Mistral Vibe: Recovery From Cut-Off Answers, Failed Compaction and Tool Errors
Mistral VibeMistral Vibe 3.0.1 made long sessions more resilient. An answer cut off at the output-token limit is discarded and re-requested with feedback, up to 6 times per turn, and a cut-off tool call is retried in smaller pieces. Compaction falls back to a simpler request instead of failing, and tool and edit errors now give actionable hints.
Key Takeaways
- Answers cut off at the output limit are re-requested up to 6 times per turn instead of ending the turn.
- A truncated tool call, such as a big file write, is retried in smaller pieces.
- Compaction falls back to a simpler request when the summary is malformed or reasoning eats the budget.
- Tool argument errors now name the faulty field and suggest the likely intended one.
- A failed edit shows the closest matching text, which helps the agent fix its own mistake.
- New settings such as
output_limit_retriesandcompaction_max_output_tokenslet users tune the recovery behavior.
Fewer dead-end turns
Mistral Vibe 3.0.1 added several recovery paths so a session keeps going when the model hits a limit or makes a malformed request.
Output-limit recovery
An answer cut off at the output-token limit no longer ends the turn immediately. Vibe discards it and asks again with feedback, up to 6 times a turn. A tool call cut off at the limit, such as a large file write, is requested again in smaller pieces instead of failing with a JSON error. Requests to a model with a declared max_context_length also no longer overflow its window: the output cap is clipped to the room the input leaves, less a 4096-token reserve. New settings (output_limit_retries, output_token_reserve) tune this behavior.
More robust compaction
Compaction no longer fails when a provider's cache repeats a malformed summary, or when the model spends its whole output budget reasoning. Vibe falls back to a simpler request. The settings compaction_max_output_tokens (80000), compaction_thinking, and compaction_reasoning_limit_tokens (32000) let users tune compaction requests.
Clearer tool errors
Tool argument errors now name the field at fault, list the allowed fields, and suggest the one the model most likely meant. A failed edit now shows the closest matching text in the file, or the lines where a repeated match occurs, which helps the agent correct itself on the next try.