mistral-vibe 3.0.1

v3.0.1

on Python PyPI

Added

• New enable_auto_approve setting: when the admin-managed config sets it to false, the auto-approve (YOLO) mode is forbidden — the auto-approve agent (and any custom agent with safety = "yolo") is hidden and --auto-approve/--yolo plus bypass_tool_permissions are ignored. Per-tool permissions still apply.
• compaction_max_output_tokens (80000), compaction_thinking (unset), compaction_reasoning_limit_tokens (32000), output_limit_retries (6) and output_token_reserve (4096) settings tune compaction requests and recovery from answers cut off at the output limit.
• The app server has experimental session/subagents/wait, session/subagents/interrupt and session/subagents/read methods: a client can wait for the subagents a turn left running, interrupt them, and read each one's status, store, model, usage, cost and steps
• vibe -p waits for the subagents its turn left running before it exits; --time-limit and other stops interrupt them, and the --output-dir export lists each subagent with its status, model, usage, cost and steps, and copies its journal
• Each subagent's cost is priced at its own model's rates; when its model is unknown, the root's rates price it and the row says so (pricingModel, pricingFallback)
• An --agent-socket server's tools/call names its caller, so the server can tell a subagent's call from the root's
• A custom tool that fails to load now shows a warning naming the file and the error, instead of disappearing silently.
• Experimental vibe run-server command: runs many vibe -p --agent-socket sessions from one process, on shared app-servers (--runs-per-app-server, default 16) that exit once idle, each run with its own config, credentials, limits and export
• Vibe Desktop sessions support scheduled loops: the assistant can schedule a recurring prompt, fired prompts show as loop cards in the conversation, and active loops are listed (and can be deleted) in the session info.

Changed

• /model and the /config model list now group models by where they come from: your organization, experiments, the model catalog, then your own config.
• Resuming or continuing a session keeps the agent it last ran on, even if your default agent changes; switch agents in the session to change it.
• ACP usage updates now report the session's prompt-cache hits as cached read tokens.
• A tool output past about 16,000 characters is saved to a file, and the model sees its start and end instead of only its start.
• A bash command that times out returns what it printed, and the default timeout is now 120 seconds, up to 300 when asked.
• tools.bash.default_timeout and max_timeout_seconds are unset unless set in config.toml; the native shell tools still default to 300 seconds, at most 600 or the default when that is longer.
• A shell tool's default_timeout above its max_timeout_seconds is now a config error naming both, instead of being lowered to the maximum.
• read_file reports the file's line count and where to read on, and an offset past the end is an error.
• read_file on a directory lists its entries, up to 500 at a time, instead of failing.
• read_file and file mentions return up to 64,000 bytes a call instead of 51,200, the same as the Agents SDK.
• On Linux, a command that runs the machine out of memory is the first process the kernel stops, before Vibe itself.
• A shell's login profile runs before a session's first command and outside its timeout, and one that fails or takes over 10 seconds is skipped with a warning.
• Bash commands can call the functions and aliases the login profile defines, for bash and zsh, and a change to a profile file is picked up by the next command.
• Vibe now supports Python 3.12 and later instead of requiring Python 3.14.
• All Vibe skill descriptions now support up to 1,024 characters.

Fixed

• Background process results from the app server keep their process ID and exit code.
• A model response kept open by keepalives without sending any data now times out after api_timeout instead of hanging the session.
• An unparseable VIBE_* environment variable such as VIBE_MODELS is now ignored and logged instead of stopping Vibe at startup.
• Sessions now record the agent they run on, so a session started with --agent reopens on that agent.
• Code the agent runs with run_typescript now sees an MCP or connector tool's own error text instead of a generic message.
• A failed /clear keeps the conversation on screen, and failed /clear and rewind requests show the server's reason.
• Picking a model your administrator enforces explains why the model does not change.
• Resuming a session from the picker restores its queued prompts.
• A malformed background-process call now fails that tool call instead of ending the session.
• When you have not chosen a model, a new session's first message now waits up to 5 seconds for the model assigned to your account.
• Editors that sign in on every connection over ACP, such as CodeCompanion.nvim, no longer open a browser sign-in page when Vibe already has an API key for that console.
• With enable_system_trust_store = true in config.toml, prompts now also use the operating system's certificates, so they work behind corporate proxies that only the OS trusts.
• Turns no longer fail when the model announces a tool call but sends none, or sends one without a name; Vibe retries once or ends the turn cleanly instead.
• The welcome banner text no longer shifts when you switch to or from a Mistral Large model.
• Connectors that need no login, such as DeepWiki, now show as not connected in /mcp instead of error - try refreshing, and pressing Enter connects them without visiting the dashboard.
• The CLI no longer hangs when it talks to an app-server that has already exited, such as after cancelling setup during a slow startup
• Hovering a subagent row with the ansi-dark theme no longer turns its text black on black
• Clicking a hovered subagent row no longer flashes the highlight back to the previous row
• Installs and launches through uv now force a GIL-enabled build of Python, fixing failures on machines with a free-threaded (no-GIL) Python 3.14 installed.
• Compaction no longer fails when a provider's cache repeats a malformed summary, or spends its whole output budget reasoning; Vibe falls back to a simpler request instead.
• An answer cut off at the output-token limit no longer ends the turn at once: it is discarded and asked for again with feedback, up to 6 times a turn.
• A tool call cut off at the output-token limit, such as a large file write, is asked for again in smaller pieces instead of failing with a JSON error.
• Tool argument errors now name the field at fault, the allowed fields and 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.
• Requests to a model with a declared max_context_length no longer overflow its window: the output cap is clipped to the room the input leaves, less a 4096-token reserve.
• Bash commands no longer lose PATH entries that a login profile resets, and the profile runs once per session.
• Bash commands still run after a command deleted /dev/null.
• The git context names the repository's default branch instead of always main, and no longer hedges the instructions a session starts with.
• In a sandbox, the git context is read in one round trip, and no longer times out on a slow sandbox.
• Cached input tokens now reach the session's token usage and runtime stats, so vibe -p reports and prices them
• Plugins that set a displayName in their ai.mistral.vibe extension now load their knowledge, hooks, agents and tool namespace instead of reporting an extension error.
• vibe -p exits only once its app-server has, and on Linux a killed vibe -p or vibe run-server takes its app-servers down with it
• vibe -p whose stdout closes early, as when piped into head, now ends the run as infrastructure_failure instead of crashing in text output
• In Vibe Desktop, a session with no new turn no longer moves in the sidebar session list when you open it or restart the app.