The codebase adds robust error handling and safety throughout chat.py: tool execution is wrapped in try/except with clear assistant-directed error messages; apply_patch now tolerates a missing apply_patch.md; message content is guarded before access; browser citation normalization only runs when the browser tool is enabled; tool outputs are printed safely; and token generation is protected against runtime failures. Readline history setup is resilient to missing files and other errors. Tensor parallelism is now configurable via --tensor-parallel-size/--tp (overriding env TP) with validation. The developer message flow is finalized, and the system prints a clear, human-readable configuration header in non-raw mode.