1d96b09875
When the LLM generates a <tool_call> block followed by a self-generated <tool_result> block in the same response, the fake result must be stripped before appending to message history. The real tool result will be injected separately by the system. This fixes a React Hallucination bug where models could fabricate tool results that didn't actually come from tool invocations. |
||
|---|---|---|
| .. | ||
| app | ||
| scripts | ||
| tests | ||
| pyproject.toml | ||
| requirements.txt | ||
| run.py | ||
| uv.lock | ||