fix: rewrite response model field back to request model#2645
Conversation
When the proxy maps models (e.g. claude-sonnet-4-6 → mimo-v2.5-pro), the upstream API response contains the mapped model name. This causes Claude Code's /context to display the wrong model name, since it reads from local config, not from the API response. This fix rewrites the `model` field in the response back to the original request model, so the client always sees the model name it configured. Three code paths are covered: - Non-streaming passthrough (response_processor.rs) - Streaming passthrough via SSE message_start rewrite (response_processor.rs) - Non-streaming transform mode (handlers.rs)
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: fbf2295584
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| // 尝试解析并记录完整的 SSE 事件 | ||
| let mut chunk = bytes.to_vec(); | ||
| while let Some(event_text) = take_sse_block(&mut buffer) { | ||
| let event_start_in_buffer = buffer.len(); |
There was a problem hiding this comment.
The rewrite index is derived from buffer.len() after take_sse_block drains the event, so it points to the remaining tail rather than the event's position in the current chunk. When a network chunk contains a full message_start event plus trailing bytes (for example, the start of the next SSE event), line_start_in_buffer is shifted forward and chunk.splice can replace the wrong byte range, producing malformed SSE/JSON output for clients. This breaks streaming responses intermittently depending on chunk boundaries.
Useful? React with 👍 / 👎.
| if let Ok(mut json_val) = serde_json::from_slice::<Value>(&body_bytes) { | ||
| if json_val.get("model").and_then(|m| m.as_str()) == Some(ctx.request_model.as_str()) { | ||
| // model 未被改写，无需处理 | ||
| } else if json_val.as_object_mut().is_some() { | ||
| json_val["model"] = json!(ctx.request_model); | ||
| if let Ok(new_bytes) = serde_json::to_vec(&json_val) { |
There was a problem hiding this comment.
The response body is rewritten to ctx.request_model before usage/model extraction, so TokenUsage::from_*_response and the fallback json_value["model"] now observe the rewritten model rather than the upstream provider model. In log_usage_internal, pricing can still be based on model when pricing_model_source != "request", which means mapped-model accounting becomes inaccurate for non-streaming calls after this change.
Useful? React with 👍 / 👎.
1 participant
Summary
When the proxy maps models (e.g.
claude-sonnet-4-6→mimo-v2.5-pro), the upstream API response contains the mapped model name. This causes Claude Code's/contextcommand to display the wrong model name, since it reads from local config rather than the API response.This PR rewrites the
modelfield in the response back to the original request model, so the client always sees the model name it configured. The proxy's model mapping remains fully functional for routing — only the response metadata is adjusted.Changes
response_processor.rsmodelfield toctx.request_model, re-serialize, and stripContent-Lengthheadermessage_startevents and rewritemodelfield in themessageobject before yielding bytes to clientcreate_logged_passthrough_streamsignature to acceptrequest_model: Option<String>handlers.rsmodelfield before serializationrequest_modelMotivation
Users who switch between providers (e.g. from Anthropic to Xiaomi MiMo) via cc-switch want Claude Code's
/contextto display the correct model name. Currently,/contextalways shows the original Claude model name (e.g. "Sonnet 4.6") because Claude Code reads from its local config, not from the API response. With this fix, the responsemodelfield matches what Claude Code expects, so the display is always correct.Test plan
modelmatches request modelmessage_starteventmodelmatches request modelmodelmatches request model