bug fix 113 - #122
Conversation
![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: 4b8765fc35
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".
| int hold = HoldBack(buf, "<|channel>", "<|tool_call>", "<channel|>"); | ||
| if (hold > 0) | ||
| { | ||
| string emit = buf.Substring(0, buf.Length - hold); |
There was a problem hiding this comment.
When a Gemma 4 checkpoint emits the reported thought...<channel|>answer shape during a streaming request, Add receives the reasoning incrementally and this branch emits every chunk that does not currently end in a partial marker as Content. By the time the closing marker arrives, only the small fragment still in _buffer can be reclassified, while the preceding chain of thought has already been sent to OpenAI/WebUI clients as answer text. The new streaming test misses this because it checks only that the marker is absent and the result ends with the answer; it does not assert that Deliberating quietly. is absent. Buffer or classify the prompt-primed region before emitting it as content so the fix also covers streaming, not only batch parsing.
Useful? React with 👍 / 👎.
Engine comparison — TensorSharp vs llama.cpp (PR smoke)No report artifact was produced — the benchmark failed before generating results (see the workflow logs). |
1 participant
No description provided.