[RGB-loop]

Summary

When a Codex request is converted to Anthropic Messages or Chat Completions, a response that stops early (max_tokens, length, refusal, truncated stream) gets status: "incomplete", but the terminal SSE event is always response.completed. Codex (rust-v0.162.1, codex-api/src/sse/responses.rs) only reads incomplete_details on a response.incomplete event, so it treats these turns as successful. The existing code comments say the status is there so Codex does not take truncation or refusal for a normal completion; that intent was not being met.

• codex_responses_sse::response_terminal picks the event name and type from response.status. Both converters use it.
• The usage parser, Codex OAuth SSE aggregation (handlers.rs) and the Chat history recorder now accept response.incomplete. Without this, the converter change would drop usage and Chat history for these responses.

This also fixes two of these consumers for native Responses upstreams, which already send response.incomplete: usage was lost, and Codex OAuth aggregation failed with No response.completed event in upstream SSE.

Behavior change for Codex users

Turns that used to end silently with a truncated result now end with an error, matching native OpenAI behavior:

• max_output_tokens: Incomplete response returned, reason: max_output_tokens
• refusal (content_filter): Codex's content filter error

This may show up more often than with native OpenAI. On Codex → Anthropic routes, when neither the request nor the provider sets max_output_tokens, the proxy injects max_tokens=8192 (request_builder.rs), so long outputs can hit the limit. Users can raise it with the provider's meta.max_output_tokens. This PR does not change that default.

Codex may retry on this stream error, which would replay the same limit and use more tokens. I did not confirm its retry policy.

Not changed: Chat finish_reason=content_filter still maps to completed, and truncated streams report max_output_tokens on the Anthropic path but stream_truncated on the Chat path.

Testing

• 7 new or updated tests cover both converters (max_tokens/length, refusal, truncation), the usage parser, OAuth aggregation and Chat history. All 7 fail with the old behavior restored and pass with this change.
• cargo fmt --check, cargo test --lib proxy:: -- --skip services::proxy and cargo test --test proxy_claude_streaming pass on Rust 1.95.
• Run in parallel, some services::proxy tests fail on temp HOME races, also with the change reverted. With --test-threads=1 they pass 77/77.
• These local runs are on Rust 1.95. The repo's pinned 1.91.1 is covered by Rust CI.
• Not run end to end against a real Codex CLI or a real Anthropic/Chat upstream. Codex's handling is from reading its source.

Fixes #546

🤖 Generated with Claude Code

[SaladDay]

Thanks for the fix!