Dodaios is an open-source agent OS for your desktop: it runs your coding agents, browser and apps by voice or chat, on your own machine. Formerly Personal Jarvis.
Changelog
All notable changes in Dodaios (called Personal Jarvis until October 2026).
Format based on Keep a Changelog, versioning per SemVer.
[Unreleased]
Fixed
- The README contributor wall shows everyone who has shipped work. People whose git email GitHub never linked to an account are named, and a pull request author appears even when the contributors index has not caught up yet.
Changed
- The small-machine vision download is
qwen3-vl:2b-instruct(1.9 GB, refreshed in the library). The previous tagqwen3-vl:2bis still published but Ollama shows it as a year old, so a new install no longer offers it. - The
personal-jarvispackage on PyPI is now a thin pointer tododaiosinstead of a second copy of the app: it installsdodaiosat the same version, maps every extra (personal-jarvis[full]installsdodaios[full]) and keeps every earlier command (jarvis,personal-jarvis,jarvisctl, …) working, includinguvx --from personal-jarvis. To move an existing pip install to the new name, runpip uninstall -y personal-jarvisand thenpip install --force-reinstall dodaios. Ifpip install -U personal-jarvisleaves an install that no longer starts, the command prints the one-line repair (pip install --force-reinstall --no-deps dodaios==<version>). - The Windows setup is built as
Dodaios-Setup-x64.exeand shows the product name Dodaios in its file properties.PersonalJarvis-Setup-x64.exeis still published with identical bytes for installs from before the rename. - The Linux package file is named
dodaios_<version>_amd64.deb, after the package inside it. - The app identifies itself as Dodaios (User-Agent
Dodaios/1.0,Dodaios-Updater) to GitHub and the plugin services it calls.
[4.1.2] — 2026-10-27
If your install is on version 3.1.1 or older, the in-app update no longer sees new releases. Run the one-line installer again to update.
Security
- The in-app updater of an installed app now installs an update only when the
release's update manifest is signed by the Dodaios release key (Ed25519) and
names exactly the version being installed, so neither swapped release files
nor an older release served as a newer one get installed. Releases built from
now on carry
installers-update-manifest.jsonand its signature (#810).
[4.1.1] — 2026-10-26
4.1.1 is the first published release of 4.1: the 4.1.0 build was held back by a security check on its dependencies. Everything listed under 4.1.0 below ships with it.
Added
- Redesign Built for you so anyone gets it at a glance (#865)
Fixed
- Raise pillow, pypdf, gitpython, filelock, certifi and opentelemetry past published advisories
[4.1.0] — 2026-10-26
Added
- Realtime 3D decks with third-person, first-person and bird's-eye views, plus a 1v1 duel mode (#866)
- Open a printed path with a plain click (#853)
- Add per-agent code sandboxes (#498)
- Grounded initiative without unrequested actions (#466)
- Add the Dodaios pet, show it as the Mascot style and give it a voxel model in the Verse (#542)
- Need neither Python nor Git before the one-line installer (#818)
- Start and update the app from any terminal (#368)
- Bill the default provider's subscription for the assistant's own learning loop (#801)
- Close third-person camera on 192 px/m decks with close-up characters (#808)
- Ctrl+click opens files outside the workspace in the editor (#799)
- Steer the commander in third person, bird's-eye on V (#791)
- Show why unattended routine runs and reviews were skipped or dropped (#790)
- Prove a drafted skill by running its steps read-only before it goes live (#793)
- Open a code build's coding thread from its build detail (#788)
- Mine the person's pane prompts and agent chats; measure memory use by reliance (#792)
- Snap the area picker to the parts of a window, not only the window (#785)
- Richer operations deck, tailored uniforms, trousers by rank tier (#784)
- Rebuild the Verse as a photoreal fleet capital ship with a Blender content pipeline (#782)
Changed
- Keep Hermes and OpenClaw agents warm between turns (#817)
- Rename every jarvis-named file and folder to Dodaios (#798)
Fixed
- Keep sherpa-onnx below 2.0, which ships no macOS Intel wheel
- File only the pane prompts this app announced (#881)
- Show every agent of the folder in the Agents tab by default (#876)
- Keep the terminal grid inside its frame when the side panel opens (#874)
- Observe the prompts typed into panes and keep tests out of live data (#867)
- Show session changes an agent made in its own worktrees (#870)
- Keep the plain download hint for a Whisper model never downloaded (#860)
- Repair taskbar pins left on an old install (#861)
- Refresh the contributor wall through a bot pull request (#857)
- Report cancelled nightly gate failures (#825)
- Run Claude subscription agents through the official Claude CLI (#856)
- Pin setuptools above advisory range (#842)
- Repair PATH at Codex spawn boundaries (#844)
- Use native audio for Gemini in WebKitGTK (#843)
- Align Whisper cache readiness (#841)
- Preserve model-less local provider discovery (#838)
- Avoid invalid AltGr escape warning (#827)
- Accept valid probe after open wait race (#823)
- Isolate Windows platform simulation (#819)
- Avoid conflicting Anthropic credentials (#834)
- Restore native controls and minimized detached windows (#452)
- Say why a voice start was refused instead of guessing (#552)
- Cross a stuck provider instead of losing the words, and ship a default vocabulary (#815)
- Keep a browser OAuth grant when its capability check fails (#814)
- Release a turn that ended without a result; close settled approval cards (#314)
- Reuse the host Python only for the exact requested minor (#813)
- Sandbox every inline artifact download, not just HTML and SVG (#812)
- Never kill a writing git, heal a stale lock, stop bundle line-ending noise (#807)
- Show a closed panel window again when the page refuses to close (#805)
- Links clicked in the IDE open in its Browser tab (#804)
- Ctrl+click on a drive-less path opens the file, not the folder tree (#803)
- Codex on its plan without the app's keys, Linux idle fallbacks, live-run fixes for the code builder (#797)
- New Claude Code panes keep the person's own renderer (#800)
- Decide when the person is quiet in Dodaios, not when the computer is idle (#796)
- Stop a release cut early when its version is already tagged (#795)
- Offer only the big parts of a window when hovering in the area picker (#794)
- Keep tokens out of logged tracebacks (#789)
- Exact pane scrollbar for Claude Code and Codex (#787)
- Count habits on this computer's zone until the UI reports one, decide only while the person is idle (#786)
[4.0.1] — 2026-10-20
4.0.1 is the first published release of 4.0: the 4.0.0 build was held back by a security check on a dependency. Everything listed under 4.0.0 below ships with it, including the move to the Dodaios name for the program, its folders and its stored keys.
Fixed
- Raise mcp to 1.30.0 for CVE-2026-59951
[4.0.0] — 2026-10-20
Dodaios all the way down
Personal Jarvis became Dodaios in 3.2.0 on the outside. 4.0.0 finishes the job on the inside: the program, its folders and its stored keys now carry the Dodaios name too. Nothing is lost on the way, and nothing needs to be done by hand.
What moves on the first start of 4.0.0
- The Python package is
dodaios(it wasjarvis). Old start commands such aspython -m jarvis,jarvis,jarvisctl, autostart entries and MCP registrations in other tools keep working. - The data folder becomes
%LOCALAPPDATA%\Dodaioson Windows and~/.dodaios-dataelsewhere. It is renamed once, in place, and a link stays at the old path, so an older version still finds everything. If the folder is in use, the app keeps running where it is and tries again next time. - Saved keys move to the
dodaiosentry of the system credential store. The old entries stay readable and stay in place, so a downgrade keeps working. - The one-line install lives in
~/.dodaios(it was~/.personal-jarvis); the next run of the installer moves it. - Windows: the app is
DodaiosApp.exe, the Start menu entryDodaios, the logon taskDodaios Autostart. Existing pins and the existing autostart task keep working. - macOS: the app is
Dodaios.app. It keeps its bundle id, so every permission you granted stays granted. - Linux: the desktop entry is
dodaios.desktop, the Debian packagedodaios(it replacespersonal-jarvis). - The configuration file stays
jarvis.tomlon existing installs; new installs start withdodaios.toml.
Downloads and packages
- Installers are published as
Dodaios-*and, for installs from before the rename, also under their oldPersonalJarvis-*names (byte-identical copies). - On PyPI the project is now
dodaios.personal-jarvisis published alongside it for a transition period.
If you installed before 3.2.0, run the installer once more (see the 3.2.0 notes); from 3.2.0 on, updates arrive in the app.
Added
- Show the user's own Chrome, signed in with their profile, in the browser panel (#747)
- Show the agent done gate's last verdict on each pane (#748)
- Keep the desktop app on the newest main (#742)
- Move the remaining spoken phrase tables into the catalog, translate turn-error codes (#737)
- Wrap long lines at the editor's edge, like VS Code (#738)
- Fold the repository row, and show what one agent session changed from the Git tab or the pane header (#732)
- Blue binary-digit cursor field on the empty front page (#728)
- Source control panel in the Git tab (#727)
- A hard-surface courier and flying made for fun (#688)
- Pick each member's seat, model and effort; one seat twice; swarms of up to eight (#692)
- Rebuild the Memory section as a linked, path-addressed vault reader (#678)
- Follow mode reads one file across several columns (#686)
- Redesign the coding floor as a tech campus floor (#689)
- Side chat upgrades: dictation, live thinking, code blocks (#682)
- A sky pad off the agents floor and a courier to fly over the city (#683)
- The coding floor as a professional office in the tower (#681)
- A team room with a live collaboration canvas and the chats beside it (#670)
- Rebuild the coding floor as the Foundry deck (#674)
- The city in the singularity outside the agents floor (#671)
- Ask a named agent and hear its answer in the same call (#676)
- Show exact content on screen, and open one thing in the app (#641)
- Open the plugin section full-window with new area icons (#669)
- Replace the Prompt tab with a plain AI chat (#647)
- A planner-and-builder mode with a living, versioned plan (#667)
- Report private memory, peak, threads and an app total in the live snapshot (#660)
- Rebuild the plugin section as a directory with a sidebar (#639)
- Show a team thread as a live board, a lane per member and a band per step (#653)
- Make 3D agent shapes livelier (#643)
- A calm working animation for agent faces (#652)
- Find wiki pages by meaning, next to the word search (#625)
- Team threads, several coding agents carry one task as peers (#613)
- Move every technical name to Dodaios with a lossless migration (#750)
Changed
- Move the Python package from jarvis/ to dodaios/ (#733)
- Rename the Jarvis-named sub-folders to Dodaios (#661)
- Ask for named checks only where the Git tab shows them (#663)
- Load the wake runtime without its verifier-training stack (#651)
- Show how many checks passed and failed with the first GitHub answer (#645)
Fixed
- Open a session's file in the editor's diff, and filter the Git tab only once per pane request (#744)
- Stop the busy cursor flashing while the wake word listens (#740)
- Keep the conversation language through errors, default system phrases to English (#731)
- Turn the binary cursor field off over the composer (#730)
- Write the startup entry in-process, without spawning PowerShell (#724)
- Resolve Vosk models from the install data directory (#722)
- Match the Start-menu shortcut when the install path has a short name (#699)
- Fly the courier by mouse aim, with cursor steering when the mouse cannot be captured (#698)
- Report finished work to a live call on an install without a speech pipeline (#697)
- Finish the visible rename to Dodaios (#694)
- Fit every named tool into a small model's budget, and watch the one busy agent (#695)
- Keep highlighted code on its code block's background in dark mode (#691)
- Keep an answer the call already said from being announced again (#687)
- Give the call every tool its rules name, under readable names (#684)
- Show a pane's branch status at once and spin while it is confirmed (#668)
- Always give the call the tools its rules name (#677)
- Remove a startup shortcut left beside the logon task (#664)
- Let one pane renew an expired login before the others start (#665)
- Address the review of the rebuilt plugin directory (#666)
- A green nightly passes the strict gate, and the alert names the gate's reason (#658)
- Open the Agents section with roster and teams already loaded (#642)
- Polled reads of changed files no longer rewrite the index (#649)
[3.2.0] — 2026-10-13
Added
- Bring 3D agent shapes to life at rest (#624)
- Show everything running for the user on one board (#635)
- Make a team's Lead tab the chat with its lead (#594)
- One set of conversation rules for every live engine, and a report-back duty for agents and missions (#634)
- Record where a tool call's target came from (#631)
- Answer "what is running, what is done" in one call and watch hand-started agents (#629)
- Report how often each tool actually ran (#628)
- Turn the Git tab into a control center for git and GitHub (#600)
- Show a model the tools it needs and keep the rest reachable (#618)
- Let the stars of a starry design twinkle on 3D agent faces (#615)
- Take several orders at once in a live call and in the chat (#595)
- Keep the motion of designs and cigar smoke on 3D agent faces (#606)
- Write the io.github.dodaios manifest namespace and read both (#583)
- Adopt every DODAIOS_ option in the programs that run before the package (#603)
- Pixel creatures as the default agent look, shapes as 3D figures (#545)
- Add the dodaios and dodaiosctl commands beside the earlier names (#587)
- Always show the branch a pane works in, with its pull request (#584)
- Editor groups side by side and stacked, with tabs that drag between them (#564)
- Accept the DODAIOS_ prefix wherever JARVIS_ is read (#591)
- Refuse every move of local main to a commit no remote has (#585)
- Discover plugins under the dodaios.
entry-point groups (#579) - Offer the app's MCP servers as dodaios and keep reading jarvis (#581)
- Let a release tell the previous version where its files live (#567)
- Keep local main a mirror of origin/main (#576)
- Rebuild the agents floor as a team campus (#561)
- Group the sidebar chat list under one heading per day (#517)
- A drill file that reproduces a slow start in a running app (#573)
- Add the install identity migration engine (#571)
- Add European Portuguese (pt-PT) across UI, replies and voice (#551)
- Explorer right-click menu and a second editor to the side (#547)
- Rename the product to Dodaios in the Python runtime (#553)
- A lead for every agent team that plans, assigns and reviews on its own (#537)
- Show a loading loop on the bar and pets while a realtime call connects (#492)
- Rebuild the library as a wall of full-bleed tiles (#538)
- Make the hooded figure the Dodaios logo (#535)
Changed
- Give numpy and scipy one BLAS thread instead of one per core (#580)
- Give every identity name one source and pin the copies (#570)
Fixed
- Branch off instead of committing on the default branch, and never hide the local state (#632)
- Give the session branch list room for a full branch name (#612)
- Say when an account is not offered subscription voice (#627)
- Stop the avatar wall job from marking every merged pull request red (#623)
- Cap a merge group at two macOS jobs (#633)
- Show the lead under the name the user gave the assistant (#620)
- Finish the visible rename in the app, the vault templates and the Publish form (#616)
- Never let a local server shorten a prompt unnoticed (#611)
- Quarantine a hung Automation probe when the wait ran out (#609)
- Start a Windows worker command without PowerShell module discovery (#608)
- Fire a shortcut again while one of its keys stays held (#598)
- Label the assistant's lines by the name the user gave it (#577)
- Give a session its CLI left unnamed a real title (#593)
- Give a file that ran out of its budget one longer attempt (#589)
- Keep the app's PATH short enough to start programs (#575)
- Suppress duplicate desktop pets during boot measurements (#502)
- Make bypass permissions the project-wide default (#486)
- Wait for GitHub to compute mergeability before the merge train plans (#565)
- Link the frontend to the Dodaios/Dodaios repository (#563)
- Recover a lost team task from the member's chat and show a goal's text once (#559)
- Point the installer at the launcher entry and name the agents Dodaios-Agents (#560)
- Keep the bot token out of errors, status, logs and tracebacks (#554)
- Keep unattended agent runs off per-token API keys (#555)
- Title a pane whose session opened with a slash command (#550)
- Keep installs cloned before the rename on the update channel (#544)
- Preserve licenses and correct OAuth disclosures (#491)
- Name new agents from their chat role (#489)
- Point the remaining live URLs at Dodaios/Dodaios (#530)
- Point the installers at Dodaios/Dodaios (#526)
[3.1.1] — 2026-10-10
Fixed
- Never report a stopped event loop as stalled (#523)
- Follow the registry's GitHub Pages hosts to the Dodaios organization (#525)
[3.1.0] — 2026-10-09
Added
- Integrate the remaining local desktop work after 3.0.0 (#514)
Fixed
- Enqueue the release candidate; fix macOS non-terminal installs (#513)
[3.0.0] — 2026-10-09
Changed
- Built-in plugins that have not yet passed a full end-to-end sign-in test now show a Preview badge, and releases no longer wait for every plugin to pass that test. A plugin is marked verified only after its complete journey was observed.
[2.9.0] — 2026-10-05
Added
- Share agents as reusable templates without running a publishing server.
- Updated desktop interfaces and bundled assets for the integrated release.
Fixed
- Preserve browser tabs and stable native capture when handing control between a person and an agent.
- Keep delegated results attached to their originating chat and serialize direct chat sessions.
- Require a separate confirmation before ending a voice conversation.
- Bound Windows clipboard reads so dictation cannot remain blocked on delayed clipboard data.
- Retry brief file locks when saving dictation history so completed restorations are retained.
- Restore macOS desktop controls and request native permissions at the point of use.
- Defer optional browser setup on headless installations and repair cross-platform lifecycle behavior.
- Strengthen execution, credential storage, refresh and release-admission boundaries.
Known limitations
- Full end-to-end plugin acceptance is deferred for 2.9.0 under a one-time maintainer-approved exception. The 46 outstanding plugin journeys remain unverified, and some sign-ins still require publisher setup. Existing audit results are unchanged; this exception does not apply to later releases.
[2.8.0] — 2026-10-03
Added
- An editor for appshots, opened from the corner card
- Draw traces as quiet prose lines
- Change floors with the elevator doors and a real button panel
- A clickable, draggable appshot card in the screen corner
- Make the area picker's magnifier a round lens
- Ride the elevator from inside the car
- Put the assistant's files up front as pages with their content
- Plain magnifier the mouse wheel zooms
- Add isolated ChatGPT subscription access and in-app sign-in (#322)
- Elevator panel and ride that feel like a real elevator
- The assistant's profile section is named after the assistant
- Bring back the quiet session counts in the project tree
- Drop the workspace pane limit and the 4x4 grid bound
- Drop the running marks from the project tree
- Restyle the project tree as a quiet folder list
- Arcade floor with ten playable retro cabinets
- The SOUL.md section shows the assistant's whole profile
- Drag a stack of side-by-side seams as one line from a seam's ends
- Record the appshot shortcuts and default the area to both Shift keys
- Give the area picker a cleaner region-capture look
- Draw turns as a timeline
- Select an area of the screen on its own shortcut
- Tell a finished turn's work as a written report
- Jarvis follows the person as their chosen pet, in 3D
- Log the JavaScript a hung window is stuck in before recovering it
- Save "remember this" to MEMORY.md during the conversation
- The live call keeps SOUL.md current with update_soul
- Draw the user's pet everywhere the static Gigi image was
- Replace the thinking comets with thought waves
- SOUL.md gives the assistant one self-maintained identity
- Mark voice and typed chats apart in the history
- Start the local engine when its card is selected
- Right-click menu with delete on gallery cards
Changed
- Share one TLS context across every httpx client
- Import transport plugins off the loop during boot warm
- Run worktree git work off the event loop
- Probe openai stream_options lazily, not at import
Fixed
- Strip preview comments with a plain scan instead of a regex
- Keep pet lookups inside their folders and strip HTML comments fully
- Hold every automatic reload during a call or a streaming reply
- Verify local speech readiness and contain failed workers
- Keep the async-route and public-docs gates green after the local landing
- Explain the quiet exception handlers the local commits added
- Centre the cat torso above its stable floor anchor
- Keep coding floor agent labels expanded
- Keep lead pet centred on its floor marker
- Hold every agent's first prompt until its terminal questions stop
- Rebuild the arcade floor as clear rooms without lag
- Keep mission heartbeats out of calls that did not start them
- Make each shortcut one tidy field like the select beside it
- Keep the workspace menu off the switching spinner
- Let Continue through once a wake word is saved
- Keep the settings labels readable in the Settings dialog
- Link stacked seams up to 24 px apart and cover the grab zones
- Continue a reopened voice chat in its own history row
- Wait for Codex's colour question before typing a prompt
- Keep the mascot and bar out of captures from their first show
- Report a crashed area picker and serialize shortcut reloads
- Sweep a mission at once when its owner process died
- Stop terminal link detection walking past the buffer end
- Hide the area controls until the backend knows about them
- Seat a chat opened by a call on the person's chat model
- Let take_appshot capture the whole screen on an explicit request
- Make the size slider glide and resize the pet live
- Capture the app window, not the mascot or bar overlay
- Play the shutter flash and sound when the live model looks
- Show one pet while the answer streams under a live trace
- Keep tool results readable and long enough to narrate
- Start spawn-point agents without waiting for a viewer
- The pet plays its working row while Jarvis thinks or works
- Read the install dir from sys.path when the prefix is lost
- Find the worker interpreter when the branded exe stands alone
- Resolve misheard agent names and route Jarvis agents vs coding panes
- Give Jarvis and agents their own conversation chats
- Brief spawned agents to do the work, not a read-only audit
- Keep shared Google grants alive and flag dead OAuth clients
- Offer only tool-calling models on the Local voice card
- Retry the engine copy while Windows holds the new folder
- Never reload the window under a live voice call
[2.7.1] — 2026-10-02
Fixed
- Bind federation signatures to their endpoint
- Close first-contact bot pairing after a setup window
- Sandbox worker HTML even when opened in a browser tab
[2.7.0] — 2026-10-02
Added
- macOS asks when a feature needs a permission. The first time you use a feature that needs a macOS permission, Personal Jarvis asks right then and macOS shows its own dialog: the microphone at the first dictation, push-to-talk, voice conversation, wake-word switch or microphone check; Screen Recording at the first screen capture you start; Accessibility when Jarvis first has to type, click or focus a window for you; Input Monitoring when you save a global shortcut (or choose the Call shortcut in setup); Music and Spotify control when you switch on "Mute music while dictating" with a player running. Nothing is asked at launch, and Personal Jarvis draws nothing of its own around macOS's dialog. If you say no, only that feature stops, and one short toast says so with one action ("Open System Settings"). Checked against faked macOS frameworks and macOS CI runners, not yet on a Mac of your own.
- One short toast after a macOS "no", in English, German and Spanish. macOS never asks twice and says nothing after a denial, so when a feature you started is blocked by a permission, the app shows one sentence in its normal toast with one action: "Open System Settings", "Ask macOS now" (a development run outside the installed app) or "Quit and reopen" (access is on but needs a restart). It shows once per problem and again whenever you try the feature again after dismissing it, in the main desktop window only, and never for background work; the one exception is the wake word, which you switched on and which tells you once per session when it cannot listen. There is no permissions page, banner, card or inline note.
jarvis permissions reset <permission>. Forgets what macOS recorded for the installed Personal Jarvis app (the downloaded app or the installer's app, or--bundle-idfor the other one) so macOS asks again on next use. It runstccutilon your Mac behind--yes, never touches another app, and prints one line and exits with an error on other systems. The rawtccutilcommands are in the troubleshooting guide.- macOS permission dialogs in German and Spanish. New builds of the downloaded app and fresh installs of the installer's app carry German and Spanish versions of the text macOS shows in its permission dialogs (the Apple Events and folder sentences are shorter and name no single app); an installer's app already on your Mac gets them the next time it is rebuilt for another reason, because the bundle version is deliberately not bumped (a bump would reset its permissions). Whether a Mac shows them is not yet tried on a Mac of your own.
- Honest checks for silent macOS failures. A flat wallpaper-only screen capture is refused instead of handed to the model, and five seconds of exact-zero microphone audio produce one "denied or muted" notice. Computer Use pauses while a macOS permission dialog is in front, so an agent never answers one.
- A dispatch-only macOS hotkey spike. A manually started workflow records, on Intel and Apple Silicon runners, whether the native Carbon hotkey calls crash the process. It changes no shipped behaviour.
Changed
- Global shortcuts need only Input Monitoring on macOS. The listen-only shortcut tap no longer also requires Accessibility, is not started at launch unless already allowed, and re-arms in place when you grant access. A restart is suggested only after real typing produced no events, never forced.
- macOS permissions are checked silently and read live. Features check the grant without prompting and act only on a live "granted"; a native request's result is never treated as proof. Identity now decides only who may ask automatically and who may reset, so a terminal-launched run asks for nothing on its own and offers an explicit "Ask macOS now" confirmation in the toast.
- The macOS bundles carry one table of usage strings. The camera, speech and system-administration keys and the camera entitlement, which nothing in Jarvis uses, are removed; Screen Recording, Apple events, Desktop, Documents, Downloads and volumes are described in one place for both the
.dmgapp and the managed app.
Removed
- The Settings > Privacy page, the permission card and every inline permission note. The floating card, the notes in dictation, the wake-word panel, mute-music, Shortcuts, Appshots and voice chat, the "Enable global shortcuts" buttons and tip, the "Mic blocked" line in the sidebar and the Privacy entry in Settings are gone (about 170 translated strings per language); macOS's own dialog and the one toast above replace them. The global-shortcut status that served the Shortcuts note is removed from the keybinds API too.
- The permissions banner, "Set up everything" and the setup permissions step. The app-wide warning banner and its "Not now", the Optional and wanted-row bookkeeping, the guided wizard with its automatic restart, the refresh event and the onboarding permissions step are gone, along with the Music and Spotify consent file and the hidden launch of those apps, the "macOS treats the app as new" identity note and the global restart-required flag. Old leftovers on disk are never read and never prompt; the source installer deletes the two leftover state files on its next run, the downloaded app leaves them in place (harmless).
Fixed
- macOS: the wake word, dictation and Computer Use no longer sit dead behind a permission check that never asked macOS. Protected features used to refuse until the app's own check said "granted", so on a Mac that had never answered a dialog macOS was never asked from the feature, and only a banner, Settings buttons or setup could raise it. The features now ask macOS themselves (see Added). Checked against faked macOS frameworks, not yet on a Mac of your own.
- macOS: the downloaded app is recognised as an installed app. The
.dmgapp was never accepted as one, so (checked against faked macOS frameworks, not yet on a Mac) its microphone, shortcuts and Computer Use stayed disabled even with every permission granted, and no Allow button was offered. It is now recognised,jarvis permissionsfinds it, andjarvis permissions resettargets its own permission records. - macOS: the downloaded app is built to be able to use its permissions. The v2.5.0 image (read on Linux, not run on a Mac) was missing the microphone framework in its frozen archive and was marked background-only in its
Info.plist. The build now includes the framework, clears the flag, and new checks fail the macOS job if either comes back — including one that boots the built app and asks it for its microphone permission. Built and checked on both macOS CI runners (Apple Silicon and Intel); not yet tried on a Mac of your own. - macOS: the release workflow is prepared to sign and notarize the image.
packaging/macos/build.shnow imports the Developer ID certificate secrets the workflow already passed along (rehearsed and run against a stand-in keychain command, not yet end to end). Until the Apple secrets are added the image stays ad-hoc signed: the first launch needs Open Anyway and an update can ask for the permissions again.
[2.6.0] — 2026-10-02
Added
- A quiet front page: one greeting, one composer, and a voice mode that lives inside the chat with a flowing aurora glow.
- Open any agent's chat on the front page from the sidebar;
@in the Jarvis chat names your agents and connected plugins. - Reasoning traces drawn as a threaded rail, with a wider, calmer column for agent chats.
- A restyled sidebar with search, your most-used agents, every recent chat, and the update button in its footer.
- Zoom the whole app window with Ctrl + Plus / Minus / 0.
- Artifacts open on a gallery with live previews.
- Pets gain new characters (Ember, Pip), idle acts, action states, a pixel-art control strip, thought bubbles, notification cards and a phone button to call Jarvis.
- Onboarding: the pet walks you through the app right after setup, then offers ten first steps.
- Routines and chat channels keep running after the app window closes.
- Jarvis can drive the Agentic IDE by voice and open coding agents in a named workspace.
Changed
- Scheduling lives on agent routines; the separate Automations section is retired.
- Routines can be created, paused, resumed or test-run by voice without a second confirmation.
Fixed
- Dictation works during a live call instead of hanging it up.
- A new agent's first prompt in the Agentic IDE is reliably submitted.
- Light-mode terminal panes are crisp and readable.
- A hung main window recovers on its own instead of waiting forever.
- Live voice: no hang-up unless you ask, a spoken yes confirms the pending action, and call transcripts are stored while the call runs.
- Credentials stay out of the app-action catalog.
- Restore the installation-guide link on the PyPI package page and unblock package publication.
- Build Linux installers through the official Ubuntu archive when the runner's default mirror stalls.
- Publish the Debian installer alongside the AppImage and other desktop packages.
[2.5.0] — 2026-10-01
Added
- Desktop pets with six built-in characters, animated reactions, a compact control strip, and configurable appearance and size.
- A redesigned profile page and a dedicated Keyboard Shortcuts page in Settings.
- Quick section switching, an Agentic IDE key menu, and multiple workspace shell tabs.
- Move running coding terminals between workspaces in the same folder without restarting the agent.
- Remote agents retain approved Jarvis tools through scoped SSH forwarding.
Changed
- First-run guidance helps users connect an agent subscription and configure their wake phrase.
- Agents continue requested work, recover incomplete tool steps, and use question cards when essential information is missing.
- Background learning and wiki work use the connected agent seat and respect subscription billing boundaries.
- Provider health indicators read actual call outcomes without sending paid test requests.
- Built-in routines no longer start scheduled work without the user's setup; edited skills are preserved during migration.
Fixed
- Native application updates and restarts preserve the installation and return the app to the desktop.
- Dictation delivery, retained terminal rendering, and delegated result reporting are more reliable.
- Stricter per-chat approval settings survive changes to agent defaults.
- Release publishing requires a tag and the newest successful CI result; version bumps keep the dependency lock consistent.
- Cancelled CI runs release their queued aggregation work, and live realtime tests require explicit opt-in.
[2.4.4] — 2026-09-30
Added
- First run inside the real app. Onboarding is one morphing card plus a guided tour of the app, and the first-run setup happens in the real interface instead of a separate flow.
- Minimal IDE tiles. Agentic IDE panes default to multiplexer-style tiles with a slim title row; the working pane always wears a blue edge, and the classic cards are one click away in Workspace options.
- Jarvis Verse spawn point in the middle of both floors: drop files or open a new workspace from there. Pane nameplates show the coding CLI's logo on two lines, desk monitors draw a pane the way the IDE does, and the pane window wears the minimal tile look.
- Windows computers. A connected Windows machine now works for agents and the IDE, with a Windows-aware setup prompt, readiness check and quit offload; a cancelled turn ends its CLI there.
Changed
- Background work stays on subscriptions once one is connected; nothing the user did not start bills an API key on its own.
- Jarvis' learning notebooks stay small and compacted.
- AGENTS.md is the only agent instructions file.
Fixed
- Reopening Jarvis right after closing it no longer stalls for minutes.
- Remote panes come back safely, fork whole, and never read a stale local transcript aloud.
- The wake word keeps its primary decoder when a slow sibling is demoted.
- An agent's Chromium starts only when it is needed.
- The loop watchdog says when a stall is the machine running out of memory.
- Clicking into a coding agent's output selects its pane; the repository picker closes as soon as a pick is saved.
[2.4.3] — 2026-09-30
Fixed
- Raise the urllib3 security floor to 2.8.0 (CVE-2026-97687..97689, published after 2.4.2 was tagged). The strict dependency audit refused 2.4.2's signed install scripts on it; 2.4.3 ships everything 2.4.2 contains plus the fixed dependency, the desktop installers and the signed install scripts.
Changed
- The README hero video shows Agents, Jarvis Verse and the Agentic IDE.
- Desktop sections load on navigation intent, so switching sections starts sooner.
[2.4.2] — 2026-09-30
Fixed
- Waiting for a process tree to exit no longer fails on Linux when a process
is caught mid-shutdown. psutil 7.2 raised
OSError: [Errno 22]for such a process; it stopped the Linux desktop installer build for 2.4.1 and could break stopping the local voice server, a mission worker's commands or the uninstaller. 2.4.2 ships the 2.4.1 application with this fix and the desktop installers for Windows, macOS and Linux.
[2.4.1] — 2026-09-30
Fixed
- Raise the pyjwt security floor to 2.14.0 (locked 2.15.1). pyjwt 2.13.0, pulled in by the MCP SDK and Twilio, carries ten advisories; the sign-installer workflow's strict dependency audit refused 2.4.0 on it, so 2.4.0 reached PyPI without signed install scripts. 2.4.1 is the same application with the fixed dependency.
[2.4.0] — 2026-09-30
Added
- Jarvis Verse. The Jarvis agents now live on a walkable 3D office map: toy-style figures with an outfit wardrobe, a lead executive suite, a team meeting room, a break lounge, Mission Control with live Spend, Agents and IDE monitors, and a coding floor (reached by elevator) where every Agentic IDE pane appears as a figure with a live terminal you can type into, opened as a movable, resizable window.
- Computers. Settings > System > Computers connects servers and local VMs over SSH — the connect dialog asks only for the address and works out the way in (key, password or a provider's API; 14 provider presets). IDE panes, workspaces and Jarvis agents can run on a connected computer, chosen when the agent or workspace is added.
- Agentic IDE. Git tab with branches, pull requests and CI state; an explorer with git changes and diffs; fork a coding agent's chat into its own worktree; coding agents keep running across app restarts and resume after a power-off; a live agents side panel; drag-and-drop ordering; up to 16 agents per workspace.
- Voice readback. Jarvis reports back when a pane or agent finishes a job, and the live voice model can reason about those reports on the next call.
- Agent chat. Agents can ask short multiple-choice question series; group chats of several agents; per-agent approval modes (Bypass, Ask, Always ask). New agents start on Bypass; blocks, denied capabilities and explicit require-approval rules still apply.
- App actions. Jarvis can run any app action under a per-action allow / ask / block policy (Settings > Jarvis actions).
- Appshots. Show the assistant the front window on a shortcut, a button or a spoken request.
- Model discovery through a public feed, so subscription model pickers gain new frontier releases without an app update.
- Jarvis keeps its own self-learning loop in memory.
- A restrained, milestone-driven boot splash.
Changed
- The settings hub opens as a centred dialog over the app; API Keys shows provider cards in two columns on wide windows.
- The lead agent carries the wake-word name on every surface.
- Local realtime voice setup: one "apply, test and use" step that starts the selected models with Jarvis.
- README and product description rewritten around "your computer becomes an AI agent".
- CI rebuilt as a lane-gated orchestrator with one required
CI gatecheck; a release tag is only admitted afterCI gatepassed on that exact commit.
Removed
- The activity recorder (awareness layer) and its recall tools: it put a
snapshot into every turn and summarised work in the background at a token
cost the product no longer wants. Old
[awareness]config tables are ignored.
Fixed
- The GPT-Live API key is reserved for the voice call; background work runs on subscriptions.
- Around 110 fixes across the Agentic IDE (glyph soup after maximize, shredded panes after minimize, blank repaint areas), the office map, agent chat, Computers, voice and the desktop shell.
- Passphrase-protected SSH keys can be added (bcrypt is now a dependency).
- Silent exception handlers and blocking async routes introduced since 2.3.2 are logged, explained or moved off the event loop; native select popups are replaced by the branded dropdown.
Removed
- Remove wallpapers. The app always paints its flat theme colour; the Wallpaper page in Settings, the bundled pictures, the wallpaper library download and uploads are gone. The marketplace no longer lists, installs or publishes wallpapers — it carries plugins and skills only.
[2.3.2] — 2026-09-23
Fixed
- Raise the AnyIO and GitPython security floors and refresh both install locks. The previous versions blocked the sign-installer workflow during its strict dependency audit.
- Preserve the v2.3.1 application and README while making the same source available through the installer release pipeline. The validation limitations noted for 2.3.0 still apply.
[2.3.1] — 2026-09-23
Fixed
- Restore the current ecosystem README, architecture illustration, demo media, and their reproducible source files in the published source download.
- Align the latest release tag and resumable source archive with the current
GitHub
maincheckout. The validation limitations recorded for 2.3.0 still apply; this patch changes documentation and release metadata only.
[2.3.0] — 2026-09-23
Added
- Continuous Live conversation history, startup audio buffering, media feedback, and deferred tool discovery for the selected reasoning model.
- Agent companions and vector avatars, persistent Mars outpost navigation, routine conversations, and agent communication context.
- Per-agent profiles and memory notebooks, roster actions, and clearer memory change receipts.
Changed
- Publish the current desktop runtime, settings, onboarding, and bundled web interface together so source downloads match the local feature set.
- Retain local runtime telemetry, planning material, test worktrees, and private verification captures outside new commits.
Fixed
- Voice startup, native window lifecycle, browser shutdown, and conversation recovery paths accumulated since 2.2.1.
- Shopify catalog manifest recognition in credential protection and explicit withholding of private agent notebook routes from the remote MCP surface.
- Runtime restart and announcement test fixtures now follow their current lifecycle contracts.
Validation status
- Production frontend build and 328 focused Live/realtime checks passed.
- The broad non-slow Python suite has unresolved failures. Full-suite and physical macOS/Linux audio qualification remain open; live API testing also encountered exhausted provider quota. This release preserves the current desktop feature set with these known validation limitations.
[2.2.1] — 2026-09-16
Fixed
- Signed installer assets are back.
pypdfmoved to 6.19.0 (PYSEC-2026-3910/-3911/-3913);pip-audit --stricthad blocked the Sigstore-signedinstall.sh/install-verify.shrelease assets on every tag since 6.15.0 picked up the advisories. - Quiet failure paths of the macOS signing identity and Automation probe now log what did not answer (silent-exception ratchet).
- Spotlight health is judged by behaviour, not by the data volume's
mdutilstate alone.
[2.2.0] — 2026-09-16
macOS permissions stop churning: the installed app carries one signing identity for the life of the install, the Music/Spotify consent becomes a real permission row, and one button sets everything up with a single restart. Around it, the chat gains slash commands and durable goals, routines gain event hooks and new trigger sources, Jarvis agents get managed live browsers, and the Society world, model menus and agent memory mature.
Added
- Set up everything (macOS). One button in onboarding, the Settings card and the app-wide banner walks through every missing permission — native dialogs first, the System Settings switches last — waits for each grant and restarts the app once at the end.
- Automation (Music & Spotify) permission row. The Apple Events consent the ducking scripts need is asked up front (a closed player is opened hidden for the dialog and closed again), recorded, shown, and resettable like every other row — no more prompt in the middle of a dictation.
- Chat slash menus, durable commands and goal supervision. Slash commands with suggestions above the composer, persistent goal controls, native goal supervision, and plugin chips in the typing line.
- Routines: human, stream, provider and workflow trigger sources; durable event hooks and authenticated webhooks; timezone-aware schedules configured through chat; routine details with schedules and execution history on the agent card.
- Managed live browsers for Jarvis agents. The original Chrome window is streamed into agent views, real agent clicks are visualised, Codex browser approvals route through Jarvis, and cold starts stay connected.
- Society world. Authored city and metro workstations, an isolated future-city transit reference, a modern pixel diorama with safe locomotion, a searchable subscription model menu, model switching from chat, read-only conversations between agent pairs, and supervised coding agents inside the current chat.
- Agent Mode launch film rebuilt with native 60 fps motion.
- Chat work trace rebuilt chronologically (RUB-51), tool receipts shown clearly, provider-neutral image/video/audio outputs, native generated images kept as chat artifacts.
- Reference-first game art pipeline for world and character assets.
python -m jarvis --doctorgainsmacos-spotlight.
Changed
- macOS app bundle signing. The source installer creates a per-user
code-signing certificate once (macOS asks for the login password that one
time) and signs the app with it. Privacy grants are pinned to
identifier + certificateinstead of the per-build hash, so updates and rebuilds keep every grant; the previous ad-hoc path remains the fallback. - Sidebar simplified with tools and profile navigation; chat history flattened; compact chat titles; Jarvis Voice promoted.
- Agent conversations simplified: completed activity collapses, agent-to- agent messages fold to one line, pair messages align on opposite sides, outbound bubbles lifted, professional written style without emojis.
- Plugin catalog opens compactly with original logos, one row per connected plugin, skills and MCP tabs in the plugin dialog; Slack scopes completed.
- Provider status discovery runs in parallel and is shared; model menus prepare their DOM before opening and persist their data.
- README: Jarvis Agents instead of Automations, screenshots from the current UI.
Fixed
- macOS permissions asked again after every rebuild (BUG-217). The ad-hoc signature made every rebuilt bundle a stranger to TCC; the installer then had to reset every grant. Closed together with the still-open notes of BUG-159/161.
- The installed macOS app missing from Spotlight (BUG-216). The bundle is
now imported with
mdimport; a broken Spotlight store is reported with the repair command, judged by behaviour rather thanmdutilstate alone. - Voice: cold-start freezes, abandoned native dictation workers, failed dictation warmups, listening after long tool responses, explicit language corrections, archive context in realtime sessions, agent replies retained until playback completes.
- Browser agents: preserved scopes, initialised CLI model catalogs, direct preview input, Jarvis windows marked.
- Marketplace: hardened browser auth, blocked-plugin audit recorded, AMD hardware detected before enabling connect, credential-free public manifests recognised, explicit marketplace tool requests honoured.
- Society: internal query replies stay inside the society channel, specialist mentions route through internal messaging, voice history stays on the voice surface, ordinary agent memory stays personal, routine tool approval delegated to the app, agent cards stay open when dismissing the model menu.
- Subscriptions: accurate Codex and Grok usage, Google login, scoped Grok subscription logins, subscription seats preserved without model catalogs.
- Desktop: startup recovery follows the replacement instance, no freezes during agent review fallback, macOS modifiers recorded even when the WebView swallows keyup, X logo contrast follows the theme.
- Costs: coding costs filtered from trend charts; short tool durations kept in milliseconds.
[2.1.0] — 2026-09-07
The Society becomes a living world: agents walk an island, talk to each other, learn skills, run routines, and propose changes the person confirms. The Agentic IDE gains a chat view for every coding session, the local-model surface wires Ollama end-to-end, and the UI settles into its neutral design-token system.
Added
- Chat tool picker. Search and select tools, skills, plugins and connectors for each message, with capability filters and provider logos.
- Agent Studio. Edit agent instructions, models, appearance, permissions and routines from the agent card.
- Internal agent messages. Preserve sender identity and delivery status when agents communicate through their chats.
- Agent Society island. A 3D world where agents live, walk between buildings, talk to each other, and carry out quests. Sixteen biomes at ten heights, graded roads, shore-driven water, clouds, and a toon-shaded M3a look.
- Agent card and foundry. A model card shows each agent's profile, tools, approval rules, focus, and routines. The Agent Foundry crowns the mountain — creating an agent walks it out of the forge.
- Agent chat and tool calls. Each agent gets its own chat surface, a contained shell, file hands, and its own browser driven by browser-use out of process.
- Agent capabilities. Agents learn active skills from finished tasks, run routines whose results arrive in chat, propose self-changes the person confirms, and obey per-agent approval rules.
- Quest Board. A posted job routes to one agent, forges a new one when nobody fits, and reads back off the board monument in the market square.
- World Kit buildings. Plugin Docks, Skill Forge, Relay Tower, Terminal Cantina, and the Memory House — each built by script in Blender and wired into the island with their own drawers.
- Shared memory. One memory for every agent, visible on the island as the Memory House building.
- Society composer completions. The composer completes teammates,
capabilities, skills, plugins, MCP servers and tools with
@. - Agent MCP surface. Drive the whole agent ecosystem from any MCP client.
- Per-client MCP credentials, quests and portable ecosystems.
- Dictation on-device pass. The final recognition pass runs on-device, out of process, in front of the cloud chain.
- Profile rebuild. Profile becomes the file the assistant keeps on you, with a two-tab layout.
- Recording strip in every composer. One visible microphone across all input surfaces.
Changed
- Neutral design tokens. The UI drops brand hue for an ink-on-paper system with Inter type scale, shell redesign, visible rims, and brighter ink.
- App icon. Redrawn as paper Gigi on an ink squircle.
- Settings cards span full width. No more capped containers.
- Orbit camera. The island can be looked at from any side, with deep-linkable camera parameters.
- Agent subscriptions. Agents run on CLI seats with pinned accounts and model switching.
Fixed
- Socket pool exhaustion on wake (BUG-215, AP-33). Reconnects are jittered and pay a shared connect budget so N panes × M windows cannot empty the OS ephemeral-port pool.
- Agent-chat shell calls dying on the first call.
- Boot-budget guard was blind, not passing.
- Agent card description editor wiping focus and approval rules.
[2.0.0] — 2026-08-29
The release that stops being one voice channel with a settings screen behind it. Nine coding agents now run in the Agentic IDE and can be read as chats instead of terminals, the local-model section sets a whole machine up in one click and proves that it worked, the interface drops its brand hue for an ink-on-paper theme that reads the same in both modes, and the license moves to Apache 2.0, the change that makes this 2.0 rather than 1.7.
Added
- Nine coding agents in the Agentic IDE, not two. Claude Code, Codex, Cursor CLI, OpenCode, Kimi Code, GLM Coding Plan, Grok Build, Antigravity and DeepSeek Harness are all registered entries, alongside a plain terminal that runs this machine's own shell and any CLI you register yourself. Each one brings its own detection, its own install command, its own trust seeding and its own conversation resume, so picking one in "Open beside" is the whole setup. Entries that keep a separate login per seat (Claude Code, Codex, Grok Build) say so; entries whose credential layout has not been verified against a live install honestly offer a single login rather than pretending to switch.
- Cursor CLI in the Agentic IDE. Cursor's official terminal agent
(
agent/cursor-agent) is a pick in Open beside, the chat's Coding CLIs list, and the install button. Windows runsirm https://cursor.com/install?win32=true; macOS, Linux and WSL runcurl https://cursor.com/install. Sign-in isagent login(orCURSOR_API_KEY) inside the pane. - A coding session can be read as a chat instead of a terminal. The Agentic IDE still opens on its grid of terminals; switching a session to its chat view now draws the same run as a conversation. One session fills the stage, every other session sits in the sidebar under its real CLI logo, and each row is titled by what the work is about rather than by a state label or a command line. A row says whether that agent is still working or finished, the badge holds through a pause and follows the socket, the header carries the same recap card the grid draws, and the terminal is one click away and one click back. Right-click a session to archive it or close its terminal.
- Every connected coding CLI is a chat seat. The chat's provider picker
lists the Agentic IDE's own CLI registry under "Coding CLIs" — OpenCode,
Kimi Code, GLM Coding Plan and DeepSeek Harness join Claude Code, Codex,
Cursor CLI, Grok Build and Antigravity — instead of the sub-agent missions'
provider map, and each one answers as a chat: the reply, the reasoning where
the CLI shares it, and every tool call with its result as rows in the
conversation, never as a terminal. OpenCode's model list comes live from
opencode models; a CLI that keeps its own login reads as available once it is installed; a CLI the registry drops leaves the picker. - A typed conversation that behaves like a real chat. A sent message is on screen at once, the composer grows with the text, the stage follows an answer that outgrows the window, the model's thinking reads as a scratchpad where it happened instead of a hidden block, a code change reads as a diff, and a finished run of steps folds itself away. Both chats take files again — dropped, pasted or picked — and say what is in them.
- The composer knows what you can name. Typing "/", "@" or "$" lists the skills, commands, plugins, agents and files that particular seat actually has, read per runner rather than typed from memory.
- Local models: one click sets everything up. "Set up everything" on the overview starts the local server when it is stopped (and names the install on the button when it is not installed), then fills every role — chat, voice, tools & screen, deep & coding — with the best download already on the server, fetches only what is missing, writes the suggested settings and reads back what it did, role by role. A slot served elsewhere is left alone and named; a refused slot does not stop the others. When another brain answers, the one decision left — making the local server the active brain — is offered right under the summary.
- Local models: the installed models are on the overview. Every download as one line — name, size, the capabilities a role cares about, which roles use it, which role it is the pick for, whether it sits in memory — with "Manage" opening the full ledger. Each card also says, in a sentence, whether its model can do the job it is being considered for. Nothing hides behind the Advanced switch any more.
- Local models: the set-up is proven, then saved for next time. "Set up
everything" ends with four real round trips — the server answers, the chat
pick produces an answer, the voice pick calls a tool, the tools & screen pick
reads an image — shown one line each with how long they took
(
POST …/local-models/verify, also the sidebar badge's new source). A role that is not configured reports as not run rather than as a pass. It then switches "Start with Jarvis" on. The Server tab carries both: the switch, with the backend's own sentence on what the next start would do, and "Run a check". From a terminal:jarvis local-models server verifyand… server autostart on|off. - Local models: the server starts with Jarvis. A few seconds after the web
server is ready, an installed-but-stopped local server is started and the
chat pick is loaded into memory, so the first answer does not pay the load
time — only while local models are actually in use (the active brain, or a
configured role), and only while
[brain.providers.ollama].autostart(on by default) says so. Switching the brain to the local server readies it the same way at once, from every switch path (REST, voice, CLI, brain tool). Nothing is ever installed without the click that names the install. - Artifacts draw what a run made. An output page renders instead of listing
filenames: an HTML run executes in its own sandbox,
htmlandsvgfences render in place, and a script becomes a card with its source folded underneath. A fence with no language tag is treated as a block, not as inline code. - The Profile section is one screen. Rebuilt to fit a single viewport, with
USER.mdrendered in place of the old sigil rail — and a writer that actually keeps that file current instead of leaving it to go stale. - Prompt Mode turns a spoken sentence into a written request. Dictation can read the whole transcript through first, name the language, draft, and read itself back before it writes — producing the professional request a coding agent needs rather than a raw transcript, while keeping your familiar form of address. The bar shows what it produced, it knows when not to write, and the sparkle pauses the mode for the moment instead of switching the setting off.
- The voice bar draws the assistant's own voice. The waveform is measured on the native path and blooms from dots into capsules, scrolling rather than mirroring, so what you see is the reply that is actually being spoken.
- Hotkey backends answer whether a chord is physically down. A held dictation now follows the real key rather than an assumption about it.
- The folder picker opens a typed path. It also makes a new folder and shows hidden ones, so starting a workspace somewhere unusual is not a detour through the file manager.
- The sidebar folds. The Chat row folds its recent chats out beneath it, the Agentic IDE's workspaces get a panel of their own, chat mode shows only the chats with every section one button away, and the column starts at a width the panes were designed for.
- An in-app update for a native installer build. The one "Update Now"
button now covers both supported installs: a git checkout from the one-line
installer updates the way it always did, and a build that came from
Setup.exe, a.dmgor an.AppImagedownloads the next installer asset from the GitHub Release, proves it against that release'sinstallers-SHA256SUMS.txt, and hands it to the platform's own upgrade mechanism. Progress is reported live while it runs instead of a spinner that says nothing. Anything else — a maintainer checkout, a fork, a plainpip install— is still reported as unmanaged and the button never appears.
Changed
- License: Apache 2.0.
LICENSEis the Apache License 2.0 andNOTICEsits beside it — the same freedoms MIT gave, plus an explicit patent grant from every contributor and a written trademark carve-out. It takes effect with this release, 2.0.0; a license change is a breaking change, so there is no further 1.x. Every release up to and including 1.6.0 stays MIT permanently — that cannot be revoked, and a copy taken under MIT keeps those rights, including the right to fork it.docs/licensing.mdsays why, what it changes for users, forks, and contributors, and what nobody has to go back and correct. - Ink and paper: the brand hue is retired from the interface. Primary is now ink on paper in light mode and paper on ink in dark mode, the dark surfaces got a real ramp so the navigation column stops being the same black as everything else, full white is reserved for actions, and state stopped being a colour — warning and success map onto the ink token instead of a hue. The voice orb's material, the run-graph categories, code diffs, every mascot and every mark were re-rendered in black and white. The design skeleton follows suit: no shadows, display type at weight 400, one radius scale. Provider logos, the ANSI slots in a terminal and the destructive colour keep their own colour on purpose.
- Local models: recommendations are installed-first. A role's pick is the best qualifying download already on the server — the largest one that fits this machine's graphics memory, preferring the role's capability (tools for a chat, thinking for deep work; voice stays in the fast class under 6 GB) — and the curated catalogue only stands in when nothing installed qualifies. Each row says why in one sentence. Eleven models on disk no longer end in "download qwen3.8:27b".
- The front page's chat runs on API keys, not vendor CLIs. The typed chat on the home surface answers through Jarvis' own harness on the providers you have keys for; coding-CLI seats live in the Agentic IDE, where a folder and a terminal exist for them to work in.
- Publishing points at what is still alive. The hosted storefront that served personaljarvis.ai was shut down, so both publish endpoints default to empty, which hides the Publish button rather than firing at a host that answers nothing. The external buttons now open the community registry on GitHub and this repo's docs. Browsing the marketplace is untouched — the registry's index never depended on that site.
- The Google Workspace CLI entry is
gws. It described GAM, an administrator tool for a whole Workspace domain, and probed a binary nothing here drives, so an installed Google Workspace CLI reported "not installed" forever.jarvisctlwas in the same state and now installs from PyPI. - The Agentic IDE's Install button installs. An entry with no install method used to invent a pseudo-method whose "command" was a documentation URL, so Install opened a GitHub page and installed nothing.
- README: every screenshot is a fresh capture of the 2.0 interface. The home view, Local models, the automations catalogue and the wallpaper gallery were re-shot on the ink-and-paper theme, and the Skills list is pictured for the first time. The recorded demos are unchanged.
Fixed
- The app stopped freezing in every section after a cold boot. The event loop stalled fifteen seconds at a time behind a lazy SDK import inside a provider call and behind a worker thread being started on the loop, while twenty-two CLI probes, a provider health sweep and sixteen resumed coding-CLI panes fought a cold disk. Brain client construction and registry instantiation now run on a worker thread, the thread pool stays resident and is warmed once voice reports ready, the CLI prober works in bounded lanes instead of sweeping the whole catalogue at once, and a cold-starting pane holds its slot until it has painted.
- Sections that read a file or a database no longer block every other
request. An audit found 234 route handlers declared
async defwithout a singleawait, so their reads ran on the one thread serving the whole app. The measured offenders — the chat list the sidebar polls every five seconds, the agent-status route with its four PATH scans and credential-store reads, the wiki's subscription-login probe — were moved off the loop, and sections repaint from cache while a fresh read is in flight. - Spend: the coding-CLI bill no longer shrinks, and it counts every
subagent. Three things made the Spend section show less than was spent —
$8 000 for a month that had read $12 300 the day before. The index of the
coding CLIs' transcripts deleted a session's turns when the CLI deleted its
log (Claude Code does after
cleanupPeriodDays), so the lifetime total sank a little every night; a change to a reader's rule dropped the whole index and rebuilt it from zero, and for the hours that took the page presented the half-read index as the bill — two views a minute apart disagreed by thousands; and the transcripts of every subagent (<session>/subagents/**, the Agent tool and workflows) were never read at all — one fifth of every Claude Code call on the reference machine. The index is now a ledger: a vanished transcript keeps its rows, a rule change re-reads the transcripts still on disk and corrects rows in place while the table stays complete, subagent transcripts are indexed under the conversation that spawned them, and while the index is still catching up the section says so — "{n} of {m} transcripts read, {x} MB to go" — and reads again every few seconds instead of once a minute. (BUG-198) - Spend says "loading" instead of showing zeros, and stops freezing the app. The section's own reads left the event loop, its numbers are already in hand by the time you open it, and an empty state is now the skeleton it always should have been rather than a confident row of noughts.
- "I can't reach my provider" no longer means "the tool list was too big". A typed "Hallo" on a local 32k model failed because 221 registered tools sent nearly 68 000 tokens of JSON schema with one word. The tool surface is now sized against the model's own context window, and a refusal that names a size is reported as a size problem instead of a network one.
- Dictation: a second dictation into the same field starts with a space. Inside Jarvis's own window a finished dictation was written at the caret exactly as transcribed, so two dictations a few seconds apart read "…Agentic IDEDas war…". The seam now gets the space a person would have typed — after a word, not after a space, an opening bracket or in an empty field, and on both sides when the caret sits inside a word joint.
- Dictation: "Agentic IDE, Agentic IDE, Agentic IDE" no longer lands in the text — from the middle of a dictation either. Whisper recites the dictionary's bias prompt over a pause and then carries on with the speech that followed, and the field dictation pastes from was checked only by what the provider's filtered copy had lost. Each copy is now judged on its own, and a recited run is dropped wherever it stands; three distinct terms in one breath stay.
- Dictation no longer ends in a recited word list. Whisper answered a silent stretch by reading the STT dictionary's own words back ("…, Claude, Agentic IDE, Claude, Agentic"); such a run is now dropped from the end of a transcript before the dictionary applies. The dictionary's near-miss repair also became strict: a word is rewritten only when it sounds like the registered one ("Klaude" → "Claude"), never because a common word is one letter away ("grob" stays "grob"). (BUG-185)
- Every stop gesture ends a running dictation. A hold recording follows the physical key rather than an assumed key-up, so a hold that never ended is no longer possible. (BUG-191)
- The Jarvis bar stopped freezing while voice kept working. The bar's UI thread could park with every timer armed and never wake, so the voice channel answered but the bar stayed collapsed. It is now woken explicitly. (BUG-202)
- Local speech-to-text stopped vanishing into a cloud provider. A local recogniser stays local; the transcription is no longer pinned to the CPU on every machine, and the local voice server no longer cancels the model load it is waiting for.
- Local models: a changed role showed its old pick for up to 15 s. The overview's in-memory memo was not dropped on a write, and the section never asked for a live rebuild after its own change. Every write now forgets the memo, and the section re-reads the overview live right after.
- Local models: an offline sweep no longer poisons the cache. A sweep that found nothing used to overwrite the good snapshot; every job keeps its picker instead, a provider switch is what starts and stops the local server, a starting server says "Starting" rather than reading as stopped, the switch applies what it persists, and an AMD or Intel card is no longer invisible to the fit verdicts.
- The model picker stops offering Jarvis's own derived models. Tags this app created for its own roles were being offered back as if they were downloads.
- A frozen desktop window is recovered rather than raised. Bringing a dead window forward does nothing; the app now detects that and starts a live one.
- Gemini: one MCP extension key no longer kills the turn. The tool schema is allow-listed before it goes out.
- Music: a blind media key is an attempt, not a result. The voice no longer announces music that never started.
- Artifacts about your own mail or calendar are built from facts Jarvis reads, never from what a worker imagined they might be.
- The Wiki's voice-fact bridge actually leaves the bus when it stops.
- A failed config write no longer ends first run at step 04.
- nanoid 3.3.18 closes the two open npm advisories.
Removed
- UltraWiki, the semantic memory mode. The store, its pipeline and
connectors, the three route modules, the CLI group, the router tool, the
whole Wiki-section Ultra body and its settings surface go, along with every
test and fixture — about 90 800 lines. The Wiki section keeps the
Obsidian-compatible vault it always had, and the either-or mode switch is
gone, so the vault answers unconditionally again. Three format services
(
extract.py,document_text.py,media.py) had no UltraWiki dependency and moved tojarvis/documents/unchanged, because the Agentic IDE's file preview needs them. Restoring the feature is a revert of one commit. - The Ollama "embedding" role. It wrote to UltraWiki's setting and nowhere else, so with UltraWiki gone there was no slot to fill: the role, its two curated catalogue entries and the verify step that could no longer run are removed. A model on disk still shows "embedding" as a capability in the ledger — only the role is retired. (BUG-207)
- The Local models setup helper. The "Setup helper" tab, a chat that
proposed a setup, ran it through its own
lm_*tools and tested the result, is gone from the app, along with its "Help me set up" and "Something is not working" buttons on the overview, thelocal-modelschat surface and thejarvis local-models assistant ...commands. Its/api/providers/{id}/local-models/assistant/*routes are still registered and answer, but nothing in the app calls them any more; they come out with the panel in a later release. The section keeps everything else: the overview, the catalogue, the server tab, the Tune sheet, and the quiet six-hourly self-check behind the sidebar badge. - The Agentic IDE's prompt bar. The composer under the grid is gone: a pane is typed into or spoken to, never written to from underneath.
- The links to the retired storefront. The buttons that pointed at personaljarvis.ai were removed rather than left to open a dead host.
[1.6.0] — 2026-08-25
The release that can be double-clicked. Personal Jarvis has always been a terminal line away; from this version it is also an installer for each of the three desktop systems, built on CI from the same scripts a maintainer runs by hand, published with the digests to check them against.
Added
- A native installer for every desktop system.
PersonalJarvis-Setup-x64.exe(a per-user Windows wizard, no administrator prompt),PersonalJarvis-macOS-arm64.dmgandPersonalJarvis-macOS-x64.dmg, andPersonalJarvis-Linux-x86_64.AppImagewith a.debbeside it. Each carries the whole application — the windowed launcher and thejarviscommand line — and each release publishesinstallers-SHA256SUMS.txtso a download can be checked before it is opened. The install scripts are unchanged and stay the supported path for anyone who prefers them.
Changed
- A voice call answers on the provider's own timing again. The Thinking pause now starts at Automatic, and automatic means Jarvis sends no turn-detection override at all: a realtime model ends the turn exactly when its own API does, and the classic pipeline keeps its 1.5 s rule. The fixed 1.5 s window added on 2026-08-18 sat on top of turn detection the provider already does, so every finished sentence waited twice and a call felt noticeably slower than the vendor's own client. Waiting longer is still one click away: any value on the slider sends both the window and the "read a pause as a pause" sensitivity that protects it.
Added
- Grok Build in the Open-beside menu. The official
grokCLI is now a workspace terminal like Claude Code and Codex: pick it when splitting a pane, resume the same conversation, and sign in through SuperGrok or X Premium+. The xAI API-key card stays a separate thing.
Fixed
Restart no longer shuts the window and stays down. Clicking Restart launched a helper that watched the first Python process. On Windows that process hands over to the branded copy and exits, so the helper thought the start had failed and launched two more copies. Those three then fought over the lock, and sometimes none of them kept the window. The helper now waits until the app is actually answering, starts the branded copy directly, and a copy that is only seconds old is waited on instead of offered as stuck. (BUG-181)
A launch that never reaches a window is no longer mute. The desktop log now starts from the launcher's first millisecond, so an "already running" bounce, a crash on the way to the window or a lock held by a stuck earlier instance is written down instead of vanishing. When the running instance has no window to bring forward, the app asks — one native Yes/No box on Windows, macOS and Linux — whether to stop the stuck process and start fresh; a restart that fails says so the same way. (BUG-180)
A dictated prompt reaches a Tauri/Electron coding-agent terminal. On Windows the transcript is now offered with delayed clipboard rendering, so Jarvis sees who reads it: a proven paste restores the old clipboard at once, an unanswered Ctrl+V falls through to Ctrl+Shift+V, Shift+Insert and finally typing (line breaks as Shift+Enter), and on a host with a clipboard watcher (Remote Desktop, clipboard history) one chord goes out and the old clipboard is restored after a 2 s grace instead of 120 ms — the timer that used to hand a slow, async-reading WebView the previous clipboard (BUG-178).
A web-searched answer in a live call is spoken to the end. Gemini Live closes every tool call with an "interrupted" edge before its turn boundary; that edge wiped the evidence the boundary guard needed, so every search looked like a mute provider, the retry it triggered cut the answer already streaming ("… dass sie ihre Plattformen" — end of reply), and the model's second answer was thrown away as a phantom. Both halves of the tool boundary are now recognized and withheld.
A Gemini brain that cannot create a context cache stops paying for the attempt on every call. A free-tier key (no cache storage) or a Vertex route without default credentials failed the cache create ~0.7 s per call, four times inside one delegated voice turn; a failure now holds for ten minutes.
Every connected coding CLI resumes its conversation, not only Claude Code and Codex. Kimi Code's current generation files its sessions under
wd_<folder>buckets with the folder recorded inside each session; the Agentic IDE only knew the previous generation's MD5 folders, so a Kimi pane always came back empty. Both layouts are read now. Grok Build and Kimi Code also write a session file the moment a pane opens, so a pane that was never prompted is no longer "resumed" into an empty conversation — a real user turn is what counts.Dictation delivers its text about as fast after a long recording as after a short one. The final read used to start only after the key was released and work through the recording one window at a time, so a one-minute dictation waited several seconds for its text. Windows are now read while you are still speaking, and whatever is left at release is read side by side on providers that allow it; a local engine still takes one call at a time.
Dictation no longer loses the words after a thinking pause. The recognizer used to be handed long pauses inside a window and sometimes stopped there, silently dropping everything after it. Every pause in an upload is now cut to half a second before it is sent, a transcript that ends well before your speech does gets only its missing tail read back in, and a hardware overflow on the microphone is counted as the real loss it is.
The local recognizer says when it cannot reach your GPU — and fixes it from inside the app. On a desktop without the CUDA runtime libraries the on-device recognizer quietly ran on the CPU (about eight times slower; the cloud fallback took 48 s for a 70 s recording). The provider card now shows what the config asked for, what really runs, and why, with a one-click install of the missing libraries. They are also available as the
cudainstall extra.Dictation text is tidied while you speak. Each finished stretch is formatted as soon as it is recognized, so after a long dictation only the last few seconds are left to format when you let go.
The on-device recognizer runs at full GPU speed. On cards where the int8 kernels crawl, it is built as float16 automatically — the local model no longer takes a minute for three words.
The wake word no longer fires on a stray single word. Since the 20 Aug crash fix, the check that decides between "Hey George" and whatever was actually said had quietly lost its strongest comparison, so a lone word like "power" or "Pedro" could open a voice session by itself. The comparison is back, and the app now says so in the log if it ever has to run without it.
[1.5.3] — 2026-08-22
Added
- The marketplace is a section of the app now. Everything the community published — plugins, skills and wallpapers — sits behind one sidebar entry with one search across all three, instead of being split over a tab, a second list and a third screen. Opening an entry shows the publisher, the source repository and the actual published files before anything installs, and once it lands the app says which section now holds it and takes you there. "Open the marketplace" works by voice, in all three languages.
- The storefront drawer says where an install would reach. A file listing tells you what a plugin ships, not where your access token ends up. The drawer now names the destination before anything installs — the hosted URL requests go to, or the command that would run on this computer — in the same words the consent dialog under Skills & Tools has always used, and it warns when an entry would replace a plugin the app already ships. Skills say whether they are portable and which agents they also run in; wallpapers name their licence.
- A feature request files as a real issue, right beside a bug. The report
section had a bug/idea/question switch nobody could see, and its GitHub path
opened a blank, unlabelled issue. A report now picks the matching issue form
and arrives filled in and labelled
bugorenhancement, with its title prefix and structure intact. A report too long for a link is put on the clipboard in full, so nothing anyone wrote is lost. GitHub is the normal path now rather than a fallback, and the account requirement is said before the login wall — not after it.
Fixed
- "Open the modes" was an unknown section to the assistant. The character screen shipped as a real section but never reached the navigation tool's list, so the spoken command went nowhere.
- The end of what you said stops disappearing (BUG-162). Dictation sometimes dropped its last sentence and still reported a clean success. 797 recorded dictations named three separate causes: the queue depth written to defend a long recording never applied on installs with a wake word, the truncation guard was calibrated so far out of reach it fired once in all 797, and a second gate believed continuous speech whatever came back. All three are closed, and the report now says when it lost frames.
- The app costs a fraction of the machine while it sits idle. Two things
ran flat out for nothing: the blank-window watchdog rebuilt a whole HTTPS
client — certificate bundle and all — once a second for a plain loopback
call, and the wiki map kept painting at 60 fps behind other windows because
its sleep gate waited for a
blurevent a never-focused window never sends. Measured on a four-core laptop, the idle cost drops from roughly a third of the machine to about a tenth. - The report button works before the backend has restarted. Straight after an update the view talks to an old backend for a while. It then had no issue form to open and the button did nothing at all — from the outside, indistinguishable from a broken section.
[1.5.1] — 2026-08-21
Added
- The board's centre is a stage, not a reticle. The mascot stands in the middle of the deck now, drawn as a vector measured off the original artwork instead of sketched by eye, and framed on the figure rather than on the bitmap's empty margin. The bottom row of cards shows its own scale at rest instead of five shrunken apologies, and the rest strip no longer carries orphans the board never loads.
- The Jarvis Bar thinks with the deck's sweep. While a turn is being worked on, the bar runs the same sweep the board uses, instead of the old orbital core — one motion language across both surfaces.
- The deck's sweep runs as one transform instead of sixty redraws a second.
- A contributor wall on the front page. The README shows the people who have contributed, and it refreshes when a pull request merges instead of waiting for the next Monday.
Fixed
- macOS stops asking for permissions it has already been given. The app
demanded microphone, screen recording and accessibility access that System
Settings showed as enabled; granting them again changed nothing, and neither
did a restart. Four causes stacked up. The app rebuilt its own bundle on
every start whenever an identity probe refused it — the rebuild changes the
ad-hoc signature, and macOS answers a changed signature by discarding every
recorded grant, so a probe that failed for a reason no rebuild touches wiped
the permissions on every single start. A copy in
/Applicationswas not recognised as this app at all, because only~/Applicationscounted as canonical: dragging the app there made that slot look empty, every start built a second bundle into it under the same bundle id, and the signature change stripped the grants of the copy that was running. The permission view applied the same rule, so that same move hid every button and failed every runtime check with nothing left to click. And a Screen Recording grant given in System Settings stayed invisible, because the preflight macOS offers is frozen for the life of the process. Now the running process is taken as proof of a valid install wherever it lives, a rebuild that would reproduce the identical app is skipped and logged instead of repeated, both Applications folders are canonical, and the Screen Recording grant is proved by reading a window title only that grant allows. - The memory map hands its WebGL context back. The wiki's 3-D map leaked its rendering context when the view was left, and a browser only grants a handful of them — enough visits and the map came back as a white canvas with a sad face. It now releases the context on the way out and survives losing one, and a CI gate fails the build for any scene that does not.
- The page loads the bundle files that are actually shipped. Some entry
chunks lived only in
index.html, not in the tree, so a fresh clone could open onto a blank window. - A busy Vertex Live is spoken, not silently retried. A 1011 resource exhaustion used to disappear into a retry loop; it now says what happened.
- A question is no longer answered with "Erledigt." In a live call, asking something could come back as that one word — a report that a job was finished, for a job nobody had given, while the actual answer was thrown away seconds later. Two faults stacked up. Gemini and Vertex Live end their generation when they hand out a tool call and start a fresh one for the spoken answer; the adapter meant to swallow that in-between marker but looked for the tool call in the wrong message, so it never once did — 39 of those markers in a single day's log, every one of them read as "the provider has gone silent". On top of that, the recovery for a silent provider treated a finished web search like a finished errand: the search carries no sentence to say, only the material for one, and the recovery said "done" and dropped the material. Now the adapter waits for the generation that actually speaks, and a lookup that came back with something says what it found instead of reporting a completed task — with "I have the information but couldn't read it out just now" as the honest floor, and "I didn't find anything on that" when the search really was empty. A wordless action still reports done, as it should.
[1.5.0] — 2026-08-20
Added
- Hybrid tool mode: the live voice model runs the tools itself. A voice
turn that needed a tool used to hand off to a second model, and the wait
was audible — measured over three days, first audio came after 0.9 s when
the live model answered and after 7.2 s when it delegated. The live model
now receives the whole tool catalogue, everything except the computer-use
vehicles, which stay with the Tool Model. The declaration fits a token
budget (
[voice].realtime_tool_declaration_budget_tokens, 20k by default) and drops whole tool families in a fixed, logged order when it is tight. The new[voice].realtime_tool_modeships ashybrid;delegateanddirectremain for anyone who wants the old behaviour (ADR-0035). - A skill is recognised from what you mean, not from a trigger phrase.
The skill roster the live model reads now carries each skill's
when_to_use— the sentence that says WHEN it applies, in the words people actually use. The first roster capped descriptions at 70 characters and cut exactly that away. It now degrades in three steps and never drops a name, because a listed name is callable and a folded one is not. A near-match reaches the model as a ranked suggestion instead of silence. Both are suggestions: no skill can take a turn over that it did not clearly win. - The whole orb moves with your voice. One loop computes the orb's level and writes it to a single CSS variable, so the sun, its corona, the rays and the ring all move as one thing — the real microphone level while you speak, a speech-shaped envelope while Jarvis does, a double heartbeat while it thinks. The reticle stays alive at rest, and the board now takes over from the start screen with a launch instead of a fade.
- Grok Build joins the subagents. The official
grokCLI is driven over a SuperGrok / X Premium+ login, the same shape as Codex, Claude and Antigravity. Connect, disconnect, test and set-active live on the Agents tab; the xAI API-key card stays separate. - Higgsfield joins the marketplace as a hosted MCP plugin.
Fixed
- A failure you hear now names its cause. "That didn't work" told you nothing you could act on. Every spoken failure — a direct tool call, a delegated one, a confirmed action that then broke — is built in one place and carries the reason the tool reported, in the language of the turn even when the rephrasing model is unreachable. A turn that broke in two places names both causes.
- The wake word survives a busy machine. The confirmation pass now streams during the tail instead of after it, audio intake keeps flowing while it runs, and the first stage has a CPU budget — on the weak reference laptop the wake word was being missed under load (BUG-150). Native Vosk calls are serialized, which ends a crash that took the whole process down (BUG-151).
- A skill can actually be started by voice. A realtime call named
run_skillnever reached therun-skilltool, so a spoken skill request quietly did nothing (BUG-158). A draft authored by voice can now be switched on, and your own morning routine wins over the built-in one. - Plain conversation stops being handed to the Tool Model. A greeting or a remark was being delegated like an action, which cost seconds and tokens for nothing.
- Music does what you meant. A play request with no service named goes to the one you have connected, "what's playing" answers in hybrid mode (BUG-156), an in-flight play is no longer aborted as a stalled turn (BUG-157), and a spoken "I'm playing that" that never called a tool is recovered instead of counted as done (BUG-154).
- A dead speaker recovers instead of going quiet. Native voice is also labelled as requested rather than as a fallback (BUG-108, BUG-086), and progress lines are spoken in the session's own voice (BUG-155).
- The Test button in Dictation probes the model that will really run. It used to test a different one, so a broken setup could pass.
- Picking a file with
@in the agentic IDE searches the workspace tree instead of stopping at the first five name hits. - The app starts lighter. Every byte of the startup bundle is parsed before the window paints, and 212 KB of it belonged to screens the first view never shows: the wiki graph library (pulled in by three tooltip helpers), the agentic IDE's whole API client (pulled in by two hooks calling one endpoint each), and their dependencies. They now travel in the chunks that use them. What remains in the startup path is what the mission deck actually needs.
- A window that never received a page is caught too. The blank-window
watchdog shipped in 1.5.0 lives in
index.html, so it can only act once a page has arrived — and the window the maintainer kept seeing had none: a frame, a title, the ground colour, nothing else. Three routes lead there, all logged on one day: the backend's single event loop frozen inside a native call, so the socket accepted the navigation and never answered it (15 s, 20 s, 42 s, 61 s and 188 s measured, the long ones inside PortAudio's stream close during a microphone restart); the backend thread dying after the port was already up; and a web view that never navigated at all. A guard now runs in the window process, outside the loop it watches: it waits out a slow boot, reloads the moment the server answers again, and when it cannot fix it, it puts the reason on the window with a button — in the browser's language, since the bundle is exactly what is missing. That page then polls the server and returns to the app by itself. Closing the microphone no longer runs on the event loop either, which removes the biggest freeze at its source: the loop now keeps serving while a wedged native handle is shut down on a worker thread. - A question about your own machine gets an answer, not a refusal. Asking whether the PC is overheating or overloaded matched none of the planner's question rules — such a sentence carries no question word and no possessive — so the turn was never allowed to read the machine, and the execute guard judged the model's shell call by the worst command that tool could ever carry. Jarvis answered "I can't run actions right now", twice in a row. The planner now knows the machine's own vocabulary and the indirect question that hides a lookup inside a statement, a call whose arguments only read passes the guard, and a turn whose every call was gated says what actually happened instead of claiming the assistant is broken.
- Dictation translation reads like the target language now. The
translate pass ran on the same tiny model the formatter uses, which
punctuates well and translates flatly — measured on real German
transcripts it dropped whole clauses. A family may now declare a
translation tier, and the pass uses it: Groq translates on
gpt-oss-120b, Gemini ongemini-3.7-flash, OpenRouter ongoogle/gemini-3.7-flash. A model you pinned yourself still wins. - A long dictation is no longer cut off by the wording ceiling. The
1200 ms budget was fixed while the work grows with the transcript, so a
tenth of all wording passes expired on one measured install — and they
were the long ones. The ceiling now ramps with the transcript length up
to
[dictation].polish_timeout_max_ms. Short dictations are unchanged, and a healthy provider never reaches either number. - Two wording defaults were years out of date. OpenAI ran on
gpt-4.1-nanoand Gemini on a 3.1 lite build; both now use the current generation their own catalog serves. - A Start-Menu click that cannot boot now says why. A missing
click(uvicorn's import-time dependency) killed the backend thread, the window waited 45 s for/api/health, and pythonw swallowed the traceback — so Personal Jarvis in the Start recently-used list looked dead. The launcher now refuses up front, a dead backend thread no longer waits out the timeout, and the window's relaunch command points atPersonalJarvis.exelike the shortcut. - A Gemini Live greeting no longer dies mid-sentence. Vertex Live fires
an empty
interruptedfor our own steering text, then speaks the real reply. The session used to treat the quiet after that reply'sturn_completeas a confirmed barge-in and cancel the speaker drain — full text in the log, half a sentence in the air. The same settle could also fire before first audio and withhold the whole greeting (BUG-152). - Vertex Live native tools work again. Vertex reports hybrid function
calls as
default:run_shell. The bridge treated that as an unknown tool and the model said the tools were down (BUG-153). - A spoken "I'm playing that" is no longer treated as done. Hybrid voice could announce a playlist (or any other action) without calling the tool; the user had to ask again. The session now recovers those false completions, music capture rematches to the connected service, and an in-flight native tool no longer looks like silence (BUG-154).
- The Jarvis Bar keeps thinking while the Tool Model works. After the short "I'll play that" ack it used to look ready even though YouTube Music was still running. Progress speech now returns to thinking until the result is spoken.
- A window that shows nothing now says why, and heals itself. Rebuilding
the frontend while the desktop window is open could leave it empty — the
ground colour and nothing else, until the app was killed. The single
automatic reload the app spends on a missing chunk hands the failure to a
view's error card, and a failure above that has no card to land in; the
bundle watch that would have picked up the fresh build lives in the bundle
that is not running.
index.htmlnow carries a watchdog that runs whatever the bundle does: an emptied page reloads once within seconds, a splash that never becomes an app after twenty, and a second failure paints what went wrong plus a Reload button instead of staying dark. The guard is released as soon as the app is up, so a later rebuild in the same session still heals. - A granted macOS permission stops reading as missing. Every start warned that permissions were off, and the pane the warning linked to showed them on. Three things kept that loop alive. Opening System Settings for Screen Recording, Accessibility or Input Monitoring flagged a pending restart even when nothing there needed changing, and nothing but quitting the app ever cleared it — while it stood, Computer-Use and the global hotkeys were really switched off with every grant in place. "Ask again", which clears this app's own record so macOS asks once more, appeared only for a permission macOS reported as denied; the two probes that strand most often report a lost grant as plain "not granted", so the one way out was invisible, and the app-wide banner carried no such button at all. And a request macOS silently swallowed still promised that a restart would help. A pending restart now dies the moment a probe sees the grant, the reset is offered wherever it can help and explained in words, and a swallowed request says so. Underneath, the app is only ad-hoc signed, so any rebuild of its bundle makes it a stranger to macOS and drops every grant: that rebuild now logs why it happened, says in the permissions view why it is asking again, and — when rebuilds recur — no longer wipes the permissions a second time before you have answered the first (BUG-159).
- The deck opens on the board, and the waiting screen is gone. Starting the app left you on a standby ring under a big "Say 'Hey Nova'", with nothing to look at until you spoke. That screen has been cut. The start is now the boot sequence and its launch: the four gates light as they turn true, and a blink later the SAME hand-off a spoken word triggers hands over to the board — orb travel, shockwave, instruments assembling, the scan. Speaking, the hotkey or a press on the orb still gets there sooner, and an install whose voice stack never reports (no microphone, voice off, a headless box) is no longer held on the start screen: with the link up the board opens anyway. The wake phrase still greets you — on the board's own headline, where it does not block the view.
Removed
[voice].background_heavy_turnsis gone from the config. It has been declared since 1.4.0 and read nowhere — the classic-pipeline half of ADR-0034 is not built yet, so setting it changed nothing. It returns with the change that actually reads it. Nothing to do if you never set it.
[1.4.0] — 2026-08-18
This release turns the front page into a mission deck, gives the assistant a shelf of characters, and stops making you wait: a heavy turn answers within seconds, you keep talking while the tool model works, and speaking interrupts thinking. Music plays by voice on Spotify or YouTube Music, Google Cloud Vertex AI is a provider family of its own, the marketplace grew a community registry with install-by-name, and your own skills actually fire. Underneath, a long list of voice, realtime, Agentic IDE and launcher fixes.
Added
A mission deck as the front page — "what is going on" on one surface, with the classic chat one click away. Opening the app used to show an empty chat: a consequence, not a state. The deck shows every section on the left with the live ones lit, what the assistant is doing right now on the right, and the Jarvis orb in the middle; a header switch moves between deck and classic view and the choice is remembered. Every figure comes from a payload the bus already publishes; nothing is estimated.
- The orb is the Jarvis orb itself, cut out of the hero artwork. It breathes with the voice state (reduced motion gets the still picture) and is ringed by one arc per running reasoning step. Clicking it does what the wake phrase does: it starts or ends the conversation.
- One card per signal, each reading its section's own data: a LOG terminal (one line per thing the assistant heard, thought, did and said, with clock time and duration); a RESPONSE instrument (hear → think → act → speak, a stopwatch, first-token and first-audio marks); runs; outputs; a coding-workspace crew roster (one row per agent — a row jumps INTO that terminal); terminals; the wiki as the same 3D memory map the Wiki section draws; computer use; screen capture; API this session (tokens and cost per model); a live word counter.
- The pictures the deck shows. Screen Context stays "one capture, then
gone"; a mirror keeps AT MOST ONE frame and forgets it after
[screen_context].deck_preview_sseconds (default 120, 0 = off). Computer-use frames are served by hash only (GET /api/deck/cu-frame/); every image response isno-store. - A HUD frame language — corner brackets, chamfered outlines, tick
rulers, status lamps in every title — with a halo on every hairline in the
theme's ground colour, so gold lines hold on both appearances over any
wallpaper. One icon rail (
DockRail) serves the deck and the collapsed sidebar, with a live pip where something is going on and the sidebar's own signals (API-keys error, plugin reconnect, the Skills → Plugins shortcut). - A boot sequence and a listening ring before the first word. The deck no longer shows nine instruments saying "nothing yet" from the first second. Three acts, forward only per session: boot — one big ring around the orb, the four gates a voice turn needs (link, voice, brain, wake) drawing in as each turns true, a console typing one line per gate with clock time and measured duration; standby — everything up, nobody has spoken: a sweep turns only while the wake word is really being listened for, the headline names the phrase to say, "Open the board" is one press away; board — from the first turn (wake word, hotkey, typed message, a press on the orb) the ring folds into the orb, the orb travels to its place and the instruments power on from the centre outward. Nothing on the stage is invented — every gate is a fact the header lamps already show, and a fact that never arrives is called absent instead of spinning forever. Reduced motion gets the same facts without the movement.
Modes — the assistant gets a shelf of characters, one of them active. A mode is how the assistant behaves: the difference between a butler who answers in one line and a friend who asks how your day went. Five built-ins ship — Assistant, Friend, Coach, Focus and Coding — and picking one applies from the next turn in voice AND in chat, no restart. A mode is a LAYER on the base persona, never a replacement, so the rules that keep the assistant honest survive every character.
- A Modes section: the shelf, and a workshop where you make another one.
A card opens with the full character text verbatim, the knobs and the
voice; every card carries an explicit "Use", "Edit" loads any mode into
the form, a built-in keeps "restore original". "Talk it through" opens a
real voice interview — the assistant asks what kind of company you want
and writes the mode itself. The mode in force and the mode you chose are
shown apart (the payload gains
chosennext toactive). In en/de/es. - By voice: router tools
list_modes,switch_modeandsave_mode— "be my friend for a bit" works. Switching changes tone only and is undone in one sentence; saving is ask-tier because it writes a file every future turn reads (ADR-0011 amendment). - CLI-first:
GET /api/modes,PUT /api/modes/active,POST /api/modes,DELETE /api/modes/{slug},POST /api/modes/{slug}/restore, and ajarvis modesgroup (list, show, activate, create, delete, restore).[persona] active_modeis the sticky choice; a screen-scoped override lives in memory only. - A mode can name the voice it speaks in. A friend should not sound like a butler: a mode may carry a TTS voice id; an empty one keeps what you configured. A realtime session pins its voice when it opens, so there a switch takes effect on the next call; a voice the provider does not have falls back to the configured one — going silent is not recoverable.
- A Modes section: the shelf, and a workshop where you make another one.
A card opens with the full character text verbatim, the knobs and the
voice; every card carries an explicit "Use", "Edit" loads any mode into
the form, a built-in keeps "restore original". "Talk it through" opens a
real voice interview — the assistant asks what kind of company you want
and writes the mode itself. The mode in force and the mode you chose are
shown apart (the payload gains
Instant acknowledgment: on a heavy turn the first sign of life arrives within seconds, and it speaks to the request (ADR-0033). Turns that go to the Tool Model or a sub-agent left the user 5–30 s without a word. A shared core now decides at dispatch, from the deterministic turn plan, what kind of work starts and how soon to speak: research, screen, mission and connected personal lookups immediately; actions and local lookups after a 3 s grace and only if the turn is still processing. The line is request-specific ("I'm looking up ."), model-composed and accepted only by a structural validator — never a stock "on it"; closed de/en/es pools naming the KIND of work are the instant fallback. Both engines get it, realtime and classic pipeline; when the work outlasts the first line by 8 s, ONE more line grounded in the tool actually running ("Still searching."). The chat shows the same first line as a muted pre-ack bubble. Switches:
[ack_brain].instant_ack(kill switch),[ack_brain].instant_ack_compose_all(on).You keep talking while the tool model works (ADR-0034). A heavy turn no longer blocks the conversation in the realtime engine, provider-neutral. Parked results never expire: a late answer waits until the session is at rest — the 30 s bound that dropped the answer whenever you kept talking is gone — and, spoken after other exchanges, its opening ties it back to what was asked. The wire is freed the moment you move on: a new turn while an earlier order's provider function call is still open answers that call at once with a closed "still executing" payload (Gemini Live blocks on function responses; NON_BLOCKING is unsupported on Vertex and the 3.1 Live model),
[voice].realtime_unblock_pending_tool_calls(on). "How far are you?" is owned by the orchestrator before the planner can read it as a new order. Honest scope: the realtime engine ships this now; the classic pipeline's half —[voice].background_heavy_turns— is declared in the config in this release, but its pipeline reader lands with the pipeline change that follows; until then a heavy turn on the classic pipeline still waits inline.docs/adr/0034-non-blocking-heavy-turns.mdnames the cells covered, emulated and degraded.The Thinking pause decides when a turn is taken — in both engines, and you may keep talking. Jarvis no longer takes the turn on every short pause; it waits for a clear one, and words spoken after the pause are appended to the same request instead of being submitted twice. The Settings "Thinking pause" (
speech.vad_silence_ms) is now ONE value for both voice engines, and which lever applies is a transport capability: a transport that answers on its own boundary (gemini-live,vertex-live) gets the pause folded into its native turn detection; a transport whose responses Jarvis requests itself (openai-realtime, local, third-party) waits the pause out on the session's own microphone before it asks for the answer, and a transcript that lands while you still talk joins the open turn — ONE response answers the whole request. On the classic pipeline the VAD already waited; new is the hold when the final transcript lands while you are speaking again: the text is held and joined with the next utterance, so no instant ack talks over the second half.Speaking while Jarvis thinks or works now interrupts him — no command word needed. Barge-in during PLAYBACK always worked; during the silent THINKING wait and a running delegated action it did nothing, and a spoken "stop" was even answered with "I am still working on it" (BUG-135). A second local Silero detector is fed during the silent wait and fires the same
barge_incontrol the playback path uses; a deterministic check (regex only, de/en/es, whole-utterance anchored so "do not stop the music" is inert) then tells a stop, a redirect and a continuation apart. A stop cancels every running delegate and confirms out loud; a redirect cancels and routes the replacement, so Jarvis holds the NEW context, not silence.YouTube Music as a marketplace plugin — music by voice, on the account you already pay for. "Play Radiohead", "play my running playlist", "play my liked songs", "pause", "skip", "what song is this", "like this", "add this to my chill playlist", "create a playlist called Late night". Google publishes no YouTube Music API, so the plugin rides the official YouTube Data API v3 through the same Google OAuth client as Gmail, Drive and Calendar — one Cloud project, one more API to enable — and never a reverse-engineered private client.
- Playback happens in YouTube Music, not in Jarvis — the same model as
the Spotify plugin. Playing opens a
music.youtube.comdeep link, so Premium, recommendations and history stay in the user's own account; Jarvis watches the system's media session and says whether playback actually started — including the honest "press play once" when the browser withholds autoplay. - Pause, resume, next, previous and "what is playing" come from the OS
media session — the same channel the keyboard's media keys use: Windows
through WinRT (
winrt-*in the[desktop]extra), Linux throughplayerctl, macOS throughnowplaying-cli; without the tool the answer names the install command instead of a fake success (docs/os-parity.mdP-33). - Honest limits, said out loud. Google allows 100 searches a day per
project (repeats are cached, own playlists cost none); "queue next" does
not exist in YouTube's API and the skill says so. Setup:
docs/marketplace/youtube-music-setup.md. - A background player instead of a browser tab (default). A small player window of its own (a pywebview companion process with a persistent profile — log in once, keep Premium) starts minimized, loads each song in place, and answers pause / skip / "what is playing" — and, new, volume by voice, which no OS media session offers. Where the player cannot run (headless, no desktop extras) or when Settings → Music says Browser, the browser path stays.
- Two connectors, one domain — a preference instead of a coin toss.
With Spotify AND YouTube Music connected, "spiel Musik" went to whichever
music skill was registered first. New
[music] preferred_service(Settings → Music,PUT /api/settings/music): a named service always wins ("on YouTube Music"), else the preference, else on Automatic the only connected one. ONE resolver feeds both the deterministic skill capture and the two tools' descriptions, so the router and the capture never disagree.
- Playback happens in YouTube Music, not in Jarvis — the same model as
the Spotify plugin. Playing opens a
Spotify by voice — a native plugin, and the "no active device" problem solved rather than reported. Play, pause, skip, back, volume, "what is playing". Spotify publishes no MCP server, only the Web API, so this is a native REST tool behind the keyring-backed Connect button; risk tier
monitor. The Web API commands Spotify Connect: with one open device the command is re-aimed at it instead of failing withNO_ACTIVE_DEVICE; with several, the assistant asks; with none, it says so and offers to open Spotify. Playback control needs Premium — the refusal names Spotify's rule. Bring your own OAuth client: Spotify's Development Mode allows five users per app, sooauth_client_family: spotifymakes your own client the documented path (PKCE, no client secret); refresh tokens are capped at six months, so the card saysprovider_limited. The Client ID field had nowhere to save to; a parity test now closes that seam for every OAuth family. Setup:docs/marketplace/spotify-oauth-setup.md.Google Cloud Vertex AI as a provider family of its own, on every tier. Vertex serves the same Gemini models as Google AI Studio, but bills a Cloud project — and until now it was reachable only by accident, through an express-mode key pasted into the Gemini field. Vertex now has its own cards for the brain, the tool model, voice input, voice output and realtime, and is selectable as a subagent; all of them read ONE shared Vertex credential.
- Two ways in: a Vertex AI express-mode key (
AQ.), or a Google Cloud project with no key at all ([google].vertex_project,vertex_location, optionalservice_account_path, signed with Application Default Credentials). Measured live: a Google Cloud API KEY does not work with Vertex at all; the cards and docs say so, and anAIzakey pasted into a Vertex field draws a warning. - The endpoint is decided by the card, not guessed from the key, and Vertex publishes its own Live model ids. Vertex is a distinct credential family with no cross-read to the Gemini slots: the two accounts run out of quota independently, which makes crossing between them a real fallback.
- Worth picking for voice: the AI Studio text-to-speech preview model is capped at 100 requests per day no matter how you are billed — what made the voice go quiet mid-day. A Cloud project has no such cap.
- Two ways in: a Vertex AI express-mode key (
Gemini speech-to-text got the card it never had. Selecting it meant hand-editing
jarvis.toml; both Google recognizers now appear in Voice Input with a model picker.Marketplace: a community registry — browse, install and remove plugins, skills and wallpapers other people published, by name, from the app, a terminal, or by asking. New third tab in the Plugins view; every card carries a "Community — not reviewed" badge with publisher and version. Trust model:
docs/marketplace/community-registry.md.- The consent dialog is the trust boundary: before anything is fetched it shows verbatim where requests and your token would go (hosted MCP URL) or which command would run locally (stdio argv). Every registry rule is re-enforced client-side by the Agent Plugins v1.0.0 loader, and bundled package cards win over community cards, so no community entry can shadow a shipped plugin's vocabulary.
- Browsing works offline and never touches boot. The compiled index is
fetched with short timeouts, validated tolerantly and cached atomically;
states are honest — fresh, stale copy, unreachable, disabled — and an
empty
[marketplace].community_index_urlturns the section off.GET /api/marketplace/community, install/uninstall routes underneath. - Install by name — one route for every kind.
POST /api/marketplace/community/install/{item_id}resolves the kind from the index and answers in ONE shape: what landed, whether it is usable right now (a valid skill is; a plugin is not until connected) and what is still missing.jarvis marketplace install <name>— the line every marketplace page prints — works now: it shows what the entry is BEFORE fetching, asks once, then states the outcome ("ready to use — no restart needed", "on your list, but NOT connected"); piped or--json, no prompt can block a script. By voice,marketplace-browsereads the index so the exact name is found, andmarketplace-installcarries theasktier — a plugin brings an outside MCP server with it. - Wallpapers are a published kind, downloaded over https under a size ceiling and put through the SAME mill an upload is.
- Read an entry before installing it:
GET /api/marketplace/community/{name}/contentsshows a skill's instructions, a plugin's two manifests, a wallpaper's picture. The card shows the install as a terminal line to copy (skills also offernpx skills add), mirrored with the storefront. - What came from the marketplace says so, where it landed — one
MarketplaceBadgein the Installed tab, on a skill's row, and on a wallpaper filter chip; the origin lives in a sidecar beside the thing itself. An install from the terminal, by voice or from the storefront now reaches the open window (MarketplaceItemInstalledon the bus; the window reloads exactly the lane that changed, no restart).
Bring your own skill or plugin — drop a folder onto the window, or import it from a path or link. For months the backend could import a skill and no screen reached it; the honest answer to "how do I add my own skill?" was "copy a folder and restart".
- Skills: an Upload button in the Skills view. Drop a folder, pick one,
or hand over a .zip; the server reports what it found — name, the paths
that will actually be installed, and every blocker at once — and only then
is there anything to confirm; a skill that trips the safety lint is
announced as landing in draft BEFORE the install. Routes:
POST /api/skills/upload/inspect,POST /api/skills/upload; staging is fail-closed. From a terminal,jarvis skills import <path-or-url>routes local paths to the newPOST /api/skills/import-local. - A SKILL.md written for another agent reads now. A file in the open
Agent Skills format — the one
npx skills addinstalls into Claude Code, Cursor, Codex and the rest — used to die on a single foreign frontmatter key. A second, tolerant reading is tried only after the strict schema rejected the file: only descriptive fields are adopted, by whitelist; nothing that grants behaviour crosses over, and a foreignstatecan only hold a skill back. Every dropped key is listed on the skill.docs/marketplace/portable-skills.md. - Plugins: drop a folder holding
plugin.jsonand the preview shows the catalog card it will produce, most importantly the authentication mode. A local upload getssource: "local"with its own mark — never the "community" badge and the review it implies.
- Skills: an Upload button in the Skills view. Drop a folder, pick one,
or hand over a .zip; the server reports what it found — name, the paths
that will actually be installed, and every blocker at once — and only then
is there anything to confirm; a skill that trips the safety lint is
announced as landing in draft BEFORE the install. Routes:
Plugin store: search across every field, filter by connection status, and the whole card connects. The search box matched the display name only — "payments" found nothing; free text now matches name, description, category, catalog id and OAuth client family, and a connection-status filter joins the category menu. Clicking anywhere on a card starts the connect flow (ONE lock, so two clicks can never race into two OAuth flows); a connected card stays inert, because disconnecting remains a deliberate click.
Agentic IDE: a split tree with hand-resizable terminals, branded title bars, and a typed prompt bar that composes exactly like the spoken one. Splits stay local and dragged sizes persist (
POST /api/agentic-ide/layout/weights); "Even out" measures TERMINALS, so one click lands every pane at the same width. The focused pane's call-sign wears a filled brand plate (signal-yellow on dark panes, gold on light ones), and all title-bar ink is keyed to the PANE's own ground, so a light pane inside a dark app stays readable. Typing into the prompt bar sends one request with compose on: Jarvis writes the briefed task with@filereferences and types THAT into the pane, and the composer's progress beats ride the event bus (AgenticIdeComposeProgress), so 10–30 s of real model work no longer looks like a wedged spinner.Agent accounts show how much of each subscription is left. Holding several seats of the same coding CLI made the switch possible but left the number the choice is made on invisible. Each row now carries its own meters — the rolling 5-hour window, the weekly one, and any per-model weekly budget the provider scopes separately; readings are live where a provider offers them, else what the CLI last wrote to disk, and every block says which. No token is ever returned or logged.
GET /api/agent-accounts/usage.Your own skills actually fire now. Fourteen days of live telemetry showed zero model-initiated
run-skillcalls — the "my skill never fires" complaint, measured. Naming a skill captures deterministically (an imperative use-verb, the word "skill(s)" and an installed skill's spoken name together fire it; a question about a skill is vetoed); the relevance layer leaves shadow mode ([skills].relevance_shadowdefaults tofalse); the model sees EVERY active skill (cap 48 instead of 20, user-authored first); a clear NARROW match carries the skill's full instructions inline, so a fast router model does not turn a hint into a tool round trip.Create a skill by voice — and it is a real skill, not your sentence in a template. "Erstell mir einen neuen Skill: Morgenroutine, jeden Morgen um 6 Uhr E-Mails, Tickets und Kalender vorlesen und dann ein 80er-Klassiker auf YouTube Music" now ends with a written skill: the assistant itself authors the whole card — name, the spoken phrase that starts it, a schedule trigger (
0 6 * * *) when you named a time, numbered steps that each name the connector actually attached right now, and a spoken answer format — and files it as a draft you switch on in the Skills view. New router toolcreate-skill(one bounded call, never a background worker), the same author behindPOST /api/skills/creator/authorand the newjarvis skills create "<what it should do>" [--name] [--trigger] [--schedule]. The author is a provider LADDER, not one model: your active provider first, then the pinned Tool Model, then the API quality tier, then the frontier chain — and if nothing can author, the assistant says so and writes nothing. The UI's skill-creator dialog andjarvis skills draftsee the same live tool inventory;draftgained--trigger,--schedule,--language.- The rest of the lifecycle by voice: "welche Skills habe ich",
"aktiviere den Skill Morgenroutine", "deaktiviere den Spotify-Skill", "lösch
den Skill Abendroutine" are first-class app commands now (
skills-listover a new leanGET /api/skills/brief,skill-enable,skill-disable,skill-delete— the last one asks first). Enabling the draft the assistant just wrote is one sentence away, and it happens only when you ask.
- The rest of the lifecycle by voice: "welche Skills habe ich",
"aktiviere den Skill Morgenroutine", "deaktiviere den Spotify-Skill", "lösch
den Skill Abendroutine" are first-class app commands now (
The assistant knows its own CLI. The
cli_jarvisctltool now carries the completejarvisctlcommand tree with argument hints, parity-tested against the real CLI — no more turns spent reading--help.The memory map is a solar system that turns around YOUR page. The Wiki section's 3D map used to turn as one rigid body and read as a flat disc. The hub is the sun and it is your own entity page (
[memory.wiki.session_rollup].user_entity_slug); pages sit on soft shells by hop count, each turning at its own Kepler speed; the camera looks THROUGH the network. Every wiki write hands each page its previous place, so the network no longer explodes and re-settles on each change;prefers-reduced-motiongets the still layout.Run visualization draws a run as an n8n-style workflow. Every card carries a category — trigger, reasoning, command, file, search, web, integration, agent, result, deliverable — with its own glyph and per-theme hue; parallel workers branch into lanes that merge at the result.
?run=<slug>deep-links a detached window to the run it talks about.A plainly heavy request delegates without a magic word. Background work could only start on a delegation word or a confirmed offer, so "Build me a Flask app with a login and a start page" produced an inline answer — never work. A second route in the spawn gate now runs beside the vocabulary one: the turn must read as a request, name a deliverable, carry no small-scope word, and clear a scope threshold — AND the model must have chosen the tool itself.
[brain.routing].force_spawn_modegainsbalancedbetweenstrictandpermissiveand ships as the default; "build me a website" stays inline, add Flask and a start page and it goes.Background work reaches the user. The Conductor scheduler had been running from every boot and writing every result into its own database without a line to the bus, the voice or the notifications. Its Runner now emits structured facts and the app decides what is said and in which language. News means a state change and nothing else: a job that starts failing, one that recovers, a first-ever failure; the 288th healthy run is silence. The morning briefing is on the clock on a fresh install and follows the configured output language instead of hardcoded German.
Two plain wallpapers: black and white. The Wallpaper section now opens with four pictures that ship inside the app — the night original, its daylight twin, pure black and pure white — under two chips, "Original" and "Plain". Black is a dark wallpaper, white a light one, so adopting either switches the mode along with the ground; "Default" still returns to the original of the mode you are in. The plain grounds are drawn, not downloaded — present on a machine with no library and no backend.
Changed
- The router prompt shrank from 27.5k to 3k characters, and it now says
"do Y" instead of "never X". It made up 48 % of the system prompt on every
turn and carried 143 prohibition markers against almost no positive
instruction — the classic recipe for a model that answers instead of acting.
Gone: rules already enforced in code, rules the persona already owns, and
four contradictions resolved in favour of the acting side; kept verbatim:
the decision table, the
spawn_workerargument format, and every dated live-failure rule. The prompt also no longer orders silence — an empty reply reaches TTS as nothing, so a spoken command looked ignored. - Real questions go to the capable model. A keyword list decided the model tier: fifteen of sixteen substantive requests took the cheap model ("list the tradeoffs between …" matched a bare "list"). The capable model is the default now and no keyword can demote; only a whole-utterance greeting, thanks, farewell or clock question and a verb-first command of at most six words stay fast — where both tiers answer identically.
- The model is told which tools it actually has. The prompt rendered 25 names from a static seed list under "complete list — no others exist"; against a real surface of 84 attached tools not one was a callable name — one reason it said "I don't have a tool for that" while the tool sat in its surface. The block is now rendered from the live surface, and the delegated voice budget grows from 6 rounds / 20 s to 12 / 45 s.
- Agentic IDE: the brief writer answers soonest, and a readback that names
the wrong pane is caught and recorded. A spoken "T2, do a deep dive on …"
reached the pane 26 s after the sentence ended, 19 s of it the coding CLI's
own process start. Under
autothe order is now Tool Model → API tier → coding CLI (~5 s warm); a hedge that dies is replaced by the next rung instead of a 90 s fallback (BUG-145). The composer no longer spends 10–30 s of reasoning on a spec the receiving agent opens anyway: thinking is off, the brief is lean, and a terminal brief is written in 1–5 s. Separately, twice the live model reported the pane the USER asked about instead of the one the trusted result opened; a mismatch is now published as a recoverablereadback_identifier_swap. Observation-only for now: how often it would fire is not yet measured. - Gemini worker runs are recorded as structured streams. The Gemini CLI worker ran in text mode, so the Critic graded blind and the Visualization drew every Gemini run as bare Request → Result. It now uses stream-json.
- The marketplace storefront lists only registry-published entries. Until
2026-08-16 it listed all 21 built-in connectors as store entries under an
"Official" tick — and none was installable. The rule is in the contract now
(
docs/agent-contract.md§3) and binds app, registry and storefront alike: a listing is a registry-published entry with an Agent Plugins 1.0.0 manifest; built-ins belong in the "built in" wall and reach the store only through a registry PR like anyone else's; "Official" names an author, never a review — every entry keeps its "Not human-reviewed" mark, ours included. - Provider & mode parity is a binding rule of the contract. Every feature
— and every plan before it — targets the full provider × mode × OS matrix
from the first sentence: every Brain/Tool-Model family incl. local, every
realtime transport, every STT/TTS/Vision/Wake provider, realtime engine AND
classic pipeline, chat, channels, CLI. A provider-specific mechanism is one
declared capability; the rest get the generic path, an emulation, or honest
degradation (
docs/agent-contract.md§3).
Fixed
- Windows: the app is findable again — Start-Menu entry, Desktop shortcut,
and a launcher that leaves the Store Python's container (BUG-138). The
Start-Menu shortcut pointed at
pythonw.exe, and Windows never lists a shortcut whose target is a generic host as an app — it now points at a brandedPersonalJarvis.exeinside the venv. A Store Python runs inside an MSIX container, so every write to%APPDATA%was diverted into a private per-package tree while the shell saw an empty folder — the launcher is now published out of the container through a helper with no package identity. A Desktop shortcut is installed, repaired and removed alongside the Start-Menu entry, and only a real window run claims the shortcut. Same defect shape on Linux (update-desktop-database) and macOS (lsregister), fixed in the same change (docs/os-parity.md). - A start that cannot happen now says so, on every OS — and bare
jarvisopens the desktop app it always promised. Underpythonw, a macOS.appor a Linux.desktop, a boot-killingModuleNotFoundErrorwent nowhere. The one import that decides it (pywebview) is checked right after argument parsing (not in--headless), logged, and shown in a native dialog when stderr is the null device — a message box on Windows,osascripton macOS,zenityorkdialogon Linux. Separately,jarviswith no arguments ran a tray icon with no backend and no window, blocking forever; it now uses the same launcher asrun.bat, and the tray loop survives behind--tray. - One spoken order stays one turn — even past a thinking pause, even a long
one. A realtime provider commits the turn on ITS server VAD; live
2026-08-13, ONE order was chopped into three turns at ordinary hesitation
pauses and every fragment dispatched an executor — the same coding pane was
briefed twice with a quarter of the sentence (BUG-131, BUG-137). The
microphone is now the evidence: voiced input frames stamp a local "the user
is talking" state, never consulted while Jarvis speaks. A boundary that
lands mid-voice appends to the turn instead of splitting it; a delegate
holds its dispatch until the mic goes quiet; a background result from an
ALREADY ENDED call is never injected into a sentence. No added latency on
an ordinary utterance. New session knob
end_of_speech_sensitivity(defaultlow) asks the provider to be patient too. - A cough no longer ends the answer, and the realtime role stops flipping.
Every interrupted edge cancelled the running answer — and Gemini reports
background noise the same way, so a cough mid-answer left half a statement.
The final input transcript now does the actual cut, because a real barge-in
produces words and a cough does not (a settle backstop commits after a
second of silence). The tool directive is one standing role plus one short
per-turn line instead of three full-text versions per turn. Stated, not
fixed: Gemini's
start_of_speech_sensitivityis still unconfigured. - A delegate reply and a readback are each spoken once (BUG-143, BUG-148),
and Vertex Live gets its emergency voice. On Gemini Live one Jarvis turn
carried two turn-ending inputs on the server and both were answered back to
back ("Ich habe work geöffnet: T eins." and then "…: T1."). The session
cannot cancel that second generation, so it refuses to play it: a bounded
guard discards a generation that begins with no open turn, no new
transcript, no local voice and no injection. The guard is armed after the
desktop's speaker queue has drained — before, a long readback was spoken
twice and three grounded answers in a row were then never heard. And
vertex-livenow maps to Vertex TTS for the no-audio fallback, where the fallback used to stay text-only. - Realtime: "Let me check that…" is no longer cancelled mid-sentence. The
unbacked-promise detector ran on every streaming delta, and its "fewer than
24 characters follow the commitment" test is true by construction on a
growing prefix. The judgement now happens at
turn_complete, with a bounded backstop after 0.35 s of provider silence. Cost: an answer opening with a commitment is held for two or three deltas (roughly 100–500 ms). - Gemini Live: per-turn directives reach the model instead of being thrown
away.
update_sessiondiscarded every per-turn instruction — the turn directive, the language pin after a switch — while behaving as if it had been delivered, so the model answered from a frozen prompt. Changed directives now travel as a small out-of-band text turn (the channel BUG-104 established), delta only, on an open user turn; a prompted retry speaks instead of staying silent until the 20 s stall watchdog. - Vertex: a Google Cloud project boots as configured, a Live session opens
in about a second instead of six to twelve, chat and Live get different
regions, and the key field saves (BUG-139..141). The pre-boot key check
dead-listed
vertexon every boot ("no key") because the project path stores no key — Application Default Credentials sign; the keyless rescue now consults the same credential-probe table every card uses. Credentials and the TLS trust store are resolved once per process instead of inside every Live handshake — session open 5.7–12.2 s → 1.1–1.5 s — and the Live adapter declares the 20 s handshake budget it needs. Measured live,globalopens NO Live session while named regions lag a generation on chat: new[google].vertex_realtime_location; only aglobalvertex_locationfalls back to a named region, with a warning, because moving where audio is processed is a data-residency decision. On the cards, "Unknown secret key: realtime_vertex_api_key" — the Vertex key slots existed everywhere a credential is READ and nowhere it is WRITTEN; and the key-format hint no longer calls anAIzakey a "Google AI Studio key" — the label is endpoint-neutral, and a correctly pasted Vertex key gets its green confirmation instead of silence. - "Look here, close that window" now does both — the tool surface is narrowed, not emptied. Three gates removed almost everything the model could act with: a smalltalk turn kept one tool, a screen-image turn got an empty dict, and any turn whose verb missed a regex lost computer_use, both spawn tools and all five write tools — which is why "play music" had no vehicle. Each gate now hides a specific tool class on positive evidence (screen turns still cannot start a background agent from injected on-screen text); an image turn goes from 0 tools to 5. German separable verbs count as operations now: "Mach das Fenster zu" put its particle at the end and registered as a pure look.
- A refusal checks the tools actually attached, not just the capability registry. The evidence gate, the unsupported-intent check and the local-action gate consulted the registry alone — a cache that lags reality — so a freshly connected plugin, CLI or MCP server was callable and invisible at once: "I have no calendar access" while a calendar tool sat in the tool surface. All three now consult the live tool surface as the last step before speaking; coverage is decided on registered tool NAMES, and a Telegram tool does not answer a WhatsApp request.
- Three guards no longer block on everyday German words, and a question
opener no longer cancels the command behind it. A turn opening with
"warum" or "wie" blocked every side-effect tool, so "Warum ist Spotify
nicht offen, mach es auf" was guaranteed to do nothing; the guard stands
down when a real imperative follows. The meta-debug keyword list held
"Fehler", "Log", "Provider", "Bug" and "Phrase" and ended the turn with a
canned line even after another tool had run; those words count only
alongside a reference to the assistant itself. A research verb blocked
every
cli_*and MCP tool, so a question about your own cloud costs could not read your own data; it stands down on a true possessive. - A promise no longer swallows the answer it frames. An answer that opened with "Ich schaue kurz." and closed with "I'll tell you when that changes" matched the commitment-and-defer check on both ends, and everything between them — the actual answer — was replaced by a canned phrase. The text is split into clauses and each is classified by structure, the same in German, English and Spanish: with substance present the promise clauses are stripped and the answer kept. Two siblings: a streamed answer opening with "[" was dropped whole (a reply starting with a markdown link was lost); and the unverified-answer backstop deleted correct answers — it now replaces only a concrete claim the model could not have known.
- A tool call the model wrote inside prose runs, and its envelope is stripped whole. The recovery path required the ENTIRE response to be valid JSON, so a model that wrote "Ich öffne Spotify." in front of its tool call executed nothing. Envelopes are found by scanning for balanced braces; four gates still fail closed (an explicit envelope shape, an odd fence count means truncated, a quote against the envelope means quoted, short prose free of example vocabulary). No bare "}" is spoken.
- Six voice filters no longer destroy correct answers, and a scrub can no
longer speak an error mid-answer.
scrub_for_voiceis the last thing that touches the model's words, and six of its patterns ate the answer: a single shell token replaced the WHOLE text with the generic error phrase (every PowerShell how-to came out as an error); the narration stripper left nothing; the tool-args pattern deleted up to 400 characters after the words "action is"; the fn-call pattern matched ANY word followed by parentheses; the jargon list deleted "Provider" and "MCP"; every URL was deleted. Each keeps its protection with a narrower match. The streaming path filtered per sentence, so a short sentence that lost its only noun was spoken as an error inside a healthy answer — it is dropped now, and the fallback is spoken exactly once, only when the whole turn is empty. The sentence splitter no longer breaks "Am 1. Januar" into "Am eins." and "Januar", nor "z. B.". - Dates, versions and IP addresses are no longer read as one giant number. The number speller stripped the thousands separator before spelling, and in German and Spanish that separator is the dot, so "17.08.2026" came out as seventeen million and change and "Python 3.11" as three hundred eleven — on every voice turn. Dotted tokens are classified first (date, IPv4, version, thousands/decimal): dates are spoken as dates in each language, versions and addresses group by group. English dates reached TTS as raw digits and were never spelled at all; fixed in passing.
- One language resolver decides the turn's language, as the contract says. Three layers re-derived it with their own German-or-English guess, and each lost user-facing output: the ack generator scored a short German line as English and DROPPED it; the evidence gate refused in a sniffed language and knew no Spanish; the Cartesia and Inworld voices picked German on three bare articles. Now the ack asks the canonical validator, the brain passes the resolved language to the gate, the voice plugins take the caller's language code and warn instead of guessing, and the turn resolves its language once.
- No more waiting a minute for an approval nobody can give. An ask-tier tool blocked for sixty seconds on a confirmation and counted the silence as a refusal; only the voice paths ever passed the flag that lets a human answer, so the CLI, the REST API, the chat surface and every unattended runner walked into that stall and came out denied. A call now declares which approval channel it actually has, in the config snapshot the surface builds in code — never in the arguments the model produces. Desktop chat uses the two-turn confirmation the voice path already had; an unattended runner returns immediately with "approval was impossible", not "the tool was refused". Timeout and denial are distinct outcomes now.
- Scheduled workflows can call tools and drive the desktop. The workflow
runner was built with no tool registry and no harness manager attached, so
every
harness_dispatchortool_callstep raised "No HarnessManager available" by construction. Tools now arrive through a live view onto the brain. Known and deliberately left: an ask-tier tool in an unattended cron workflow blocks on approval and times out honestly — silent auto-approval is a safety decision, not a wiring fix. - Background work that dies says so out loud, and a finished mission never
goes unannounced.
spawn_workercreated the mission, found no Kontrollierer and returned after a bare log line (the mission stayed PENDING forever); a failed cron trigger, a crashed scheduler tick and a failed workflow step were log-only. All announce through the readback path scheduled tasks already use, so "exit 1" becomes a plain sentence; one sharedFailureAnnouncerspeaks the first failure at once and keeps a repeat silent for an hour. A mission whose summary did not survive the voice filter used to fall silent; the readback retries with the summary-less phrase. - Computer use: no acknowledgment for a screen action the machine cannot
run, the 120 s mission guillotine is lifted, a mission may wait for a slow
screen, and the queue wait is no longer charged to its budget. The
deterministic fast path spoke a commitment without checking that
computer-use is wired at all; it now runs the preflight before the
acknowledgment and refuses honestly with a
cu_not_wiredphrase in all three locales. A desktop mission was cut off after a hardcoded 120 seconds while the step budget allowed 100 steps; the ceiling is[computer_use].mission_timeout_snow, default 600. Waiting for a page or an installer looked exactly like being stuck, so the model re-clicked a control that was already working — which IS what trips the dead-end guard; the model is now told that waiting is work, bounded (six consecutive waits with no visible change, two minutes per mission). A mission queued behind an active one used to burn its whole budget on the desktop lock; the work clock now starts when the lock is actually held. - Agentic IDE: fleet briefs by voice land on the panes you named in the size
you said, hanging up ends a pane's brief, the folder step reads like
cd, and panes are always shown.- Six deterministic repairs to fleet briefs, each from a live 2026-08-12/13 call: "prompt the five claudes to X, the two codexes to Y" yields a per-CLI task map instead of every pane receiving the whole enumeration (BUG-132); counts in the TASK half of a fleet request never add panes — "open two terminals and the two should hunt bugs, one takes macOS and one takes Linux" opened FOUR billed panes for a spoken "two" (BUG-134); a spoken self-correction no longer opens the RETRACTED fleet (BUG-130); a garbled "Claudes" earns the "did you mean Claude Code?" question instead of two silently inherited panes.
- Composing a brief takes 16–30 s while a delegated voice turn is force-answered after 20 s, so one order produced four deliveries and a pane typed into 20 s after the user had hung up (BUG-136). Voice teardown abandons briefs still being written; REST and CLI keep the detach.
- A split follows the subscription switch unless its anchor's seat was chosen deliberately: only a pinned seat propagates.
- The folder step (BUG-144): a typed "cd project" went through verbatim to a
"Not a folder" review step; a leading
cdand quotes are dropped, a relative path resolves against the folder on screen, and a typed path becomes the selection only after the backend confirmed it is a folder. "Browse" opened BEHIND the maximised app; it opens in front now. A dropped folder used to be searched for by NAME; inside the desktop shell the host now reports the real path (WebView2 verified;docs/os-parity.mdP-34, macOS/Linux pending). - Panes are always shown, and the wizard advises from ten terminals instead of asking (BUG-146). Six panes on an ordinary window measured 56 columns each and every one was covered by a "needs about 60 columns" card; both that card and the wizard's block were built against a repaint bug measured at ~13 columns but fired at three to four times that width. Both are gone; the terminal is exactly as wide as its tile, and the thresholds (10 terminals, 40 columns) are one muted sentence of advice. Trade-off stated: the ~13-column blank pane is announced beforehand, not prevented.
- The mode you picked is read back, and coding mode stays scoped to the
Agentic IDE. The configured-slug reader imported a
get_configthat does not exist and a blanketexceptturned the ImportError into "the default mode", so every switch was written tojarvis.tomland never read: the assistant kept answering as Assistant. It reads throughload_config()now, and a built-in overlaid by your own file is markededited. Opening the IDE used to switch coding mode on and nothing switched it off, so the assistant stayed in coding mode in Chats and by voice until restart; leaving the section hands the mode back — only a mode this SCREEN switched on — through an in-memory override the registry derives from its own state at every transition, so it can never overwrite the mode you chose. - A request to CREATE a skill is no longer captured by a skill it merely
mentions. Voice session 2026-08-18: "… einen neuen Skill erstellst … ein
Lied mit YouTube Music …" was taken by the YouTube Music skill because its
brand trigger matched the words INSIDE the description, the model spent the
whole tool budget reading
jarvisctl --help, and the answer was "Das hat gerade nicht geklappt." A deterministic authoring resolver (jarvis/skills/authoring_request.py) now decides "the user wants a NEW skill" before any brand trigger runs — in every spoken conjugation ("Skill erstellst", "skill zu erstellen", "create a skill", "crea un skill") — and keeps force-spawn and the evidence gate out of such a turn. Theskill-creatorbuiltin's own card is a short "call create-skill once" now; Anthropic's long guide moved toreferences/anthropic-skill-creator.md. The same rule now covers every word the product uses for "a skill" — routine ("Morgenroutine"), automation, workflow, Ablauf — and lifecycle requests: "deaktiviere den YouTube-Music-Skill" no longer RUNS the music skill. - The wake word stops firing on random words and no longer goes deaf
mid-sentence (BUG-133). Four stacked causes in the Vosk keyword spotter:
the acoustic competition re-scored the whole 3 s ring instead of the
span-trimmed audio, a leading-isolation gate hard-rejected any wake with
speech shortly before it, the verify re-score pooled phrase words from
anywhere in the decode, and a stage-1 hit during backoff died behind a
stale candidate. New
scripts/vosk_wake_bench.pyreplays synthesized de+en voices through the real detector: false fires 10/48 → 0/48. - Dictation's polish pass may no longer cut the verb out of a sentence.
Measured on 68 live polished dictations, 33 had lost at least one word and
exactly one was ever rejected. Prompt v7 states the floor (a filler is a
SOUND; a verb, auxiliary, modal or negation is corrected in place, never
removed) and a new positional
lost_verbdrift guard enforces it. - Push-to-talk heals itself when the key-up edge is lost. A lost release (focus change, UAC prompt, RDP reconnect) left the recording pill open for the full 60 s max-hold while every later press was swallowed as key-repeat. A press arriving past the repeat grace window is treated as the release that never arrived and submits what was held.
- Marketplace security: a community plugin no longer inherits your keys, an
https-only download stays https-only past a redirect, launcher arguments
cannot walk around the allowlist, and a skill import refuses path-shaped
names. A connected stdio plugin was started with a copy of the whole
environment — where
OPENAI_API_KEY,ANTHROPIC_API_KEY,GITHUB_TOKENand every other exported credential live; an untrusted server now starts with an allowlisted environment plus its own declared token (not covered, and written into the trust model: the delegated worker path hands the server definition to theclaudeCLI). The https guard now travels with the request past every redirect. The stdio rules checked exactly the first word ofnpx,uvxordocker, while each has options that hand the command line back to the publisher; arguments are now an allowlist per launcher and the package must be a pinned name. And a skill's frontmatternamebecomes its install folder and was only checked for non-emptiness, so../../evilwas an arbitrary-file-write primitive; both import routes enforce the slug charset and containment. - Models the static price table never listed are priced, and the Live
socket's real model is metered (BUG-142). The deck's "API this session"
card showed 272k tokens as $0:
gemini-3.7-flashand the Vertex Live default were absent from the table, and the realtime session reported an empty model id whenever the card pinned none. Rates now go static table → cached provider feed (OpenRouter pricing rides onModelInfo.pricing) → $0 with a warning; realtime sessions expose the id they really connected. The table gainsgemini-3.7-flashandgemini-live-2.5-flash-native-audioand correctsgemini-2.5-flash/-liteandgemini-3.6-flashto the vendor's list price (verified 2026-08-18). - The desktop keeps its browser profile across restarts, and light mode has
an original wallpaper of its own. pywebview starts in private mode by
default: the WebView2 profile lived in a fresh temp folder per launch, and
everything the frontend keeps in localStorage died with the process — the
chosen wallpaper, the deck/classic surface, pane sizes, the theme cache —
which is how a restart landed on light chrome over dark artwork. The
profile now lives in
data/webviewper checkout (docs/os-parity.mdP-35: only WebView2 has been run live). And a mode WITHOUT a pick fell back to the one bundled night scene, so a switch to light put light chrome on a scene where nothing could be read; each mode now falls back to the picture authored for it. - Deck: the signal colours read on paper, the log shows what the live
session heard, and the crew roster stacks when narrow. The deck was built
on the dark stage only — tints that glow on black and vanish on ivory; every
tint is a pair now. The realtime session never publishes
TranscriptFinal— what you said arrives asTranscriptionUpdateflagged final — so a live conversation showed every SAY, THINK and DONE line and not one word of your own; the log and the word counters read both events. - The background music player: no per-turn config parse, an honest request
deadline, a player that always dies. The music tools read the preferred
service from a briefly cached read instead of parsing
jarvis.tomlon every router turn; the player client's request wait uses one fixed monotonic deadline; the host tears its window down in afinally, so a read fault can no longer leave a ghost player, and the free-form eval command is gone from the protocol.
Removed
- The typed composer left the mission deck. You speak on the deck, you type on the classic surface; the message bar duplicated the chat and cost the cards a row.
- The approval preview in the Agentic IDE's typed prompt bar is retired. Its fallbacks delivered raw typed text into a pane, which is what the composing bar exists to rule out; the composed brief is typed directly.
- The plugin store's "coming soon" strip is gone — every name on it has a real catalog entry.
- Privacy: two PII scrub manifests are no longer published. The scrub pattern files were tracked in the public repo, so the files defining what must never be public were themselves an inventory of it. They are untracked and gitignored; the CI docs scan falls back to a generic manifest.
[1.3.2] — 2026-08-12
The five hundred generated wallpapers are now one click away on any machine, the Contacts section grew into a real address book, and Gemini keys from Vertex express are routed automatically everywhere a Google key is used.
Added
- A "Download library" button in the Wallpaper section. It fetches the 500-piece generated collection (~190 MB) from the project's asset release, shows live progress, unpacks it safely, and fills the grid without a restart. The app stays fully usable without it: the bundled original remains the default, and a failed download is an honest message with a retry, never a broken section.
- Contacts became an actionable live master-detail section: profile fields, vCard import/export (UI and CLI), a deep link, and a layout that survives narrow panes.
- Google keys are probed once for their home (AI Studio or Vertex express) and every Gemini client — chat, realtime voice, wiki embeddings — follows that route automatically; the key form and cards explain it.
- The in-app feedback form is back beside the Discord invite.
- The voice bubble in the Agentic IDE shows a subtitle-style live transcript.
- Dictation keeps soft word onsets and lifts quiet utterances before they reach speech recognition.
Fixed
- The emergency voice keeps streaming when a TTS quota blocks the primary.
- Provider errors say "free tier day limit" instead of 500 characters of JSON.
- macOS: the input-monitoring probe reads
IOHIDCheckAccess's 32-bit enum correctly; on Apple Silicon the old 64-bit read could see garbage high bits and silently degrade the permission check. - Agentic IDE: a pane follows the size its agent really got, a briefed terminal position no longer spawns a fleet, and background tasks are reaped on any early exit.
Security
- pypdf lifted past PYSEC-2026-3655 and PYSEC-2026-3656.
[1.3.1] — 2026-08-11
This release teaches Jarvis to draw. Ask for a picture in plain conversation ("draw me a flowchart of this") and it renders one on the spot, files it in the Outputs gallery, and opens the new view that shows it. Runs themselves now appear as node graphs whose edges trace where the run actually went. The rest is a broad polish pass: self-repairing terminal panes, wallpapers you bring yourself, a memory that finds short names again, and a stack of macOS fixes.
Added
- On-demand pictures. Ask for a drawing and Jarvis renders the thing under discussion as a flow, hierarchy, comparison, timeline, or bar chart. The model supplies only labels; the app draws the markup itself, so nothing model-written is ever served as HTML. The tool is offered only on turns that actually ask for a picture.
- A Visualization section. Every run is drawn as an n8n-style node graph, a live run's timeline follows the mission while it works, and a run's gallery page leads with its mission map.
- Wallpapers, properly: bring your own picture, mark favourites, and give dark and light mode each their own wallpaper. The mode toggle swaps them.
- Your own coding CLIs in the Agentic IDE: register and edit extra terminal CLIs right from the terminal step, with documentation on how.
Changed
- The API-keys view answers "am I done here?" at a glance: a saved key says "Key saved" next to a green check, and the get-your-key link names the site it points to.
- A fleet spawn that repeats an instruction already running on open panes is refused, and the refusal names who is already on it.
- The Command Deck no longer announces finished agents out loud. The bell and the on-screen report lane stay.
- Outputs show the user's real request instead of the internal quality directive, and light mode got a readability pass with minimum-contrast floors in agent terminals and over bright wallpapers.
Fixed
- Agentic IDE panes behave: overlapping panes detect and repair themselves, a pane too narrow for its agent says so, a clamped pane scrolls instead of clipping, and nothing changes shape under the pointer or during a half-drawn screen.
- The wiki memory finds short names again. "Joy", "Uwe", or "BMW" fell below the old keyword floor and were unsearchable; new pages also get their multilingual search aliases the moment they are written.
- macOS: terminal panes spawn as login shells, the first Terminal launch survives the Automation consent dialog, audio ducking honors never_mute and recovers from failed restores, and split UTF-8 characters in terminal output no longer garble.
- The managed local realtime server recovers from failures it previously could not see, reports honest boot progress, and delivers crash forensics with a crash-loop verdict instead of a silent retry.
- A wake call stands alone: speech leading into the wake word is hard-gated away from the command that follows (BUG-127).
- Three Win32 defects in the window-focus watcher, and Linux autostart entries are now spec-escaped.
[1.3.0] — 2026-08-10
This release makes fully-local voice a plug-and-go experience: the app now installs and supervises everything itself — including Ollama — connects in milliseconds instead of seconds, and lets you choose the brain model that fits your machine. It also adds a desktop wallpaper picker, gives every dropdown the app's own theme, and withdraws the unfinished ChatGPT subscription voice option until it works dependably.
Added
- A desktop wallpaper picker: choose the desktop surface's background from a thumbnail catalog, reachable from the sidebar and by voice ("wallpaper", "Hintergrundbild", "fondo de pantalla").
- The app installs and starts Ollama itself when local models are selected: runtime detection distinguishes not-installed / stopped / running, each with the right one-click fix (winget or the official installer on Windows, Homebrew on macOS, the official script behind non-interactive sudo on Linux — every other case gets an honest refusal naming the fix).
- One click on the self-hosted realtime card now covers the whole local chain: Ollama, the brain model download, the server environment, the vendored patch, and the proving smoke boot.
- A brain-model picker on the voice-server card: every installed and curated model annotated with an honest fits/does-not-fit verdict for this machine's accelerator; switching never needs a reinstall and an explicit choice that does not fit is refused instead of silently substituted.
- Local model cards reach the whole public Ollama library, not just the curated shortlist — any published model can be searched, downloaded, and used.
- A lifecycle supervisor for the managed voice server: prewarmed at app start, pidfile ownership (PID-reuse safe), start/stop from the card and the CLI, and an Ollama keep-alive so the brain model stays resident.
- Honest connect errors on every voice surface: a failed start attempt now names the provider and reason instead of silently returning to idle, and a deleted install fails fast with the repair action instead of retrying for two minutes.
Changed
- Local realtime connects in milliseconds: "localhost" is pinned to
127.0.0.1 (the OS resolver's dead IPv6 attempt cost
2 s per connect), and the SDK client and model probe are reused across calls (430 ms more per call before). - Self-hosted brains get a compact instruction profile (about a third of the size, prefix-cache-friendly ordering), cutting per-turn thinking time on small local models by several seconds.
- Delegated replies wait for the local server's own readback long enough (a declared per-provider budget) instead of muting the answer through a text-only fallback.
- Preflight tells the truth on unsupported GPUs ("no supported accelerator" instead of a false "0 GB"), and the managed install works with venvs created by a different Python.
- The curated local-model catalog was refreshed to the current generation and is verified against the live Ollama library.
- Every dropdown in the app now uses the shared themed select instead of the operating system's native control, so menus look the same on every platform (a guard test keeps native selects from coming back).
Removed
- The "ChatGPT subscription (Codex)" voice card no longer appears in the provider settings: subscription voice over Codex's experimental Realtime protocol is not reliable enough to offer yet. An already-selected configuration keeps working, and the option returns once the transport holds calls dependably.
Fixed
- A deleted managed install no longer leaves every call chasing a nonexistent server for two minutes; stale installs are detected and offered a repair.
- Killing the voice server no longer strands orphan processes: process groups are terminated with escalation on POSIX, and ownership is verified before any kill.
- Embedding models (BGE, GTE, E5, MiniLM and friends) and cloud-hosted tags are no longer offered as the local voice brain — the first would answer a spoken turn with a vector, the second would quietly break the "runs entirely on your machine" promise.
[1.2.3] — 2026-08-07
This release makes live voice calls survive crashes and restarts, lets the assistant actually use its personal memory in conversation, and polishes the Agentic IDE and the desktop orb.
Added
- Added a headless live-call probe and per-session postmortem forensics for the realtime voice path, so a broken call can be diagnosed from its record instead of reproduced by hand.
- Added the connection handshake to the visible call state — and stopped paying for it twice on the metered channel.
- Added terminal zoom helpers to the Agentic IDE and let the voice bubble headline wrap to two lines, so the sentence naming the agent stays readable.
- Added a light mode to the agents board.
- Added a self-hosted realtime option hardened for Windows machines without the symlink privilege, whose model downloads previously died mid-fetch.
Changed
- Personal memory now works retrieval-first: every substantive turn searches the local knowledge vault (single-digit milliseconds) and a strictness verdict decides what may ride along — after an audit found the old refuse-to-search default produced two full live days without a single injected memory. Questions about your own past ("when was I last…") are now recognized and answered from memory instead of guessed at.
- The desktop voice orb now carries the in-app bubble's look, controls, and energy, and shows its state inside the sphere instead of a speech bubble.
- The Agentic IDE workspace's chat rail became the project sidebar, and the chat view now reads the agent's conversation from the CLI's own record instead of transcribing the screen.
Fixed
- A crashed self-hosted voice server no longer ends the call: the reconnect window is sized from measured cold revives (including GPU TTS warm-up), the transport pump survives waking mid-rebuild, and the server is revived with fault diagnostics armed.
- Live-call transcripts are honest again: hallucinated user turns, subtitle outros, and the echo of the assistant's own barge-in cut are no longer recorded as real speech, and a transcript that arrives whole instead of streamed is accepted.
- The microphone is freed fast after a call ends, a dying socket can no longer hold it shut, and a silent line is no longer punished.
- The one-speaker rule is re-asserted every turn, taming the opening monologue on subscription voice calls.
- OpenAI models that only speak the Responses API are served instead of answering with a fake network apology.
- An ordered pane delivery in the Agentic IDE survives the caller's hangup, CLI renames and closes are announced to every open view, and "cannot be read" is told apart from "not written yet".
- On macOS, raising an app window via accessibility works again on Sonoma and Cocoa apps; on Linux/macOS a cleanly exited terminal child no longer reads as crashed.
- Blocking desktop calls were moved off the async event loop, removing UI freezes.
- Saying "look at the screen" followed by a content word is no longer mistaken for a screen-context request, and a reported "it spawned" is no longer misread as a request to spawn an agent.
[1.2.2] — 2026-08-05
This release makes realtime subscription voice and the Agentic IDE more dependable while adding richer voice controls and knowledge-map views.
Added
- Added native realtime voice transport with in-app readiness diagnostics, input-level feedback, browser microphone controls, and a complete action path.
- Added 3D UltraWiki memory maps with 2D/3D switching on both knowledge surfaces.
- Added Agentic IDE chat-or-terminal workspace choices, voice-orb context, prompt receipts, persistent terminal sizing, and clearer working/done state feedback.
- Added a real light theme selectable in Settings, replacing a switch that previously changed nothing.
- Added a draggable voice bubble to the Agentic IDE that replaces the fixed voice side column and speaks status updates from wherever it sits.
- Added UltraWiki word search over a meaning-neighbourhood of terms, so a query finds pages that use related words, not only the literal ones.
- Added the voice orb as a fourth on-screen display style: the glowing sphere from the Agentic IDE now also runs as a free-floating desktop window, so it can be dragged onto any monitor instead of living inside the app. Switching between the mascot and the voice orb applies immediately, without a restart. On a Linux session without per-pixel transparency the orb window stays hidden with one actionable log line rather than showing an opaque square.
- Added a REST-mounted project and chat library behind the Agentic IDE chat surface, so conversations and projects are reachable from the CLI too.
- Added an icon-rail sidebar: the navigation collapses by default and has an explicit toggle, giving the workspace more room.
- Added the Agentic IDE chat surface itself: a starting screen, a composer and quick actions, so a conversation can begin without opening a terminal pane.
- Added in-app downloading of local brain models, so the Ollama card fetches what it needs itself instead of sending you to a terminal.
Fixed
- Prevented subscription voice from hearing itself, inventing user turns, fragmenting replies, losing action responses, or going silent after teardown.
- Preserved the beginning of deliberate push-to-talk and call-hotkey speech, including quick follow-ups during the wake-word echo lock.
- Kept screen-context questions out of Computer-Use and routed provider failures across available families instead of leaving core paths silent.
- Hardened Agentic IDE pane recovery, scrolling, writer failures, terminal links, copy shortcuts, and live-tail switching.
- Repaired autostart, window focus, input capture, and audio ducking on macOS and Windows, and made every pickable key bindable as a hotkey on macOS.
- Moved desktop log writing to a dedicated thread so logging can no longer stall the thread that emitted the record.
- Stopped the wake-word shape gate from rejecting genuine wakes over its own spelling assumptions; verification now relies only on word-agnostic evidence.
- Made subscription realtime calls honest end to end: the surface shows a real connecting state during a provider's cold start, the reply language is pinned when the call opens, a stalled turn recovers instead of hanging, per-call cost is reported truthfully, and the transport pre-warms early and works on macOS and Linux, not only Windows.
- Stopped a config write from silently dropping every section it was not asked to touch, which could erase provider pins and most of a working configuration.
- Kept the Agentic IDE workspace on one screen with sideways scrolling, made a pane drop land where the pointer aims, and stopped a provider card that cannot be activated from stacking a wall of identical warnings.
- Made the onboarding local-brain button select the local brain instead of a cloud one, and stopped the sidebar painting rows at stale positions on macOS.
- Delivered the opening words of a spoken sentence instead of the fragment that survived the recognizer's warm-up, and let a delegated turn own its own answer rather than losing it to the turn that handed off.
- Stopped a local model that is still loading from being reported as broken, and stopped a text-only local model from advertising vision it does not have.
- Stopped the realtime transport from re-sending the whole instruction block on every turn, leaving a finished turn unclosed, or honouring a direct-speech clearance after its audio must have ended.
- Made the pre-push secret guard scan with built-in patterns when the privacy gate is unavailable, instead of loading nothing and reporting a clean push.
Changed
- Labelled the Agentic IDE and the Codex subscription realtime route as Beta, so their maturity is visible before you rely on them.
[1.2.1] — 2026-07-31
This corrective release reconnects the public 1.2.0 line with the complete maintained source tree and closes the release, install, and UI verification gaps found after 1.2.0 was published.
Added
- Added an experimental ChatGPT subscription path for realtime voice, with honest capability checks and fallback behaviour.
- Added the Agentic IDE chat view and a consistent coding-CLI picker when a workspace or chat rail opens another agent pane.
Fixed
- Restored the complete current implementation to the public release line, including the UltraWiki Python package and its frontend source.
- Preserved interrupted Agentic IDE pane recovery without mistaking the restored pane's own repaint for new work.
- Restored portable
[full]dependency resolution across supported Python, operating-system, and CPU combinations, including Windows on ARM64. - Fixed the README's package-page asset URL and aligned stale frontend tests with the shipped UI.
[1.2.0] — 2026-07-30
This is the first public release since 1.1.5 and it is a large one: a voice-driven workspace for coding agents, a personal knowledge base, voice that can run entirely offline, and a dictation pass that tidies your wording.
Added
Voice input and voice output can now run entirely on your own machine — no API key, no cloud account, nothing leaving the device. Three new providers appear in the API-Keys view, each marked Local · no key needed:
- Whisper (on this machine) — OpenAI's Whisper
large-v3, the full multilingual model, for the highest accuracy. One-time download of about 3 GB; noticeably slower than a hosted provider on a machine without a graphics card, which is the price of keeping everything local. - Nemotron (on this machine) — NVIDIA's Nemotron 3.5 streaming model. Covers 40 languages including German, downloads about 690 MB instead of 3 GB, and transcribes several times faster than real time on a plain CPU. No NVIDIA hardware required despite the name.
- Piper (on this machine) — the established offline voice engine, about 200 MB, with one voice each for German, English and Spanish. It speaks faster than real time on a CPU. It sounds good rather than indistinguishable from a person; a hosted voice is still the more natural option.
Everything is installed from inside the app. Each card says honestly whether its engine and model are actually on this machine, and offers a single button that fetches what is missing. A provider whose files are not there cannot be activated at all — instead of being switched on and then failing silently on the first sentence.
It degrades honestly. If you select a local provider and later start Jarvis on a machine where it is not installed, voice input crosses to whichever cloud provider you do have a key for rather than going dead. Local speech recognition also tells the rest of the app that your words never left the device, so the optional dictation clean-up stays local too.
- Whisper (on this machine) — OpenAI's Whisper
Dictation now tidies up your wording, and this is ON by default. After a dictation is transcribed, a fast model rewrites the structure of the text — punctuation, capitalization, filler words and sentence breaks — before it lands in whatever you were typing in. Your exact words and their meaning are kept; the pass is not allowed to rephrase you, and a rewrite that drifts too far from what you said is discarded and your own text delivered instead.
Where your text goes. While the pass is on, the finished transcript is sent to the model you selected. If your speech recognition already runs on your own machine, the pass stays there too and never crosses to a cloud provider on its own — picking a cloud model in the dropdown is the deliberate exception. On an install with no text-model key at all, nothing happens and the raw transcript is delivered exactly as before.
Your raw text is always kept. Every dictation stores what was actually recognized alongside what was delivered, and the history shows both, so a rewrite you dislike is never a loss. Each row also says what the pass did to it, including when it did nothing and why.
Where the switch is: the Voice section → Language tab → Clean up my wording. Turning it off is immediate and needs no restart. The same tab chooses which model family answers, and a Test button runs one fixed sample through your own setup so you can see the before and after before trusting it with your words.
Agentic IDE — a voice-driven workspace for coding agents. Open several named terminal panes, each running a different coding CLI (Codex, Claude, Gemini and others), then split, drag and resize them freely. Address a pane by its name, spoken or typed, and the instruction goes to that agent; one request can brief a whole fleet at once and reports back which agents actually received it. Drop files onto a pane and they are analysed before they become part of the prompt. Workspaces and sessions survive a restart and offer to resume. When a spoken call-sign is unclear, the assistant asks once instead of guessing — and says so plainly when an instruction reached nobody.
Screen Context — private, one-shot visual context for voice and chat. Jarvis can capture the active screen only when explicitly asked, shows a visible capture indicator, redacts password fields and sensitive text, and never stores the image. Windows, macOS and Linux/X11 are supported; Wayland and headless systems fail closed with an honest explanation. Screen observation remains separate from desktop-control permission.
UltraWiki — a personal knowledge base built from your own material. A staged import pipeline, hybrid keyword and semantic search, and an Explore view with topics, moments and a graph. Readers cover local folders, GitHub issues and pull requests, cloud-storage attachments, phone media (with image description and audio transcription), generic HTTP and RSS sources, and Obsidian vault import and export.
What it needs: semantic search requires one embedding backend — Ollama runs locally with no key at all, or use a Gemini, OpenAI, Voyage, Mistral or Cohere key. Without one, search runs on keywords only and the built-in health checklist says so rather than failing quietly. Storage is SQLite by default and needs no setup; Postgres and Supabase are optional (
pip install "personal-jarvis[ultrawiki-postgres]"). Your own content never leaves your machine except to the providers you configure. UltraWiki can now answer questions with real citations, returnsinsufficient_evidenceinstead of inventing a source, keeps configured folders fresh automatically, and lets users edit source settings in-app.Local and subscription brains. A generic provider for any server speaking the OpenAI chat API (llama.cpp, vLLM, LM Studio, HF serve), a keyless local Ollama provider, and an Anthropic-subscription option — a fully offline or self-hosted setup is now a real path.
Marketplace connectors, including Home Assistant, plus a self-hosted server option for supported services.
Changed
- Wake-word verification no longer assumes the wake phrase is spoken in German; it follows the language actually in use.
- Skill routing is deterministic and relevance-scored, and mission workers now reach the same knowledge the voice assistant does.
- A coding CLI is a registry entry rather than a hardcoded path, so a newly supported agent is reachable by voice the day it lands.
- UltraWiki's background import paces itself against a configurable share of the machine instead of monopolising a core.
- The in-app documentation now covers 50 maintained guides, including Dictation, Agentic IDE, Screen Context, UltraWiki, local AI and Home Assistant.
Fixed
- A dictation that lost part of its audio no longer reports itself as a success. It is marked as partly transcribed, the recording is kept, and the history offers to transcribe it again.
- Dictated text lands where you were actually typing. The transcript now carries the target the pipeline resolved, so a dictation meant for another program is no longer written into whatever Jarvis field last had focus — invisibly, in a section you were not even looking at.
- A finished terminal pane now rings the bell. The notification used to wait for something a particular coding CLI prints — first an interrupt hint, then a running clock — and neither is printed by every product in every phase, so panes finished all around you and announced nothing. A pane is now judged by whether its screen is still moving, which is true of any terminal and of a CLI nobody has taught it about yet. A pane that ends says why.
- A skill imported from a link no longer activates itself. Imported skills are stored as drafts and stay inactive until explicitly enabled, and a downloaded file can no longer declare itself active.
- A misheard agent name can no longer deliver an instruction to the wrong terminal.
- A denied tool call inside a mission is treated as a recoverable setback rather than ending the whole mission.
- Sub-second audio gaps mid-answer, a realtime fallback misreported as a hangup, and a turn-planner failure ending a live call.
- A spoken mention of an unrelated product name could unlock desktop control; dictation and auto-type tools went dead while the app ran elevated.
- Disconnecting a marketplace plugin now revokes the grant at the provider instead of only forgetting it locally.
- A dictation is no longer punished for being corrected. The safety check that watches for words going missing counted a repair as a loss, so "deskto" becoming "desktop" was treated as a deleted word and the tidied text was thrown away — taking the corrections you wanted with it. A word that reappears spelled correctly now counts as fixed, while a word that genuinely vanishes still stops the pass.
- Dictation recognises the spoken language reliably, and translation works out of any supported language rather than only out of English, with the two ends no longer swapped.
- The on-device voice speaks the right language. Switching to the local Piper voice inherited a pinned German setting from whatever provider came before, so every answer came out in German — including the ones meant to be English or Spanish. A downloaded voice you picked yourself is also kept across a switch instead of being reset.
- A terminal pane no longer shows a screen full of garbled characters when you return to it: the replay was being drawn over what was already there instead of on a cleared screen. The pane scrollbar also works in every CLI now, rather than only the one whose wording it was reading.
- The app starts and responds faster. A catalogue of installed extensions was being re-read from disk on every lookup — a sweep across hundreds of packages that blocked the app for seconds at a time. It is read once. A stall in the interface now also names the code responsible, instead of reporting that something, somewhere, was slow.
- Startup now recovers cleanly from malformed structured environment values instead of leaving the desktop app unable to open.
- Chat history once again prefers real voice sessions, hides phantom text threads with no user message, and persists new text conversations without duplicate websocket output.
- UltraWiki no longer imports linked worktrees or retains raw content in deletion tombstones. Legacy tombstones self-heal, stale source timestamps are corrected, and Gemini capability probing no longer repeats a rejected request for every background summary.
[1.1.5] — 2026-07-26
Added
- Personal Jarvis is on PyPI.
pipx install personal-jarvis(orpip install personal-jarvis) now installs it from the package index, so no clone and no Git URL are needed. The one-line installer stays the recommended path; this is the isolated, any-OS alternative. - The project page carries the full README with its images, the license, the supported Python versions and operating systems, and links to the website, the repository, the changelog, Discord, and the issue tracker.
Changed
- Every
v*tag now publishes to PyPI automatically through GitHub Actions using Trusted Publishing — no API token and no repository secret is stored anywhere. A tag whose version disagrees with the packaged one fails the build instead of publishing an untraceable release. - The README's images and document links are absolute URLs, because PyPI renders the README on its own domain and does not resolve repo-relative paths — they would otherwise be broken images and dead links.
[1.1.4] — 2026-07-23
Added
- The bar is now yours to size. A new "Bar size" slider in
Settings → Appearance live-resizes the JarvisBar and every surface it owns,
remembers your choice, and defaults to a larger, physical-size-consistent
135% so the bar looks the same on any screen. It is also reachable over the
API (
GET/PUT /api/settings/bar-size). - The bar follows you across monitors. It moves to the monitor your mouse is on, and dragging it from one screen to another now works correctly.
- Two more ways to run speech-to-text on a single key. OpenAI Whisper and Gemini cloud STT plugins let a downloader whose only credential is an OpenAI or Google key dictate without a separate transcription provider.
- A discreet, honest hint under the engine switch explains that single-mode is about which API keys you hold, not a lesser mode.
Changed
- Opening an app now fills the screen.
open_appmaximizes a freshly launched window, and also maximizes an already-running one when it is brought to the front.
Fixed
- Google Drive works again. Drive now talks to Google's REST API directly with the full scope instead of the hosted Drive connector, which only served Workspace preview accounts and returned "403" for everyone else.
- A single plugin no longer takes the main brain down. One tool's schema
used an object-key constraint (
propertyNames) that Gemini rejects, which had been failing the whole request and cascading through the provider chain; those constraints are now stripped before the call. - Hanging up from the bar can no longer freeze it or deafen the wake word. A realtime session's teardown is now time-bounded, so closing a call with the bar's X no longer stalls on a slow provider or leaves the bar stuck "listening".
- Escape and hang-up abort an on-screen action instantly, even mid-step while Computer-Use is still thinking, instead of only between steps.
- The wiki stops crying wolf. A content judgment during curation is no longer misreported as a provider chain failure, so the big red banner no longer appears when nothing is actually broken.
- Background text jobs cross to a different provider family when the default Claude API path has no key, instead of dead-ending.
- Telegram and Discord user-facing strings are now English.
- On Linux, Computer-Use honestly reports characters it had to drop instead of claiming success.
- The custom wake-word onboarding copy is corrected and now provisions the real Vosk fallback.
- Delegated action turns answer in the same language as the rest of the conversation.
- Computer-Use recovers from Gemini's "thinking required" errors and no longer falls back to a blind last-resort vision pass.
[1.1.3] — 2026-07-21
Added
- The wiki now learns who you are from how you talk, not only from literal
statements. Captured facts carry an evidence tier (explicit / behavioral)
and a personal-salience score: "I love being out on golf courses with my
buddies" now yields an
*(inferred)*profile bullet, while low-value world knowledge stays out of the vault. A one-shotrecurate-profileCLI (dry-run by default, snapshot first) re-judges an existing profile, and an expandable memory-map view shows the vault's structure (ADR-0029). - Spoken failures now name their real cause, and calendar trivia ("what day is tomorrow?") is answered natively instead of being delegated.
Changed
- The assistant acts only on an explicit ask. Background agents spawn only when you ask for one (deterministic gate + a notice in the Agents and Outputs views), and Computer-Use missions start only when you explicitly ask for an on-screen action — a knowledge question is answered directly or via web search, never by driving your browser (BUG-107).
- One voice per call. The short-lived escalation that rendered delegated replies through the surface TTS was reverted: the native realtime voice speaks every reply in a session (BUG-086).
- The API-Keys view uses a compact layout on laptop screens, tells the truth about shared/fallback keys, and warns before deleting a key other features still depend on.
Fixed
- Realtime calls survive provider transport resets. The transport is rebuilt proactively inside Gemini's GoAway window, the rebuild's conversation seed is accepted again (BUG-104), and a raced disconnect no longer kills the call mid-sentence.
- Realtime answers are honest and audible. The live model no longer asserts stale pre-cutoff facts as current, no longer invents niche figures or drifts to a misheard sound-alike entity (BUG-106), no longer replays an already-delivered answer, and the 1-2 s post-turn deaf window on the desktop microphone path is closed. Mid-sentence provider pauses are de-clicked, speaker echo no longer confirms as barge-in or doubles the answer in a second voice, and the voice bars move in sync with the voice.
- Computer-Use missions know what they are doing. A corrective follow-up ("do it in my Chrome browser") now carries the original task and its constraints instead of executing the correction's literal words (BUG-105), and recent missions are visible as context to the next one. macOS action loops are much faster: bounded accessibility-tree walks, a capped focus probe (kills a 15-second open-app stall), and batched actions no longer refused by foreground-signature churn.
- macOS Keychain dialogs are quiet again. Secrets collapse into one vault item accessed through the Apple-signed security CLI — no more password-dialog storms at boot (BUG-103).
- Updates preserve your data on every path. The wiki vault is salvaged and restored across updates and resets, with a snapshot taken first and no silent failures.
- Providers and credentials behave on every setup. Brain tiers can read a
realtime-scoped key as a last resort, native Claude subscription login is
honored, tool models honor
reasoning_effort(fixes an OpenAI empty-output crash), rejected optional parameters are remembered per endpoint instead of retried every step, and realtime quota exhaustion falls back across provider families. - The wiki background curator no longer loops. Judge-rejected batches are bisected, repeated rejections back off and park the poisoned row, a dead session's vault lock is stolen immediately, and template/code scaffolding no longer pollutes the memory map.
- Boot and wake are snappier. The multi-second Vosk wake spawn delay on a busy CPU is gone and heavy imports left the boot path — voice-ready wall time on the reference Mac dropped from ~28 s to ~14.5 s.
- Fresh installs prefetch complete local voice models, and
open_apprecognizes installed macOS/Linux apps the whitelist misses.
Removed
- The desktop push-to-talk shortcut has been retired. Voice Keybinds now contains only Call and Hangup. Legacy push-to-talk values remain readable so existing configuration files still boot, but they are no longer registered or exposed through the Settings API.
[1.1.2] — 2026-07-20
Fixed
- Audio input and voice-output devices now stay accurate after hardware changes. The settings picker refreshes safely when microphones, speakers, headsets, or virtual devices are connected or removed, shows the actual input/output devices on the current OS, and preserves a valid selection without requiring an app restart.
- Wake, push-to-talk, and realtime voice recover instead of going silent. Wake-microphone reopen failures and stalled reads now trigger bounded recovery, push-to-talk release is reliable, microphone-send stalls rebuild cleanly, stale rebuild timeouts are ignored, and novel follow-up speech is preserved while self-playback and barge-in are cancelled atomically.
- Fresh-install model choices and fallbacks are respected. A model selected through the provider UI is no longer overwritten when no explicit router section exists, same-provider fallback models remain available, coding-CLI readiness/login state is truthful, and tests no longer inherit a maintainer machine's private provider configuration or credentials.
- Mission results and desktop workflows fail honestly and remain usable. Mission outcome handling, connected coding-CLI flows, saved-file drag-out, and macOS Computer-Use indicator behavior were hardened across success, retry, and unavailable-capability paths.
- Wiki capture uses one canonical entity taxonomy. Bundled vault templates, graph ingestion, and desktop rendering now agree on entity types, while newly captured facts surface consistently in the graph.
- Cross-platform setup and validation are more portable. Non-Windows Registry operations return an explicit unsupported result, subprocess and stdio checks recognize the running virtual-environment interpreter without relying on PATH, and simulated Windows paths keep Windows separators on macOS/Linux CI. Skill-draft traversal checks now reject both separator styles.
- macOS: Jarvis Bar hover controls now react and remain stable. The non-activating Qt panel could miss mouse-move events, while replacing its alpha input mask emitted a false leave as the pill expanded. The result was an unresponsive or flickering hover state whose close and microphone controls could not be used reliably. The panel now explicitly accepts mouse movement and reconciles the real cursor against a stable pill footprint, preserving distinct mouse-out and mouse-over visuals during idle, listening, thinking, and speaking states (BUG-095).
- macOS: the Jarvis Bar now follows the Dock's real visibility instead of an invisible work-area boundary. Fullscreen Spaces can hide the Dock while Qt continues to reserve its 57-pixel strip, which prevented the bar from being dragged to the true bottom of the app. The bar now uses the complete screen edge while the Dock is hidden, retreats above it when it returns, and restores the user's preferred position when it hides again. Menu-bar and multi-display safe areas remain respected (BUG-094).
- macOS: the Jarvis Bar is transparent and animation frames no longer pile up. Aqua-Tk 9 kept an opaque black Canvas backing and composited every new RGBA frame over the old one, producing a rectangle at rest and concentric red/green/gold outlines while speaking. The macOS companion now uses Qt's translucent surface with full-frame alpha replacement; Windows and Linux keep the established Tk color-key path unchanged. Bar clicks are also executed in the parent process again, transparent window padding passes clicks through to the app underneath, the companion no longer steals macOS foreground focus every 500 ms, and parent TTS loudness now reaches the companion equalizer (BUG-093).
- macOS: the repeated Keychain password-dialog storm is stopped. A Control key created by an older direct Python launch could make macOS ask for the login-keychain password again on every protected request — often dozens of identical dialogs, with Always Allow unavailable because that Python executable had no verifiable signature. Jarvis now performs one serialized credential read per process and, after the one necessary approval, safely re-creates that legacy item under the verified installed app identity. Normal restarts and source updates then reuse the app-owned item without asking again; direct development launches cannot weaken or claim its access rules.
- macOS: the uninstall command works again. On a Mac, the documented
one-liner
bash ~/.personal-jarvis/install/uninstall.shprinted a syntax error and did nothing at all — no prompt, no removal. macOS still ships a 2007 version of the shell the script is written for, and one line of the uninstaller was written in a way only newer versions understand, so the file could not even be read, let alone run. Affects installs on 1.1.0 and 1.1.1; Windows and Linux were never affected. If you are stuck on an affected Mac and want to uninstall before updating, this does the same job and skips the broken script:~/.personal-jarvis/.venv/bin/python -m jarvis --uninstall. - Shell scripts are now checked against the shell macOS actually ships. Nothing in the automated checks had ever done that, which is why the dead uninstaller shipped twice while every test stayed green. Every shell script in the project is now verified to be readable by that older version before a change can land.
[1.1.1] — 2026-07-19
Fixed
- macOS: realtime voice no longer talks to itself. On built-in speakers next to the built-in mic, the assistant's own playback could come back as a "user" turn and be answered — spiralling into an endless two-voice self-conversation (BUG-089). Realtime sessions now recognize their own recently spoken words (including every canned error phrase) and drop the echo before it can start a turn; a genuine answer that adds anything new always gets through.
- Outage apologies stop repeating themselves. When no language model is reachable, the spoken "I can't reach my language model" notice now comes at most once per half minute instead of on every turn — repeats complete silently and honestly in the log.
- The emergency fallback voice keeps the caller's voice profile. When the realtime voice dies mid-call and a reply is re-rendered locally, the substitute voice now matches the session voice's gender instead of hard-flipping to a male default — no more "second assistant" joining the call.
- The fallback voice also stops re-rolling its delivery mid-answer. The local re-render now speaks a reply as ONE take (honoring the configured voice-consistency knobs), so a long answer can no longer audibly change character between sentences (BUG-090); session records now label each turn with the voice that actually spoke it.
- Smoother replies while you can interrupt. The local interrupt detector's per-frame inference moved off the audio loop — one less stutter source on slower machines.
- macOS: the status bar comes back after a crash. The out-of-process bar host now respawns itself (bounded, with honest logging) instead of staying invisible for the rest of the session.
- The session build warns when the realtime provider and every configured brain provider share one credential family — the setup in which a single quota error silences both at once.
[1.1.0] — 2026-07-18
Added
- Desktop control is on by default for fresh installs. New installs can ask Jarvis to operate the computer out of the box; every action still runs through the safety tiers, and the switch remains in Settings.
- Live "Test" button on the Claude / Codex / Antigravity agent cards — verify a connected coding CLI actually responds, right from Settings.
- A release-safety gate against half-shipped UI bundles. Commits and pushes are now blocked automatically when the shipped web UI references files that were never added to the repository — the failure previously produced a permanently blank window on every fresh install.
- macOS: Keychain access is a first-class permission. The permissions view now surfaces it with guidance instead of leaving credential saves to fail silently.
- Wiki memory keeps up with realtime voice. Conversations held in realtime mode are swept into the wiki by an evidence-safe background backfill, and a profile update can create its missing topic page in the same batch.
Changed
- Conversation mode is now the default. After Jarvis answers, the mic
stays open for a natural follow-up; one-turn-per-wake becomes an explicit
opt-in (
[trigger].single_turn_mode = true). The developer speech CLI now honors the same setting. - Field-tuned routing defaults (spawn / smalltalk / marker lists) now ship for every install instead of living only in the maintainer's local config.
Fixed
- Jarvis no longer answers its own voice. Speaker output picked up by the microphone could start a phantom turn (BUG-084); the echo is now suppressed at the source.
- One-click updates got honest and resilient. A transient release-check failure no longer breaks the update button; a staged-but-unfinished update is surfaced as "finish the update" instead of silently starting over; a failed install after restart reports the rollback instead of pretending nothing happened; the status overlay never offers a non-newer version.
- The uninstaller stops the running app first instead of failing to delete files that were still in use.
- The model picker no longer crashes on fresh installs that have no
[brain.providers]section yet. - Installed coding CLIs are detected reliably even when the desktop app's environment lacks their install directories on PATH.
- macOS: the transparent window backing is re-asserted on every reveal, with loud diagnostics when the pyobjc layer is missing (BUG-075 follow-up).
- Dependency security floors:
mcp>=1.28.1,json-repairdeclared explicitly; the Windows-ARM64cryptographyexposure is documented. - Realtime voice sessions survive connection rebuilds intact. A rebuilt provider transport now inherits the running call's transcript and keeps one voice identity across every native rendering order; the frozen turn is mirrored to the surface, and the session-end event fires from every surface, not only the browser (BUG-085/086/088).
- Wiki capture got failure-proof. Recently-failed providers are demoted to the end of the fallback chain, a slug collision no longer demotes valid links (old demotion scars self-heal), and a failing companion topic page no longer blocks the primary fact.
- macOS uninstall no longer triggers a Keychain password-prompt storm during credential removal.
[1.0.12] — 2026-07-18
Fixed
- macOS permissions no longer look "auto-denied" after an app update. Updating rebuilds the app bundle with a new ad-hoc code signature, and macOS then orphans every previously recorded permission: Microphone falls back to "not asked" while Input Monitoring and Input Control read as silently DENIED without ever showing a prompt. The installer now detects the signature change and resets its own stale permission entries, so macOS asks fresh instead of inheriting a dead denial (BUG-083).
- "Open Settings" now always lands on the requested privacy pane. macOS System Settings ignores the pane deep link while it is already running and just raises whatever pane was open last; Personal Jarvis now closes a running System Settings first so it relaunches on the right pane.
- Screen Recording no longer shows a stale "Not allowed" with a dead Allow button. macOS freezes that permission's status until the app restarts and never re-prompts after the first request. The permissions UI now shows "Restart pending" instead of the stale state, hides the request button that could never prompt again, and keeps the restart call-to-action.
- Background agents start only on an explicit request. A deterministic gate now enforces the delegation contract at every model-chosen spawn site; a plain conversational remark can no longer start a background agent. A blocked spawn instructs the model to answer inline and, for genuinely heavy tasks, offer delegation — a clear yes then unlocks exactly one spawn.
- The installer self-heals a stale or broken install directory instead of aborting with an error.
- Dependency resolution no longer fails on ARM64 Linux: the on-screen indicator's optional Qt dependency is excluded where no compatible wheel exists; the indicator degrades to a logged no-op there.
Changed
- The installer banner mascot was redrawn as hand-drawn pixel art, crisp on light and dark terminals.
[1.0.11] — 2026-07-18
Consolidation release. v1.0.6–v1.0.10 were cut from a separate macOS-focused line; this release unifies both lines into ONE repository history, so it carries every fix from both sides. The repository also moves to the standard shared-history workflow: GitHub secret scanning with push protection is enabled, and releases now always ship the entire current state.
Fixed
Voice calls no longer end on their own right after Jarvis answers. Three independent causes fixed: the provider dropping its Live WebSocket after a long reply now triggers an in-place transport rebuild instead of a hang-up (BUG-071); a "hello?" probe while a delegated answer is still being computed gets a deterministic wait answer instead of derailing the session (BUG-070); and the hang-up gate is re-checked at the moment of speaking, so a stale preamble can no longer play into an already-ended call.
Delegated realtime answers are ~3.7× faster (live p50 15.6 s → 4.2 s): no thinking on tool rounds, stable per-turn caching, and text-leaked tool calls eliminated (BUG-072).
Realtime sessions could hang forever on shutdown when the pump's single cancellation was lost mid-await; the bounded wait now re-cancels (BUG-081).
Fresh installs no longer show "Model unavailable" for the default model. The Gemini default pointed at a model id the API no longer serves; model health checks now probe the exact model the runtime would use, and switching models in the picker takes effect reliably.
Wake word reliability: a shape-only offline confirm can no longer win against a stronger acoustic candidate, and the boot storm no longer starves the wake-model load (wake was deaf for the first minute after boot).
macOS JarvisBar renders correctly — the bar appears without the opaque grey box (Tk 9 paints systemTransparent opaque; the native window backing is now cleared) and survives Tk 9 init order; the wake engine no longer crash-loops on comma-decimal locales; fresh macOS installs no longer crash at first launch.
Computer-Use: desktop missions are serialized behind a global actuation lock, every mission gets an id with per-id cancel, and silently refused guard actions are surfaced instead of swallowed (BUG-082).
Sidebar logo shows reliably — missing assets return an honest 404 and the image self-heals instead of rendering broken.
Realtime voice no longer pauses or stutters mid-reply. When the live provider's output transcription lagged its audio (Gemini Live: routinely 3-22 s), the voice-scrub gate held the audio back — first as multi-second dead stops mid-word, then (after an interim 400 ms bounded hold) as rhythmic block-wise stutter. Mid-reply holds are now removed entirely: once the reply's opening transcript has been vetted clean, audio flows unconditionally and the scrubber acts as a trailing kill switch that still cancels the response on a detected leak. The strict fail-closed turn opening is unchanged (BUG-080).
macOS 15 no longer kills the app when hotkeys arm or Computer-Use types ("Personal Jarvis quit unexpectedly", SIGILL). pynput resolved the keyboard layout through HIToolbox TSM calls on background threads, which modern macOS aborts with an uncatchable illegal-instruction trap. Two-layer fix: global hotkeys now use a TSM-free Quartz event-tap backend, and a main-thread keyboard-layout snapshot is captured at boot so any remaining pynput keyboard path (e.g.
keyboard.Controller()in Computer-Use actuation) reuses it instead of touching TSM off-main — degrading to the pyautogui fallback, never crashing (BUG-077).macOS installer no longer dies with a bare exit code on uv-provisioned Pythons. The app bundle's native launcher is now an in-repo compiled C stub linked against the exact Python runtime in use — replacing the py2app alias stub, which only worked on framework Pythons and left Intel Macs (uv standalone bootstrap) with an unlaunchable bundle. Desktop-integration failures now write
data/logs/install-desktop-integration.logand print the actual error instead of discarding it (BUG-076).Voice endpointing degrades honestly without onnxruntime. The WebRTC VAD fallback tier is now actually wired (Silero ONNX → WebRTC VAD → RMS energy); previously the middle tier was documented but never imported, so onnxruntime-less systems (e.g. Intel Macs) silently fell back to bare energy endpointing (BUG-061 follow-up).
The installer's speech-model report is honest — it reflects which models/runtimes are actually usable on the machine instead of assuming the full stack, and can always be produced without raising.
Windows-only dev scripts refuse to run on other platforms with a one-line message instead of an ImportError traceback.
Added
- Optional browser lock. The local web UI opens without the Control Key by default; the lock is now an opt-in setting.
- Per-voice audio previews for realtime provider voices in the settings.
- Computer-Use screen indicator — a gold glow border while a desktop mission runs, with Esc-to-cancel, backed by an in-process run registry.
- Screen-adaptive JarvisBar sizing — small screens shrink the bar, large monitors keep the approved look.
- Real macOS menu-bar icon. The tray icon is hosted on the AppKit main
thread (pystray
darwin_nsapplication+run_detached), completing the BUG-056 follow-up — macOS gets the same tray surface as Windows/Linux. - Mascot/orb on macOS. The mascot/orb overlay now renders in its own subprocess host (BUG-057 follow-up), with Aqua-Tk alpha transparency.
- macOS audio ducking. Music and Spotify are ducked via AppleScript for the duration of a voice session and restored afterwards, with an opt-in master-volume fallback.
Removed
- The stale root-level
install.sh(legacy clone-then-run script). The one advertised one-line installer remainsinstall/install.sh.
[1.0.10] — 2026-07-16
Fixed
- The one-line installer works in non-interactive environments again.
The welcome gate's terminal probe relied on
test -r/-w, which passes on CI runners and headless automation where/dev/ttyexists but cannot be opened (no controlling terminal) — the piped install aborted before doing anything. The gate now probes with a real open and quietly skips the question when no terminal is available.
[1.0.9] — 2026-07-16
Fixed
- Headless first boot: onboarding answers again. The full server's
security boundary now serves
/api/onboarding/*without a credential, matching the serve-first bootstrap. A fresh install on a headless box could otherwise never complete onboarding — the first-boot contract broke the moment the full app took over from the bootstrap (caught by the fresh-install smoke workflow on the v1.0.8 release commit).
[1.0.8] — 2026-07-15
Added
- macOS permissions surface. Runtime TCC permission probes (microphone,
accessibility, screen recording) with a settings panel, an onboarding step,
REST routes, and a
jarvisCLI command — degrading to a quiet no-op on other platforms. A dedicated macOS desktop CI workflow guards the path. - In-app docs system. Authoring pipeline, docs overview, sidebar and full-text search UI, plus a public-docs CI check.
- Wiki grounding. Extraction now records bounded, secret-redacted evidence excerpts (migrations 0006–0008) with an audit trail and a backfill for existing entries.
- Computer-Use foreground target guard — actuation tools verify the intended window is actually in the foreground before clicking or typing.
- Brain tool-call recovery — malformed provider tool calls are repaired instead of failing the turn.
- Persistent realtime voice sessions stay open between turns, with a voice-mode badge in the session UI.
Changed
- Unified web-surface security: one cookie/bearer policy as route-level defense in depth, an auth gate in the frontend, and authenticated mission WebSockets and terminals.
- Transactional self-update: the relauncher applies updates as a guarded transaction with rollback on failure.
- Scoped provider credentials, realtime scrub-gate refinements, onboarding / socials / settings polish, and WS schema updates.
Fixed
- Desktop installs remain discoverable after an in-app update. Managed installs now register and repair the Windows Start-menu and Installed Apps entries, the macOS per-user app bundle, or the Linux application-menu entry. Installer, updater, first desktop paint, and uninstaller share one guarded lifecycle, while headless and developer checkouts remain untouched.
[1.0.7] — 2026-07-14
Fixed
- macOS first boot works. Three native first-launch aborts ("Python quit unexpectedly") fixed in one forensic series: the tray status item, the Jarvis bar/orb Tk windows, and the virtual-cursor overlay were created off the main thread (AppKit/Aqua-Tk abort natively, BUG-056/057); PortAudio re-initialization is now serialized single-flight and the global-hotkey event tap preflights the Accessibility grant instead of letting macOS kill the process (BUG-058). macOS runs with the desktop window + Dock icon; the menu-bar icon and on-screen bar return once main-thread hosting lands.
- Local speech pack install no longer blames your internet. A missing prebuilt wheel (e.g. Python 3.14 + av) is now diagnosed honestly, pip runs wheel-only on end-user machines (never a source build), and the installer prefers Python 3.13/3.12 until the native stack ships 3.14 wheels (BUG-059).
- Grounded wiki answers. "What is in my wiki" is answered by a new deterministic listing tool in one round instead of blind probing; contract pages carry a provenance warning; delegated voice turns get a hard wall-clock deadline with a forced final answer (BUG-055).
- Realtime stability. A benign cancel race no longer ends the call, and German (any-language) capability verbs reach connected tools directly instead of always spawning an agent.
Added
- OAuth token refresh lifecycle for marketplace plugins (scheduler, PKCE/token-store hardening, Gmail/Calendar REST updates).
- Sessions view rework: richer session detail and turn cards.
[1.0.6] — 2026-07-13
Added
- Realtime voice engine — first public release (previously withheld): low-latency speech-to-speech conversations with tool delegation, plus the tool-model pick and a broad reliability wave.
Changed
- A missing Node.js no longer blocks the one-line installer. Node only powers the optional coding-agent worker (Claude Code / Codex) and a few Node-based integrations; the installer now notes its absence and continues, pointing to the in-app path for adding the worker later — instead of turning new users away at the door.
[1.0.5] — 2026-07-09
Fixed
- The wake word now works for non-English speakers. Wake detection routes to a model that matches the language you actually speak — the right-language local keyword model, or multilingual Whisper as a fallback — instead of silently going deaf on an English-only model. The missing language model is fetched automatically on a language switch or boot, wake stays pinned to the CPU, and a new language selector (with a "Test wake word" readiness check) lets you set the spoken language directly in Settings.
- The Windows taskbar button shows the Jarvis mascot instead of the generic Python logo. The app re-launches through a mascot-branded executable that owns its window, and it self-heals a stale Start-Menu shortcut. Best-effort and fully guarded: on a read-only or Store-Python install it degrades cleanly with no change in behavior.
Changed
- Clearer API-keys screen. NVIDIA NIM is now flagged with a "not recommended" caution badge (its free tier is slow), Inworld TTS is no longer mislabeled as a realtime provider, and the Pipeline / Realtime voice-engine switch was redesigned for clarity.
[1.0.4] — 2026-07-08
Fixed
- Custom wake words work out of the box on a fresh install. The per-language
Vosk keyword-spotting model is now provisioned automatically — installer
prefetch, an off-boot self-heal on first run, and an in-app "Download wake
model" button — so a freely chosen wake phrase resolves to the reliable
any-word engine instead of silently degrading to the transcribe-and-match
path that cannot recognize a hard proper noun. Works for every supported
language (
en/de/es), with no training and no GPU, on any OS including Apple Silicon. The word-agnostic openWakeWord backbones now ship in the package, an unservable custom phrase degrades loudly (with a one-click fix) instead of failing silently, and onboarding verifies the microphone level and the spoken wake word before marking setup complete.
[1.0.0] — 2026-07-03
First public release of Personal Jarvis — a voice-driven meta-orchestrator that turns one spoken request into a fleet of self-checking AI agents.
Highlights
- Voice-first pipeline — wake word → speech-to-text → multi-provider Brain →
text-to-speech, fully streaming, with honest, language-aware readbacks
(
de/en/es). - Provider-agnostic by design — every tier (router, ack, STT, TTS, worker, critic) degrades or crosses provider families on a missing or dead key. No single provider is load-bearing, and credentials are managed entirely in-app.
- Cross-platform core — the base install boots on a headless
python:3.11-slimVPS; Windows-desktop and local-voice features live behind optional extras. - Jarvis-Agents mission system — isolated
git worktreeworkers with a self-healing critic loop. - Plugin marketplace (OAuth + MCP) and a cross-platform control CLI
(
jarvisctl). - In-app "Update available" button — managed desktop installs get a one-click "Update Now" control in the top bar when a newer version ships.
Fixed
- Desktop app could hang forever on "Getting ready to listen". The startup banner and the top-left voice status cleared only when the speech pipeline published its one-shot ready signal. If pipeline construction crashed or an un-timed model load wedged warm-up, that signal never fired and the UI stayed in "starting up" indefinitely — even though typing already worked. Added two fail-safes: the pipeline-construction crash handler now publishes an honest degraded-ready signal so the UI is released immediately, and a pipeline-independent watchdog in the web server force-releases the UI after a generous deadline. The banner can no longer stick forever.
- Local "Faster-Whisper" appeared as a ready STT provider even when not installed. The provider list never checked whether the local-voice extra was present, so the card always showed as configured on a base install, and its model dropdown listed all Whisper checkpoints regardless of what was downloaded. Local Faster-Whisper has been removed as a user-selectable speech-to-text provider; cloud STT (Groq / OpenAI / OpenRouter) is the supported dictation path. The wake word (which uses its own local Whisper) and the key-free STT resilience fallback are unaffected.
- Declared
clickas an explicit dependency so thejarvisctlCLI imports on a clean install (it no longer arrives transitively viatyper), restoring a green CI. - Restored
TRADEMARK.mdto the published tree and removed a dead documentation link from the README. - Stopped defaulting the archival store to the removed
chromabackend.
Changed
pyprojectmetadata for the public release: English, cross-platform description and a[project.urls]block.
[v1.0.0-board] — 2026-04-25
First consolidated release of the Jarvis Board: Phase A through D.
Added — Phase A (Personal Dashboard)
jarvis/board/aggregator.py:BoardAggregatorparses FlightRecorder JSONL fromdata/flight_recorder/, groups it per day, writesdaily_stats+personal_recordsintodata/board/personal.db.jarvis/board/store.py:BoardStoreas a read-only query facade for the API.jarvis/ui/web/board_routes.py: GET/api/board/personal/{summary, heatmap,tools,records}+ POST/refresh.- Frontend:
BoardViewwith<HeatmapGrid>,<ToolBarChart>,<StatsCard>,<PersonalRecordsList>. React-Query polling 30 s. rechartsas a new frontend dependency.
Added — Phase B (Achievements + AI-Bio)
- 10
AchievementSpecs injarvis/board/achievements.py: 7 Mastery (first_mcp,tool_dabbler/journeyman/master,triple_combo,sub_jarvis_summoner,ten_x_engineer) + 3 Reflection (centennial,kilo_club,one_year_with_jarvis). AchievementEvaluatoras anEventBussubscriber. Idempotent viaINSERT OR IGNOREonachievements.id.AchievementUnlockedevent injarvis/core/events.py.BioGeneratorwithBrainLikeprotocol. Anti-cliché test gate against 12 forbidden words. Brain outage → the old bio is kept.BioScheduler: asyncio tick 60 s, Sunday 18:00 + master-achievement trigger. Date guard viaaggregator_meta.- API: GET
/api/board/achievements, GET/api/board/bio, POST/api/board/bio/regenerate. - Frontend:
<AIProfileCard>,<AchievementGrid>with a live unlock toast viapushToastin the WebSocket handler.
Added — Phase C (Federation Backend)
- New subproject
board-backend/(FastAPI + SQLAlchemy + SQLite + PyNaCl). - Routes: POST
/api/v1/identity/register(admin-token), POST/api/v1/sync(signed), GET/api/v1/me, GET/healthz. - Ed25519 crypto + canonical JSON in
crypto.py. - Replay protection:
ts_ms± 5 min. Constant-time token comparison. - In-memory rate limiter: 10/min/IP on
/identity/register. - Pydantic
extra='forbid'as the central PII wall. - Multi-arch Dockerfile (amd64 + arm64) + docker-compose + healthcheck.
- README with three deploy scenarios (Localhost, Raspi, Hetzner+Caddy).
- Local:
jarvis/board/sync.pyas a background push client (60 s). - New jarvis.toml section
[board.federation](defaultenabled = false).
Added — Phase D (Friends + Federation)
- Backend routes: POST
/pair/{initiate,accept}, GET/friends, PATCH/friends/{pubkey}, POST/activities, POST/stories, POST/reactions, GET/federation/feed?since=..., POST/federation/reactions/inbound, DELETE/federation/identity/{pubkey}. - Tables:
friends,pair_tokens(10-min TTL, single-use),activity_items(visibility: private/friends/public, optionalexpires_atfor stories),reactions(UNIQUE constraint). interesting_score = reactions × exp(-age_h / 24)— deterministic, hardcoded halflife.FederationPulleras an asyncio task per friend (offline ≠ blocking).StoriesCleanup1 h tick, deletesexpires_at < NOW.- Frontend:
<FriendsView>with tabs "Feed" + "Manage",<PairDialog>with QR code (qrcode.react),<StoryComposer>(max 280 chars),<ReactionBar>(🚀 🧠 🔥, owner-only counts),<FriendsList>with a per-friend pull-interval stepper. - Settings page: new section "Backend Connection" (Disconnect, URL, copy pubkey).
- Local federation proxy (
jarvis/ui/web/federation_proxy_routes.py) with whitelist paths — the browser frontend does not sign itself, the privkey stays in the local backend.
Added — v1.0 Release Prep
tools/board_demo.py— bootstrap 2 backends + 5 identities + 30 days of activity.tools/board_perf.py— aggregator + federation-pull benchmark.tools/board_pentest.py— 19-vector pen test against a live container.docs/jarvis-board/ARCHITECTURE.md— for backend forkers.docs/jarvis-board/FEDERATION_PROTOCOL.md— wire-format spec v1.docs/jarvis-board/PERFORMANCE_AUDIT.md— aggregator 2.94 s / 365d, federation pull 17 KB / 10 friends.docs/jarvis-board/SECURITY_AUDIT.md— 19/19 pen test PASS.docs/jarvis-board/MIGRATION_v1.md— 4-stage migration for existing users.- README.md extended with a "Jarvis Board" section.
Fixed
httpxmoved from dev-only to a production dependency inboard-backend/ pyproject.toml—routes/pair.py,reactions.py,background.pyimport it on every container start. Bug uncovered during the first Phase-D container rebuild for the pen test (Phase D had until then run with the Phase-C image).
Security
- Three independent PII filter layers: aggregator whitelist
(
export_all_for_federation()), sync-client whitelist (_build_payload), serverextra='forbid'Pydantic wall. - Ed25519 sigs on all federated routes with re-canonicalize (reverse-proxy-resilient).
- Constant-time admin-token comparison + per-IP rate limit.
- 19-vector pen test: auth bypass, replay (past + future), tampering, PII leak, malformed body, SQL-injection regression, forget-me path mismatch — all PASS.
Not in Release
- Phase E (Public Aggregator / Strava-style segments): deliberately not built. Plan §0 requires ≥ 2 months of Phase-D burn-in first, so that anti-cheat mechanisms can be designed evidence-based. First Phase-E decision: ~ 2026-06-25.
- Bundle splitting: frontend JS is 444 KB gzip (Vite warning).
Functionally uncritical (<500 ms initial load on modern devices),
but code splitting for
recharts+@tanstack/react-queryis a follow-up PR.
[Pre-board] — before 2026-04-24
Phases 0–5 (skeleton, speech, plugin system, risk tier, memory,
harness dispatch, vision/computer-use/admin/async/control/telemetry)
are in the repo, documented in docs/phase{1a,1c,2,4,5}-*.md and
ADRs docs/adr/0001-0008. This CHANGELOG only starts with the v1.0
Board release — pre-board history is reconstructable via git log.