Mistral
Mistral Vibe CLI
2 days since its last release · 3.1.2 ·
Mistral Vibe is an AI coding tool from Mistral. Its latest release is 3.1.2, released on 10 Nov 2026. It shipped 8 releases in the last 30 days, and we have tracked 101 since 9 Dec 2025.
- Latest
- 3.1.2
- 10 Nov 2026
- Last 30 days
- 8
- releases
- Usual gap
- ~4d
- between releases
- Tracked
- 101
- since 9 Dec 2025
- Homepage
- github.com/mistralai/mistral-vibe
Releases per month
last 12 months- 18 Dec
- 5 Jan
- 4 Feb
- 9 Mar
- 11 Apr
- 9 May
- 11 Jun
- 9 Jul
- 7 Aug
- 8 Sep
- 7 Oct
- 3 Nov
Release history
101 releasesReleases 80
-
Turn plugins on or off, and manage skills from a new /skills list
- Installed plugins can be disabled with the disabledplugins setting, and plugins can ship off by default and be enabled with enabledplugins
- New /skills command lists installed skills with search, a Space toggle to turn each on or off, and a preview of the selected skill
- With registry skills on, /skills adds a Shared tab to add shared skills, pin a version or alias per scope, preview versions and remove skills
- Sessions open faster because the model catalog is cached on disk and refreshed in the background
- Headless --max-tokens and --max-price limits now stop the run before another completion starts, and messages sent during background waits arrive at once
-
New lifecycle hooks, a /pr list, Vim keybindings and setup import
- New preagent, notification, sessionstart and sessionend hooks can block or rewrite messages, add context and react to session events
- /pr lists the session's pull and merge requests with status, and PR tracking is now on by default
- Built-in /vibe:import skill brings instructions, skills, MCP servers, hooks and permissions over from other coding agents
- Vim Normal and Insert modes in the prompt input, plus a pinned prompt at the top while you scroll
- The /teleport command, & prompt prefix, --teleport flag and /remote-project are removed
-
Context window breakdown, prompt history search and image reading
- New /context command shows what fills the context window, with token estimates per category
- Ctrl+R now searches prompt history, and voice recording moves to Ctrl+T
- The agent can open and see image files with readfile on models that read images
- Footer shows monthly Vibe Code usage, and the model picker moved to the input box border
- Errors and warnings now appear in the conversation, replacing floating pop-ups
-
Release notes browser, faster session startup and custom Python tools dropped
- New /release-notes command (alias /whatsnew) browses every version's release notes in a searchable full-screen view
- Sessions open faster, and organization settings now load from cache and refresh in the background
- Custom Python tools in ~/.vibe/tools/ no longer load and show a startup warning; Vibe can migrate them to skills
- Headless mode adds --model to pick a model for one run and --system-prompt-file to replace Vibe's instructions
- Fixes subagent questions getting stuck, idle CPU use, and --continue picking the wrong session
-
PR and MR tracking in the bottom bar, stricter config checks
- Experimental session artifacts detect, sync and show linked pull and merge requests with CI status in the bottom bar
- Vibe warns about unused or misspelt config keys, strictconfig refuses to start on them, and vibe config schema prints all settings
- Breaking config changes: autocompactthreshold = 0 is an error, maxoutputtokens no longer caps replies, compactionthinking removed
- Footer shows the active model and thinking level, and skills edited on disk load without /reload
- Messages no longer wait for connectors or MCP servers with cached tools, plus many fixes to compaction, images and shell timeouts
-
Higher subagent limits, visible reasoning and a mouse capture toggle
- Default subagent limits raised to 32 alive and 32 running at once (from 8 and 4), with new maxsubagents and maxconcurrentsubagents settings
- The model's reasoning now shows by default as a collapsible Thought row
- New enablemousecapture setting leaves text selection and copy to your terminal
- Images sent to models without image support no longer fail the turn; the model gets [image] text instead
- Custom subagents now respect their tool permissions, skill limits and hooks, plus many session, clipboard and organization settings fixes
-
Admin control over auto-approve mode, sturdier output-limit recovery
- New enableautoapprove setting lets admins forbid auto-approve (YOLO) mode; per-tool permissions still apply
- Answers or tool calls cut off at the output limit are retried with feedback or in smaller pieces instead of failing
- Compaction falls back to a simpler request instead of failing, with new settings to tune it
- Vibe now supports Python 3.12 and later, and vibe -p waits for running subagents before exiting
- Bash timeouts return partial output, long tool outputs keep start and end, and readfile lists directories
-
Rust-only CLI with Unified Harness; Python 3.14 now required
- The Rust TUI is now the only interface, and every session type uses the Unified Harness; the legacy Python TUI, plan mode and old execution runtime are removed
- Vibe now requires Python 3.14 or newer (previously 3.12), and certificate checks follow stricter RFC 5280 rules that may reject some private CAs
- New Rust CLI features: /branch forks a conversation, a debug console (/debug, Ctrl+\), clickable options, and a more compact welcome banner
- Headless mode (-p) gets turn, token, price and time limits, graceful stops and a JSON run export with an outcome exit code
- Admins can block all MCP server operations with enablemcp; Desktop can import projects and sessions from another coding agent
-
Rust CLI gets plugins, narrator and proxy setup; headless runs gain limits
- Rust CLI adds /plugins and /reload-plugins, /proxy-setup, /whoami, and a narrator that reads each turn's summary aloud
- Model and thinking pickers now open together; press s to apply a choice for the current session only
- vibe -p reads prompts from file or stdin, adds --time-limit and --output-dir, and uses distinct exit codes
- New settings: enablebackgroundprocesses, enablesubagents, cursorblink, topp, maxoutputtokens, apistreamidletimeout
- Bundled local Devstral model removed; default model now has a 256k context window and off/high thinking levels
-
New Rust CLI onboarding, /loop scheduling, teleport to Vibe Code Web
- Rust CLI gets a first-run setup wizard with theme choice, keyring API key storage and browser sign-in; --setup reruns it
- Rust CLI can send a session to Vibe Code Web with /teleport, &<prompt> or --teleport; VS Code extension gains project links and teleport
- Rust CLI adds /loop for recurring natural-language prompts, /voice settings, composer undo/redo and /skill-name mentions in prompts
- Vibe now always runs on the Unified Harness and exits with a clear error if the runtime is missing, instead of falling back to the legacy engine
- Session titles are on by default for the CLI, and [utilitymodels] lets you pick the model behind titles and smart approve
-
Archive local sessions, plus many Rust CLI and security fixes
- Local sessions can be archived to hide them from active lists, and managed worktrees are cleaned up afterwards
- Rust CLI pins the live todo plan above the input; Cmd+\ opens the full plan in a right sidebar
- Rust CLI supports configuring Vibe Code remote projects
- Vibe Desktop shows unseen indicators for local Code sessions that need attention or finish a turn
- Path allowlist and shell wildcard checks tightened, so some outside-workdir accesses may ask for approval again
-
Teleport works with Vibe and workspace API keys, plus Windows session fixes
- /teleport now works with Vibe and workspace API keys, including free accounts, with no internal feature gate
- Clearer message explaining which API keys to use when teleport rejects a Codestral key
- Warns when managed configuration contains removed Vibe Code settings
- Windows: opening an already-open session says 'Session is already open'; idle-session cleanup no longer mistakes it for idle
- Automatic project context no longer runs repository-configured Git clean or process filters
-
Unified Harness leaves experimental; tighter shell approvals and hooks
- The Unified Harness is no longer labeled experimental; --legacy-harness is the documented way back to the legacy Python harness
- Subagent transcripts can be viewed and switched to from the prompt, with live subagent status
- Skills can be turned on or off from the /skills browser, and authors can mark skills explicit-only
- MCP and connector tool calls now follow the permission system and ask before running by default
- Shell approvals no longer widen to different programs, and smart approve and hooks behave more reliably
-
Security fixes for shell permission bypasses, plus connector and smart-approve fixes
- Shell permission checks now require approval for risky syntax and command options that could bypass workspace and denylist controls (five CVEs fixed)
- Smart-approve now asks for confirmation when a risky action is requested again, and shows why it is asking
- Auto-approved calls show the smart-approve reason as a note instead of a warning
- One invalid connector no longer stops the others from loading; oversized tool schemas are dropped individually
- @ file mentions now resolve against every workspace root, including --add-dir ones, and outside-workspace mentions stay plain text
-
New /branch command forks a conversation into a resumable session
- New /branch command forks the current conversation into a new resumable session and leaves the current one unchanged
- Resume the forked copy in another terminal with vibe --resume <id>
- @ file mentions now use Git-aware discovery and accept pasted files and folders
- Retained chats stay readable and restore their worktrees when work resumes
- New session logs are no longer readable by other users on POSIX systems
-
VS Code debug menu with session stats panel, plus worktree and permission fixes
- VS Code extension gets a debug submenu: log level picker, open log file, and a togglable session status panel
- Status panel shows agent stats: steps, tokens, tokens/sec, context and cost
- Session picker opens faster, and no longer drops a session that is touched while paging
- Selected permission mode is now kept when a session starts in a worktree
- A failed "always" permission save to ~/.vibe now warns instead of failing silently
-
Smart approve fixes, folded tool call groups and worktree plugin
- Tool call groups fold into one compact summary line by default; click to expand or press Ctrl+O to toggle all
- New built-in worktree plugin with a skill to create, reuse and clean up git worktrees
- Risky calls go back to the model to self-correct, escalating to an approval prompt only after a streak
- Smart approve fixes: classifier always uses the fast Mistral model, and secret paths behind cd or in git diff are no longer auto-approved
- Removed an unauthenticated debugpy listener on localhost:5678 that DEBUGMODE=true enabled under vibe-acp
-
Skills browser, plugin MCP servers in /mcp, and queued prompt controls
- New experimental /skills browser with search and filter, off by default; shared skills sync on session start
- Plugin-declared MCP servers now appear in /mcp and can sign in via /mcp login
- Prompt queueing gains select, edit, remove, pause and resume controls
- vibe update checks for and installs a newer release; vibe --yolo now selects the auto-approve profile
- Devstral Small removed from the default models list
-
Faster startup and fuzzy search in the MCP servers browser
- Fuzzy search added to the MCP servers and connectors browser
- Faster startup through additional lazy imports and deferred schema builds
- Trust prompt is skipped when --worktree or --trust already grants access
- Themes now apply correctly at startup
- Links are normalized before opening to prevent broken clicks
-
AI-generated session titles and editable message queue
- Sessions and the terminal window now get titles generated by the LLM
- Queued messages can be selected and edited before they are sent
- Git info now appears in the session header
- Model connections are reused across tool turns for faster responses
- Rate-limited calls retry with adaptive backoff, and /mcp login now routes to the connector auth flow
-
Slash commands and pickers now work while the agent is running
- Slash commands and setting pickers can be used while the agent or a command is running
- Models are shown by display name instead of the raw config alias
- Connector bootstrap failures are now surfaced instead of reporting no connectors
- The process is named Vibe CLI and the bottom bar shows its PID
- Python dependencies updated to clear known security advisories
-
New in-app session picker, plus faster startup and resume
- In-app session picker with live preview
- Faster startup, as the TUI now renders before session initialization, and faster session resume
- New /log-level command to control log level at runtime, and /retry is now exposed to ACP clients
- Worktrees appear in the transcript and are cleaned up when their local session is deleted
- Fixes: new worktrees branch from the remote default, configured temperature is respected for reasoning models, and tool-permission bypass reaches child agents
-
Ghost-text skill completion, OpenTelemetry tracing and many fixes
- Inline ghost-text completion for skills typed mid-prompt
- OpenTelemetry tracing configuration is now available
- /clear and /new accept an optional first prompt and show the resume id
- Default agent renamed to ask and now defaults to accept-edits
- Compacted conversation history is preserved and resume is faster
-
Admin config layer, an unpinned default model option and faster /resume
- New admin config layer for shared, enforced settings that apply over user config
- New Default (unpinned) model option in /model and /config tracks the recommended model without pinning
- User and trusted project config now compose, with project config overlaying user config instead of replacing it
- Explore subagent can load skills, and local sessions start in the selected worktree
- Fixes for file drag-and-drop in iTerm2 and Ghostty, subagent handoff after compaction, and a faster /resume listing
-
New /retry command and choice of fork or rewind in place
- New /retry command retries an interrupted response
- /rewind now lets you choose between forking and rewinding in place
- LLM retries are now shown in the UI and logs
- /status shows provider prompt-cache usage and discounts cached tokens from the session cost
- Fixes for copy formatting, duplicate approval widgets and context token display on session resume
-
Skill-creator, searchable /config and cache-hit token stats
- New built-in skill-creator: a guided flow to author, update and delete your own skills
- /config is now a searchable, full-screen settings browser with typed edit modals and per-layer origin display
- Browser sign-in can target a custom login domain for self-hosted or private-gateway deployments
- Session stats now track cached tokens, which were previously always reported as zero
- App-server protocol extended and shipped as a standalone vibe-app-server binary for desktop integration
-
Simpler MCP server setup, auto theme and faster scrolling
- New vibe mcp add and vibe mcp remove commands manage remote MCP servers without editing TOML
- Auto theme follows the terminal or OS appearance
- Tool calls render in a denser, grouped layout in the TUI
- Scrolling in long conversations is smoother and faster, with lag spikes removed
- Trust dialog is simplified and defaults to Trust folder; tool descriptions are compressed to cut context use
-
New /new command and review of agent changes from ACP clients
- ACP clients can now approve or revert agent changes, with live refresh
- New /new command works as an alias for /clear
- readfile no longer asks for approval on plan files in plan mode
- Configuration handling fully migrated to ConfigOrchestrator; config.toml now stores only user-set values
- Teleport context summary flag removed
-
Hooks are now stable, with renamed hook types (breaking)
- Hooks graduate from experimental: hooks.toml is stable and hooks now load whenever declared, with no flag needed
- Breaking: hook types renamed to postagent, pretool and posttool, and the hookeventname payload field matches; update existing hooks.toml files
- The enableexperimentalhooks config flag and VIBEENABLEEXPERIMENTALHOOKS env var are removed
- @file mentions now inject as readfile tool calls
- MCP server connection errors now show in the TUI, and MCP discovery failures show in ACP sessions
-
Smarter Windows shell handling, double-Esc rewind and richer teleports
- Rewind now triggers on a double Esc when the input is empty, with improved keybindings
- Teleports now include summarized context
- Windows command prompt is now shell-aware, with hardened bash tool permissions and cross-platform shell handling
- askuserquestion selection works with the mouse and supports free-text entry
- MCP OAuth tokens are no longer dropped on transient refresh errors
-
Managed bash tool, tool-disabling flag and smarter compaction
- Experimental managed bash tool for running and tracking shell sessions
- New --disabled-tools CLI flag turns off specific tools for a run
- Per-tool description overrides via prompts/<name>.md in the tools directory
- New setting to clear context when a plan is accepted
- Discovered AGENTS.md files are now surfaced during readfile calls; compaction is more reliable with a summary fallback
-
New --worktree option and ACP rewind support
- New --worktree NAME option creates or reuses a git worktree and runs inside it
- Rewind is now supported over ACP for both the agent and the host
- The teleport URL in the CLI is now clickable
- Higher context limit for Leanstral
- Fixed a crash when submitting an empty multi-select question
-
Edit diffs show whole lines; auto-approve works in lean mode
- Edit diffs now show the whole line of content
- Auto-approve now works in lean mode
- Declined or skipped tool calls render as a muted square
- Long tool call titles wrap instead of being cropped
- Fixes for --continue resume, a MarkupError crash on square brackets, duplicate mcpservers names, teleport remotes and ACP defaultagent
-
j/k list navigation, exit confirmation option and more askuserquestion options
- j/k navigation in selectable lists across the TUI, including questions, theme picker, rewind, voice and MCP panels
- New askconfirmationonexit config option prompts before quitting
- Project-level .vibe/config.toml now persists config option changes
- askuserquestion now supports more than 4 options and questions
- No longer prompts to log in for disabled MCP servers; teleport diff no longer mutates the real git index
-
Crash reporting added; fixes a crash on square brackets in tool output
- Adds Sentry crash reporting for unhandled exceptions in the TUI, only when telemetry is enabled
- Recoverable tool-call errors are now muted in the UI; only terminal errors show as hard failures
- Fixes a MarkupError crash when tool output contained square brackets
- Restores missing OpenTelemetry chat and LLM spans after the Mistral Python SDK 2.4.10+ telemetry opt-in change
-
Add OAuth MCP servers with /mcp add, plus a login crash fix
- New /mcp add slash command for adding OAuth MCP servers
- Bare exit, quit, :q and :quit now act as slash commands instead of being sent as prompts
- Fixed MCP OAuth login crash when the keyring backend can't be loaded
- Petit chat animation now pauses when idle
-
Clickable URLs, token usage display and macOS image paste
- Web fetch and web search output now has clickable URLs
- The TUI shows context window usage in tokens at the bottom
- Images can be pasted from the clipboard into the TUI on macOS
- Better MCP OAuth sign-in in the /mcp panel, faster UI startup and quicker connectors
- Fixes for macOS keychain access, heredoc commands wrongly denied, copying selected text and ACP hangs
-
OAuth login for MCP servers, API keys moved to OS keyring
- New /mcp login, /mcp logout and /mcp status commands authenticate OAuth-backed MCP servers from the TUI
- API keys are now stored in the OS keyring instead of plain text
- Edit diff view shows all replaced occurrences, not just the first
- New vibe --check-upgrade forces an immediate update check; --yolo is an alias for --auto-approve
- ACP accepts inline image content blocks; stdio MCP connections now persist across tool calls and 5xx responses are retried
-
Faster /resume and a gentler update-failure message
- /resume now opens much faster
- Startup update-failure message is gentler and no longer alarms when an update can't be applied
- ACP workspace trust is now gated behind the workspace-trust client capability
-
Syntax-highlighted diffs and fuzzy slash command search
- Slash command autocomplete now uses fuzzy search
- File edit diffs are syntax-highlighted, line-numbered and follow the theme
- /resume now limits the session picker to the current folder
- Clipboard copy confirmations appear inline instead of stacking toasts
- Fixes for max-token responses, context-too-long errors, SSE parsing crashes, the Free plan banner and reasoning content sent with thinking off
-
Experimental tool hooks, a message queue and collapsed tool output
- Experimental beforetool and aftertool hooks in hooks.toml can deny a call, rewrite tool inputs or append context to the output
- Messages typed while the agent or a bash command is running are now queued and shown above the input
- Tool result output is collapsed by default, and common read-only shell commands like ls and cat run without approval
- Compaction now re-injects prior user messages so the agent keeps the original task goals
- Breaking for experimental hooks: postagentturn retries must exit 0 and return a deny decision on stdout, not exit code 2
-
ACP gets /teleport; fixes for JetBrains and Zed integrations
- The /teleport slash command is now available over ACP, matching the TUI command, for IDE integrations
- Vibe now prompts at startup to install a pending update before the session continues
- JetBrains IDEs no longer trigger a preemptive auth prompt over ACP or drop terminal arguments
- Zed now registers slash commands like /help instead of rejecting them, because the initial advertisement is delayed
- Built-in tool prompts now point to edit instead of the removed searchreplace name
-
Image attachments via @-mentions and safer file writing
- Attach images with @-mentions in the TUI for vision-capable models
- Session deletion is now available over ACP as the session/delete extension method
- Browser sign-in shows a copyable fallback URL, so SSH and remote sessions can finish onboarding
- writefile is now create-only and refuses to overwrite existing files; read and edit use a new tool-call format
- LLM calls now retry on network errors and timeouts
-
System trust store option for corporate TLS and private CAs
- New enablesystemtruststore config flag switches the shared SSL context to the OS trust store for corporate TLS and private CA setups
- MCP HTTP transport now uses the shared SSL context, so it honors SSLCERTFILE, SSLCERTDIR and the system trust store opt-in
- Compaction summaries are now injected into the conversation instead of replacing it
- Fixes a crash during initialization
- API key onboarding and plan-upgrade prompts now link to the new Mistral Vibe Code extensions page
-
/teleport moves to Vibe Code Web sessions; banner link fixed
- /teleport now uses the new Vibe Code Web sessions
- The VS Code extension promo link in the CLI banner now points to the renamed mistralai.mistral-vibe-code extension
-
Custom compaction prompts, safer -p mode and browser sign-in by default
- Programmatic mode (-p) no longer auto-approves tool calls by default
- Custom compaction prompts via the compactionpromptid setting, loaded from ~/.vibe/prompts/ or .vibe/prompts/
- Browser sign-in is now on by default for Mistral providers, with no experimental flag needed
- New --max-tokens flag for programmatic mode
- /teleport to Vibe Code Web now carries the working diff and last commit
-
Shared skills from ~/.agents/skills, theme picker and in-CLI connector auth
- Skills now load from ~/.agents/skills so they can be shared across agents
- Unauthenticated connectors show in /mcp with needs auth or needs setup labels, and OAuth starts from inside the CLI
- Theme selection is back, with a theme picker during onboarding
- Newly discovered connectors are now disabled by default; existing choices are kept
- Current date is injected into the system prompt, and a minimal system prompt variant was added for eval and training
-
Friendlier session titles for @mentions, plus ACP terminal and MCP menu fixes
- Session titles now format @mention syntax into human-readable text
- Session title is sent on the first prompt so IDE session pickers show it immediately
- Parallel bash tool calls in ACP each get their own live terminal instead of racing on shared state
- MCP menu no longer shows D/E shortcuts in the detail view when there are no tools to enable or disable
- New layered config read path built on TomlFileLayer, ConfigBuilder and ConfigOrchestrator
-
Add extra repo folders to a session, plus editable plan mode
- New --add-dir flag pulls additional repository roots into a session
- Plan mode is easier to read, with a live-editable plan you can open for editing with Ctrl+G
- --continue is now scoped to the current terminal (TTY)
- Subagents now share the session's permissions
- Tool approvals are debounced so they no longer interrupt you while typing
-
Syntax-highlighted diffs in IDE agent, plus SSL and allowlist fixes
- Syntax-highlighted file diffs for writefile and searchreplace in the IDE agent webview
- Spinner and loader now shown for ! (bang) command output
- Bash allowlist patterns are saved without the wildcard suffix
- One combined SSL context is now used for all outbound HTTPS requests
- Non-interactive tools are hidden from the LLM in ACP and programmatic mode
-
New /loop command to run prompts on a recurring interval
- New /loop command runs a prompt or slash command on a recurring interval
- New defaultagent config setting
- enabletelemetry now takes precedence over enableotel
- Logging environment variables are now documented in --help
- Fixes for retry handling in Temporal activities, history file clobbering across parallel instances, compacted session IDs and TUI scroll sensitivity
-
New /rename command and persistent "Always allow" tool permissions
- New /rename command renames the current session
- "Always allow" tool permissions now persist across sessions
- Bang commands (!command) run via an async subprocess for better latency
- /clear no longer chains parentsessionid to the previous session
- Agent-loop warmup is now eager, so the ready telemetry event fires sooner
-
Default model moves to Mistral Medium 3.5; OAuth for connectors in /mcp
- Default model is now mistral-medium-3.5
- Connectors can sign in with an OAuth flow from the /mcp menu
- New extraheaders field on ProviderConfig for custom provider headers
- ACP tool results now carry structured metadata
- More telemetry: user-cancelled actions and a new-session event on every session reset
-
Hooks, scratchpad for subagents, and /copy command
- Experimental hooks system with a post-agent-turn lifecycle
- Scratchpad directory for temporary working files shared with subagents
- New /copy slash command and custom compaction instructions via /compact
- Enable or disable MCP servers and individual tools from the /mcp menu
- autoapprove config renamed to bypasstoolpermissions; --continue now only finds sessions from the current directory
-
Built-in skills system and MCP server working directory setting
- Adds a builtin skills system, starting with a self-awareness skill
- MCP stdio servers can now be given a cwd (working directory) setting
- /connectors now works as an alias for /mcp, and R refreshes the MCP browser
- The /terminal-setup command is removed
- Fixes the /debug command throwing an error and a race condition in banner startup
-
1M context window and max thinking budget for Opus, plus streaming fixes
- Opus now gets a 1M context window and the maximum thinking budget
- Fixed markdown fence context loss that caused streaming rendering problems
- Fixed proxy chain URLs in apibase parsing
- Added an "Other" task type for non-code requests in the CLI prompt
- Removed the Alt+Left and Alt+Right key bindings from chat input
-
Trust folder dialog shows detected files and risks; faster startup
- Trust folder dialog now displays detected files and LLM-assessed risks
- Text-to-speech through the Mistral SDK, with telemetry tracking
- MCP and git I/O deferred to a background thread for faster CLI startup
- Telemetry URL is now configurable
- Fixed encoding detection fallback for non-UTF-8 files
-
Console View for debugging, plus a new /mcp command
- Console View added for enhanced debugging and monitoring
- New /mcp command shows MCP servers and their status
- Manual command output can now be forwarded to the agent context
- Faster startup through lazy-loaded dependencies and quicker folder parsing
- Fewer redundant permission prompts for parallel tool calls needing the same permission
-
New /data-retention command to view data retention and privacy settings
- Adds a /data-retention slash command that shows Mistral AI's data retention notice and privacy settings
-
Word-wise cursor shortcuts in chat input, plus Linux build and MCP fixes
- Alt+Left and Alt+Right now move the cursor word by word in the chat input
- Fixed a broken build on Linux
- Errored MCP servers are no longer counted in the banner
- Better bash denylist matching and clearer error messages
- Command messages are skipped when navigating with rewind
-
Reasoning effort setting for supported models, plus ACP message IDs
- New reasoning effort parameter for models that support it
- ACP message-id support for reliable message boundary identification
- Non-standard HTTP 529 errors are now handled gracefully and retried
- Scrolling is smoother, with sensitivity reduced to 1 line per tick
- Fixed text selection errors when copying from unmounting components
-
Rewind mode to navigate and fork conversation history
- New rewind mode lets you navigate back through conversation history and fork from an earlier point
- Streaming LLM chunks now keep their messageid when aggregated
- Better error handling for SDK response errors
-
Dependency pinned back; ACP context usage updates removed
- Context usage updates via ACP are removed
- Agent-client-protocol dependency pinned back to 0.8.1
-
OTEL tracing, text-to-speech and a Skill tool for task lists
- OTEL tracing support for observability
- Text-to-speech (TTS) functionality added
- New Skill tool for managing task lists and workflows
- Standalone --resume command opens a session picker, and /model now uses a list-based picker
- Fixes for bash allowlist/denylist on Windows, shift+space, agent switching during streaming and scrolling performance
-
Theorem-proving agent, voice mode and parallel tool execution
- New dedicated theorem proving agent powered by leanstral, set up with /leanstall
- Voice mode with real-time transcription
- Parallel tool execution for improved performance
- AGENTS.md now works at user level (~/.vibe/), in subfolders and in parent folders
- Bash allowlist and denylist now active on Windows; find removed from default allowlist to prevent -exec abuse
-
Skills now extract arguments; update toast stays until seen
- Skills now extract arguments when invoked, improving parameter handling
- Update notification toast no longer times out, so the restart prompt is always seen
- Auto-compact threshold falls back to the global setting when not set at model level
- Session ID added to telemetry events for better tracing
- Smaller Vibe Code payload after removing filecontentbefore
-
SDK-friendly harness file loading and web search config fix
- New HarnessFilesManager loads harness files selectively, so the SDK can be used without touching the file system
- Web search tool now infers its server URL from the provider config instead of a hardcoded production API
- askuserquestions tool is disabled in prompt mode
- Space key fix now covers all input widgets, including question prompts and proxy setup, in the VS Code terminal
-
Reasoning effort setting, per-model auto-compact and faster agent switching
- Reasoning effort can now be configured, with a thinking blocks adapter
- Auto-compact threshold is now set per model
- Faster agent switching: expensive file scan removed from the system prompt and git operations cached
- Plan mode improved and the CLI banner now shows your user plan
- Fixes for Space key in VSCode 1.110+, arrow-key history at wrapped lines and an update notifier crash
-
Web search and fetch tools, /resume command, and MCP sampling
- New web search and web fetch tools for retrieving and searching web content
- New /resume command to choose which session to resume
- MCP sampling support: MCP servers can request LLM completions
- Agents, tools and skills are now discovered in subfolders of trusted directories (monorepo support)
- Terminal bell and window title change when action is required or a task completes
-
Better clipboard copy and Ctrl+Z backgrounding, plus faster streaming
- Clipboard copy tries OSC52 first, then falls back to pyperclip when a system clipboard is available
- Ctrl+Z puts Vibe in the background
- Streaming and scrolling performance improved
- File watcher is now off by default and can be enabled in config
- Fixes for the skills count banner, token percentage after /clear, and Vertex AI credential caching
-
Google Vertex AI support and Agent Skills discovery
- Adds Google Vertex AI support
- Discovers skills from .agents/skills/ (Agent Skills standard) as well as .vibe/skills/
- ACP gains session/load and session/list for loading and listing sessions
- Sends user interaction and tool usage telemetry, configurable via enabletelemetry
- Adds a Proxy Wizard (proof of concept) for CLI and ACP, plus proxy setup and JetBrains ACP registry docs
-
Redesigned CLI interface and faster loading of long sessions
- New look and layout for the CLI, replacing the separate onboarding theme selection screen
- Long sessions now load incrementally: 20 messages at a time, with Load more for older ones
- ACP support for thinking, using agent-client-protocol 0.8.0
- Env file path can now be a FIFO
- Fixed agent prompt overrides being ignored and inaccurate token counts in the UI
-
Dotenv overrides, custom rate limit messages and session fixes
- Environment variables can now be overridden by dotenv files
- Custom rate limit messages are shown depending on plan type
- Fixed duplicate messages in persisted sessions
- Fixed ACP bash tool so chained commands are passed as the full command string
- Fixed global agent prompt not loading correctly; no longer offers to resume when there is nothing to resume
-
Subagents, custom slash commands via skills, and auto-update
- Subagent support added
- AskUserQuestion tool lets the agent ask for interactive user input
- User-defined slash commands through skills
- Auto-update feature and a what's new message shown after updating
- Modes refactored into Agents; instructions.md and the workdir config setting are removed
-
User-definable reasoning setting and a spinner rendering fix
- Adds a user-definable reasoning field
- Fixes a rendering issue with the spinner
-
Agent Skills and reasoning support, plus native terminal themes
- Adds agentskills.io support
- Adds reasoning support
- Adds native terminal theme support
- Fixes -p mode so it auto-approves tool calls
- Fixes a crash when switching mode and some cases where clipboard copy failed
-
Modular mode system and trusted folders for local .vibe directories
- Adds a modular mode system
- Adds a trusted folder mechanism for local .vibe directories
- Documents public setup of vibe-acp in Zed, JetBrains and Neovim
- Adds a --version flag
- Fixes a segmentation fault on exit and extra spacing around assistant messages; UI and performance tweaks
-
Clipboard and chat scrolling improvements, plus a security fix
- Fixes a command injection issue in the GitHub Action prompt handling
- Adds key bindings to scroll the chat history, with better scrolling while output streams
- Adds more copy-to-clipboard methods to cover all cases
- Update checks now run at most once a day and use PyPI
- Fixes tool permission handling for the allow always option in ACP, and fixes issues with vLLM
-
New config option to turn off commit signing
- Added includecommitsignature in config.toml so you can disable signing commits
-
Context length doubled to 200k, plus a copy-paste crash fix
- Context length improved from 100k to 200k
- Fixed a rare crash when copy-pasting
Smaller updates 20
-
Fixes a crash when @-mentioning a file
-
/help sorted alphabetically; /teleport always available
-
Fixes a Textual version problem
-
Teleport shows the latest GitHub connection status
-
Fixes the changelog and what's new display
-
Looser agent-client-protocol version requirement
-
Fixes text encoding issues on Windows
-
Fixes the bash tool not being discovered by vibe-acp
-
Bug fixes for safety, permissions and long-session speed
-
Fixes config desyncing issues
-
Fixes a crash when continuing a conversation
-
Bug fixes and internal cleanup
-
Clearer error and no trust prompt when run in home directory
-
ACP terminal-auth gated on client support; user-agent header fix
-
Faster Textual UI, bash timeout config fix, and README improvements
-
Fixes streaming with the OpenAI adapter
-
Zed extension packaging fixes and cleanup
-
Packaging fix for the Zed extension release process
-
Fixes the setup flow for vibe-acp builds
-
Fixes the update notification
Showing the newest 100 of 101.