fix(responses): support OpenAI Responses tool loops

[saxyguy81]

Summary

• Normalize OpenAI Responses function_call item IDs to fc_* while preserving call_id as call_*.
• Add true OpenAI-compatible Responses route coverage for /responses, /v1/responses, and /codex/v1/responses.
• Add local previous_response_id continuation state for Codex Responses tool loops, scoped by client auth context with TTL/max-entry cleanup.
• Accept Droid/Factory Responses streaming fields such as stream_options.include_usage, plus Droid-sent prompt_cache_retention and safety_identifier, while stripping Codex-backend-unsupported fields before upstream dispatch.
• Preserve valid Responses SSE for streaming tool calls, including response.function_call_arguments.*.item_id == fc_*, completed function_call.id == fc_*, and function_call.call_id == call_*.
• Normalize provider/FastAPI 4xx failures into OpenAI-style JSON error envelopes so clients see rejected parameters instead of empty bodies.
• Rebuild sparse Codex response.completed payloads from streamed output items when the final upstream completed event omits output.

This remains parallel to #64; it is the Responses/OpenAI-provider compatibility PR. The latest commit folds in the Droid/Factory streaming and continuation compatibility work on top of the original fc_* fix.

Validation

• git diff --check
• uv run ruff check ccproxy/plugins/codex/plugin.py ccproxy/plugins/codex/config.py ccproxy/plugins/codex/adapter.py ccproxy/plugins/codex/routes.py ccproxy/plugins/codex/responses_state.py ccproxy/llms/models/openai.py ccproxy/llms/streaming/accumulators.py ccproxy/llms/formatters/common/identifiers.py ccproxy/streaming/deferred.py ccproxy/streaming/buffer.py ccproxy/streaming/errors.py tests/plugins/codex/integration/test_codex_basic.py tests/plugins/codex/unit/test_adapter.py tests/plugins/codex/unit/test_responses_state.py tests/unit/llms/streaming/test_accumulators.py tests/unit/llms/test_openai_responses_request_models.py tests/unit/streaming/test_buffer_parse_responses.py
• uv run ruff format --check ccproxy/plugins/codex/plugin.py ccproxy/plugins/codex/config.py ccproxy/plugins/codex/adapter.py ccproxy/plugins/codex/routes.py ccproxy/plugins/codex/responses_state.py ccproxy/llms/models/openai.py ccproxy/llms/streaming/accumulators.py ccproxy/llms/formatters/common/identifiers.py ccproxy/streaming/deferred.py ccproxy/streaming/buffer.py ccproxy/streaming/errors.py tests/plugins/codex/integration/test_codex_basic.py tests/plugins/codex/unit/test_adapter.py tests/plugins/codex/unit/test_responses_state.py tests/unit/llms/streaming/test_accumulators.py tests/unit/llms/test_openai_responses_request_models.py tests/unit/streaming/test_buffer_parse_responses.py
• uv run pytest tests/plugins/codex/unit/test_adapter.py tests/plugins/codex/unit/test_responses_state.py tests/unit/llms/test_openai_responses_request_models.py tests/unit/streaming/test_buffer_parse_responses.py tests/unit/llms/streaming/test_accumulators.py tests/plugins/codex/integration/test_codex_basic.py -q - 58 passed
• uv run pytest tests/plugins/codex/unit/test_adapter.py tests/plugins/codex/unit/test_routes.py tests/plugins/codex/unit/test_responses_state.py tests/plugins/codex/integration/test_codex_basic.py tests/plugins/codex/integration/test_codex_websocket.py tests/unit/streaming/test_buffer_parse_responses.py tests/unit/llms/streaming/test_accumulators.py -q - 67 passed
• uv run pytest tests/plugins/codex tests/unit/streaming/test_buffer_parse_responses.py tests/unit/llms/streaming/test_accumulators.py tests/integration/test_streaming_converters.py -q - 91 passed, 1 existing Pydantic deprecation warning
• uv run mypy ccproxy/plugins/codex/adapter.py ccproxy/plugins/codex/routes.py ccproxy/plugins/codex/responses_state.py ccproxy/plugins/codex/plugin.py ccproxy/plugins/codex/config.py ccproxy/llms/models/openai.py ccproxy/llms/streaming/accumulators.py ccproxy/llms/formatters/common/identifiers.py ccproxy/streaming/deferred.py ccproxy/streaming/buffer.py ccproxy/streaming/errors.py tests/plugins/codex/integration/test_codex_basic.py tests/plugins/codex/unit/test_adapter.py tests/plugins/codex/unit/test_responses_state.py tests/unit/llms/streaming/test_accumulators.py tests/unit/llms/test_openai_responses_request_models.py tests/unit/streaming/test_buffer_parse_responses.py - success, 17 source files
• ./Taskfile check - all checks passed; mypy success on 592 source files; 616 files already formatted

Live local smoke

• Local ccproxy /health returned pass after restart.
• Live curl verified streaming /responses, /v1/responses, and /codex/v1/responses with stream_options.include_usage.
• Live curl verified streamed function tool call + previous_response_id continuation, with fc_* item IDs and call_* call IDs.
• Live curl verified unknown previous_response_id returns OpenAI-style JSON error body.
• Droid/Factory target passed:

droid exec \
  --cwd /Users/smhanan/Projects/inline \
  -m custom:ccproxy-chatgpt-5.5-0 \
  --auto medium \
  'Use the Execute tool to run exactly: pwd && git branch --show-current. Then reply beginning GPT_DROID_EXEC_OK.'

Output began with GPT_DROID_EXEC_OK and returned /Users/smhanan/Projects/inline plus branch codex/inline-typora-repair.

[saxyguy81]

Coordination note for reviewers:

This PR is intentionally parallel to #64 and #66.

• fix: sanitize compacted tool history before Anthropic dispatch #64 fixes Anthropic compacted tool-history sanitization and context truncation behavior.
• fix(responses): support OpenAI Responses tool loops #65 fixes OpenAI Responses protocol compatibility for function-call item IDs (fc_* item IDs while preserving call_* call IDs).
• feat(models): add provider model and effort aliases #66 adds provider/model/effort aliases for Claude and Codex-compatible clients.

There is no intended dependency between these PRs. They can be reviewed and merged independently; if one lands first, I will rebase the others only if GitHub reports a conflict.