Mistral Vibe: Long Conversations Compact Instead of Failing
Mistral VibeMistral Vibe 3.0.3 made context-window overflows recoverable. Long conversations now compact instead of failing when the summary or rebuilt context does not fit, and a session resumed on a model with a smaller window compacts a large tool result on the next turn. Model settings also show each model's compaction threshold, and a stalled model reply now retries.
Key Takeaways
- Oversized conversations compact instead of erroring when the summary or rebuilt context will not fit.
- Resuming on a smaller-window model no longer fails over one large tool result; the next turn compacts it.
- True overflows now surface as a clear context overflow, not a generic model error.
- Model settings show each model's compaction threshold.
- A reply now waits up to 12 minutes for first output and retries once after 5 minutes of silence.
auto_compact_threshold = 0is now an error, and two compaction output settings were removed.
Fewer dead ends in long sessions
Long agent sessions eventually outgrow the model's context window. Mistral Vibe 3.0.3 changes how that situation is handled so it ends in compaction rather than an error.
Compaction that copes with tight limits
- Long conversations now compact instead of failing when the summary or the rebuilt context does not fit.
- A session resumed on a model with a smaller window is no longer refused because one large tool result will not fit. The next turn compacts it.
- A conversation larger than the model's window fails as a clear context overflow rather than a generic model error.
- A session resumed on a model whose window cannot hold the conversation at all is refused up front, instead of failing later.
- A reply stopped by a full context window counts as cut off, and Vibe shows a warning when the model refuses a request as too long or a compaction has to shorten its inputs.
Visibility and timeouts
Model settings now show each model's compaction threshold, including the default one. A model reply waits up to 12 minutes for its first output and retries once after 5 minutes without output.
Config changes to know about
Two changes may need attention. auto_compact_threshold = 0 is now a config error: set a positive value or remove it. Old -1 values are removed on load so the model's default applies. The compaction_max_output_tokens and output_token_reserve settings were also removed. Context window sizes now come from the user's config, then the provider's model list, then a built-in table, with a warning and a 262,144-token assumption for unknown models.