Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Mistral Vibe 3.1.0: see what fills the context

Written November 9, 2026

Mistral Vibe 3.1.0 (Nov 4) adds /context: what fills the window — system prompt, tools, skills, conversation, free space — per category.

What /context shows

Mistral Vibe 3.1.0, released November 4, 2026, adds a /context command that answers a question every long coding session eventually asks: what is actually filling the context window? Run it and Vibe breaks the window down by category, with a token estimate for each: the system prompt, the tool definitions, the loaded skills, the conversation so far, the free space left, and the auto-compaction buffer reserved for squeezing the session down when it gets full. The same view opens from the token counter in the footer, so the breakdown is one click away at any point in a session.

The reason the breakdown matters is that the window is not one thing. A session can feel short on room for several different reasons, and each one has a different fix. If the conversation is the bulk, the session has simply run long and needs compaction or a fresh start. If the tool definitions are large, the agent is carrying tools it is not using. If the skills are heavy, a loaded skill is eating space. If the system prompt dominates, the configuration is the problem. Before 3.1.0, a Vibe user facing a "forgetting" model had no way to tell which of those it was; the model looked forgetful either way. Now the breakdown names the filler, and the fix follows from the name.

What else 3.1.0 brings

The footer now shows monthly Vibe Code usage spent, turning yellow and then red as it runs out, or showing pay-as-you-go usage where that applies — the same kind of glanceable budget signal as the context breakdown, for money instead of tokens. File completion gets a switch: the enable_file_autocomplete setting, and the VIBE_ENABLE_FILE_AUTOCOMPLETE environment variable, turn off @ file completion and the background index that powers it, for anyone who does not want Vibe indexing their files in the background. The agent can now open an image file with read_file, so screenshots and diagrams in a repo are readable in-session instead of being opaque paths.

Two smaller changes round out the release. Ctrl+R searches prompt history, so a command from an earlier session is recoverable without retyping it. /whoami shows the last four characters of the API key, which is enough to confirm which key a session is running under without exposing it. On macOS, new worktrees of a trusted repo now clone from a template checkout by default, which speeds up worktree creation; set worktree_copy_on_write = false to turn that off. And behind a flag sits an experimental Mistral MCP connector, enabled with experimental_enable_mistral_mcp_connector — experimental, which means it can change or disappear in a later release.

How to use it

  1. Run /context before blaming the model. When Vibe starts losing the thread — forgetting a file it read, dropping an instruction from earlier in the session — open the breakdown first. Most "forgetting" is a full window, and the breakdown says what filled it.
  2. Fix the category, not the session. If tools dominate, drop the tools the session does not need. If skills dominate, unload the heavy skill. If the conversation dominates, compact or start fresh. A generic "start a new session" throws away the conversation's value when the problem was a tool definition.
  3. Watch the footer, not the invoice. The usage indicator turns yellow, then red, as the monthly allowance runs out. A session that burns through the allowance on long prompts is visible in the footer before it shows up as a bill.
  4. Turn off file autocomplete if you do not want the background index. Set enable_file_autocomplete to false, or VIBE_ENABLE_FILE_AUTOCOMPLETE to false, and the @ completion and its index go away. It is a preference, not a fix — the index is the feature, and the switch is there for people who do not want it.

Verdict

Run /context before blaming the model — most "forgetting" is a full window, and the breakdown names the filler. For long Vibe sessions, the command is the first diagnostic, not the last resort; a model that cannot see the instruction is not a model that forgot it, it is a window that is full.

What can go wrong

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Kiro CLI 3.0: V3 is the default, Classic is gone · Copilot CLI 1.0.106: /code, /settings, GPT-6.1 Terra · Grok Build: SpaceX's $8-a-month coding agent