke regressions.
- Point DeepSeek at api.deepseek.com/anthropic with x-api-key headers
- Native request builder, DeepSeek-specific thinking/block sanitization
- Drop deepseek from OpenAI-chat server-tool preflight; update tests and docs
- Default smoke model deepseek-v4-pro; re-export dump_raw_messages_request
- Add ReasoningReplayMode and top-level reasoning replay in OpenAI conversion
- DeepSeek/NIM request bodies use reasoning_content when thinking is enabled
- NIM retries without reasoning_content on 400 from upstream
- Per-provider smoke models (FCC_SMOKE_MODEL_*) independent of MODEL mapping
- Fix smoke model override parsing for owner/model names with slashes
- Live smoke: reasoning tool continuation uses synthetic thinking+tool history
- Tests and docs updated
Consolidates the incremental refactor work into a single change set: modular web tools (api/web_tools), native Anthropic request building and SSE block policy, OpenAI conversion and error handling, provider transports and rate limiting, messaging handler and tree queue, safe logging, smoke tests, and broad test coverage.
Alishahryar1