Added
Ctrl+Sstashes the input box draft, which comes back after you send another message/goal <objective>sets a goal Vibe keeps working on across turns until it is done./goal edit,/goal pause,/goal resumeand/goal clearmanage itmax_goal_token_budgetinconfig.tomlcaps the token budget a goal may set, and is the budget of a goal set without one- The
/modelpicker has a/fuzzy search row that filters the model list as you type. Arrows and Enter keep working while searching, andEscleaves the search, clears the filter, then closes vibe -p --thinking LEVELsets the model's thinking level for one run, without saving it/raw [on|off]shows agent messages as raw Markdown, so a selection copies their source textpre_llmhooks run before each call to the model and can stop the turnpost_llmhooks run after each model response, tool calls included, and can ask the model to try again, within a budget of 3 retries per turn shared withpost_agenthooks/loopon its own opens a loop manager that lists the scheduled loops with their schedule and next run, deletes one withd, and adds one from a title, an interval or cron, and instructions withn. While the session is idle, the row above the input shows the next scheduled loop and its title- Inside tmux with
mouse off, a startup warning explains how to scroll the conversation, andCtrl+Nhides it for good /trusted-folderslists trusted and untrusted folders, flips or forgets a decision, and trusts a new folder with path completion- The
post_agenthook payload includesturn_idandresponse_text, the turn's final answer show_dangerous_directory_warning = falsehides the startup warning shown when Vibe runs in the home directory or another risky folder
Changed
- When Vibe quits, it waits for
session_endandnotificationhooks still running for 2s at most, instead of 5s - A project's
.vibe/config.tomlor an agent profile can turn usage telemetry off but no longer turn it back on: itsenable_telemetry = trueis ignored with a warning post_agenthooks now share their 3 retries per turn withpost_llmhooks. A retry the turn has no step left for fails the turn, naming the hook, instead of an internal errorvibe -pprints the reason when a hook ends the turn without the model's answer- Image paths are only turned into image placeholders when pasted or dropped on their own, never while typing or when recalling a prompt. Pasting a path again right after it became an image placeholder or
@mention keeps it as plain text - Picking a model's default thinking level no longer pins it in your config, so the model follows later changes of that default
- A pause inside a script the agent runs now reads "Script in progress, next step in 5s" instead of "Sleeping"
- When your API key does not use your Vibe Pro plan, a box under the banner now says so on every start and explains how to switch keys, and
/whoamiexplains it too. The Vibe Pro upgrade offer now shows in its own box under the release notes - Vibe no longer shows a warning when a provider's model catalog has no entry for the model or cannot be reached, since it falls back to its built-in model table. The warning is only written to
vibe.log - When you return to an answered question, a green
›marks the answer you gave - On a question you answered with your own text, ←→ keep switching questions and
eedits the answer - After interrupting a turn, ↑ selects the queued messages so you can edit or remove them before sending the queue
- When your Mistral provider's
api_baseis outsidemistral.aiand its subdomains (a self-hosted deployment), Vibe no longer contacts Mistral's experiment service and no longer sends crash reports to Mistral, even with telemetry on - Voice dictation and read-aloud now use the server of your Mistral provider's
api_baseinstead ofapi.mistral.ai(exceptapi.eu.mistral.aiandapi.us.mistral.ai, which keepapi.mistral.ai), unlesstranscribe_providersortts_providersin your config set another address - A model's own
auto_compact_threshold, such as the one the model catalog sets, now wins over a top-level one, including one from your organization's admin config. To apply one value to every model, an admin adds a[[models]]entry per model with itsname,providerandauto_compact_threshold - The welcome banner cat plays one animation round on start and stops. A click on it replays the round
- Voice features are now installed by default on ARM64 Linux, so the
voiceandvoice-aarch64extras are no longer needed - Shell commands that read protected files now ask for approval even when the reading command is otherwise allowed, in POSIX shells and PowerShell. When Vibe cannot tell exactly which files a read touches (variables, brace expansion, recursive or piped reads, a change of directory, an interpreter wrapper), the approval covers only that exact command
/resumeshows aResumingspinner while the picked session loads- Mistral Large 4 is now the default model. To stay on Mistral Medium 3.5, or if your own Mistral server does not offer Large 4, set
active_model = "mistral-medium-3.5"
Fixed
- With tool calls made together, each
pre_toolhook line shows above its own tool instead of all of them above the first tool - Undo now restores a prompt draft cleared with
Ctrl+Cor a doubleEsc - Approving permanently no longer fails after the directory Vibe was started in is deleted
vibe.logno longer reports a model provided by an experiment as unavailable at startup, and reports a model that is really missing only once per process- Allowing a tool for the session or permanently now also approves the parallel calls it covers, instead of asking for each one
- Reaching
max_subagentsno longer blocks new subagents: the one that has waited longest is closed to make room, up to 128 subagents per session - On a first run,
--modelnaming a model from Mistral's model list no longer fails with "Unknown model" while that list is still loading - Installing or upgrading Vibe on Python 3.15 no longer fails building
cffiorrpds-pyfrom source, and the documented install commands now pass--python 3.14 - A pasted image that cannot be attached now shows why, such as a full disk or a missing file, instead of failing with 'Image file is outside the workspace or session attachments'
- Third-party providers that add
nameto streamed chunks after the first one (such as OVHcloud AI Endpoints) no longer fail every call - On Windows, auto-approved file reads and writes on Git Bash paths like
/c/Users/...no longer land in a strayC:\c\...directory - Sessions with many images or large tool results no longer become impossible to resume with "stored JSON document is at capacity: runtime-state.json"
- Smart approve no longer asks for approval on every tool call when the session uses a model that rejects temperature 0, such as some reasoning models
- Subagents no longer fail with "child Session does not retain the requested generation" when a
pre_agenthook is configured - A subagent's view now names the model its
active_modelselects instead of the main session's model - Resuming a session that used many subagents no longer takes minutes to show the prompt
- First run on Linux without an API key no longer hangs at "Starting the app server…" waiting on a keyring unlock prompt
- A
VIBE_*environment variable Vibe cannot use now shows a warning naming the wrong fields, instead of being ignored silently - Tool filters naming a former tool, such as
greporweb_fetch, now show a warning saying what replaced it, rather than silently matching nothing - The install script and
vibe updatenow work behind TLS-inspecting corporate proxies (uv 0.11+): the install script makes uv trust the system certificate store, andvibe updatedoes so whenenable_system_trust_storeis on - A third-party provider that ends its answer without any content is now retried twice, then fails the turn with a clear error instead of being retried for minutes
- A folder whose only project config is
.vibe/hooks.tomlor.vibe/skills.tomlnow asks to be trusted, andvibe -pwarns that it is ignored until--trust, instead of skipping it silently - On Windows, creating a new file in a git worktree or a repository with
.gitattributesno longer hangs - On Windows, Enter and Ctrl shortcuts keep working after a command run by the agent changes the console's input mode
- Question tabs wrap onto several rows in narrow terminals instead of being cut off
- Removing the oldest of several queued messages no longer shows the remaining one twice once it is sent
@file completion now finds files anywhere in large repositories, not only those sorted among the first 32,000 pathsvibe --setuponboarding screens no longer go blank in small terminals: the theme picker and the sign-in options render in short terminals, with an enlarge-terminal notice only when nothing fits- A trusted folders file that cannot be read or parsed is now kept instead of being reset to empty at startup
- Plugins are no longer listed under "Not loaded" in
/pluginsby mistake: plugin knowledge no longer fails with a permission error, and plugins that loaded are no longer listed there or warned about at startup because of informational notes - A turn that was running while the computer slept now continues after the wake instead of failing with an invalid API key error: it waits up to two minutes for your Mistral sign-in, and retries once with a refreshed one
- A config list entry missing its
name, such as an empty[[mcp_servers]], now shows a configuration error saying what to fix instead of an internal error mistral-large-latestis treated as the model the Mistral API serves under it, today Mistral Large 4: a 1,048,576-token window, low, medium and high thinking (offuses low), and earlier reasoning kept in its requests- Payment refusals now say why: a turn refused because the account has no credit left says so instead of a generic backend error, and an HTTP 402 refusal, such as a reached spending limit, shows the reason without the raw error body and no longer offers
/retry - Requests no longer fail with
process() takes no keyword argumentswhen an oldbrotlipackage is installed, as in some conda environments - Connectors exposing more than 128 tools now load with all their tools instead of being dropped
- Moving a local session to another project right after its turn ends no longer fails with "Session is already open"
- Quitting while the session is still shutting down no longer crashes the app server
- Plugins that ship manifests for several agents, such as Superpowers, now load instead of being rejected as ambiguous
- Under OIDC sign-in, model calls now use the session's own organization and workspace instead of the first one your account can reach, which could be an organization without access to the model
- A model your plan does not include no longer makes every later message fail with "credential was rejected" until a restart
- When the provider refuses a model call for a reason other than the API key, such as the model not being in your subscription tier, Vibe now shows the provider's reason instead of calling it an invalid API key
- Pull request status keeps syncing between Vibe sessions on Linux when the file watcher limit is reached
- Typing
/teleportor/remote-projectnow shows that the command was removed instead of sending it to the model
Removed
- The "Hello , how can I help you?" greeting under the startup banner, and its
show_greetingsetting. Configs that still setshow_greetingload without a warning