fix(codex): strip metadata and harden error converter (#51)#58
Closed
fix(codex): strip metadata and harden error converter (#51)#58
Conversation
…er (#51) When Claude Code CLI points at ccproxy's /codex endpoint, the anthropic.messages -> openai.responses converter copies Anthropic's metadata.user_id into the Responses payload as "metadata". The Codex upstream (chatgpt.com/backend-api/codex/responses) rejects this with "Unsupported parameter: metadata", so every Claude Code -> codex request was failing with HTTP 400. - codex adapter: add "metadata" to the unsupported-key strip list in _sanitize_provider_body so it is removed before the upstream call, same as max_tokens/temperature. - simple_converters: convert_anthropic_to_openai_error now coerces non-Anthropic error shapes (e.g. Codex's FastAPI-style {"detail": "..."}) into a minimal ErrorResponse envelope instead of raising ValidationError. Without this, upstream 400s were being upgraded to 502s in the non-streaming error path. Adds regression tests for both the metadata strip and the error converter (three shapes: Anthropic-native, FastAPI detail, and arbitrary dict).
Contributor
There was a problem hiding this comment.
Pull request overview
This PR fixes Codex /codex request failures caused by sending unsupported metadata to the Codex upstream and hardens Anthropic→OpenAI error conversion so non-Anthropic error payloads (e.g., FastAPI {"detail": ...}) don’t raise validation exceptions and incorrectly surface as 5xx errors.
Changes:
- Strip
metadatafrom Codex outbound request bodies in_sanitize_provider_body. - Update
convert_anthropic_to_openai_errorto coerce unexpected upstream error payload shapes into a minimal Anthropic error envelope before conversion. - Add regression tests for both the Codex body sanitization and error-shape coercion.
Reviewed changes
Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.
| File | Description |
|---|---|
ccproxy/plugins/codex/adapter.py |
Removes metadata (and other unsupported keys) from provider payloads sent to Codex. |
ccproxy/services/adapters/simple_converters.py |
Coerces non-Anthropic error payloads into a valid Anthropic ErrorResponse prior to conversion. |
tests/plugins/codex/unit/test_adapter.py |
Adds a regression test ensuring metadata is stripped by the Codex adapter sanitization step. |
tests/unit/services/adapters/test_simple_converters.py |
Adds regression tests for Anthropic-native and non-Anthropic error payload shapes. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
+244
to
+248
| if not isinstance(data, dict) or "error" not in data: | ||
| message = "" | ||
| if isinstance(data, dict): | ||
| for key in ("detail", "message", "error_description"): | ||
| value = data.get(key) |
There was a problem hiding this comment.
The coercion guard only runs when the payload is not a dict or is missing the top-level error key. If an upstream returns { "error": ... } in a non-Anthropic shape (e.g., error is a string or missing the discriminator type), ErrorResponse.model_validate(data) below will still raise ValidationError and can still upgrade an upstream 4xx into a 5xx. Consider try/except ValidationError around model_validate and falling back to the minimal envelope on validation failure.
The OpenAI Responses -> Anthropic and OpenAI Chat -> Anthropic stream converters were emitting ContentBlockDeltaEvent with delta=TextBlock(type="text") instead of delta=TextDelta(type="text_delta"). The Pydantic model accepts either (TextBlock is a tolerated fallback), but the real Anthropic wire protocol and the Claude Code CLI's SDK parser require type="text_delta". The effect was that Claude Code CLI pointed at ccproxy's /codex endpoint received a 200 OK stream, parsed message_start/content_block_start/content_block_stop/message_stop correctly, but silently dropped every text delta — the user saw nothing. Adds two regression tests pinning the on-the-wire type to text_delta for both the Responses and Chat converters, including a check against model_dump(by_alias=True) so the serialized payload can't drift.
When Claude Code CLI targets /codex, conversations with history and
tool-use cycles were dropped or mangled, and tool streaming events did
not match the official specs. Rewrite the Anthropic -> Responses input
translation and align tool streaming with Anthropic/OpenAI specs.
- anthropic_to_openai/requests.py: translate the full message list
into Responses API input items (message / function_call /
function_call_output), preserving interleaved text and tool_use
ordering within assistant turns. Add a deterministic
_clamp_call_id so tool_use/tool_result pairs stay intact when ids
exceed OpenAI's 64-char limit. Accept LegacyCustomTool alongside
Tool in the custom-tool mapping.
- common/streams.py: emit tool_use content_block_start with empty
input {} and stream arguments via input_json_delta.partial_json,
per Anthropic streaming spec. Official SDKs ignore input attached
directly to the start event.
- openai_to_openai/streams.py: tool_call continuation chunks no
longer re-emit id/name. Per OpenAI Chat streaming spec, those
fields only appear on the first chunk for a given tool call.
- models/openai.py: FunctionCall.name is now Optional to support
the continuation chunks above.
- services/adapters/delta_utils.py: identity fields (index, type,
id, name, call_id) are overwritten instead of merged. Without
this, providers that re-send these per chunk (e.g. the Codex
Responses->Chat adapter) produced "shellshell..." /
"fc_abc_fc_abc..." and broke downstream validation.
Tests cover the full tool cycle, interleaved assistant ordering,
list-form tool_result content, pending user text after a tool
result, long call_id clamping, LegacyCustomTool acceptance, tool_use
streaming events, and delta_utils identity handling.
Owner
Author
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Test plan
Known follow-up (not in this PR)
Even with this fix, there's a separate streaming-display issue when Claude Code CLI targets `/codex`: the server emits valid Anthropic SSE chunks (`message_start`, `content_block_delta`, `message_stop`) and returns 200, but the CLI shows no visible response. Worth tracking separately — it's orthogonal to the metadata 400 this PR fixes.