The most important Orkestrai changes, organized by delivery.
01
Orkestrai 0.53.1
The terminal never holds your typing again.
After an automatic message's Enter, what you type goes straight to the terminal, with or without a receipt. Only keys typed in the second between the paste and its Enter wait for it; if that Enter was lost, the automatic message is sent first and your text follows within about 4 seconds, never mixed in. 0.53.0 held typing until a receipt that could fail to arrive, leaving a leader idle for hours and Canvas bar and Conversation messages queued.
Each agent's queue keeps moving: after 30 seconds of an idle agent without the receipt of its last automatic message, the next ones are delivered without retyping anything. If Codex or Claude puts a mid-turn message back into the composer and nobody has typed since, Enter submits it once.
A receipt is recognized when Codex or Claude merges the automatic message with other composer text, or records it among several prompts of the same turn. This applies only to inbox messages, which carry a unique id.
On Windows and WSL, the Codex paste is closed with Ctrl+E instead of the End key. Both only move the cursor to the end of the line, but Ctrl+E is never read as Esc, which would interrupt the turn.
A routine no longer retries a step whose delivery is uncertain, since it may already be in the CLI and would be typed twice.
Deliveries that stay unconfirmed are matched against the transcript at growing intervals instead of every 1.5 seconds. On Windows they had piled up and kept re-reading large transcripts, slowing the server down.
The Canvas queued-message notice now says Codex and Claude take messages mid-turn, as soon as they can.
When replaying owner text held during an automatic paste on Windows or WSL, close the Codex key-stream burst before restoring its Enter. A delay alone could leave the owner's line visible but unsent when Codex read the keys late.
02
Orkestrai 0.53.0
Mid-turn messages and Windows Codex paste submission.
Busy Codex and Claude agents receive inbox messages during their turn, as when you type into a working CLI: the agent answers what it can and continues its task instead of waiting for a 20 to 60 minute turn to end. Exact receipts, separation of your draft and one automatic message at a time still apply. Messages wait while a question or approval is open. Other CLIs and notices that only restart a stopped agent still wait for the turn boundary. If a Codex question opens after the paste, Enter is not sent and the text stays as a visible draft for you, never retyped.
Terminal mouse reports and late or duplicate cursor replies no longer create a false typing warning or block Enter and pending-paste recovery. Real keys and text remain separate from automatic messages.
A manual Enter that is lost does not disable recovery of the unchanged automatic Codex message. Recovery preserves its exact receipt and never retypes the body.
A terminal whose session disappeared can recover from input or resize errors as well as initial attachment, including a newly created terminal. Stale session errors and old buffered keys never target its replacement.
After a terminal disappears, check its saved exact receipt before requeuing the original message. Interrupted leader-to-specialist and specialist-to-leader handoffs retain their saved content.
On native Windows and WSL, automatic messages to Codex are submitted even when a loaded machine makes Codex read keys late. Previously Enter could become a newline inside the paste, leaving [Pasted Content] unsent and the agent's queue stuck. An End key, which only moves the cursor, now precedes each automatic Enter.
03
Orkestrai 0.52.8
Windows terminal input and automatic message recovery.
Terminal mouse reports and late or duplicate cursor replies no longer create a false typing warning or block Enter and pending-paste recovery. Real keys and text remain separate from automatic messages.
A manual Enter that is lost does not disable recovery of the unchanged automatic Codex message. Recovery preserves its exact receipt and never retypes the body.
A terminal whose session disappeared can recover from input or resize errors as well as initial attachment, including a newly created terminal. Stale session errors and old buffered keys never target its replacement.
After a terminal disappears, check its saved exact receipt before requeuing the original message. Interrupted leader-to-specialist and specialist-to-leader handoffs retain their saved content.
On native Windows and WSL, automatic messages to Codex are submitted even when a loaded machine makes Codex read keys late. Previously Enter could become a newline inside the paste, leaving [Pasted Content] unsent and the agent's queue stuck. An End key, which only moves the cursor, now precedes each automatic Enter.
04
Orkestrai 0.52.7
Windows message delivery, separate terminal typing and resilient updates.
Typing during a confirmed automatic delivery waits for the exact prompt receipt, including when Enter fails. Your draft returns before another automatic message can use the terminal.
The terminal shows when your typing is waiting. Ctrl+C discards that staged draft and takes control; Enter alone can help a pending automatic message when no owner text is staged.
Raw bridge controls refuse occupied input without reporting successful delivery. Terminal protocol replies remain immediate. A reconnect restores the waiting status; redraws, timeouts and cancellation do not insert your text into an unconfirmed paste.
Supported agent CLIs started inside Shell nodes use the same receipt protection. Explicit bridge replies return without waiting for a transcript; the Shell keeps its existing provider and account settings.
Short Codex messages painted as literal text can recover a missed Enter too, using the entire unchanged owned body. Real persisted-inbox tests cover an idle leader, external callers and Kanban assignments with 300 unrelated image nodes.
Code Intelligence is disabled by default, including legacy modes that were enabled without the new explicit opt-in. Activate Manual or Assisted in workspace settings; existing graph history is retained and other workspace operations stay available.
Update checks retry a transient feed failure once, then consult bounded, validated metadata from the official GitHub release API. If newer installers exist while the feed is unavailable, the app offers the official download page without bypassing installation verification or mislabeling Windows as an unsigned Mac.
Restored owner input keeps text and Enter separate on native Windows, so the restored line does not become another unsubmitted paste. Ctrl+C and PTY exit cancel its delayed Enter. Short Codex input also recognizes verified footer redraws below its complete unchanged body.
05
Orkestrai 0.52.6
Recover automatic messages after Windows terminal redraws.
Cursor shape and blinking controls are recognized as terminal formatting, keeping them out of the pasted draft check. Actual owner edits and working turns still prevent automatic Enter.
Codex can redraw its model, path and warning rows or clear blank cells to the right of a freshly painted paste badge. Absolute-position and line-based redraws that preserve the draft no longer block Enter-only recovery. When ConPTY splits one write into literal text and a smaller wrapped paste block, recovery also verifies the painted text against the exact prefix implied by that block’s size, including a full-screen redraw using line breaks. Owner edits, active turns, erased composers and exact prompt receipts remain protected.
Terminal restoration excludes answered or expired cursor-position queries while retaining fresh requests awaiting their first viewer. Their duplicate replies could be mistaken for owner edits and prevent recovery of an unsent automatic paste. Live cursor replies, actual owner drafts and exact delivery receipts remain protected.
Confirmed automatic messages wait for the initial native input-ready boundary on Windows/WSL before writing the paste, using the existing quiet fallback for older CLIs. A loading-screen prompt glyph alone does not start delivery; native macOS input and owner drafts keep their behavior.
06
Orkestrai 0.52.5
Windows agent delivery and stable Canvas navigation.
The native Windows terminal can clear its footer after showing a pasted message. This no longer prevents Enter-only recovery of an unchanged automatic paste. Owner drafts, working turns and exact delivery receipts remain protected.
Panning keeps node contents and terminal screens visible. Terminal parsing pauses during navigation without detaching the stream or replaying the entire history on every movement; pending output remains bounded.
Task dispatch and inbox routing no longer load unrelated Canvas image payloads to resolve names. Optional learning recall has a 400 ms delivery budget and bounded shared reads, so a stalled lookup cannot keep an assignment out of the agent inbox.
Codex paste recovery remains available after heavy terminal redraws trim the original write mark from retained history; a fresh matching badge and unchanged input are still required.
07
Orkestrai 0.52.4
Restore Codex paste recovery and external agent replies.
Requested cursor-position reports from the terminal no longer disable recovery of an unsent Codex paste. Real owner edits and working turns remain protected.
External agents receive saved replies through a private receipt, including after MCP or Core reconnects. Retry an uncertain ask only with its original receiptId; read or wait for the recorded result without resending the message.
08
Orkestrai 0.52.3
Windows delivery recovery, bounded image previews and safer updates.
Unconfirmed bridge sends remain queued. Codex messages wait for working turns to finish, and receipts from the exact saved conversation remain valid when Windows delays updating the file timestamp. Native done/error states accept input; explicit blocked states stay protected.
Codex idle reports and composer echoes guide Windows/WSL submission. New conversations can recover a stuck first paste, including an Enter absorbed as a newline, while working turns and owner drafts remain protected.
Messages without an exact receipt remain pending after ten minutes, and tasks only become running after confirmed acceptance. Legacy running states with no receipt are repaired without retyping prompts or releasing task custody.
Canvas image previews load near the viewport with two decode jobs, a bounded pixel cache and released offscreen canvases. Creative dialogs open on demand; original files and agent references stay intact. Animated originals play on selected nodes within the pixel budget. Image connection handles stay mounted across viewport changes. Focus keeps full image controls and previews independently of Canvas visibility and zoom. Accessible image names remain available. Full editing controls appear on hover or selection; readable unselected images keep lightweight titles instead of rebuilding toolbars on zoom. Overview nodes reuse composited surfaces during pan and zoom, releasing that hint when the Canvas returns to readable zoom. Cancelling an asset preview keeps Core running: cancellation is attached after stream error handling, and active consumers still receive genuine read failures.
Automatic Codex messages use explicit paste boundaries on Windows/WSL. An unchanged pending paste can recover with Enter after a proven idle turn, without retyping text or submitting owner drafts. Slow transcript scans are shared instead of overlapping.
Closing the Windows app no longer starts a silent update. Verified downloads wait for Restart and update; updater stages and errors are saved in the bounded, redacted diagnostics log.
Release validation installs the published Windows version, upgrades it with the new NSIS installer, checks executable hashes and shortcuts, and starts the installed Core twice. Every push and pull request also tests delivery with the real Codex TUI on native Windows, using local model responses without credentials. Production publication requires this check; the optional 65-minute run covers the reported one-hour boundary.
09
Orkestrai 0.52.2
Clean replies from agents launched manually inside a Shell.
When Claude or another recognized agent is started manually in a Shell, match its reply to the exact question in its own saved conversation and current directory, or use its authenticated bridge reply. An explicit reply returns to the waiting caller without a duplicate in its inbox. Terminal redraws, prompts and silence never count as a confirmed model answer. Normal Shell commands keep their output capture. Also recognize agents on Linux and WSL when the OS reports a runtime thread name.
A workspace Shell keeps its own temporary bridge identity and original workspace after changing directories. Process inspection is on demand; queued prompts are cancelled if the agent exits or changes, and no native provider or resume configuration is written to the Shell.
10
Orkestrai 0.52.1
Reliable Codex inbox delivery, preserved terminal input and navigation across grouped images.
Hold Space and drag over a group, image or node to move only the Canvas camera, including at high zoom. Release Space to move and edit nodes normally. Typing in inputs, editors or terminals keeps its usual behavior; Alt+Space dictation stays available. Navigation temporarily overrides a creation tool without canceling it. Losing focus releases the shortcut.
Keep Codex inbox delivery and activity status ready after a turn ends, even when background commands from that exact turn finish later. A genuinely newer or unidentifiable turn still blocks delivery; completion never invents a reply or delivery receipt.
Keep the current terminal cursor when another selected terminal finishes starting or recovering. Delayed startup does not steal input from another terminal, composer, dialog or application.
Keep eligible terminal output moving when Chromium delays animation frames: a 50 ms deadline preserves the shared two-write budget. Hidden windows suspend parsing, and disposal cancels deadlines and visibility listeners.
11
Orkestrai 0.52.0
Canvas file controls, repeated dictation, bulk attention reads, Windows Shell restoration and bridge repairs.
Reveal the file used by an image, document, media or editor node in the host file manager, selected where supported. Leaders and specialists have the same action through fs_reveal and orkestrai fs reveal, with existing workspace and repository boundaries.
Show or hide the quick message bar under Canvas terminals globally in Settings → Terminal, without stopping agents or disabling Conversation or terminal input.
Keep one persisted bridge token during concurrent startup and repair stale, incomplete or mismatched workspace configuration before agent launches. Preserve existing valid tokens and reject redirected configuration paths.
Preserve cached interface preferences when Settings reads fail, and report unsuccessful saves without claiming success.
Validate the packaged JavaScript over HTTP on each native target, including lazy pages and workers, instead of relying on Core health alone.
Mark all Attention Center items as read at once without resolving pending questions or approvals. Leaders and specialists can acknowledge their own workspace through the CLI/MCP.
Keep the microphone available beside Send even after dictation or typing, so another recording appends to the same draft in Conversation, the Canvas bar, leader panel and agent conversation.
Remember and restore Shell directories in Windows PowerShell, Command Prompt and WSL in Canvas and Focus without editing shell profiles. Configure terminal rendering from the host ConPTY backend instead of Unix assumptions.
12
Orkestrai 0.51.3
Conversation synchronized with agents and more responsive large projects.
Show public terminal progress and final answers automatically in the Conversation after owner-message delivery, without requiring a bridge post or another model turn. Recover the exact saved conversation after reopening, deduplicate updates across concurrent readers/restarts, and exclude reasoning, tool output and other profiles.
Canvas Git indicators and lists no longer hash every changed or untracked asset during polling. Concurrent display reads are shared; reviews and Git mutation guards retain exact content checks with bounded subprocess concurrency.
Workspace file operations are asynchronous even with Code Intelligence disabled. Changes to the same file stay serialized without blocking unrelated files. HTTP previews and downloads stream with bounded buffering; internal buffered image operations accept files up to 64 MiB.
A search timeout never falls back to a blocking project crawl. Missing-runtime fallback searches have time and entry budgets; Unity root caches and nested Floors are excluded without deleting assets or authored definitions.
Slow Core requests and HTTP 500s now record their operation class, including UI module downloads. Diagnostics are rate-limited and omit URLs, paths, bodies, headers and credentials.
13
Orkestrai 0.51.2
Bounded graph updates, safe Floor snapshots and responsive workspace folders.
Look up incremental relationships by the affected files instead of scanning a whole graph revision. Preserve independent workspaces, fileless relationships and caller rebinding across batches.
Keep Unity-generated Library, Temp, obj, Logs and UserSettings roots out of code scanning and file watchers. Preserve authored Assets/Packages and ordinary source folders in non-Unity repositories.
Prevent nested Floor/internal-state copies even when previously tracked. Repair Git exclusions for linked checkouts, reserve snapshot/checkout space and preserve source, artwork and the owner index. Unsafe existing branches stay intact and are rejected before checkout. Count modified files relative to nested workspace roots before staging snapshot data.
Open and change workspace folders without synchronous project file I/O. Limit instruction reads to regular UTF-8 text files up to 1 MiB and preserve unsupported existing files; a slow project read no longer blocks another workspace.
Preserve versioned roles, skills, native designs and other authored .orkestrai content. Omit only the reserved runtime paths from Floor snapshots; migrate identifiable generated ignore blocks while preserving independent owner rules.
14
Orkestrai 0.51.1
Responsive conversations and workspace startup, with accurate activity.
Open Conversation without waiting for Canvas preparation, saved-role files, routines or external channel bindings. Failed reads offer retry and preserve visible history.
Keep the message feed alive, reconnect silent sockets and recover stored messages after reconnect, foreground return or a missed notification, independently of workspace metadata.
Prepare directories, PATH/registry, Codex trust and saved roles asynchronously. Preserve account isolation and logical command identity, share concurrent role reads and keep live launch settings unchanged during polling.
Find queued owner messages before the colleague-message page limit and prioritize a direct owner reply before further coordination, without automatic reminder turns.
Show an agent as working while its exact provider turn is still running, even when the terminal is silent. Reconcile silent completions in the background with bounded asynchronous reads; preserve explicit blockers, replacement sessions and independent composer readiness.
Recognize Codex reasoning and tool work when the turn start has left the bounded transcript tail. Settings and usage updates do not count as new work.
15
Orkestrai 0.51.0
Durable task execution, dependency-aware dispatch and terminal recovery without blocking independent work.
Persist each task execution attempt, admit one implementation at a time per terminal, expose queued/running/review states and start dependent work only after its prerequisites finish.
Notify the coordinator once when a proven worker turn stops without a result; preserve deliberate waits, restart recovery, independent work and the leader-only coordination role.
Retry terminal viewer connections continuously with bounded backoff, retain failure counts across brief reconnects and tolerate renderer suspension without killing agents.
Read Claude/Codex session discovery and rotation metadata asynchronously; remove synchronous Windows process-tree termination and WSL path conversion from the Core event loop.
Provide a compact shared agent guide, with version-matched reference sections available through the CLI and MCP. Keep all capabilities and the focused delivery guidelines available to every agent.
Recover disconnected agents with unfinished authorized work through a bounded startup queue, while preserving conversations, explicit manual sleep, workspace pauses and provider limits.
Return paginated summaries of open tasks to agents, with filters and explicit task_read details, preserving the complete UI board and history.
Return unfinished cards to the backlog when their agent disappears, repair legacy dangling assignments without deleting work, and avoid treating identical status reports as new supervisory progress.
Count all message metric pages without loading message bodies. Scope token totals and daily charts to linked workspace sessions and explicitly label incomplete local coverage.
Apply focused coding guidelines to leaders and specialists: state material assumptions, keep changes minimal, preserve unrelated work and verify usable outcomes.
Clarify that informational replies do not require another acknowledgement; preserve concrete questions, explicit requests and existing routing.
Keep task actions below the title without covering editing or joint review; allow clearing prerequisites even if a referenced task no longer exists.
Keep leader coordination independent of the specialist implementation slot, while preserving task attempts and prerequisite checks.
Deliver pending inbox messages with failed bridge tool results too, so a Git conflict or validation error does not delay owner messages until the next successful call. Preserve the original operation error without retrying the mutation; a failed inbox read leaves messages queued.
16
Orkestrai 0.50.1
Reliable leadership conversations and faster dispatch and recovery for large teams.
Keep send receipts, history failures, pending questions and draft clearing scoped to their original conversation and workspace, including leaving and returning to the same agent. Refresh agents and the leader profile after leadership changes, recruitment and reconnects while preserving the direct conversation you explicitly selected. Queued metadata refreshes follow workspace switches. Returning to the original channel shows a persisted send even if its live update was missed, without clearing the newer draft.
Let other agents receive messages while one native or WSL transcript is slow to read or confirm. Keep six terminal writers, release each after Enter, and preserve exact receipt checks, ordering and protection against duplicate prompts.
Cache unchanged prompts, replies and turn boundaries for teams of 30 agents with limits on both entries and memory. Changed files invalidate the result; temporary read failures are retried instead of keeping the inbox blocked.
Restart persistent agents through at most four concurrent starts, prioritizing queued leaders and keeping supervision responsive. Avoid duplicate starts and respect later workspace pauses, mode changes and supervisor shutdown.
When dispatch stalls, desktop logs include bounded numeric queue counts, event-loop delay and process memory, at most once per minute. Prompts, conversation contents, node identities and credentials are excluded.
Read the scheduled queue health through the supported aggregate monitor API, fixing the health task that failed every five minutes. This framework task is separate from agent supervision.
17
Orkestrai 0.50.0
Keep code indexing independent of conversations, expose measured graph usage and unblock Adventure progression and storage.
Transfer workspace leadership between providers with a durable context handoff, retaining each agent’s saved conversation, profile and model. Choose Leader, optional Deputy or Specialist when creating an agent or configure existing agents from their menu, CLI or MCP. Creating the first leader uses its existing launch contract without an empty handoff or another model turn.
Require independent primary and deputy approvals on the same delivery revision when a deputy is configured. Findings keep the card open, corrections reset both approvals, and confirmed completion recovers after restart without blocking unrelated cards.
Run graph scanning, parsing, persistence, semantic indexing and historical comparisons off the Core event loop, with an isolated database and committed read snapshots. Owner messages do not join graph transactions. Bundle and verify the target-specific ripgrep executable for code indexing, including ARM64 packages built on Intel hosts. Packaged checks index and query real TypeScript before publication.
Skip unchanged stale indexes, retain physical rows across incremental revisions, rank scoped FTS candidates before joins, process semantic symbols in batches and deduplicate new history chunks. Bounded parser and semantic caches avoid repeated work without changing vector scores. Automatic resource failures back off instead of retrying continuously. Source excerpts are included only when a bounded file read matches the indexed hash.
Show workspace-scoped query counts, latency buckets and estimated returned context in Usage and performance, CLI and MCP. Migrate and back up graph storage automatically without importing provider secrets.
Restart Torment from 1-1 after its final boss with auto-advance enabled, preserving XP and clears; Paragon continues to begin at hero level 100. Announce the new circuit and explain endgame progression in the map.
Grant one soulstone directly for each credited act-boss kill, even with full bags or chest storage. Repeat reports never grant it twice; harder stones still open earlier boss chests. Existing saved boss chests receive their missing soulstones once on upgrade, accounting for stones you already own and first-free chests. Only bag space is needed to open that backlog. Reserve enough bag slots for the chest's maximum loot, including a fifth item from a Torment boss; smaller chests can still use the remaining space.
Add Store all and Withdraw all in the Camp, CLI and MCP. Transfer unworn items atomically up to destination capacity, preserve locks and overflow, and report moved and remaining counts.
Contain socket and WebSocket errors to the failed collaboration/model-access relay connection, clean up its room membership and keep unrelated rooms available.
Keep unsent Canvas prompt text per workspace in the current tab when changing views or reloading. Returning from a focused Design through browser Back closes the editor first, preserving the team and board. Send acknowledgements cannot clear a newer draft or another workspace.
Preserve recorded inbox answers and pending reply forwards when delayed dispatcher receipts arrive; state transitions are atomic and a receipt cannot rewind a replied message.
Update Joi, follow-redirects, Minimatch and modern Glob to patched releases. Preserve the legacy matcher API used by packaging with an exact-hash compatibility patch and real glob/archive tests, keeping the full moderate-severity audit enabled.
18
Orkestrai 0.49.1
Reliable received access, live controls and guest message delivery.
Queued messages can advance after an Orkestrai guest agent explicitly finishes a tool-only turn without answer text. Unfinished turns and unsent terminal drafts remain protected. Guest inference remains marked as working during silent model/tool work, and pending inbox messages join bridge tool results during that turn and remain in managed history.
Approval and revocation update in the open invitation dialog and both access lists without reopening the page. Model and image request counts update live; accepting a device alone does not consume a model request. Long invitations wrap inside the dialog. Model restrictions use provider catalog dropdowns and effort controls list compatible levels. The Orkestrai CLI is shown as included, independently of received connections.
Received-model agents on Windows start through the bundled console runtime and CLI entry instead of relying on CMD or a global orkestrai command. WSL startup checks include the workspace launcher in PATH and distinguish missing commands from unavailable distributions and directories.
The model access panel now shares the same centered, responsive page column as the overview and provider cards, with matching card spacing and corners.
19
Orkestrai 0.49.0
Time-limited Claude and Codex access with guest-side tools and files.
Lend selected Claude/Codex account profiles to a device you approve, with expiry, per-account model/effort restrictions and request/concurrency limits. This is independent of workspace collaboration.
Stream model responses while every workspace tool, file and command stays on the guest computer. Native bridge/CLI/MCP tools use the same registry, and the managed guest conversation can resume locally.
Approve Codex image generation separately: guest references and generated files stay on the guest, and image requests consume the same access limits. Invitations and private device keys use the desktop vault; provider credentials never travel to guests.
Build and verify the inference relay independently of desktop lifecycle scripts, with a restricted container context and both relay protocols checked in CI.
20
Orkestrai 0.48.0
Switch Claude accounts without replacing the saved conversation.
Switch a Claude agent between your own account profiles while retaining its exact saved conversation, tool results, role and tasks. Only that conversation is copied; credentials stay in each account. Native and WSL transcript reads follow the selected profile. Conflicting destination histories and unsafe paths are rejected before the old terminal is stopped.
21
Orkestrai 0.47.0: reliable inbox, shared lessons and chest storage
Queued messages resume, approved lessons become searchable, and boss chests remain usable.
Queued messages can reach idle Claude agents after a scheduled turn even when the original prompt has left the bounded transcript tail. New provider work, human drafts and uncertain prompt receipts remain protected.
Active approved lessons appear in the knowledge graph and the agent’s Remember profile. Owner messages and task briefs receive bounded relevant memories with sources, without an extra agent turn. Slow memory reads do not block messages; review/off modes and workspace scope remain authoritative, and private provider files are not imported. The profile updates live and shows loading failures separately from an empty memory.
A soulstone from a harder difficulty also opens act boss chests from easier ones: a chest spends a soulstone of its own difficulty first and, without one, a soulstone of the nearest harder difficulty.
Red chests are stored apart, up to the same limit as the rest, so those waiting for a soulstone no longer take the place of white, blue and quest chests. Each boss's first chest, which opens free, is always kept.
When chest storage is full the strip warns that new chests are being lost, and opening red chests in the Camp skips those waiting for a soulstone instead of stopping at the first one.
Update qs to 6.16.2 to prevent comma-formatted query serialization from crashing on untrusted objects, preserving the current Postman runtimes and the dependency audit gate.
Fix creating and editing sourced memories from the interface: workspace route identifiers no longer invalidate the strictly validated memory body.
Recover a native Claude automatic paste whose first Enter was missed during startup: confirm only the existing pasted draft, with a bounded retry and no intervening human input or active provider turn. Never retype the message or change its delivery receipt. Terminal focus, device and color reports do not count as human draft edits.
Keep bold text, links, table headings and linked-node counts readable on every note paper in light and dark themes. Glass notes follow the selected theme, and linked counts appear immediately when the canvas opens.
22
Orkestrai 0.46.2: Adventure heads into Nightmare
Beating the last boss of a difficulty takes the party to the next one.
With Advance after a clear on, a party that beats the last boss of a difficulty heads to Act 1 of the next one instead of fighting the same boss forever. The strip announces the opened difficulty, also with the option off, and Torment's last boss remains the end of the road.
23
Orkestrai 0.46.1: Adventure desktop input
Restore hero and HUD clicks without depending on forwarded mouse movement.
Windows and Linux X11/XWayland recover Adventure hero and HUD interaction when mouse forwarding is missing. Cursor sampling stays bounded, stops with the strip, and leaves macOS on its existing path. Native Wayland does not expose the required global cursor API; use the Camp in the main app or an X11/XWayland backend.
24
Orkestrai 0.46.0: Adventure above the Dock
Your team becomes a pixel-art party that levels up, finds gear and fights bugs while it works.
Adventure production builds use browser Web Crypto for battle report identifiers, avoiding a Node-only cryptography import in the renderer.
The splash shows database preparation from the start, including when preparation begins before the splash finishes loading.
Listen in the sound settings plays the channels as they are set on screen from the first click, and the soundtrack no longer bunches up missed beats after the computer was busy or asleep.
Battle fought before the first report ever reached the server is now paid in full when it arrives.
An agent's Adventure actions now come from its own terminal: it cannot act as the leader or as another teammate by naming them.
Adventure art loads faster and without the colored blocks shown while it arrived: sheets are cached for good after the first load and heroes, monsters and icons fade in.
Startup no longer stops the app's server while a large database is backed up and updated: Orkestrai waits for it, says on the splash that your data is being prepared, and opens on its own when it is done.
The camp follows links to another workspace, hero or tab without keeping the previous one's save, and a Back button returns to where you were (a conversation, the canvas…). The hero card on the strip always fits inside the strip, the armory no longer covers the stats in narrow windows, and rarity colors are readable on light themes.
The cube crafts off hands for the class you pick, like main hands.
Battle results reach the save exactly once: a report whose answer was lost is sent again unchanged and never paid twice, and battle fought while the app's server was away is sent in reports it accepts and paid in full when it returns.
New Adventure: a transparent strip above the Dock or taskbar shows the workspace's agents as a pixel-art party. The leader is the Maestro and each teammate gets one of five classes from its role (Artificer, Sorceress, Knight, Ranger, Cleric), with a choice per hero. Heroes belong to the workspace's guild and have their own names: when the leader dismisses an agent its hero rests in the guild, and the next agent of the same class takes it over with its level and gear; the Maestro always follows the current leader. While any agent works, the whole formation fights (three slots, up to eight with runes): pinned heroes first, then heroes whose agents are working, the Maestro, the other agents and the resting guild. Heroes whose agents work fight inspired (+20% damage), agents working outside the formation earn 60% of the experience and heroes far behind the strongest learn half again as fast. With nobody working the party camps.
The strip floats on the desktop even with Orkestrai minimized. Drag it by the grip on its panel — at full width it moves up and down and to another display, compact it goes anywhere — and it stays where you leave it until Back to the Dock. Settings → Adventure picks its width, size, scenery (full or minimal) and a panel that is always visible or only on hover.
Every hero levels from 1 to 100 on its own, then keeps going through endless Paragon levels (each adds 1% attack and 1% life), earning the most experience from monsters near its level, and spends one point per level in an eight-tier tree per class (passives, formation-wide auras and active skills; the last tier needs 70 points; free respec and optional automatic spending).
Every active skill has its own animated effect where it lands: meteors, frost novas, arcane storms, arrow rain, earth spikes, judgment beams and fireworks strike the monsters; piercing shots and volleys fly to their targets; rockets arc and explode on impact; healing light, sanctuaries, gears and anthems play on each ally, with green numbers for the life restored; a lion-crest dome covers the party under Bulwark; whirlwinds spin around the knight. Level-ups and revivals rise in gold, and arrows, bolts and notes now visibly fly to their target.
Every hero has a full armory: main hand and a class off hand (shield, quiver, focus crystal, relic, gadget or songbook), head, shoulders, chest, hands, legs and feet, a necklace, two rings and two trinkets. Gear drops in nine rarities from Common to Cosmic, with item levels, random affixes and sockets for gems, runestones and glyphs; every base tier comes in three variants with their own look and focus. A hero's outfit on screen follows its gear — novice, adept, heroic and celestial — and its weapon glows in its rarity's color.
Eighteen unique items, each with a power the battle plays (cleave, ricocheting arrows, echoing skills, thorns, executions, a falling star every eighth attack and more): three from each act boss, and three wandering ones on Hell and Torment.
Six class sets, each a full suit of six armor pieces: two and four pieces add bonuses, all six unlock a party power. Six elixirs, drunk at the camp, help every hero for 30 minutes. Gems come in five cuts, from Chipped to Imperial, and the cube fuses three into the next.
The camp is now a fantasy RPG screen in both themes: an illustrated banner, picture tabs, a character sheet with the hero on a pedestal between its armor and weapons (and how close it is to the next outfit), a skill tree of lit tiers with framed skill icons and a combat bar, treasure chests, a potion shelf, the cube on its altar, rune grimoires, pet dens, a painted map of the five acts and a chronicle of the journey.
New recruits join in a class the guild lacks once it has two of theirs (the healer first), and Balance the guild, in the camp, does the same for existing heroes after showing who changes: skill points come back in full and the Maestro never changes.
Agents can look after their heroes when you ask: orkestrai adventure (MCP adventure_read/adventure_act) shows the guild and the agent's own hero and runs camp actions — each specialist on its own hero, the leader on the whole guild.
Settings → Adventure has a sound mixer: switch each channel (soundtrack, attacks, spells and skills, loot and triumphs, alerts) on or off, set its volume and listen to it. Each weapon and each of the 19 active skills sounds like itself, and the soundtrack follows the party from the camp fire to the road, elite waves and bosses.
Chests open themselves as they arrive, each hero puts on clear upgrades by itself and loot nobody wears turns into gold up to a rarity you choose; locked, unique and set pieces are never touched. Switch each one off in the camp: chest opening and salvage in the Inventory, automatic gear on each hero. The Unsealing and Locksmith runes now add an item to every white and blue chest.
Chests: white ones from any monster, blue ones from the elite closing each stage, red ones from act bosses (repeats take a soulstone of their difficulty) and a golden quest chest for every finished Kanban card, paid to the hero whose agent finished it, even with the strip closed.
Five acts of ten stages on four difficulties (Normal, Nightmare, Hell, Torment), with region levels up to 110. Heroes can fall and get back up; losing a stage falls back one stage or retries it. A portal returns to any cleared stage to farm it.
The camp, in the main window, inspects and manages everything: heroes with equipment, stats and skills; inventory, stash and chests; the Orchestrion Cube (synthesis of nine items into a rarer one, alchemy, crafting, sockets, gem fusion, removal and offerings), which levels up with use; 23 gold-bought runes (formation slots, a second active skill, richer chests, more storage and more); eight pets tamed by defeating bugs; the map with travel and settings; and the loot log.
Work done while the strip is closed still counts: each hero earns experience and gold for its agent's recorded working time, at offline efficiency and without chests. All art is original, made with Orkestrai's image workflows.
25
Orkestrai 0.45.0: reliable Portals, responsive terminals and team accounts
Repeated Portal captures, bounded terminal rendering, leader account inheritance and Pi/Oh My Pi support.
New agents inherit the account profile of their same-CLI leader in Canvas, team creation, presets and CLI/MCP recruitment. Explicit choices remain authoritative, and a missing profile never silently falls back to the default account.
Release Portal window-close subscriptions on detach, reparenting and close, sharing one listener per window so repeated Chat/Canvas navigation does not retain old Portal state.
Label graph and companion-profile comboboxes explicitly and keep the Computer conversation-authorization form at a readable width. Browser checks follow current select semantics, restore settings after timeouts and stop CI early on repeated failures.
Copy full Portal screenshots through a native capture lease with explicit logical clipping, priming the attached widget before resizing it. Avoid stalled CDP captures on Windows/Linux while preserving private-field masks, dimensions, scroll and the same page/session.
Initialize the hidden Portal capture host before attaching pages and reassert native painting after reparenting. Native checks require an explicit successful completion and remove Chromium profiles after process exit, preventing false approvals on Windows.
Disable system-font discovery in the isolated PDF parser, which uses packaged or embedded fonts, avoiding a headless Electron initialization stall before Linux OCR starts.
Wait for the document parser to confirm IPC readiness before sending PDF or spreadsheet bytes, preserving extraction deadlines, cancellation and offline OCR verification.
Integrate reviewed Drizzle ORM 0.45.4 and Superforms 2.31.0 updates, preserving SQLite transactions and existing Zod v3 form contracts.
Include native canvas 1.0.13 and electron-updater 6.8.11, checking offline OCR, confined update-cache paths, malformed metadata and SHA-512 validation while retaining macOS trust and manual-download policies.
Name Portal control and authorized-agent selectors explicitly for assistive technology under the updated combobox semantics. Browser tests exercise those roles and restore their locale after failures.
Use the app's patched YAML 2.x release in Bruno's filestore and scripting packages too, avoiding their vulnerable pinned copies without changing the parser major.
Update reviewed Canvas, editor, PDF, Bruno and Computer dependencies. Explicitly keep Computer embedded in Orkestrai despite the SDK's new daemon default. Retain Svelar 0.7.9 and its verified validation compatibility patches until the schema migration is complete.
Keep Bruno's Azure Identity at 4.13.0 with the unaffected MSAL 3.x branch until newer Azure Identity versions support patched MSAL; do not force a different authentication-client major.
Update development dependencies, including Electron 42.11.18, electron-builder 26.18.0 and Playwright 1.64.1, keeping the native ABI and verified patches. Undici 7.30.1/6.29.1 and fast-uri 3.1.9 address memory-exhaustion, response-crash, integrity and URI-normalization advisories; the complete dependency audit remains enabled.
Add Pi and Oh My Pi terminals, headless JSON output, model/thinking controls and native role prompts. Both use the workspace bridge. Automatic session restoration and transcript mirroring are not advertised until their exact conversation readers are implemented.
Busy terminals return without parsing all retained histories together. Viewer queues and WebSocket buffers are bounded; parser acknowledgements share the rendering budget. Hidden or covered viewers stop output streaming while their agents keep running, and restore a current snapshot when shown again.
Closed Canvas graphs and expired workspace snapshots are released; the workspace cache keeps at most four snapshots. Regression tests use real garbage collection and repeated Chat/Canvas navigation.
Intentional terminal shutdown is a disconnect, so closing the app no longer creates misleading code-129 agent failure notices. Unexpected failures still report errors.
Repeated Portal screenshots keep working outside the Canvas viewport, after detaching the page and during Canvas movement. Captures use a hidden compositor even when saved geometry exists, serialize the shared surface across Portals, preserve the same page and session, and keep private fields masked without opening or focusing a window.
26
Orkestrai 0.44.1: local voice and agent messages
Restore voice on Apple Silicon and recover message delivery after a missing receipt.
Security: Zod 4.6.7 and a verified compatibility fix reject oversized constrained collections before parsing their elements. Existing app schemas, Svelar 0.7.9 and JSON Schema contracts are preserved; every installer tests the actual ESM and CommonJS entry points, and the full dependency audit stays enabled.
Fix Could not find sherpa-onnx-node on Apple Silicon: each installer includes the voice library for its target system and CPU, verified against the lockfile SHA-512 and the actual binary architecture before signing. Downloaded voice models are preserved.
Publication requires native voice to execute under real Node on macOS Intel, Apple Silicon, Windows and Linux. The ARM64 check runs against the signed installer on an Apple Silicon runner, without downloading models or rebuilding the app.
Queued messages resume when the exact conversation proves that a newer turn started and finished, even if the previous prompt's receipt is missing. The old prompt is never retyped or falsely confirmed, and a human draft or startup dialog stays protected. Recovery covers native and WSL terminals and reaches the waiting sender even if a background monitor sees completion first.
Agents can ask about assigned To do cards without moving them to Doing first. A closed, archived, unrelated or reassigned card still cancels a stale message.
27
Orkestrai 0.44.0: Chat, Needs you and attachments
Attachments in every message field, agent states in sync and Needs you holding only what is yours.
Security: DOMPurify 3.4.17, Electron 42.11.9, Vitest 4.1.12 and ip-address 10.7.4 address sanitizer, frame-permission, development mock-registration and IPv6 parsing advisories, without changing Electron's major or Monaco's version; the full dependency audit stays enabled.
Needs you only shows what you can resolve: a message left queued for an agent the leader already dismissed no longer becomes a stuck notice. Existing ones close the next time the app starts, and a notice with nothing to decide now has Dismiss. A question without preset answers shows Answer, which opens its conversation, and never disappears unanswered.
A question an agent asked you stays pinned at the end of its conversation, under Waiting for you, until you answer it. The number next to the agent in Chat always points to something on screen, even when the question is hours old.
Each agent's state (working, idle, waiting) is the same on the Canvas, in Focus and in Chat, live: Chat and Focus no longer need a trip to the Canvas to update. An agent that receives a queued message shows as working right away.
Attach images, PDFs and other files to messages in Chat, the Canvas message bar, the leader panel, Focus and an agent's chat face: paste them, drop them on the field or use the paperclip. Attachments show as chips before you send, sent images appear in the conversation and each agent receives every file's path in its own runtime to open it, on floors and in WSL too. Files added while a message is sending stay for the next one, and a draft's unsent files are discarded when you switch conversation or workspace; a stored message keeps its files even when delivery to the agent fails. A node's chat face also gets the microphone.
The terminal's quick field sends images: pasting or dropping an image attaches the file to the message instead of falling into the terminal or becoming a path in the text. An image pasted straight into an agent's terminal (Ctrl+V) counts as a draft when the clipboard really holds one, so a queued message waits instead of being typed on top of it.
The focus ring of the Chat message field follows the field's rounded edge; a second, square and misaligned outline used to appear inside it.
On Windows, tooltips near the top of the window (like the Focus header buttons') open downward instead of hiding behind the minimize, maximize and close buttons, and Design mode and the prototype player start below those buttons. The native buttons and Snap Layouts stay.
The usage pill in the top bar shows the account the open workspace's agents run on and names the profile when the provider has more than one account. It used to show the highest usage of every account on the computer, such as the work profile in a workspace that uses the default account. Automatic work also stops only when the agent's own account reaches its limit.
Agents can screenshot the Portal: the capture becomes a PNG in .orkestrai/portal/screenshots/ and the agent gets its path to open the image, even when the Portal is off screen. The image used to come back as huge text the agent could not view, so it ended up opening another hidden browser without the Portal's session.
28
Orkestrai 0.43.0: Canvas, Focus and Chat
The Canvas reads at a glance: who is working, what needs you and where everything is.
Security: update follow-redirects to 1.16.1 and Vite to 6.4.5 to fix cross-origin credential-header leaks and development editor command injection advisories, with the full dependency audit still enabled.
One top bar on every platform holds the Canvas/Workbench switch, search, the attention bell and the microphone. On macOS the window buttons sit inside it.
Terminal headers show each agent's live state, a stable teammate color with initials, the time in the current state and the launch model.
The Canvas dock keeps everyday tools one click away and moves the rest into Plan, Code, Devices and Workspace menus. Creation tools can still be dragged onto the canvas.
The dock, minimap and zoom controls float on a translucent surface.
Connections stay quiet until one of their nodes is selected.
Zoomed far out, nodes show their name and status in large type instead of an empty frame.
A message bar at the bottom of the Canvas talks to the selected agent, the last one you chose or the leader. Type @ to point at a teammate, note or portal; an agent that is asleep wakes up first.
A summary card in the corner tells what happened since your last visit: what needs you, what was finished and who is working.
Agents that need you show it inside their own node, with Reply, Resolve and Snooze. When the terminal itself is waiting for permission or an answer, the node says so too.
Selecting an agent shows a small bar above it with Focus, Message and more actions.
Notes that an agent keeps only for itself stack behind the newest one. Click the counter to spread them out and Stack to gather them again.
The workspace sidebar floats over the Canvas and lists the team, the ones that need you first. Group titles use a condensed display face.
Chat is a third mode next to Canvas and Workbench: a team channel where agents post progress and their conversations with each other appear, plus a direct conversation with each agent. Without a mention, a team message goes to the leader; @ sends it to the agents you name.
Agents can ask you something without stopping their work. The question appears in Needs you, in the agent's node and in Chat, with answer buttons when the agent offers choices, and your answer goes back to that agent.
Approve or refuse right where the request appears: high-risk actions waiting for permission, flows waiting for approval, tools an agent proposed, lessons waiting for review and Review Center reviews. Design proposals and usage limits also show up in Needs you.
Each agent has a profile in Chat with what it is doing, its task, what it remembers and its routines.
Workbench is now called Focus, and the top bar switches between Canvas, Focus and Chat.
Focus and Chat show Needs you at the top of their sidebar, with the same answers and decisions as the bell.
Always-on teammates: an agent's profile in Chat shows whether it runs only while the app is open, wakes on demand or keeps running in the background Core, with its recent activity. Agents marked 24/7 are highlighted in Chat.
Connect an agent to a Telegram chat or a Slack channel. Messages from the people you allow reach the agent's inbox and appear in its conversation; its answers go back to the same chat under the workspace autonomy policy.
Describe a routine in your own words in the agent's profile; the agent creates the automation and confirms in Chat.
Use Orkestrai from your phone on the same network: pair it in Settings > Phone with a QR code and add it to the home screen. The phone shows Needs you with answers and decisions, Chat and the team's status, opening the workspace that needs you most.
Remote guests allowed to view approvals now see Needs you too, and operators can answer agents and decide requests from there.
Agent nodes can show their conversation instead of the terminal: the toggle in the header switches faces while the terminal keeps running.
Scores: select part of a canvas, choose Save as score and place it again in any workspace from the sidebar. It keeps the connections, the roles of the selected agents and their routines.
Pen: draw freely on the canvas from the Plan menu; each stroke becomes a shape you can move, recolor or delete.
SSH machines: open a terminal or an agent on another computer from the Devices menu, using the hosts in your SSH config.
Opening a node from the team, search or a link centers it in the part of the Canvas the floating panels leave visible. On the phone, long requests stay short until tapped and each card says whether it is an approval, a question or a blocked agent.
The Canvas got a new look: liquid-glass chrome, a tool rail on the left, a panel rail on the right, quick-action pills over the message bar and branch and zoom pills.
Open workspaces now sit as tabs in the top bar with their own Needs you count; the switcher lists every workspace and the drawer keeps folders and editing.
Nodes outline themselves in their live state color, named groups read as team frames and talking connections glow in gold.
New Leader panel: the leader's console on the Canvas, with who leads, every agent's state and one field to brief the leader.
New team map with filters, fleet counts and a side panel to jump to any agent.
New team composer: up to twelve agents with their CLIs, created at once inside a named group.
Notes became sticky paper with nine colors and handwriting, text or mono type; the pen has ink colors and widths.
Focus now opens on Needs you, with an inspector to decide or answer, and lists the team by floor.
Chat shows the team's faces, what each agent is doing, routines, connected channels, a profile column and requests as cards.
The phone page counts what needs you in a headline and lists who is working.
Pause all and Resume all freeze the agents in place and continue them where they stopped.
The message bar opens the conversation with the selected agent above it, and named groups collapse to their label.
New Checklist and Calendar nodes in the Plan menu.
Pass an agent's task to another agent from its menu or by typing “pass the task from A to B”.
Type or dictate board commands in the message bar: spawn agents, pause everyone, open the team map and more.
Loops can fire when an agent goes idle or prints a text, and stop when the agent writes their goal; seven loop templates are ready.
The Usage panel shows tokens by day and by agent from the local session logs.
Portals have a web inspector button.
New Studios: ask an agent for a diagram or a document from your project, delivered on the Canvas.
Notes export to PDF.
New agents: Gemini CLI, Qwen Code, Aider, goose and Amp, with official install commands in Provider Center.
Each Canvas reopens at the zoom and spot where you left it.
The message bar has a microphone: talk to an agent with a live waveform, and hear the reply read aloud.
Groups have an Arrange button that lines members up inside the frame, and the frame now follows its members.
Agents return to idle as soon as their terminal goes quiet, and every screen shows what they did in words.
Images agents post in Chat show the same workspace file you see on the Canvas.
Finished image workflows leave the canvas into a new Archived panel, and their images gather beside the agent.
Notes are readable on every paper, checklists tick with a click and the editor has a formatting bar.
The top bar has a More menu, Organize keeps connected nodes together and the Canvas shows that it is opening.
Going back to the Canvas through history opens it right away instead of hanging on the previous page.
Creating an agent lets you pick the CLI, starting from the one you used last.
Floor hooks open at the top of Floors with an explanation and examples.
Focus now centers on the agent: tabs on top, a header with its state, what it is doing and its task, and a side panel with its requests, recent activity and the direct conversation with voice. The tree shows only the open workspace, with an option to see them all.
The Canvas Ground floor and the Leader panel show the whole team. Each agent keeps coding in its floor's worktree and carries a chip with the floor's name; clicking the chip opens that floor. Agents that had landed on top of other nodes move to a row below the board.
Chat has a microphone in the message field. Requests waiting for you no longer shrink into stripes once the conversation scrolls, show their whole title and, for lessons, the full text; the main button keeps its contrast on hover; and a Needs you item opens the whole request right there. An image an agent shows in Chat also appears on the Canvas, beside that agent and connected to it. Recent activity in an agent's profile shows its latest work; it used to show the oldest events and hide the current task.
Lessons an agent writes just before closing its card are now kept on their own when the task finishes, like any other; lessons from agents that were removed no longer ask for your decision, and old waiting lessons of both kinds settle when the app opens. Only lessons from agents in Review mode, and those that seem to bypass rules or store secrets, still wait for you. The notice at the top of an agent on the Canvas opens the whole message with a click and gains See all, which opens the agent in Focus with its whole terminal and conversation. A message sent from the Canvas bar stays visible with its delivery state: queued while the agent finishes its turn, delivered and replied, with a shortcut to the reply; in terminal mode the agent itself shows that a message from you is queued.
The message field inside an agent on the Canvas now sends a direct message, like the Canvas bar: the question and the reply also show in Chat, and the agent shows it as queued until it receives it. A plain shell still receives the text as a command, and so does an agent whose CLI is asking something in its terminal, such as a permission or a choice. Needs you no longer keeps notices nobody can close: messages exchanged between agents while one of them is blocked no longer appear as a reply for you, and the notices of removed agents or nodes close with them, including old ones, when the app starts.
In a team with a leader, agents' questions go to the leader first, who decides on its own anything technical, infrastructure, priority, ordering or scope within the plan. Only spending, credentials, external publication and irreversible deletions reach your Needs you, and the team does not stop to wait. Chat shows every question, who decides it and who answered, and you can still answer any of them. The leader also learns what you tell a teammate directly, including your answers.
A message you send to an agent that is running a long command (tests, a build, a load run) no longer shows as delivered before it can read it: it stays queued and arrives with the command's result. Your messages come first and ask the agent to answer in Chat before going back to work. The agent also repeats in its terminal the answer it gives you in Chat, so you see it in both places.
The mouse wheel scrolls a busy agent's terminal again after the Canvas reconnects (coming back from Chat or Focus, or after an update). Agents with a lot of output lost, on reconnect, the signal that they run full screen with the mouse, so scrolling never reached them.
Floors whose work the leader brought to main can now leave the disk: when main once held every file exactly as the floor left it, even if it changed them later, cleanup frees the floor; and a floor ported with adaptations can be archived by the leader itself after checking it. Either way the branch keeps the commits. Lessons that only wait for their task to finish no longer ask for your decision, and an agent can no longer open an empty request in Needs you.
Each agent's terminal and its conversation are one conversation: what you type in the terminal (on the Canvas or in Focus) and what the agent answers there also show in Chat, Focus's side panel and the phone, and the agent shows as working while it answers. In a team with a leader, when an agent is blocked, waiting or failing, the leader solves it: it no longer shows in your Needs you or Mac notifications unless it is spending, credentials, external publication or an irreversible deletion.
The remote companion's conversation is now the app's Chat: what you write from the phone or browser goes to the agent's direct conversation and shows in Chat, Focus and the terminal. The remote shows the team channel and every direct conversation with questions, answers and delivery, loading whole messages and older history. The remote terminal opens with more history and scrolls again while agents work.
Agents also read and change Canvas checklists and calendars (orkestrai checklist/calendar and the MCP tools), and every change applies to the latest content: two screens or an agent editing at once no longer erase each other's items, and every screen shows what changed elsewhere.
Security and reliability fixes: an agent's answer goes back to Telegram or Slack only when it replies to a message from that chat; display names no longer pass for allowed users; closing or escalating a teammate's question takes the leader's terminal; an answer that cannot be recorded or queued for the agent leaves the question open to send again, and a recorded answer is already in the agent's queue, even after a restart; Telegram and Slack messages are neither lost nor handed to the agent twice, even in a large backlog or when read twice at once; what another screen saved at the same moment is no longer lost when an answer or another change fails; a blocked teammate's approvals stay in Needs you; floors count as integrated only from what main gained after them; “Load earlier” in Chat and in remote access neither repeats nor skips messages, even ones written in the same instant; the answer shown for a line typed in the terminal is never one from an earlier turn, even when the same question is asked again, also in CLIs that record no time; Chat opens on the selected agent; and a loop stops only on goal text written after its latest prompt was submitted to the terminal.
29
Orkestrai 0.42.2: Settings section selection
Keep the sidebar selection correct at the end of the page.
Clicking a short section near the bottom of Settings selects it even when the page cannot scroll farther. Direct section links scroll after settings load. Native anchors, keyboard navigation and browser history remain available; scrolling away resumes automatic section tracking, including the Windows desktop scroll container, without idle polling.
30
Orkestrai 0.42.1: Stable simulator streaming and lighter live updates
Fix the long-session MJPEG GC storm, keep Portals clear and reduce repeated background work.
The iOS Simulator screen no longer makes the Canvas slower the longer it stays open. Chromium counted every streamed frame of an image as memory until the image closed, which after hours forced several full garbage collections per second on every core. The screen is now drawn frame by frame from a cancellable stream and stops as soon as the node leaves the view.
Portals keep the live page on screen during clicks, selection and other changes that do not move them; the movement preview appears only when a drag, pan or zoom actually starts. Previews copy the frame already on screen at full resolution instead of recapturing the page, which made it blink and look blurred. Agent screenshots are unchanged.
Mobile nodes opened while the Canvas was zoomed out now show the live iOS Simulator or Android emulator screen instead of a black box. The node measured itself through the Canvas zoom and zooming does not resize it, so the screen was fitted to about one pixel; it now uses the node’s real layout size.
Kanbans share one event-driven feed per workspace. Board pickers read only terminal/note identities rather than full notes/images, unchanged snapshots do not rebuild the UI, hidden windows suspend reconciliation and a 30-second fallback recovers missed events.
Unchanged Markdown reuses sanitized output in a bounded browser-only memory cache instead of repeatedly parsing and sanitizing the same content.
Portal commands wake immediately on scoped metadata events, with non-overlapping 10-second fallback checks and bounded requests. Transient Core startup/network errors no longer become unhandled renderer rejections. Loading pages skip movement previews without error logs; full agent screenshots remain strict.
Concurrent iOS Simulator discovery shares one request and reuses results for at most five seconds, invalidating them on simulator operations. Collapsed or padding-only Mobile viewports are treated as unmeasured instead of shrinking the screen to nearly zero.
Computer background maintenance continues for healthy workspaces when an old project folder is missing. The missing workspace keeps its unhealthy status; explicit cleanup/capture and permission, quota or symlink failures remain strict. No project data is recreated or deleted.
Code-graph symbols remain selectable immediately after canvas zoom or graph resizing. Visible and picking buffers redraw atomically in the settled resize frame instead of exposing a cleared picker until another frame; no per-viewport-frame work is restored.
Known limits: isolated renderer spikes still need workload profiling. WebGL terminal rendering remains under evaluation; this patch does not enable an unverified GPU lifecycle change or delete OS swap/unrelated data.
31
Orkestrai 0.42.0: Faster canvas navigation and native Portal configuration
Reduce connection and hidden-terminal paint work; agents configure Portal hosts and local installation is verified.
PNG/JPEG/BMP images use cached, display-only 256-pixel bitmaps at overview zoom instead of repeatedly rasterizing full-size files. Originals stay mounted and return at readable zoom, selection or review. Aspect ratio, alpha, masters, files and agent references are unchanged; no per-frame drawing or disk cache. Known animated/vector formats keep their original renderer.
Connections keep solid geometry and activity colors during pan/zoom and overview, avoiding native dash subdivision even on small canvases. Readable idle zoom restores the pattern; settled offscreen connections skip paint without deleting relationships or selected/pinned controls.
CSS-hidden terminal bodies now batch TUI DOM updates instead of redrawing every output chunk. Live PTYs and ordered output remain intact; existing periodic/backlog bounds apply, focus flushes immediately, and stable visibility restores live drawing.
Highly connected nodes share floating-anchor calculations with their connections instead of rescanning the same neighbors per edge. Four immutable snapshots bound retention; node/edge updates invalidate the calculation and regression tests drag a high-degree hub.
Portals no longer poll geometry on every animation frame or stay hidden after a lost pointer-up. Canvas transforms hide the native page once and restore the same document after geometry settles, preserving input, agent access and explicit restrictions.
Dense canvases defer unfocused node bodies and secondary header controls during movement. Overview zoom below 0.2 also defers unreadable terminal paint, not the live PTY or its output. Nodes and drafts remain mounted; motion paint restores after 150 ms of quiet geometry.
Private Portal movement previews reuse a one-second cache and are bounded to 640 pixels. Document changes invalidate it and stale captures cannot replace a newer preview. Full agent screenshots and the live native page are unchanged.
portal_configure_hosts and canvas portal-hosts let assigned leaders and specialists configure exact hosts in every workspace. Current profile revisions and serialized writes preserve grants, pauses and browser state; explicit policies still apply.
The local macOS installer verifies copy, architecture, signatures, entitlements and the new Core before retaining one compact rollback outside Applications. Targeted termination uses orderly Portal/Core shutdown.
Bounded native drag/zoom checks show the renderer returning to low CPU after gestures. Sustained-session validation remains open: this release does not claim to eliminate every long-session painting/GC spike.
32
Orkestrai 0.41.1: Portal work without visibility blockers
Background control is the default, not another permission request.
Navigation and tab updates no longer replace live grants or pauses with stale renderer settings. The Canvas sends only changed browser fields; the ORM merges serialized Portal patches with the current record. Other node types retain their existing replacement behavior.
Unconfigured Portals allow leaders and specialists to work offscreen, detached or with the window minimized in every workspace. Shared validation, Electron, settings UI and CLI/MCP report the same effective policy. Explicit visibility-only settings, pauses, agent/host restrictions, protected fields and enabled Security gates remain binding; the same persistent page is used without stealing focus.
33
Orkestrai 0.41.0: Native canvas autonomy and saved terminal choices
Agents operate the app directly; a blocked card does not park independent work.
Correct existing managed Shell commands, cwd and title through shell_update; explicitly restart with shell_restart using the current configuration and live-session revisions. Edits never type into a running process or overwrite other commands. No owner intervention or Computer setup is required.
New Portals no longer automatically restrict navigation to their first hostname. Empty host lists allow HTTP(S); explicit restrictions remain exact in Core/Electron and IPv6 hosts such as [::1] survive saving. Invalid host input stops saving instead of silently weakening the list.
Leaders and specialists gain native Canvas inventory, atomic revision-checked layout updates, idempotent Shell creation with background startup, and reversible note archival through CLI/MCP without Computer permissions.
Task-specific waits and unanswered questions no longer suppress other executable cards. Fully blocked and completed boards stay quiet; explicit pauses, policies, usage limits and uncertain-delivery safeguards remain intact.
Restore Codex model and reasoning effort confirmed through terminal /model before cold resume or reload instead of reapplying old node arguments; unrelated flags and exact conversation ownership are preserved.
Add the native canvas delivery use case and guided tour in all three languages. Future workspace features must include native agent access and contract tests, not only UI controls.
Known limitation: sustained high Electron renderer CPU after hours of use remains under investigation. This release does not claim to resolve it.
Native tool availability is not functional acceptance of project journeys. Real authentication challenges, explicit owner restrictions and macOS permissions remain separate.
34
Orkestrai 0.40.0: Continuous team delivery
An asynchronous agent inbox, Floors that land on a checkout with local changes, a machine queue for heavy runs and lighter canvas rendering.
Parse fragmented cursor, focus, mouse and color controls without mistaking them for a human draft. Preserve genuine unsent text, exact provider receipts and uncertain-submission protection.
Highlight only the communicating agent pair. Human and system prompts do not animate connected notes or images; very large automatic canvases keep activity colors without continuous dash repainting.
Agent messages go through a persistent inbox: an agent in the middle of a turn receives everything in one prompt at its next boundary or attached to bridge responses, and the asker is released at once. Late answers reach the asker inbox instead of being lost; answers are never typed again into a waiting asker terminal. New orkestrai inbox and orkestrai reply commands (and inbox and reply MCP tools).
Assigning a task to a busy or still-starting agent no longer blocks the leader or reverts the task to unassigned: the briefing waits for the agent next turn boundary. New-task and completion notices follow the same path and wait for an offline leader to return.
Late provider confirmations are reconciled from the transcript without resending the message; attempts that never reached the provider return to the queue. Transport failures no longer mark the recipient as errored or pause supervision; the owner still sees the undelivered message in Control Center.
Floors start from the current main checkout, including uncommitted changes, and share installed dependencies without sharing caches. Finishing a task commits the Floor work, and floor land integrates even with local changes on main by applying only the Floor delta with a per-file three-way merge; on conflict nothing changes.
Landing a Floor records one commit on main with exactly its files, leaving other local changes alone. Retire integrated idle Floors only when ignored artifacts are verified regenerable test reports, tool caches or logs without sensitive files; preserve build/dist/out/target, symlinks, ignored data, live terminals, running processes and pending cards. Without room for a new Floor, safely removable integrated Floors are retired first.
The leader no longer waits for the owner to commit, land, review visually or free disk: owner-only checks become non-blocking follow-ups while it keeps integrating the rest. A waiting leader is reminded again when the team finishes work it can integrate.
New orkestrai heavy -- <command>: builds, E2E and packaging wait their turn in a machine queue sized by CPU, memory and disk, with automatic renewal and release on exit. New orkestrai stats with delivery and reply latency, queues, tasks and Floors.
Leader supervision reacts to new team blockers and results, not only card changes, with bounded stall reminders. Agent conversations whose transcript exceeds 256 MB restart in a fresh session with a handoff. AGENTS.md instructions are about 70% smaller and describe the continuous delivery flow.
Off-screen terminals stop repainting output and the simulator pauses its stream while hidden; off-screen notes defer rendering. The database limits and truncates its WAL, compacts free pages at startup and writes compact migration backups.
Never attribute one agent's new conversation to another when two agents of the same provider start together in the same folder: the first delivered prompt confirmed in its own transcript binds the conversation, and ambiguous launches no longer guess.
Agents no longer inherit the session of an outer Claude Code or Codex that launched Orkestrai. Inherited, Claude agents stopped persisting their own transcripts (breaking delivery confirmation, reply capture and resume) and received that session's messaging token.
A reply always reaches someone: answers to a reply go back to its sender, and a specialist's reply to a task briefing (checkpoints, questions) reaches the agent who created the task or the leader.
Messages no longer wait forever for an agent whose previous turn was cut off when the app closed: a turn left open by a process that no longer exists does not block delivery. Asking an agent that never saved a conversation starts a fresh session, as assigning a task does, instead of failing. An agent cut off mid-task by a restart is told to continue its card when its terminal returns.
Panning and zooming large canvases stays smooth: connections, the code graph, off-screen terminals and Portal previews react once movement stops instead of on every frame.
Landing a Floor never writes through a symbolic link, keeps the Floor when a repository hook refuses the commit, and removes only regenerable ignored files such as test reports, caches and logs.
orkestrai heavy stops only its own command tree and never runs past the machine's slots after losing its reservation. Inboxes deliver independently, concurrent answers count once, and answers keep arriving after an app restart.
ask returns within seconds and the answer arrives in your inbox; answers that could not be forwarded are delivered later exactly once. Landing a Floor also carries executable-bit changes, treats paths literally so no ignored private file is committed by a pattern match, and retries refused landings that delete files.
The leader closes each card, including its own coordination cards, as soon as the work is landed and committed: the Kanban is the source of truth.
Claude agents working in a Floor keep their conversation when the app starts again, instead of starting over without context.
A specialist idle for 10 minutes with a card still in progress is reminded to continue, close it or report what blocks it: at most twice per card, never while it waits on someone.
Builds and test suites under orkestrai heavy stop when free disk space drops below 1 GB, instead of filling the disk.
orkestrai heavy stops a build or test run whose agent died with the app, instead of leaving it running outside the queue next to the run the resumed agent starts again.
Update Handlebars to the official 4.7.10 patch in Postman and Bruno without downgrading their runtimes. Reject malformed compiler ASTs and forbidden constructor lookups, escape inline precompiled script terminators, preserve ordinary API Client visualizations and keep the full dependency audit gate.
Recover every pending answer forward through cursor pagination, keep retrying after failed lookups and preserve a new pending signal raised during another recovery pass. Restore reply capture for every unanswered batch item even when the first item was a notice or already answered.
Terminals receive the canvas xterm capabilities even when the app inherits TERM=dumb or vt100 from an outer process. Preserve explicit session overrides and forward TERM and COLORTERM into WSL.
Keep completed boards quiet and prevent unchanged supervision from consuming provider quotas.
Keep automatic supervision silent with no todo/doing cards, even with active Floors awaiting cleanup or promotion. Worktrees alone no longer request model turns or keep an on-demand leader awake; maintenance requires an explicit authorized task.
Persist a semantic board fingerprint with delivery receipts. Remind once for unchanged work across Core and terminal restarts; ignore timestamp-only updates and card ordering. Meaningful changes can trigger another guarded reminder after the cooldown; cancel stale queued context while preserving completion, uncertain-delivery, human-blocker, pause and quota safeguards.
Use a compact reminder with at most five shortened card summaries while deduplicating against the full pending board. Test five hours of unchanged work, completed boards with active Floors, restarts, task changes and stale delivery cancellation.
Update http-cache-semantics 4.3.0, Joi 17.13.8 and source-map-js 1.2.2; use global-agent 4.1.3 with a tested legacy Electron proxy bootstrap to remove sprintf-js. Replace the patch-package/braces chain with a confined jsdiff 9.0.0 runner: validate versions and exact before/after hashes for all eight existing patched files before writing, preserve LF/CRLF and idempotence, and stage atomic replacements resilient to disk exhaustion. Keep all patches and the full audit gate.
Recover missing columns automatically while preserving saved data.
Repair incomplete historical workspace-group schemas by adding missing parent, ordering, timestamp and collapsed-state columns and the index after validating existing definitions. Preserve rows, later workspace columns and legacy application-managed parent links. Never recreate tables, invent identities or silently clear unresolved references.
Include bounded schema-only expected/actual details in startup errors without logging row values or arbitrary defaults. Test independent DDL from the reported database, the full migration chain, backup preservation and rollback; verify signed Intel macOS packaged startup against that fixture before uploading installers.
Reject extra nested RSA DigestAlgorithm elements in node-forge with the reviewed, not-yet-released upstream PR #1152 patch pinned by immutable commit and SHA-512, plus a local install-time check requiring empty NULL parameters. Preserve Postman and the full dependency audit gate; test valid and malformed signatures, private-key imports and the packaged dependency.
Preserve existing data and keep recovery available when startup fails.
Resume compatible interrupted workspace-group migrations after validating columns, foreign keys and indexes. Preserve saved rows; reject incompatible structures.
Commit pending SQLite migrations and history atomically after a private, verified backup including WAL data. Keep only the newest completed automatic backup; unchanged boots do not create extra copies. Explicitly install the exact Electron/CPU SQLite prebuild even when packaging strips native build descriptors.
Start update discovery before the server and offer a localized native download, retry and logs dialog on startup failure. Detect child exits promptly and safely close destroyed splash windows.
Keep Control Center activity category badges readable on tinted surfaces in light and dark themes, with accessibility tests that always include populated activity.
Reuse the portable production bundle tested by CI and E2E for the exact commit, verifying provenance, lockfile and file hashes; keep native dependencies target-specific. Sign and notarize both macOS architectures concurrently, then validate and combine their updater manifests before publication. Pin lockfile checkouts to LF on every OS and test Windows-style Git conversion while keeping exact hash verification.
38
Orkestrai 0.39.0: Faster orchestration and safe Floor retirement
Batch worktree checks and bounded waits with explicit safety checks.
Audit Floors in one native call and clean selected safe revisions in a task-authenticated batch; preserve branches, local files and active work.
Confirm removal, show accessible failure alerts and report real merge conflicts; keep pending Floors visible and enforce workspace/path boundaries.
Limit Floor Git concurrency to four processes and share overlapping scans.
Return new cards without waiting for leader notification; remove the extra 90-second response wait; never confirm partial turns or resend uncertain deliveries.
Supervise active Floors after cards finish and avoid repetitive status, usage and memory preflights.
Rebuild packaged SQLite for the exact Electron version and CPU before signing, and verify its ABI entry point to prevent incompatible Node binaries from shipping.
Patch dependency advisories with axios 1.20.0, DOMPurify 3.4.16, fast-uri 3.1.8, ip-address 10.7.2 and moment 2.31.0; keep brace-expansion fixes within its existing 1.x, 2.x and 5.x branches.
39
Orkestrai 0.38.0: leader supervision and filesystem stability
Guarded reminders for pending Kanban work and more resilient Code Intelligence watchers.
The Core supervises idle leaders with pending todo/doing cards, persisted delivery receipts, cooldowns and human/automation safeguards. Completed boards stay quiet.
Code Intelligence excludes nested internal Floors/reviews, uses polling on macOS and recovers conservatively from filesystem exhaustion without silently serving an unwatched stale graph.
Keep SPA route, query and hash changes inside the same live Portal document. History navigation no longer triggers a detached preview capture that makes the whole page flash; explicit navigation still loads normally.
Update reviewed development dependencies, including Electron 42 patch releases, Svelte, Paraglide, icons and YAML tooling, while keeping the existing desktop ABI.
Integrate reviewed production dependency updates for storage, parsers, UI components, queues, native graphics, the Cua SDK and embedded voice.
Patch vulnerable transitive dependencies: fast-uri 3.1.7, ip-address 10.5.1 and undici 6.28.1, without changing the existing undici 7 runtime.
40
Orkestrai 0.37.0: Create music, narration and sound effects
Native audio workflows through fal.ai in Canvas and Workbench.
Add native fal.ai audio workflows for music, narration, dialogue and sound effects with model-specific fields, workspace references and Canvas/Workbench playback.
Share encrypted accounts, model grants, budget reservations, durable queue, agent CLI/MCP and workspace output files with existing creative workflows.
Keep literal speech separate from production notes; estimate character/duration and per-audio units (including MiniMax's audios unit) and reject unsupported raw audio before submission. Existing image and video workflows remain compatible.
Use audio-specific generation controls and identify denied output downloads without resubmitting paid jobs or making private media public.
Add owner-controlled private or one-hour link delivery for new fal outputs, pin privacy to cost previews and queued runs, and show provider request IDs for failed downloads.
Keep leader implementation delegated to specialists; explicit coordination/review assignments remain available. Approved plans and routine owned development-server work do not require repeated approval.
Keep terminal input responsive with asynchronous cached transcript lookup, bounded delivery queues, fewer Canvas refreshes, same-size PTY resize suppression and WebSocket recovery. Failed quick prompts retain their draft and show an error.
Confirm automatic terminal handoffs from the exact provider transcript on macOS and Windows/WSL; support Claude's positioned composer and pasted-content format, without stacking another task onto an uncertain draft.
Keep CLI JSON intact in pipes, expose whether assignment actually dispatched, provide explicit task_dispatch, and preserve semantic blockers until an explicit state change.
Attaching a simulator creates or reuses its Mobile Canvas node and existing session. Portal renderer/capture timeouts and failed loads remain recoverable without silently replaying expired actions.
Restore authenticated Computer inspection and isolate capture-retention errors to the affected workspace.
Restore assigned work for freshly opened recruits with free-form specialties; saved role templates are optional at startup, while explicit application of a missing template still fails.
Drain pending native Portal operations and flush browser cookies before shutdown; late requests cannot recreate tabs while the app is closing.
Allow busy providers time to persist queued prompts and recover the terminal queue on late exact confirmation, without replaying a message.
Resume saved agent conversations for structured messages after restart, including off-screen Floors. Suspended workspaces stay stopped; raw terminal commands never start a replacement session.
Keep verified deliveries pending when replies are late instead of reporting a broken transcript or asking to reload; CLI distinguishes delivery from a confirmed conversation.
Allow authorized Portal logins with write-only passwords over HTTPS or loopback, preserving redaction, private fields, OTP and explicit security gates.
Restore the previous virtual mobile device after Core restart without duplicating its node, waking unloaded workspaces, stealing another workspace device or bypassing physical confirmation.
Avoid competing compositor captures and recover transient Portal presentation failures without another resize.
41
Orkestrai 0.36.3: Portal stability and presentation
Keep Canvas controls accessible and separate Canvas zoom from website responsiveness.
Portal pages stay behind sidebars, toolbars and responsive dropdowns after dragging. Menus restore the same session when closed. Left/top origin clipping uses the in-node preview until the page is back in view; right/bottom clipping stays interactive.
Canvas zoom preserves the website viewport, responsive breakpoints, pixel density and independent same-origin Portals. Resize the node or choose a responsive viewport to change the actual page dimensions.
Portal previews and agent screenshots retain the full logical viewport, current scroll position and protected-field masks.
Prevent oversized Portal flashes during zoom and at idle. Capture outside the visible native hierarchy and restore a correctly sized frame; agent reads and title changes no longer trigger preview captures or synthetic page reloads.
Prevent a native Electron crash when opening or inspecting a new Portal. Wait for document readiness, keep native sizing separate from emulation, recover after failed navigation or renderer loss, and discard captures across document changes.
42
Orkestrai 0.36.2: Canvas media and project delivery
Keep Canvas inventory accurate and let agents deliver existing edited videos inside the configured project.
Keep nodes visible when deletion fails, including partial selections, and preserve connected edges on a storage failure.
Exclude historical media paths from default agent workflow reads. Historical generations require an explicit query; removed media is not automatically reused and project files remain intact.
Let assigned agents add existing edited project videos as playable Canvas nodes without generating, uploading or moving files. Validate format, size and optional hash, preserve project boundaries and reuse existing nodes on retries.
Expose the configured project folder, including distinct host and WSL paths. Keep briefs, prompts, montage sources, renders and final deliverables there by default; changing terminal directories does not change the authenticated workspace.
43
Orkestrai 0.36.1: Portal access and Canvas fixes
Restore team access to Portals, import local videos and keep browser surfaces aligned.
Restore Portal use without requiring an enforcing autonomy policy or an assigned task. New and unconfigured Portals allow all workspace agents, including new recruits; explicit access lists, manual/read-only settings, pauses, protected fields, enabled policies and audit trails remain in force.
Drop MP4, M4V, MOV, WebM or MKV files up to 64 MB as native Video nodes with playback and reusable workflow references. Preserve original bytes and leave provider permissions unchanged.
Keep Portal pages behind Canvas controls. Use a temporary in-node preview while dragging, panning, zooming or resizing, then restore the same live page without reloading its session.
44
Orkestrai 0.36.0: a new Canvas and Workbench experience
Refreshed native tools, direct manipulation and clearer navigation across desktop and Remote.
Drag creation tools from the Canvas dock and sidebar items into Workbench panes. Reorder flow steps, sequence clips, design layers/pages and collection-runner requests directly.
Start empty workspaces with one-click actions, check image executor readiness and drop references onto empty image nodes.
Refresh Canvas and Workbench with consistent node frames, grouped tools, readable tabs, compact usage meters and keyboard-accessible row actions.
Redesign the native Design toolbar, layer tree, selection controls and property panels, with color swatches, layout controls and multi-selection alignment/Boolean actions.
Unify menus, dialogs, focus indicators, theme-aware surfaces and reduced-motion behavior; improve compact text, contrast, selected states and recovery messages.
Simplify Settings and Provider Center with live sliders, switches and segmented selectors. Improve command search and documentation navigation.
Refine terminal voice controls, notes, Kanban, Git, portals, API collections, creative workflows, automation nodes and Workbench hubs while preserving existing data and agent contracts.
Guide Remote connections with explicit steps, status and recovery hints; keep sharing and connected workspace views consistent.
Update the website and project READMEs with real app recordings and refreshed product showcases.
Accept tool, file and character drops on the empty Canvas, keep Flow execution controls reachable and restore code graph indexing controls.
Prevent clipped collection-runner footers, Git file lists and compact device panels; keep Design selection controls and property values visible.
Restore selected permission/scope states, icon-only tooltips and readable badges across themes; distinguish unavailable usage data from healthy status.
Keep Remote responsive when a shared workspace contains a huddle, and correct misleading labels, translations and empty-state actions.
45
Orkestrai 0.35.0: Read scanned PDFs with local OCR
Local OCR makes scanned PDFs searchable by users and agents.
Read scanned and mixed PDF pages offline with bundled English, Portuguese and Spanish OCR, preserving originals and page citations.
Show OCR confidence, encrypted files and partial extraction; automatically reindex older scans and contain parser timeouts in a separate process. Normalize bundled PDF resource paths for Windows, macOS and Linux.
46
Orkestrai 0.34.0: Second Brain: connected knowledge
Turn documents and project work into cited, reusable knowledge.
Search native workspace knowledge with citations, tags, backlinks and a separate knowledge graph in Canvas and Workbench.
Import document nodes by dropping PDF, Markdown, XLSX, CSV and other files; retain originals and show extraction/freshness limits explicitly.
Retain provider-independent agent lessons with task reflection, bounded recall and automatic/review/off controls shared by UI, CLI and MCP.
Explore a colored, interactive 2D/3D knowledge network with live workspace and linked-file updates, stable camera/positions, source details and offscreen/idle resource cleanup. Keep hover details available across labels and after dragging a source. Pinch gestures also work when a finger starts over a label.
47
Orkestrai 0.33.0: more video providers
Use fal.ai, BytePlus ModelArk or Higgsfield with separate accounts, contracts and budgets.
Add explicit BytePlus ModelArk and Higgsfield video providers alongside fal.ai, with provider-specific contracts, vault accounts and persistent queue routing. Existing fal and Codex image workflows stay compatible.
Expose matching provider/account/model controls, dated public rates and Higgsfield configured account quotes. Preserve required references on explicit provider changes and block unsafe credential routing or duplicate paid retries.
Document provider upload, cancellation and reference limits in all three languages; share contracts with agents, storyboards and guided onboarding.
Stop elastic connections after settling instead of restarting their own animation. Bound non-converging physics, reuse adjacency indexes and defer offscreen image/workflow painting without unloading agents or changing source files.
Show named video references and previews for canvas nodes and standalone project files. Browse project media, expand a reference preview, and keep paths/API mappings under technical details without changing agent bindings. Align account/model controls.
48
Orkestrai 0.32.0: native video workflows
Opt-in fal.ai video generation on the existing Canvas and Workbench.
Wait for the Claude message composer before delivering initial roles, Kanban tasks or automatic Enter. Workspace trust dialogs remain under human control instead of accidentally selecting No, exit; readiness timeouts do not claim task delivery.
Separate character production notes and approval/version provenance from generation prompts; distinguish casting instructions from scene dialogue without changing frozen references.
Expose bounded provider descriptions, examples and defaults to users and agents; edit nested model inputs, flag stale parameters after switching endpoints and compare account base prices in the model picker before estimating a run.
Isolate creative forms per mounted dialog so restoring several Canvas nodes cannot reuse character, account, policy, brand or recipe form identifiers. Keep provider settings scrollable with account selection, tabs and actions fixed, including compact windows.
Keep searchable model and reference menus inside the viewport, including workflow inputs near the bottom of a dialog.
Reserve character image and voice inputs when attaching scene media so the editor offers the next available model slots without conflicting with approved identities.
Declare exact workspace media response sizes so voice reference previews can determine their duration without waiting for playback to finish.
Close the video encoder confirmation after approval so the completed download cannot leave the sequence editor blocked behind an overlay.
Search the full fal.ai video catalog, including Seedance 2.5, with model-specific contracts, nested parameters, multiple image/video/audio references and per-workspace model grants. Existing Wan/Kling drafts and Codex image workflows remain compatible.
Configure vault credentials, workspace and model permissions, agent access, outgoing-data consent, estimates, budget reservations and concurrency under existing security gates.
Recover persistent jobs without duplicate paid submissions, cancel tracked work and retry downloads into the project. Manual UI, CLI/MCP, documentation and guided onboarding share the same workflow. Native video/audio files keep their original bytes and each output rechecks filesystem and network permissions. Provider-rejected generation is a failure, not a retryable download; its budget reservation is retained.
Save and owner-lock versioned character appearance and voice references. Reuse immutable, fingerprinted masters through UI and agents; block incompatible voice bindings and silent identity replacement before paid generation. Drag approved characters between workspaces with all visual and voice references, independent local copies and grouped native nodes; credentials and permissions are not transferred. Preview master images and voice, automatically match compatible inputs, and insert named character references without silently dropping assets or changing identities.
Require a protected operating-system credential store and preserve existing credentials when the vault file is unreadable or malformed.
Find curated skills by spaced or hyphenated names even when the online registry is unavailable.
Keep video account settings read-only until loading completes so delayed responses cannot overwrite edits or validation feedback. Keep creative dialog edits and selections stable during workspace refreshes.
Add native storyboards with ordered scenes, dialogue, locked character selections, references and linked image/video drafts. Share revision guards with CLI/MCP; preserve outputs and show stale or missing dependencies without automatic paid execution.
Compare image and video variants with synchronized playback and immutable, hash-bound review history. Changed files invalidate approval; agents can propose but cannot self-approve.
Add contextual image variation, genuine-alpha background removal, annotated changes and animation drafts. Preserve originals and frozen source lineage through the existing Codex image and fal video workflows.
Save shot intent alongside scenes without losing reference or voice bindings.
Version palettes, logos, products and creative rules, then place their exact assets in any workspace.
Save versioned scene briefs with named inputs and recreate editable storyboards without copying execution authority.
Order, trim and caption existing video clips without changing their source files.
Delete workspace-bound policies, audit records, integration events and tools in dependency order without leaving a partially removed Canvas; preserve unrelated workspaces and project files.
Update the serialization dependency to fix denial of service from malformed input.
49
Orkestrai 0.31.5: agent communication and terminal selection
Reliable agent handoffs and precise terminal text selection.
Deliver queued agent messages without waiting for an entire model response. Reciprocal messages unblock waiting callers, Claude mid-turn messages retain their transcript association, and valid completion reports remain available to the leader.
Keep terminal mouse coordinates aligned with Canvas zoom for dragging, word selection and scaled displays.
50
Orkestrai 0.31.4: responsive dictation and terminal recovery
Prevent indexing stalls, cancel stuck dictation, and preserve each agent's conversation.
Exclude dependency folders and generated build output from code indexing at every depth. Select large file sets deterministically to prevent repeated database rewrites that block terminals and delay microphone startup.
Cancel microphone startup or transcription from the same control in Canvas and Workbench. Stop capture promptly and discard late results after cancellation or navigation. Startup times out after 15 seconds, transcription after three minutes; recording finishes automatically after 15 minutes.
Restore each terminal with its exact conversation. Recovery cannot adopt or stop another agent's PTY, and simultaneous wake requests cannot create duplicate writers.
51
Orkestrai 0.31.3: opening project folders
Open output folders without enabling Computer Control.
Restore agent-requested opening of project and generated-image folders in the system file manager without requiring Computer Control or screen permissions. The dedicated bridge action confines paths to the workspace or approved repositories, authenticates the agent, records an audit trail, and preserves explicit security restrictions.
52
Orkestrai 0.31.2: Canvas selection transfers
Copy and move selections between workspaces again.
Restore copying and moving Canvas selections between workspaces, including large selections of 59 nodes, while preserving internal connections and relative positions. Invalid requests are rejected before changing nodes, with clearer validation feedback.
Restore external file drops: images dropped on empty Canvas become separate persistent reference nodes, while files dropped on an agent become prompt attachments without sending automatically. Recognize protected file drags before drop and preserve every input while uploading.
Save numeric Design edits only once on Enter and wait for pending saves before applying layout, avoiding duplicate revision conflicts.
53
Orkestrai 0.31.1: incoming companion messages
Forwarded messages and media captions reach the authorized companion.
Recognize forwarded desktop messages and media captions from the authorized contact so companions receive their events instead of silently waiting. Preserve message identity and duplicate protection; downloading or transcribing attachments still requires separate authorization.
54
Orkestrai 0.31.0: continuous desktop companions
Conversation monitoring, private memory and scheduled work with owner-controlled access.
Monitor authorized desktop conversations in the background and wake the Canvas agent when new messages arrive. On macOS, optional automatic navigation reopens the approved contact.
Process incoming messages in persistent batches, including after reopening a chat, with complete-draft verification, duplicate-send protection and recovery for interrupted composition.
Configure each companion's persona, language and voice, with optional private conversation memory, sourced preferences, retention controls and deletion.
Let agents create and manage task-linked calendar schedules for reminders and reports, with timezones, recurrence and missed-run handling.
Configure native photo and file delivery, authorized incoming audio transcription, 30 local TTS voice presets and audio playback in the Workbench. Media controls depend on the supported application; synthesized audio is an attachment, not a native voice note.
Keep desktop access scoped to approved applications and contacts, with publication safeguards, auditable actions and automatic cleanup of temporary captures and attachments.
Improve Computer panel reliability, permission feedback and observation controls across Canvas and Workbench. Package the native desktop runtime for each installer architecture.
Update runtime and UI dependencies, harden integration delivery against duplicate retries, and improve the contracts and validation for agent-created tools.
55
Orkestrai 0.30.0: natural-language desktop work and signed microphone access
Ask an agent to operate your existing desktop session
Signed macOS apps and helpers now include microphone and Apple Events entitlements. Local packaging and release CI verify the actual signed permissions, not only signature validity and notarization. The STT model is unchanged.
Agents can prepare and connect the workspace Computer node, create their own assigned task without self-dispatch, and open or reuse an approved native application through CLI and MCP. Existing browser windows keep their signed-in session.
Preparation inherits only an existing bounded Computer/app grant and never reactivates an owner-paused node. Application identifiers can be authorized before an app opens, and closed authorized apps remain visible and revocable.
Computer Control uses stable native macOS window IDs, rejects stale or ambiguous windows, verifies focus before input, serializes host input across workspaces, and blocks automatic replay after a possibly partial failure.
Declared email/publication, purchase, credential, or destructive risks use the existing Security gates before native input. Approval is bound to the exact request and attempt. Pixel-only actions still require correct risk classification by the agent.
Native SecretRef delivery suppresses echoed process output and errors; screenshot storage and reads reject paths that escape the workspace through symlinks. Typed content stays out of audit metadata.
Computer state and agent captures refresh in Canvas and Workbench without manual reload. Readiness reflects activation and OS permission. Manual typing stays bound to the selected allowed window; stored captures are read-only evidence, not stale click maps.
Computer inventory collapses identical native window/display records and excludes conflicting identities. Duplicate macOS Finder records no longer crash rendering or leave the loading indicator stuck; failed loads offer retry, and refresh keeps the existing panel visible.
Flow execution waits for pending step edits to be saved in Canvas and Workbench. Failed saves prevent execution instead of starting with stale or missing steps.
Native desktop adapters correct Linux window/PID discovery, Windows PowerShell argument handling, literal text and shortcuts, and macOS display coordinates and capture bounds. Windows desktop commands load only built-in PowerShell modules and preserve argument boundaries. Launch uses registered applications; unsupported targets fail explicitly.
Microphone access is limited to the real app renderer and exact local origin, and the Settings input meter resumes a suspended audio context. Documentation and the guided desktop use case now start from a natural-language agent request, including Remote.
56
Orkestrai 0.29.0: Persistent agents, visible browser workflows, and reusable tools
Watch an agent work in your signed-in browser
Background schedulers wait for database migrations on first launch and retry transient database failures without crashing the Core.
RPM installers no longer claim shared build-id paths that conflict with other Electron applications. Linux release validation checks the generated RPM before upload.
Updated the transitive Joi and js-yaml parsers to patched compatible versions; the dependency security audit is clean.
Settings now includes Autonomy & 24/7 Core with background execution, optional start at sign-in, live uptime, process status, automatic recovery count, and a guarded restart command.
Closing every window keeps enabled active work in the system tray; Quit Orkestrai remains the explicit command that stops the Core and its child processes.
The desktop supervisor reconnects after an unexpected Core exit or system resume without creating a second provider writer.
Core discovery metadata is private to the operating-system user, and the health endpoint requires a per-process token that is never exposed to the renderer.
Automation runs are persisted before execution, claimed with renewable leases, checkpointed, retried with bounded backoff, recovered after interruption, and moved to Needs intervention after the final attempt. Closed target agents are started or resumed automatically.
Each provider terminal supports Interactive, On demand, or Persistent lifecycle modes with idle sleep, exact conversation resume, per-agent concurrency and provider-usage safeguards, manual wake/sleep, and Core supervision without a rendered Canvas.
Portal is now a Core-managed browser with persistent workspace/private profiles, host allowlists, workspace-confined files, semantic references, typed actions, durable automations, and attributable evidence without exposing login credentials to agents.
Automations now includes a workspace security plane with standing grants, high-risk approval gates, quiet hours, concurrency limits, immediate emergency stop, encrypted and bound SecretRefs, and a verified hash-chained audit export. Resolved values are redacted before any output metadata is persisted.
The Integration Center now connects Gmail, Slack, Telegram, WhatsApp, GitHub, and HTTPS webhooks through operation-level grants. OAuth and app credentials stay in the operating-system Vault while agents use typed, idempotent actions and sanitized event history.
A native Computer node can inspect and control explicitly approved desktop apps on macOS, Windows, and supported Linux X11 sessions through typed, idempotent commands, bounded screenshot evidence, and content-safe audit records.
Tool Workshop creates reusable versioned workspace tools from integrations, bounded HTTP, deterministic transforms, and confined commands. Agents propose drafts, the owner publishes immutable revisions, automations execute them idempotently, and exact SecretRef bindings keep credentials inside trusted executors.
Tool Workshop forms stay independent across Canvas and side panels, and browser executors use the correct label.
Portal movement now follows Canvas geometry changes without 100 ms polling, repeated tab hiding, or redundant page zoom resets. Codex forwards ephemeral bridge identity to MCP, and quick prompts use the confirmed-submit queue.
Saving autonomy policies, approving gates, and creating SecretRefs now validate their HTTP route parameters correctly.
Browser popups remain embedded in the Portal and preserve their opener instead of opening a separate OS window. Protected screenshots wait for a fresh frame, including during explicitly enabled background control.
Portal agents now operate the same visible native page with named-agent grants, read/interact modes, explicit background opt-in, immediate pause, protected-field masking, stale-reference checks, and no arbitrary scripts. Preauthorized agents can publish bounded browser/integration/HTTP/transform tools; gate resumptions preserve checkpoints and the original revision. Audit details, local webhook setup, and latched emergency stop are explicit in the UI.
57
Exact social image delivery without unsafe cropping
Creative workflows now target real platform dimensions and keep important content inside measurable safe areas.
Choose Instagram square or portrait, Stories/Reels, TikTok, or bounded custom dimensions in the same Canvas, Workbench, CLI, and MCP workflow.
Orkestrai preserves the untouched native master and resamples the complete frame only when its aspect ratio already matches the target within a strict tolerance.
An incompatible result returns to native ImageGen with a dedicated safe-area reframe prompt, preserving text, logos, people, products, and calls to action instead of cropping or distorting them.
Generated images copied together with their workflow to another workspace retain a copied and remapped native-master provenance.
Native Windows PowerShell agent terminals restore wheel history scrolling without zooming the Canvas underneath.
58
A clearer, more consistent workspace
Canvas, Design, Workbench, settings, and supporting tools now share a denser and more legible visual system in light and dark themes.
The native Design workspace adds a scrolling dot grid, elevated artboards, a unified multi-selection box, live size and angle readouts, stable frame labels, and a task-oriented inspector order.
Selection and hover chrome is isolated from artwork effects and stripped consistently from every export, thumbnail, and visual-review reference; ruler guides no longer invalidate an approved revision.
The light theme is restored before first paint, previously undefined semantic tokens now resolve consistently, and provider marks remain readable in both themes.
Canvas node headers share one compact button and tooltip contract, node corners and shape controls align with the active theme, and the minimap uses recognizable per-tool colours with pan and zoom.
Dense labels across Canvas, Workbench, Remote, API Client, code graph, Providers, Skills, settings, documentation, and account screens now meet a readable baseline.
Rapid node edits are kept optimistically and persisted in order, preventing an older response or component remount from restoring stale image-workflow settings.
Official Design templates are audited for clipped text, and production E2E now keeps visual baselines for Canvas and Design across three desktop viewports in light and dark themes.
59
Reliable Git Floors on WSL
Windows workspaces backed by WSL now keep every Floor operation inside the selected Linux distribution.
Floor creation, Git hooks, previews, landing, removal, status summaries, and code-graph change impact now execute through the workspace WSL runtime instead of Windows Git.
The agent bridge identifies the main . repository, its Git status, and its native or WSL runtime separately from optional additional repository aliases, so an empty alias list is no longer ambiguous.
60
A professional native Design workspace
The same Design node now provides a sharp, focused visual workspace and a complete path from concept to reviewed code.
Focus mode fills the app viewport while preserving the exact document revision, selection, tool, camera, zoom, and panel state when returning to Canvas.
Pages, hierarchical layers, precise selection, rotation, measurements, arrangement, professional color and gradient controls, responsive auto layout, prototype connections, Inspect, and agent operations now share one revision-safe document.
Canvas previews render complex designs sharply at 2048 px and reopen framed on visible artwork instead of a blurred thumbnail or empty black region.
Guided explorations require a Brand board, responsive frames, typed and bound tokens, components, prototype, applied framework code, and current visual approval before delivery can close.
Task-scoped handoffs expire instead of arriving after completion or reassignment, completed work cannot regress from late status, and resumed leaders must reconcile and finish active blocked work.
Windows and WSL bridge commands use a pinned SHA-256-verified console Node runtime, preserving stdout, stderr, MCP JSON-RPC, and structurally confirmed agent replies instead of invoking the graphical executable.
Focused Design transitions now release browser focus before hiding the previous surface, preventing inaccessible controls and Chromium accessibility warnings.
Newly recruited agent terminals reach their first stable prompt before Kanban work is delivered, and nested layer duplication preserves parent-child order.
Code imports preserve responsive dimensions, scoped selectors, inherited typography, complete CSS padding, content-hugging controls, mixed inline text, flex layout, mobile navigation, and floating actions without clipping, overlap, or compression.
Agents remain scoped to the original native or WSL workspace after entering a scratch or sibling directory, without copying its bridge token.
API Client compatibility now uses the latest reviewed Bruno filestore, script, and request runtimes, while CI uploads failure diagnostics through the Node 24-compatible artifact action. Production dependencies audit with zero known vulnerabilities after the upgrade.
61
A complete Remote workspace on every screen
Remote navigation is reliable on phones, and each agent now has a safe operational work summary.
Team shows each agent's current focus, recent semantic activity, and bounded coordination totals without exposing internal message bodies or terminal output.
The phone navigation occupies a stable viewport row with Overview, Team, Tasks, and Activity visible while More provides Huddles and Reviews with attention counts.
Task columns become a readable vertical selector on small screens, while desktop keeps the horizontal board.
Agent dialogs and Huddle rooms have explicit return actions, and safe areas no longer clip headers, composers, navigation, or the full-screen terminal.
Huddle details fit the remaining screen area, mobile text entry avoids Safari page zoom, and an open PWA reloads when an upgraded service worker takes control.
62
Confirmed agent message delivery on Windows
Inter-agent messages now start a real provider turn instead of stopping as pasted text in a terminal composer.
Windows and WSL wait for the TUI composer redraw to settle before sending Enter, preventing the key from overtaking a pasted prompt.
A terminal redraw no longer counts as delivery: Orkestrai confirms the exact prompt in the provider transcript and retries only Enter within a bounded window when needed.
Direct questions, one-way handoffs, Kanban assignments and completion notices, roles, routines, and design reviews now use the same confirmed delivery path.
WSL sessions preserve their distribution-specific transcript home and Linux working directory for both delivery acknowledgement and exact reply correlation.
63
Secure dependencies and CI-gated releases
The dependency baseline is secure again, terminal packaging is deterministic, and a failing CI can no longer publish a desktop release.
Patched dependencies resolve the Faker, fast-uri, qs, and xmldom advisories while preserving every supported Postman dynamic variable and the Bruno and Postman script runtimes.
Clean macOS and Linux installs restore the executable permission required by the node-pty spawn helper, preventing terminal startup failures caused by package extraction modes.
Shell terminals select an installed host shell instead of assuming that zsh exists, including when an older workspace is reopened on another platform.
The attention counter maintains WCAG AA contrast in the light and dark themes.
Embedded STT and TTS now use sherpa-onnx-node 1.13.7 with the matching official native packages for macOS, Linux, and Windows.
CI now runs the complete Playwright end-to-end suite alongside the dependency audit, unit tests, and production build. Local preflight and Release Desktop both require that exact commit to pass before a tag can publish platform artifacts.
64
Move complete Canvas selections between workspaces
Reuse a connected working setup without rebuilding nodes or carrying unsafe runtime state.
Select up to 100 Canvas nodes and copy or move them to another workspace while preserving relative layout and every connection whose two endpoints are selected.
Note attachments, image files, and native Design documents are copied into the destination; terminals keep configuration and saved commands but begin a clean conversation, while API credentials and execution history remain detached.
Move commits the destination before removing the source and blocks active routines, running image workflows, missing files, and duplicate mobile-device nodes with actionable feedback.
65
Operational code intelligence from question to delivery
The native code graph now connects bounded context, live ownership, explainable evidence, revision history, and traceable handoffs.
Context packages enforce a selectable token budget, preserve exact symbol and relationship provenance, bound source excerpts, and redact likely credentials before any handoff.
Leader, agent, Council, task, and review handoffs create traceable work artifacts instead of copying untracked terminal text.
Live overlays show active agent, task, and Floor ownership and warn when parallel work overlaps on paths or indexed symbols.
Named investigations restore repository scope, mode, query, selection, source location, and graph camera.
Every relationship can explain its static, inferred, or runtime origin and confidence; revision comparison covers both symbols and relationships.
Graph selection, Monaco source position, CLI, and MCP share the same persisted contracts without resetting the user viewport.
Workspace provisioning preserves existing AGENTS.md and CLAUDE.md content inside projects and updates only an explicit Orkestrai-managed block.
Assisted, Manual, and Disabled workspace modes control graph freshness and access. Assisted refreshes settled structural and semantic edits incrementally, reuses unchanged semantic vectors, and makes agent reads wait for the newest known revision.
Mouse, touch, wheel, visible canvases, and pointer picking remain aligned through outer Canvas zoom and node resizing. Scale-aware labels stay readable; click selects, centers, and opens a symbol, double-click focuses it closer, and empty-space double-click no longer zooms blindly.
Repository scopes, Change impact, lexical search, and semantic intent search explain their actions directly.
Normal Portal tab requests create another Portal node in the same Canvas, while true authentication pop-ups remain sandboxed child windows with the persistent session.
Restored MCP or CLI sessions can safely derive Control Center identity from an assigned workspace task when their environment identity is missing.
Notes created through CLI or MCP connect only to their exact author by default, keeping parallel flows isolated unless a specific agent or the complete team is selected explicitly.
Canvas connections now offer Adaptive, Elastic, and Static rendering. Dense workspaces shed idle rope simulations earlier, while Static disables all edge physics and animation for lower-power hardware.
66
Native image workflows for people and agents
Briefs, references, Codex ImageGen, and reusable outputs now form one traceable canvas graph.
Image Generation nodes combine their prompt with connected Note context, up to five ordered Image references, and a connected live Codex, then return each validated output to the selected workspace folder and canvas.
A connected Codex has full typed control: it can create and configure drafts, attach, remove, or reorder inputs, and execute one logical run with up to ten requested outputs.
The assigned Codex uses its authenticated built-in image_gen.imagegen tool, passes local references through referenced_image_paths, and follows one canonical alpha contract controlled by the node toggle.
Transparent outputs are decoded and checked for real alpha pixels. With multiple references, ImageGen first composes on a uniform white matte and removes it in a second native edit; a proven checkerboard-cutout prompt provides another native fallback across the bounded attempts. Python and local pixel manipulation remain forbidden, and abandoned runs expire.
Orkestrai never asks for or stores an image API key or calls a provider endpoint directly; only the assigned Codex can complete the run and exact paths, signatures, sizes, and public errors are bounded.
Canvas, Workbench, CLI/MCP, and the guided creative-team tour operate the same visible workflow without parallel automation state.
The creative tour now builds and validates Character → Brand → Carousel, waits for real outputs, and automatically connects each result to the next stage.
The UI states that an authenticated Codex account or plan with ImageGen is required; no OpenAI API key is needed.
New and recovered agents resume only an exact conversation ID already attributed to their canvas node. Without that ID, Orkestrai starts a clean conversation instead of using another terminal's latest session.
Interactive Codex terminals receive the MCP launcher from the current Orkestrai installation before resume, so stale global configuration cannot hide image workflow tools.
Packaged builds retain the complete WebADB runtime; unavailable Portals back off with manual retry, stale board calls return not found, and expected Windows ConPTY cleanup races no longer flood logs.
Desktop diagnostics now use AudioWorklet, current Electron and updater APIs, structured Portal script results, and dependency-aware Svelte state without emitting known runtime warnings.
67
Reliable terminal paste on Windows
Ctrl+V now pastes copied text instead of being mistaken for an image-paste command by agent CLIs.
Windows terminals detect native clipboard text and dispatch a real xterm paste, including bracketed-paste handling for interactive CLIs.
Clipboard contents stay inside Electron and are never exposed through the renderer bridge; only the focused Orkestrai window can request the paste.
When the clipboard has no text, Ctrl+V preserves the original control character so provider image-paste workflows remain available.
68
Persistent workspace pause and quiet Windows background work
Unload now keeps a workspace idle until an explicit open, while internal runtimes stop stealing focus on Windows.
Unloading persists a paused workspace state and stops its live terminals, task-driven session creation, and background automation enqueueing across navigation and app restarts.
Paused workspaces stay visible in Canvas and Workbench with their layout and exact agent conversations saved; explicitly opening one is the only action that resumes it.
Internal CLI and server child invocations that reach Electron’s Windows single-instance handler no longer restore or focus the app; intentional app launches and user actions keep their normal foreground behavior.
69
Portable Roles and faster workspace setup
Reuse specialist Roles, select the right provider account, and file a workspace into its folder during creation.
Discover Roles from any selected project folder through the native picker; imported files are bounded, validated, confined to that project, and never overwrite an existing role.
Choose a configured Provider Profile while creating an agent; the provider/profile pair is validated before persistence while credentials remain in secure storage.
Create regular or preset-based workspaces directly inside a validated sidebar folder, including from the folder header, without a partial root-level creation.
Profile-only terminal changes now reach the backend and show the correct account-specific confirmation.
Usage views distinguish multiple accounts through their unique routing identity and profile name instead of crashing on duplicate Svelte keys.
Windows and Linux install a verified update silently after restart confirmation and relaunch the app.
Windows side panels stay below the native title-bar controls and within the visible viewport.
The terminal theme menu scrolls on short screens, and enabled buttons consistently expose their interactive pointer cursor without misrepresenting disabled controls.
70
Safe Codex MCP configuration
Codex keeps the automatic workspace bridge without losing control of its global dotfile or Git visibility.
Codex receives the Orkestrai and official Figma MCP definitions through ephemeral native and WSL launch overrides; workspace provisioning no longer rewrites ~/.codex/config.toml.
The exact malformed multiline args and duplicate env structure written by older Orkestrai builds is repaired after validation, with a backup, serialized access, and atomic replacement; unrelated malformed TOML stays untouched.
AGENTS.md, provider MCP files, and opencode.json are no longer hidden through .git/info/exclude; exact legacy blocks are narrowed to Orkestrai-owned runtime and skill directories.
Bridge provisioning failures now enter desktop diagnostics instead of disappearing silently.
71
Organized workspaces, responsive Portal, and desktop diagnostics
Workspace navigation, embedded browser testing, provider setup, terminal personalization, and support diagnostics become more complete without weakening local boundaries.
Workspaces support persistent nested folders with drag and drop, renaming, collapse state, cycle prevention, and non-destructive folder deletion.
A coordinating workspace can explicitly authorize sibling repositories through safe aliases such as @api-tests/bruno, allowing agents to edit repository-owned Bruno and Postman collections while arbitrary parent paths and symlink escapes remain blocked; creation flows that omit this optional list remain compatible.
Every Portal has a persistent editable name separate from its address; agents inventory all workspace Portals with explicit connection state, target them by unique name or id, reuse repeated URLs, and get a theme-aware address rail plus the same responsive viewport with contained two-axis scrolling.
Workspace MCP servers and installed skills propagate to Cursor, Cline, Devin, Antigravity, and OpenCode through their supported native configuration formats.
The Skills marketplace opens with a curated catalog, merges live results safely, and validates bounded external registry payloads before they reach the interface or filesystem.
Terminal settings preview colors, font, and padding, use OS-specific shortcut labels, and add Monokai, Ayu Dark, Rosé Pine, and Solarized Light; selection remains accurate at any Canvas zoom.
Installed builds expose Developer tools and a direct action for rotating local logs that capture renderer and internal-server failures while redacting common credentials and excluding normal agent output.
72
Provider accounts, live status, identity, and clearer routing
Nine agent CLIs now have clearer identity, isolated local accounts, public service health, and safer terminal startup.
GitHub Copilot joins the canvas as a native provider with detection, models, permissions, resume, bridge provisioning, and headless operation.
Provider Profiles keep multiple local accounts separate and resolve their credentials only in server memory immediately before PTY spawn; secrets never enter persisted canvas node payloads.
Provider Center shows allowlisted public service status and bounded incident details without accepting arbitrary upstream URLs, while unavailable checks never report a false healthy state.
Provider names and marks stay consistent across the canvas, Workbench, floors, usage, status, routing, and the Remote companion.
Profile names are unique regardless of capitalization, legacy duplicates migrate without data loss, and profiles in active use cannot be deleted.
The PTY WebSocket accepts only the exact application origin and port, including equivalent loopback hostnames, closing cross-port localhost access.
Usage and routing opens at a useful default size, exposes Leader routing first, and contains scrolling without zooming the canvas.
Obsidian joins the terminal theme catalog with readable ANSI colors in the terminal preview and live sessions.
73
Reliable project, Portal, voice, and terminal state
A focused stability patch keeps project configuration authoritative and restores the exact desktop state users expect.
Shells and agents no longer inherit the Orkestrai server APP_KEY or other private runtime values, so Laravel encryption, cookies, and sessions use the project .env.
Portal login pop-ups remain inside a sandboxed Orkestrai window with the same persistent session; cookies and storage are flushed to disk and the node restores its last URL.
Dictation captures direct PCM through the same Web Audio route as the live meter, normalizes quiet speech, and reports when the selected microphone opened without producing signal.
Terminal font and pane geometry settle before PTY reattachment, and ANSI history finishes replaying before redraw so the cursor stays aligned after navigation.
74
Durable coordination and shared context
Messages, attention, delivery evidence, workspace knowledge, reusable teams, and focused multi-agent conversations now form one recoverable operating history.
Durable message envelopes, delivery receipts, semantic activity, and a global Attention Center keep questions, permissions, failures, and outcomes visible across restarts and workspaces.
Universal Search adds structured filters for messages, activity, attention, Workspace Memory, and Huddles; direct commands open Workstreams, Annotation Center, and Team Packs.
Workstreams connect Kanban, agents, Floors, Git, Council, reviews, evidence, tests, risks, and linked files without creating a second source of truth.
Workspace Memory keeps source-backed, revisioned decisions and constraints; Annotation Center projects canonical code-review and Design feedback with stale-revision detection.
Versioned Team Packs add release notes, immutable history, integrity validation, portable import/export, and runtime-state stripping to reusable teams.
Persistent Huddles support selected agents, targeted turns, dictation, transcript recovery, linked tasks and Workstream evidence, CLI/MCP access, and encrypted role-scoped use from the Remote PWA.
Agents can link repository-owned Bruno, OpenCollection, and Postman collections to the native API Client and synchronize guarded edits back to the original project files.
Attention items expand to the complete original request, and provider replies are correlated to their exact historical turn instead of producing false structured-transcript failures.
Maestro recruitment now starts the agent in the leader’s active Floor and selected WSL runtime; Kanban advances only after the complete briefing reaches a working terminal.
Windows and WSL Codex composers confirm long task submission and retry only a swallowed Enter, while selected terminal text copies through the native Windows clipboard.
75
Complete API test authoring for people and agents
Runtime-aware JavaScript tests, completion, and protected agent workflows now share one collection model.
The Tests tab separates structured assertions from a full-height JavaScript editor with contextual completion for Bruno, Postman, and native Orkestrai APIs.
Connected agents and leads can create, read, fingerprint-replace, execute, and export complete collections through typed MCP tools or CLI commands while local secrets remain redacted.
Bruno tests blocks and Postman test events round-trip through their matching exports, and saved runners preserve environments, iterations, data rows, flow, and failure policy.
Documentation code examples now use language-aware syntax highlighting with legible token contrast.
76
Official Postman and Bruno scripting runtimes
Imported API scripts now run through the official Postman Runtime or Bruno’s official safe QuickJS runtime.
Separate scopes, sendRequest/runRequest, cookies, collection flow, visualizers, bundled libraries and Chai, legacy helpers, and the operating-system-encrypted vault are supported.
Imported Bruno request and folder variables, post-response variables, declarative assertions, and tests blocks execute through Bruno’s official runtimes.
Collection runners expose accurate iteration data and metadata to Postman and Bruno scripts.
The scripting reference includes copyable Postman, Bruno, and native examples and identifies hosted Postman services and unsafe Bruno NodeVM access that remain outside the portable, secure runtime.
77
Reusable commands and a multi-protocol API Client
Shell startup becomes repeatable while the API Client covers editing, execution, security, responses, and synchronization in the daily workflow.
Invalid pre-request and post-response scripts now identify their exact request or collection stage and source line instead of showing a generic execution failure.
Terminal commands can be saved per terminal or globally, searched, run on demand, and optionally executed once when a pure shell is restored.
The API Client executes HTTP/REST, GraphQL, WebSocket, and gRPC with protocol-specific editors, messages, and every streaming mode.
Assisted OAuth 2.0 with PKCE, cookies, proxy, custom CA, PEM/PKCS#12 certificates, and TLS verification cover authentication and network workflows.
JSON, JavaScript, GraphQL, and XML have syntax color, search, and formatting; JSON/XML responses use expandable trees and protocols expose transcripts.
Nested folders, isolated drag, contextual actions, and multiple runners preserve order, environment, iterations, delay, failure policy, and variable chaining.
Import and export cover Bruno, OpenCollection, Postman, Swagger/OpenAPI, Postman environments, and versioned native backup with explicit fidelity notes.
78
Native RPM installation for Linux
Fedora, RHEL, CentOS, and compatible distributions now have a package-native Orkestrai installer.
Every Linux release now publishes an RPM alongside the existing AppImage.
The package includes the public maintainer metadata required by native Linux installers.
RPM files use the same stable Orkestrai product naming as the other installers.
The release pipeline verifies the RPM and its latest-linux.yml updater entry before publishing any assets.
Installed RPM builds use the package-aware Linux update path.
79
Native Design Studio and reusable workspace tools
The largest Orkestrai release so far connects visual design, implementation, agents, and daily desktop work in one traceable workspace.
Native Design Studio now covers editable vectors and SVG, Pen and Bezier editing, layout and constraints, typed tokens and modes, components and libraries, Figma interoperability, framework code generation, visual diff, prototypes, motion, collaboration, quality audits, recovery, and high-throughput typed agent operations.
Guided UI exploration creates three progressive directions and requires a revision-aware human visual approval before the selected direction expands into complete states, system, prototype, and code.
The native API Client executes saved HTTP requests, imports Bruno and Postman collections, works in Canvas and Workbench, and exposes secret-safe typed tools to connected agents.
Reusable Canvas annotations support visible duplication and complete multi-selection copy and paste; native shells preserve their folder, and agents can explicitly discover and edit existing notes.
Terminal scroll no longer zooms the Canvas at its limits, Escape reaches Vim and other TUIs without losing xterm focus, and Command/Ctrl+K searches the complete localized documentation again.
WSL sessions validate the exact distribution, path, login environment, provider CLI, and conversation before spawn or resume, while completed Floors retire obsolete nodes and edges from operational views.
Usage documents real quota capabilities for all eight providers, audio settings select and test input and output devices, and adaptive indexed Canvas connections keep dense workspaces responsive.
80
Three complete UI directions from one brief
A guided Canvas workflow now makes design exploration complete, comparable, and traceable before implementation begins.
One transaction creates a linked spec, five Kanban stages, and native Clarity, Expressive, and Efficient Design documents.
Each direction must include responsive states, typed tokens, reusable components, a navigable prototype, quality evidence, and framework code.
Users can work manually or delegate the first stage to a currently running workspace leader.
Every task dispatch includes the complete linked note, description, images, files, and links.
A human decision gate prevents automatic implementation before one direction or an explicit combination is approved.
Canvas handles and connections reuse indexed graph geometry for better dense-workspace performance.
81
Predictable audio and responsive canvas connections
Audio routing is explicit across the app, while dense workspaces keep collaboration signals without paying for full rope physics everywhere.
Settings → Voice selects and tests the microphone used by every local dictation surface and the speaker used by previews and spoken replies.
A live input meter and short output tone verify the devices before saving.
Removed devices fall back to the system default; errors distinguish permission, missing hardware, interrupted capture, and likely contention for the only microphone.
Canvas edges adapt simulation detail and frame rate to edge count and viewport visibility while keeping selected and active conversations visible.
Hidden windows and reduced-motion environments stop connection animation.
82
Design quality, scale, and recovery
Native Design documents now carry production audits, complete starting structures, and resilient recovery inside the same editor.
Quality checks naming, clipping, overlap, WCAG contrast, and accessibility and jumps directly to each affected layer.
Complete product, marketing, mobile, and design-system templates remain native and editable, including variables, semantics, and realistic states.
Automatic backups, safe schema migration, bounded revision history, explicit restore, and corruption preservation protect the source document.
Viewport-based incremental rendering keeps large documents responsive without hiding content from search, agents, export, or audit.
Connected agents use the same typed audit and template operations through Orkestrai CLI and MCP.
83
Live collaboration for native Design documents
People, agents, Council, Floors, and Remote now review one visual source with visible authorship and explicit approval.
Presence, cursors, selections, and follow mode make live Design work visible.
Short renewable layer leases reject conflicting writes without locking the complete document.
Page and layer comments keep replies, mentions, resolution, and authorship in revision history.
Visual proposals show a live preview and structural diff, then apply atomically only after explicit approval.
Proposals can open Council or create a parallel Floor without creating a second design system.
Remote devices receive independently approved Design permissions and only sanitized page, activity, comment, and proposal summaries.
Connected agents comment, propose, and decide through the same typed, revision-safe Orkestrai MCP operations.
84
Interactive prototypes and motion in Design Studio
The native design source now carries testable behavior and reusable animation without a separate prototype file.
Multiple starting flows organize distinct journeys inside one Design document.
Click, press, hover, and timed triggers support frame navigation, overlays, history, scrolling, and variable modes.
The focused player previews transitions, hotspots, fixed layers, overflow, device framing, and fullscreen.
Sharing creates a self-contained read-only HTML prototype without exposing workspace files.
Motion tokens, per-layer tracks, keyframes, easing, CSS animations, and Motion.dev output remain revisioned and searchable.
Designers, leaders, and connected agents use the same command bus and conflict protection as manual editing.
85
Native design delivery from source to visual review
Design Studio now connects editable visual documents to real project code and reproducible implementation evidence.
HTML/Tailwind, Svelte, React/JSX, and Vue structure can be imported as native editable layers without executing project scripts or configuration.
Selected frames and groups generate reviewed Svelar/Svelte 5, React, Next.js, Vue 3, or HTML/Tailwind files while compatible Code Connect mappings reuse real project components.
Preview-before-write, optimistic document revisions, existing-file hashes, workspace path boundaries, and atomic writes protect both design and code.
Generated artifacts remain linked to the Design document, appear in universal search, and open directly in Monaco or Review Center.
Live Portals and attached iOS/Android devices can be captured at frame, mobile, tablet, or desktop sizes and compared through normalized reference, implementation, overlay, and pixel diff views.
Visual evidence creates a traceable Kanban feedback task or Git-linked review. Connected agents use the same flow through typed Orkestrai MCP tools or the bundled CLI.
86
Windows and WSL in one team
Every agent can use the environment where its tools are installed without splitting the work into separate workspaces.
The workspace defines the default runtime, and each terminal can inherit it, use native Windows, or select a specific WSL distribution and Linux path.
Provider detection and models, PTY, resume, Council, recruitment, and the bridge follow each terminal's effective runtime, including mixed teams.
Changing environments restarts only the affected terminal and validates the distribution, path, and CLI without silent fallback.
The installed desktop app now ships the WSL runtime modules required to create, restore, and run these workspaces.
When creating or editing a WSL workspace, the Windows-visible folder is derived and locked automatically from the Linux path.
87
Workbench, verified decisions, and secure collaboration
Orkestrai becomes a complete operating surface for focused work, review, automation, mobile validation, and encrypted collaboration.
Workbench opens live workspace artifacts in up to eight persistent resizable panes with vertical or horizontal tabs.
A native file explorer, Monaco editing, offline previews, universal search, and provider usage make focused work complete.
Control Center tracks real agent state and message delivery; Review Center keeps Git diffs, evidence, risks, and decisions traceable.
Council compares independent provider perspectives while the human selects the final plan and any isolated floor to land.
Mobile Device controls iOS Simulator and Android emulators or approved physical devices from Canvas and Workbench.
Workspace sharing creates separate browser/mobile and installed-app invitations through the opaque production relay.
Operators hold structured conversations whose reply is bound to the exact prompt and provider session; the overview preserves the leader history and waits for the real end of turns with intermediate messages and tool use. Administrators can also start or restore an offline agent.
The Remote microphone transcribes messages to the leader, selected agent, or remote terminal through the host local STT, with encrypted audio and no relay processing; terminal text waits for confirmation before execution.
Raw terminal control is a separate per-device Administrator permission, disabled by default, limited to one responsive terminal, rate-limited, end-to-end encrypted, and audited.
Opening the terminal closes the agent conversation before taking over the screen.
The sharing dialog keeps stable dimensions across Invitation, Access, and Audit and clearly marks the active tab.
The canvas map stays within the phone viewport instead of creating page-level horizontal scrolling.
Primary buttons keep readable white text over violet in the dark theme.
Browser and mobile device approvals reach the host through the production relay with the official Remote PWA origin enabled.
The website demonstrates these capabilities with interactive Motion-powered showcases and reduced-motion support.
88
Terminals mode and ten complete themes
A focused view organizes every workspace artifact, while terminals gain more choice without crowding their controls.
The new Terminals mode provides a searchable explorer across every workspace and opens terminals, boards, notes, portals, files, flows, and usage in the full work area.
Returning to the Canvas preserves the workspace and centers the selected item.
Existing PTY sessions resize to the full focused area instead of rendering chats compressed or garbled at their smaller Canvas dimensions.
The voice orb finds and opens the active workspace leader in Terminals mode as well.
Provider, role, theme, reload, Maestro Mode, and removal stay organized in a compact header menu.
Terminals now offer ten complete ANSI palettes with visual named selection.
The locate-on-canvas action uses one centered icon, while Settings explains the standalone Fn/Globe limitation on macOS.
89
Reliable Kimi roles and visible voice controls
Kimi role profiles start correctly, while the voice orb makes its position behavior discoverable.
Orkestrai adds the required frontmatter and automatically repairs legacy or missing Kimi role files, preventing invalid-agent PTY failures.
The voice orb has a clickable pinned or movable badge that opens position controls directly, while its tooltip also exposes the platform-specific shortcut.
90
Native roles and precise quota windows
Preset instructions now enter at the correct provider level, while Usage routes against the exact period you choose.
Claude receives preset roles as a system prompt, Codex as developer instructions, and Kimi through its agent file; other providers receive a short role-file reference instead of long pasted text.
Quota routing can monitor the 5-hour, weekly, or monthly window and explains when a provider does not report the selected period.
Kimi remaining quota and Codex additional limits are normalized into one bar per reported period.
shadcn switches display their real state, the workspace editor is responsive and scrollable, and the Windows title-bar divider spans the full width.
The website now identifies Orkestrai as Apache 2.0 open source and links source code and downloads to the primary repository.
91
Dictation that can send when you are ready
Terminal speech can now submit automatically, while the model picker remains searchable and correctly framed.
A new opt-in setting appends Enter after terminal dictation.
Regular fields such as tasks, roles, notes, and forms remain insert-only.
The searchable provider model selector now follows the shadcn-svelte combobox composition and no longer clips or misaligns its search field.
92
Agent conversations that can be proven
Cross-provider messages now bind to the correct workspace, fail clearly without confirmation, and hand completed tasks back to the leader.
Claude and Codex were validated talking in both directions with real replies from their provider transcripts.
Codex uses the real workspace directory and Kimi uses its exact path hash, preventing concurrent projects from crossing conversations.
Unquoted multi-word ask messages are preserved, and timeouts or unconfirmed replies fail explicitly.
task done automatically hands completion back to the leader without overwriting an unfinished human draft.
93
A sharper canvas on every platform
Voice, recovery, model selection, Windows integration, and the light theme now behave consistently in daily work.
The voice orb targets the focused field on the first click, can be pinned or dragged, and moves away from open panels.
Large provider model catalogs are searchable and scrollable, including Devin.
Orkestrai Light keeps accessible contrast across nodes, panels, text, icons, provider marks, buttons, and hover states.
Canvas organization aligns selected nodes or the whole workspace and keeps connections behind nodes.
Usage severity colors match between the panel and node, while Skills loads initial results automatically.
Workspace recovery validates real provider conversations and avoids stale session errors or unnecessary role reinjection.
Windows gains the correct CLI launcher, DPI-aware terminal selection, and a styled desktop title bar and menu.
94
Quota-aware teams and personal themes
Provider limits become part of the canvas while the whole application gains configurable light and dark themes.
Usage can be added as a persistent canvas node for Claude, Codex, and Kimi.
The node configures source, fallback, and threshold, and the leader reads the same recommendation through CLI or MCP before assigning new work.
Settings adds three dark themes, one light theme, and a semantic token editor with live preview.
Custom themes support duplication and validated JSON import/export and persist across restarts.
Canvas nodes and the main operational pages now honor global theme tokens.
95
Your agents, without toolbar clutter
One Agents menu keeps every provider available while favorite services remain one click away.
The eight provider buttons are consolidated into one responsive Agents menu while Shell stays directly accessible.
Up to four ready agents can be pinned beside the menu in the user’s preferred order.
Pinned preferences persist globally across workspaces and app restarts.
Unavailable agents remain discoverable through Provider Center and never occupy toolbar space.
96
Devin joins the local team
The official Devin CLI now works as a native canvas agent with models, bridge tools, and exact resume.
Provider Center detects the official Devin CLI and loads the models available to the user’s account.
Devin terminals start with autonomous workspace access and support both interactive and headless work.
The Orkestrai skill and MCP bridge are provisioned in Devin’s native project directories.
Local ATIF transcripts provide clean agent replies and spoken responses, while concurrent sessions are matched by workspace.
Cursor starts with workspace trust and MCP approval; Antigravity starts autonomously with selectable effort.
97
Long dictation and reliable workspace bridges
Voice uploads, restored Portals, and concurrent Claude conversations keep their correct runtime state.
Local dictation accepts recordings of approximately 15 minutes instead of failing at the adapter’s former 512 KB request limit.
Saved Portals retry until their local dev server returns and wait for a real page before automation.
Concurrent Claude agents reserve distinct conversation ids so clean replies always come from the intended transcript.
Portal failures retain their actionable detail and provider replies never expose raw terminal redraws.
Workspace resume no longer injects roles again and submits continuation only to agents with unfinished work.
The server stays responsive while macOS waits for workspace-folder permission.
98
Terminal sessions recover after restart
Process-local PTY ids no longer prevent saved provider conversations from reopening.
Obsolete PTY ids are removed when a workspace loads while each provider conversation id remains preserved.
Recovery uses a stable protocol code and persists the replacement session before reattaching.
Team context stays visible while voice and providers become usable wherever the work happens.
Dictation inserts text into the active task, role, note, or form field and falls back to the canvas leader only when no field is active.
Agent providers can be switched in place without losing roles, floors, layout, or connections.
Preset roles now carry complete operating protocols and reach agents automatically when their PTY starts.
The leader receives every unassigned initial task with its complete description, images, and linked note before delegation.
Native notifications distinguish task completion, project completion, and attention.
Floors shows each task title, stage, and assignee; shape editing now matches large rendered typography.
100
Seven providers, one team
Cursor, Antigravity, and Cline join the canvas with the same native integration as existing agents.
The toolbar detects Claude, Codex, Kimi, OpenCode, Cursor, Antigravity, and Cline.
Models, effort, resume, and permissions follow each CLI's actual capabilities.
Skills and MCPs are provisioned in each provider's native format without erasing existing configuration.
Documentation guides agent selection by desired outcome without requiring terminal expertise.
Provider Center adds local CLI detection, OS-aware installation, official sign-in guidance, and capability details.
New installations start in English and save the user's language as the first onboarding step.
Startup now waits for the saved locale before enabling interaction, preventing mixed-language screens and lost clicks.
Preset terminals start with each provider's autonomous full-access flags, and existing empty-argument terminals are repaired without overwriting custom commands.
101
Workflows for multidisciplinary teams
Custom Kanban stages and complete presets expand Orkestrai to marketing, design, content, and product work.
Boards support up to ten stages with custom names, colors, and order.
New ready-made teams cover campaigns and launches, brand and design, and content and SEO.
The Orkestrai Contributing preset coordinates a leader, two independent oracles, and specialist agents.
The bridge keeps the whole team aware of custom columns and the selected process.
102
Presets, roles, and clearer operations
Ready-made teams become easy to find, duplicate, and adapt to real work.
Preset library with search, filters, and install flows for new or existing workspaces.
Twelve complete roles, portable skills, and ready-made teams for multiple stacks.
The Floors overview now shows agents, tasks, changed files, and branch state.
Native menus plus refreshed Settings and Documentation improve navigation and consistency.
103
Reliable messages and quiet notifications
Communication between people and agents no longer mixes drafts or silently cuts content.
Human drafts remain intact while automated messages wait in a per-session queue.
Inter-agent messages are no longer silently truncated at 4,000 characters.
Normal silence and successful reloads no longer produce false attention alerts.
Brazilian Portuguese accents were corrected across the main interface.
104
Signed and notarized for macOS
The first trusted macOS release removes the Gatekeeper block and prepares automatic updates.
Apple Silicon and Intel builds are signed with Developer ID Application and notarized by Apple.
CI validates the signing authority, Team ID, Hardened Runtime, Gatekeeper acceptance, and stapled ticket before publication.
The release remains available from the public legacy feed while the source repository transition is completed.
Mac users on 0.1.3 or earlier need one manual installation of 0.1.4; subsequent signed releases support trusted in-place updates.
105
Valid macOS package and protected updates
An emergency fix for Gatekeeper and unsafe replacement by builds without Apple signing.
The partial ad-hoc signature in 0.1.2 was replaced with a complete signature with sealed resources.
CI now verifies deep signatures, DMGs, and ZIPs for Apple Silicon and Intel before publication.
Legacy Mac updaters are blocked; the new app detects releases and guides manual installation without removing the current app.
On the first launch without a Developer ID, try to open the app, dismiss the warning, then use System Settings → Privacy & Security → Security → Open Anyway; authenticate and confirm Open. Windows was not affected.
106
Lighter Usage polling for long sessions
A maintenance fix that protects provider rate limits.
Automatic Usage refresh moved from 60 seconds to 5 minutes.
The server cache uses the same interval to avoid duplicate calls when reopening the panel or returning to the app.
The manual refresh button still fetches fresh data immediately.
Fewer requests reduce the risk of Claude returning HTTP 429 during long sessions.
107
Updates restored and tasks with complete context
A maintenance release for the desktop package and Kanban delegation.
The updater module is now included in the installed application, fixing the incorrect message that automatic updates were unavailable in the package.
Versions 0.0.1 and 0.1.0 require one manual installation of 0.1.1; later updates return to the normal flow on supported platforms.
Manual tasks reach the leader only after persisting the title, markdown description, and every image.
The leader and assigned agent receive the complete briefing in their terminals without losing descriptions or visual references.
108
Multilingual voice, Ports, and leader dictation
A release focused on local operation, speed, and workspace control.
Supertonic 3 INT8 replaced Kokoro for spoken replies, with pt-BR, en-US, and Latin Spanish presets, preview, and 0.75x to 1.50x speed.
The new Ports panel lists listeners tied to local Portals and stops them with confirmation and PID revalidation.
A global voice orb now triggers the leader microphone flow and sends transcription to the exact nodeId.
Codex MCP now uses absolute paths on Windows and starts correctly even outside a workspace.
Fixes for Claude session resume, deleted voice model downloads, and the Cmd/Ctrl+K dialog on large displays.
109
A more expressive canvas and reliable automation
Kanban, markdown, images, workflows, and the agent bridge received a broad revision.
Task composer with markdown descriptions and images before creation.
New Image node for mockups, screenshots, and diagrams connected to the team.
Workflows now start sessions automatically, surface failures, and feed other workflows.
MCP bridge audited for Claude, Codex, Kimi, and OpenCode with contracts and message sanitization.
Complete application coverage in Portuguese, English, and Spanish.
110
Global search, onboarding, and team context
Feature discovery and continuity between agents became clearer.
Cmd/Ctrl+K documentation search from any screen.
Use-case-driven onboarding with actions performed directly on the canvas.
Team presets now carry agents, roles, notes, routines, and MCPs.
Improved tooltips, keyboard navigation, and empty states.
