What’s new in PandaCode.

Connect without trusting each device, and a new mobile app look

## New

• No more trusting each device. Computers and phones signed in to the same GitHub account can now control each other while Allow remote control is on. The Trusted switch is gone from Settings → Remote access. Everything still travels end-to-end encrypted.
• Project icons on your phone. Computers now send each project's icon to the PandaCode mobile app, so the inbox shows the same icons as the desktop sidebar.
• Mobile app (mobile/, built from source):
  • The system tab bar with Liquid Glass. On the iPhone Duo there is a side rail and a split view with the inbox beside the open task.
  • The desktop's daily wallpapers, with visibility and blur settings.
  • An app-only Light, Dark or System appearance.
  • Messages sent while offline go out on reconnect.

## Fixes

• A computer that gets too many requests from one device at once now says it's busy right away, instead of making the device wait for a timeout.
• On the phone, threads that came up empty now load, and task lists and run states stay in step with the computer.

## Known limitations

• The first time you open a long thread on the phone, it can take a few seconds to load.
• The mobile app isn't in the App Store or Play Store yet. Build it from mobile/ (see mobile/README.md).

PandaCode on your phone

## New

• A mobile app for iOS and Android (mobile/, built with Expo). Sign in with the same GitHub account as PandaCode Connect, and see and control the tasks running on your computers from your phone. Tasks never run on the phone.
  • Task lists for each computer, grouped by project, with live status and search.
  • Live transcripts with Markdown, folded tool calls, and approval and question cards.
  • Send, queue, steer and stop, and change the model, effort and permissions.
  • Start new tasks in any project on a computer.
  • Native Liquid Glass on iOS 26, with solid fallbacks on older iOS and a Material look on Android.
• Desktop support for phones. A computer with remote control on now lists its tasks and projects for trusted devices. It also starts tasks in a project that isn't on GitHub, and still shows an approval or question that was already waiting when the phone opened the task.

## Known limitations

• The app isn't in the App Store or Play Store yet. Build it from mobile/ with your relay URL and GitHub client ID (see mobile/README.md).
• Each computer must trust the phone in Settings → Remote access, and the phone must trust the computer.
• Images can't be attached from the phone yet.

PandaCode on your phone

## New

• A mobile app for iOS and Android (mobile/, built with Expo). Sign in with the same GitHub account as PandaCode Connect, and see and control the tasks running on your computers from your phone. Tasks never run on the phone.
  • Task lists for each computer, grouped by project, with live status and search.
  • Live transcripts with Markdown, folded tool calls, and approval and question cards.
  • Send, queue, steer and stop, and change the model, effort and permissions.
  • Start new tasks in any project on a computer.
  • Native Liquid Glass on iOS 26, with solid fallbacks on older iOS and a Material look on Android.
• Desktop support for phones. A computer with remote control on now lists its tasks and projects for trusted devices. It also starts tasks in a project that isn't on GitHub, and still shows an approval or question that was already waiting when the phone opened the task.

## Known limitations

• The app isn't in the App Store or Play Store yet. Build it from mobile/ with your relay URL and GitHub client ID (see mobile/README.md).
• Each computer must trust the phone in Settings → Remote access, and the phone must trust the computer.
• Images can't be attached from the phone yet.

A better menu bar

## New

• A glass panel under the panda. Clicking PandaCode's menu bar icon now opens a slim panel instead of a plain menu: running tasks with a spinner, tasks that need your approval or answer, and finished tasks with a check or cross, each with its agent and project. Click a row to open the task; arrow keys and Enter work too.
• Plan usage with bars. Each agent's plan limits show as thin bars with reset times, turning orange from 80% and red from 95%. Use the refresh button to check again; nothing is checked in the background.
• Right-click the icon for Open PandaCode and Quit.

## Known limitations

• The finished list is kept in memory and starts empty after PandaCode restarts.

No Keychain prompts

## Fixes

• PandaCode no longer asks for your Keychain password after updates. macOS tied every key PandaCode saved (bot channels, cloud sandboxes, provider API keys, connections) to the exact build that saved it, so each update asked for your login password again. On macOS these keys now live in one private file in PandaCode's data folder that only your user account can read, the same way remote access has since 0.3.29.
• Your saved keys come along. Each key still in the Keychain is moved over the first time PandaCode needs it, then removed from the Keychain. macOS may ask one last time per key while it moves; after that it never asks again.

## Known limitations

• Windows and Linux are unchanged: they keep using Credential Manager and Secret Service, which never had this problem.

Menu bar

## New

• PandaCode in the menu bar. A panda icon shows how many agents are running. Its menu lists running tasks, tasks that need your approval or answer, and tasks that finished since you last looked; click one to open it.
• Plan usage at a glance. The same menu shows each signed-in agent's plan limits, such as Claude's 5-hour and weekly windows or Codex's limits. They are checked only when you open the menu or pick Refresh usage, never in the background.

## Known limitations

• The "finished, not viewed" list is kept in memory and starts empty after PandaCode restarts.
• Each section shows up to 10 tasks.

Panda bot icons and simpler bot settings

## New

• Panda icons for bots. Pick from 10 panda icons in different colors, or upload your own image (PNG, JPEG, WebP or GIF up to 5 MB, cropped to a square). The icon shows in the sidebar, the conversation header, the welcome screen and the web chat.
• Bot settings look like Settings. The bot settings window now has the same layout as PandaCode's Settings: sections down the left (Profile, Channels, Schedules, Memory, Activity) and grouped rows on the right. Conversations, rules and the activity log now live together under Activity.

## Known limitations

• The web chat page shows the bot's first letter instead of a panda icon. Uploaded images do show there.
• Bots that use an emoji icon keep it. Pick a panda or upload an image in the bot's Profile to change it.

Cloud agents

## New

• Run an agent in the cloud. In a New task, open the Run on menu (beside Local / Worktree) and pick a connected cloud provider. The agent runs on a machine in your own account: Daytona, Cloudflare, AWS EC2, Google Cloud, E2B, Modal, Vercel Sandbox, CodeSandbox, Runloop, Blaxel, Tensorlake, Freestyle or boat.
• Set up for you. PandaCode creates the machine, installs Git, Node and the agent's CLI (Claude Code, GLM, MiMo, Codex or OpenCode), copies your login before every run, and uploads your checkout, uncommitted changes included.
• Works like any task. Streaming, approvals, questions, steering, the queue and the model picker all work as they do locally. A status strip shows each stage (Installing, Running, Reconnecting, Syncing) with Sync, Retry, Open, Stop and Delete machine.
• Close the laptop. The run keeps going on the machine, and quitting PandaCode doesn't stop it. When you come back, the task catches up and finishes normally.
• Changes come home. After each run the work lands in a local worktree on pandacode/cloud-<id>, so the Changes tab, pull requests and Run work as usual. Idle machines stop on their own after about 15 minutes; the next prompt starts them again or recreates them from your local copy.
• Settings → Cloud has connect forms for every new provider, Vercel Sandbox and Google Cloud options, and an optional long-lived Claude token (claude setup-token) for cloud agents. See cloud agents.

## Known limitations

• The provider integrations were built from each provider's API docs and tested against mocks, not live accounts. Please report anything that fails on first use.
• Railway can't run commands on a machine, so it can't host agents. CodeSandbox works only on VMs that offer its Pint connection.
• Modal sandboxes last at most 24 hours from creation; Tensorlake's unattended time depends on your plan.
• Copying your local Claude login (instead of using a claude setup-token token) may sign you out on this computer if the cloud copy refreshes it.
• On providers without a file API, copied logins travel inside the setup command and could briefly appear in that machine's process list.
• Existing Cloudflare sandbox users should redeploy the Worker from Settings → Cloud.
• Images and browser captures can't be sent to a cloud task yet.

Quieter panels and bigger Oh My Pi catalogs

## Fixes

• New tasks no longer open the right panel. If you close the panel, new tasks keep it closed until you open it again. Tasks you have opened before still remember their own panel state.
• Oh My Pi loads large model catalogs. PandaCode now asks omp to send big replies in pieces (omp's RPC protocol v2) and puts them back together, so long model lists no longer fail with "RPC response exceeded the transport limit". When a catalog still fails, the model picker shows the agent's own reason (#91).
• Remote access never overwrites a secrets file it cannot read. If PandaCode Connect's file is unreadable or damaged, saving a sign-in now fails with an error instead of replacing your token and device key. Two PandaCode windows saving at once no longer lose an update (#106).

## Known limitations

• The panel choice is remembered only while PandaCode runs; after a restart the first new task opens the panel again until you close it.
• Oh My Pi's chunked replies were tested with recorded frames, not a live omp install.

No more Keychain prompts for remote access

## Fixes

• PandaCode Connect no longer asks for your Keychain password. macOS tied each saved Connect secret to the exact build that saved it, so every update asked for your login password again on each launch. Remote access now keeps its sign-in and device key in a private file in PandaCode's data folder that only your user account can read, so no Keychain prompt appears at launch or after updates.

## Known limitations

• Sign in to remote access once more after updating (Settings → Remote access). This computer appears as a new device: mark your computers Trusted on each other again and revoke the old entry.
• Keys you save for bots, cloud sandboxes and providers still live in the Keychain.

Muse Code

## New

• Muse Code is a harness. PandaCode runs Meta's muse serve and talks to it over MSP, Muse's own session protocol. Text, reasoning and tool activity stream as Muse works. Follow-ups continue the same Muse session with PandaCode's tools still attached. Approvals and questions, including multi-select, open in PandaCode. You can steer mid-turn, stop, attach images, pick a model and reasoning effort, and run Muse skills from the / menu. Standard, Accept edits, Plan and Full access map onto Muse's approval modes and sandbox. Sign in once with muse login (#94).

## Known limitations

• Muse reviews tool calls with its own AI judge before PandaCode sees them and approves the ones it considers safe; muse serve cannot turn this off, so Standard and Accept edits only ask about calls it escalates or that leave the sandbox.
• Muse has no plan mode. PandaCode's Plan runs it read-only (no file writes or shell) with a planning instruction.
• Agent usage shows no usage limits for Muse, and subagent transcripts, goals, workflows and /compact are not available for Muse yet.

Cline, and steering that never stops the run

## New

• Cline is a harness. PandaCode runs cline --acp and lists Cline's models from your account without sending a prompt. Reasoning effort, Plan, Accept edits and Full access all work, and follow-ups continue the same Cline session. Sign in once with cline auth (or set CLINE_API_KEY) (#93).
• Steer without stopping. Every harness now takes a steer mid-run: Claude Code and Codex steer the live turn, OpenCode takes a second prompt, and the other ACP harnesses send it the moment the current turn ends, in the same run. A Steer button sits beside Stop (⌘/Ctrl↩ still works), the pending message shows whether it is steering now or sends at turn end, and a failed steer returns to the queue (#92).

## Known limitations

• Cline 3.0.73 accepts only text in prompts, so attached images are passed as file paths for Cline to open. It does not use PandaCode's tools (automatic titles, Run/preview, team tools), and it has no slash commands, context usage or usage limits in PandaCode.
• A steer that is saved, later fails and is resent can appear twice in the history.

Current GLM Coding Plan models

## Fixes

• The GLM Coding Plan picker lists the models Z.AI actually serves. The new default is GLM-5.3, alongside GLM-5.3-Flash, 1M-context versions of both, and GLM-5-Turbo. The old IDs (glm-4.7, glm-5.1) were being quietly routed to newer models on Z.AI's side. Claude Code's fast background model is now GLM-5.3-Flash. Thanks @DreDay4u (#86).

## Known limitations

• Tasks that saved an old GLM model ID keep it. Z.AI routes it to a current model. Pick GLM-5.3 in the model picker to make the choice explicit.

Windows on ARM and app-wide lead guidance

## New

• Windows on ARM. A native PandaCode-Windows-arm64-setup.exe for Snapdragon and other ARM64 Windows PCs, with automatic updates. npx pandacode installs it.
• One lead guidance for every team. Settings → Agent teams has a Lead guidance field that applies to every team. A team's own guidance still replaces it for that team, and resetting a team's guidance returns it to the app-wide one. Thanks @ktappdev (#79).

## Fixes

• The agent team editor is readable again. The Filter models field is full height, and model names stay visible beside their effort dropdowns. Thanks @ktappdev (#84).

## Known limitations

• The Windows ARM64 build is new and hasn't been tried on a real ARM64 PC yet.
• After app-wide lead guidance is set, an older PandaCode version can't read the saved team library. Clear the guidance before downgrading.

Chats, connections, a design canvas and a skills library

## New

• Chats without a project. New chat starts a conversation that isn't tied to a codebase. Each chat gets its own folder under ~/PandaCode Chats, chats are listed in their own Chats section, and project-only controls are hidden. Move to project… moves a chat into a project before it gets going.
• Named connections for each agent. Add "Claude Work" and "Claude Personal", or two Pi profiles. Each connection has its own executable, arguments, environment and config folder, and secret values are kept in the system keychain. Choose a connection per task in the model picker, and tasks keep theirs.
• Design canvas. Agents can create HTML artboards. You see them in the inspector's Design tab, pan and zoom around them, click an element to comment on it, and send the comments back to the agent. Artboards are files in .pandacode/design, so they show up in diffs and reviews.
• Skills & rules library. Import a skill or rule from a folder or a GitHub repository, then turn it on for each agent, either globally or for a project. The library shows which agent has what, spots edited or unmanaged copies, and scans for secrets or approval-bypass instructions before anything is turned on. It never overwrites files it didn't write.
• Jujutsu. Jujutsu (jj) repositories get their own Changes panel: change ID, description, bookmarks, changed files, diffs, Describe, New, Commit and the log. Repositories that also use Git show a jj chip.
• Capabilities & cache. Settings → Tools & add-ons lists each agent's commands, MCP servers, hooks and skills side by side, without exposing secrets. It also shows how PandaCode's instructions are cached for each agent, with the cache tokens the agent reported.
• Workflow recipes. Save multi-step routines as recipes, such as review, check, summarise and wait for approval. A dry run shows what each agent can do natively and what needs a fallback. Running a recipe queues its steps as ordinary messages. Open recipes from the + menu.
• GitHub Copilot CLI as an agent. Thanks @238SAMIxD (#73).
• Workspace search. The inspector has an activity bar and a Search tab with find-in-files results. Thanks @SakithaSamarathunga33 (#72).
• Project checks after turns (off by default). When an agent finishes cleanly, PandaCode runs the project's own test or check command. If it fails, one editable fix message is queued; it never loops. Thanks @matkarimov099 (#68).
• Code search for agents. If codedb is installed, agents get a fast, read-only code search tool, with codedb's telemetry and background daemon turned off. Thanks @matkarimov099 (#65).
• New task from a project heading. Hover over or focus a project in the sidebar to start a task there. Thanks @jojomensah89 (#76).
• Intel Macs. A native PandaCode-macOS-x64.dmg is published with automatic updates, and npx pandacode installs it.

## GitHub view

• Keyboard navigation works everywhere. The arrow keys move through workflow runs as well as pull requests and issues, and Enter opens the focused row.
• Run details go deeper. Expanding a run shows each job's steps with durations, the job's summary and earlier attempts.
• Check workflow files asks an agent to find invalid YAML, unknown runner labels and broken uses: references.

## Fixes

• Node from fnm, nvm, mise, asdf or nvm-windows is found, so agents installed with npm under a version manager work when PandaCode starts from the Dock.
• Pi installed with the pi.dev installer works on Windows.
• The project picker no longer closes as it opens right after a new task is created.
• A board task you open while PandaCode is still starting stays open.

## Known limitations

• Jujutsu support, codedb search and Copilot have been tested against recorded output only, not live installs.
• The Intel macOS build is new and hasn't been tried on a real Intel Mac yet.
• The design canvas runs no scripts inside artboards, so interactive prototypes appear static.

Readable agent traces

## Tool activity

• Tool calls read as sentences. Instead of raw tool names, each call says what it did: "Ran rg -n createSearchIndex src", "Read search-index.ts", "Searched src for toLowerCase", "Edited search-index.ts +4 −3", "Fetched developer.mozilla.org", "Updated the plan · 3 steps". MCP tools name their server and tool.
• Groups summarise themselves. A group of calls is one line, such as "Ran 1 command and 2 searches, read 1 file and edited 1 file", with an icon for each kind. While it runs it reads "Reading 2 files…". Failed calls are counted separately and marked in red.
• Expanded rows show the work. Opening a row shows command output, file contents, matched files or a red and green diff. The raw input and output are still there, one click further in.
• Everywhere you see tools. The same traces appear in conversations, the board, the quick launcher, subagents and team workers.

## Fixes

• Raw output stays open while a subagent streams. Updating a subagent's tool row no longer closes the raw output you were reading.

## Known limitations

• The new traces have been tested in Chrome but not yet in WebKit, the engine the macOS app uses.

Stay signed in, pick models that exist

## Remote access

• PandaCode Connect stays signed in across updates. Updating no longer asks you to sign in again. If 0.3.20 or 0.3.21 signed you out, 0.3.22 picks up your earlier sign-in by itself, as the same computer, so the computers you trusted stay trusted. A brief Keychain hiccup right after an update is retried quietly instead of signing you out, and your sign-in renews itself before it expires.
• The model picker shows what the other computer has. For a task that runs on another computer, the picker lists only the providers and models installed there. If your current model isn't on the computer you pick, PandaCode switches to one that is and tells you; picking This computer again brings your choice back.

## Known limitations

• If the earlier sign-in can't be found (for example it was removed from Keychain), sign in once more in Settings → Remote access.
• Remote access is covered by automated tests; it has not been tried between two real computers yet.

Oh My Pi, Command Code and a better GitHub view

## New agents

• Command Code. Run Command Code as an agent through its ACP server. Its models, effort levels and slash commands appear in PandaCode, and Standard, Accept edits, Plan and Full access use Command Code's own permission modes. You can save an API key in the system keychain, or sign in with command-code login. Thanks @ktappdev (#56).
• Oh My Pi (omp). Run omp, the Pi fork, as an agent. It uses your existing omp login and setup.

## Teams

• Lead guidance. Team setup can carry your own rules for the team lead, for example "only coordinate, never edit files". They're added after the built-in rules. Reset to default clears them. Thanks @ktappdev (#61).
• Effort for each role. In Settings → Agent teams you can set the lead's effort separately from the workers'. For example, the lead can run at High while the workers run at Medium.

## GitHub view

• Actions make sense next to pull requests. Each run shows its event, its pull request, how many attempts it took and how long it ran. You can filter runs by event or pull request, or keep only the latest run per pull request. Each pull request has a Runs link to its runs.
• Run details. Expanding a run shows its jobs with their durations, and its artifacts with their sizes.
• Workflows that never started. A run GitHub couldn't start, for example because of an invalid workflow file or a missing runner, now reads "Didn't start". Its Fix action asks the agent to check the workflow file.
• Typing is no longer erased. Text you type in the GitHub view now stays when the view refreshes.

## Everywhere else

• Remove a project. Settings → General → Projects lists your projects and can remove one, including one whose folder was deleted. The project's tasks are removed from PandaCode; the folder stays on disk.
• Drag the window by its header again on macOS.
• Links open in your browser. Links in Markdown, provider cards and the quick launcher now open in your default browser. Thanks @jojomensah89 (#40).
• File icons. The Files tab, the @ menu and changed-file lists show Material Icon Theme icons. They load the first time they're needed, not at startup. Thanks @SakithaSamarathunga33 (#49).
• Task timers and defaults. Working tasks show how long they've been running. Settings has a default model for new tasks and an Expand tool activity option for past turns; a running turn's activity always stays open. Thanks @ExcuseMeBro (#37).
• Linux arm64. A native PandaCode-Linux-aarch64.AppImage is published with automatic updates, and npx pandacode installs it.
• Android emulators on Windows. The Simulators tab lists Android devices on Windows. Without an SDK, it explains how to install one.

## Fixes

• Agents work again from the Linux AppImage. The AppImage's own library and Python paths no longer reach agents, terminals or git. These had broken /usr/bin/env ("version LIBSYSTEMD_254' not found"), curl and python3` in agent shells.
• Grok Build is found in ~/.grok/bin, where its official installer puts it.
• A warm Claude agent that isn't reused is now logged with the reason, to help track down slow follow-up turns.

## Known limitations

• The Linux arm64 build and the Windows Android support have not been tested on real machines yet.
• Oh My Pi has been tested against recorded messages only, not a live install. Run omp with Bun, as its installer sets up.
• Command Code can't be steered mid-turn: messages you send during a turn are queued.

Run a task on another computer

## Remote access

• Choose the computer beside Local and Worktree. Once you're signed in to PandaCode Connect, the bar above the prompt box lists *This computer* and your trusted, online computers. Pick one and send: the task runs there, and everything else works exactly like a local task: the same sidebar row, transcript, streaming, queue, Stop, approvals, model and permissions. The separate Computers section and remote view are gone.
• The other computer gets the project for you. It opens the same GitHub repository if it already has it, or clones it into ~/PandaCode, then uses the same branch. With Worktree it works in a fresh worktree, and it never switches a checkout that has uncommitted changes or a running agent.
• You're warned when GitHub is behind. If this computer has unpushed commits, uncommitted changes or an unpushed branch, PandaCode tells you the other computer will start from GitHub's version and asks before continuing.
• Fast and light. The other computer pushes its agent's output about ten times a second at most, and messages are copied once per run, so nothing polls.

## Fixes

• Fixed "Could not read this computer's PandaCode Connect key from Keychain (-25293)". It happened when two PandaCode builds on one Mac shared a keychain entry. Each install now has its own entry; sign in to remote access once more after updating.

## Known limitations

• Images and browser captures can't be attached to a task that runs on another computer yet.
• Changes, diffs and commits for a remote task stay on the computer that runs it.
• Remote access has been tested with automated tests only, not yet between two real computers.

Control your other computers

## Remote access

• PandaCode Connect lets one PandaCode drive another. Sign in with GitHub in Settings → Remote access on each computer. Turn on Allow remote control of this computer on the one you want to drive, and mark the computers Trusted on each other after comparing their fingerprints. Your other computers then appear under Computers in the sidebar. From there you can read their tasks as they stream, send messages, start new tasks, stop runs and answer approvals.
• It's off until you sign in. Nothing is sent anywhere before that. After sign-in, PandaCode keeps one connection to the PandaCode relay, which runs on Cloudflare. Messages between your computers are encrypted end to end (X25519 and XChaCha20-Poly1305), so the relay forwards only ciphertext and stores no task content. Remote prompts follow the approval settings of the computer they run on.
• It's free, with up to 10 computers per account.

## Known limitations

• Remote access was tested with automated tests only; it has not been tried between two real computers yet.
• A computer must stay awake and online to be controlled. One that sleeps shows as offline within about 75 seconds.
• Very long transcripts load in pages, and a single message over 120 KB is shortened in the remote view.

Install with one command

## Installing

• npx pandacode installs PandaCode. It downloads the latest release from GitHub, checks it against PandaCode's update signing key before installing anything, and installs the right build: macOS Apple silicon into Applications, Windows x64 with a silent install, Linux x64 as an AppImage in ~/.local/bin with a menu entry. npx pandacode --dry-run downloads and verifies without installing.

## macOS

• PandaCode signs itself on your Mac. The app is signed with a certificate created once on your machine and kept in its own keychain, without any prompt. It happens on install, at startup and after every automatic update, so keychain access and privacy permissions now survive updates instead of being asked for again.

## Known limitations

• The first local signing changes the app's identity once, so macOS may ask one more time for keychain access to secrets PandaCode saved before.
• Local signing does not satisfy Gatekeeper; installs through npx avoid the first-launch warning because the download is not quarantined.
• The Windows and Linux npx installs have not been tried on real machines yet.

A smoother, more reliable iPhone Duo

## Simulators

• Folding no longer freezes PandaCode. In 0.3.16 the loading spinner on the Duo's display redrew the whole screen image on every frame and could stall the window until the new screen arrived. It now costs almost nothing, and the 3D view stops drawing when nothing moves.
• Folds happen first and always land. The 3D Duo folds the moment you click; the simulator follows in the background. Clicking several poses quickly ends on the last one, and only that one is sent. Closing is reliable again: the hinge sweeps for 0.2 s instead of jumping, waits for the previous switch to settle, and recovers by way of a half-open position if the simulator misses a fold. Ten open/close cycles in a row each finished in about 1.2 s.
• Agents tap the right thing on the inner display. With the Duo open, finding a button by its label returned the cover display's position, so taps landed in the wrong place. PandaCode now reads element positions on the inner display from the simulator itself, and agents can find app content there that was missing before.

## Known limitations

• Describing the inner display takes 2–4 seconds longer than the cover display.
• If the simulator's SpringBoard gets stuck mid-fold, folds stop switching displays until the simulator restarts.
• iPhone Duo needs Xcode 27.1 or later on macOS.

A shorter OpenCode picker and a faster iPhone Duo

## Harnesses

• OpenCode's picker stays short. With no models chosen in Settings → Agents → OpenCode, the model picker shows Default and five models: the one the task uses, then the ones you used most recently. Type to search every OpenCode model; a model you pick from search moves to the top next time. A selection saved in Settings works as before.

## Simulators

• iPhone Duo looks like the real thing from behind. The back now has the wide camera plateau with two lenses, the mic and flash, the Apple logo, a matte titanium finish and the straight hinge edge.
• Folding is instant. Closed, Book and Open move the 3D Duo the moment you click. The display shows a spinner until its screen arrives, and a fold that fails moves back and shows the error.
• Faster screens. A fold reaches the new display in about half a second instead of about 1.5 seconds, and the inner display streams compressed frames, about three times faster and twelve times smaller than before.

## Known limitations

• iPhone Duo needs Xcode 27.1 or later on macOS.
• Timings were measured on one Mac with a booted Duo.
• Browser tests that already failed in 0.3.15 still fail; they are unrelated to these changes.

iPhone Duo, GitHub Actions and Railway deploys

## Harnesses

• OpenCode lists its models right away. With nothing chosen in Settings → Agents → OpenCode, the picker shows the whole catalog; choosing models there still narrows it. Newer OpenCode CLIs that reject models --verbose fall back to the plain list. (#31)
• pnpm installs on Windows. OpenCode and Pi installed with pnpm now launch: PandaCode follows the .cmd shim only to the entry its package declares, and never through a shell. Thanks to @Crqptx. (#30)

## Simulators

• iPhone Duo. With Xcode 27.1 or later, add an iPhone Duo from the Simulators tab. The panel draws its cover and inner displays, switches between Closed, Book and Open, and agents can fold it, tap and type on either display.
• Android emulators on Windows. The Simulators tab now appears on Windows and lists Android emulators from an installed Android SDK. AVDs load when HOME is unset, the emulator starts without a console window, and newer system images such as android-37.0 show their API level. One-click SDK setup on Windows is not part of this release. (#20)

## GitHub view

• CI at a glance. Pull requests show a Passed, Failed or Pending checks badge.
• Actions tab. Each repository lists its recent workflow runs with status, branch and age. A failed run has a Fix button that starts a task, which reads the failure with gh run view --log-failed. (#33)

## Deploys

• Railway. Connect an account or workspace token in Settings → Agents → Cloud sandboxes. PandaCode uses an installed railway or a pinned npx @railway/cli@5.71.0; the token reaches only the CLI's environment and is scrubbed from output. Only deploy-related commands run. Thanks to @RimasTam. (#22, #25)
• Google Cloud. Uses your installed gcloud and its login with a project and region you set. Commands that print or create credentials are refused, tokens are scrubbed from output, and usage reporting is off. (#22)

## Sidebar

• Completed tasks leave the sidebar. Moving a task to Completed hides it from the sidebar, pinned or not; it stays on the board and comes back when reopened. Thanks to @ExcuseMeBro. (#32)

## Fixes

• macOS window dragging. The whole empty header drags the window again, including the project name and the space between toolbar buttons. (#21)
• Blank window on Linux. The AppImage no longer bundles old Wayland libraries that stopped Mesa 25+ from starting WebKit. Thanks to @zupat. (#1, #24)
• Releases from forks link their own updater files, and versioning works on CRLF checkouts. Thanks to @RimasTam. (#27)
• A repository hook test no longer depends on your global core.hooksPath. Thanks to @webdevtodayjason. (#23)

## Known limitations

• iPhone Duo needs Xcode 27.1 or later on macOS.
• Android on Windows was not tested on a Windows machine for this release.
• The Linux AppImage repack is new; please report if the window is still blank.
• GitHub Actions status updates when the GitHub view refreshes; there are no push notifications.
• Six browser tests that already failed in 0.3.14 still fail; they are unrelated to these changes.

The prompt box, properly fixed

## Fixes

• No scroll bar in an empty prompt box, for real this time. On the latest macOS the empty prompt box still showed a sideways scroll bar after 0.3.13, which only fixed the vertical one. The prompt box never scrolls sideways now; long words and links wrap instead, and long messages still scroll up and down.

## Known limitations

• Six browser tests that already failed before this release still fail; they are unrelated to this change.

A tidier prompt box

## Fixes

• No scroll bar in an empty prompt box. On the latest macOS, the prompt box showed a scroll bar until you typed something. It now scrolls only when your message grows past the box's maximum height, and a restored draft fits as soon as the task opens.

## Known limitations

• Six browser tests that already failed before this release still fail; they are unrelated to this change.

DeepSeek, a document reader and generative UI

## Harnesses

• DeepSeek is a new harness. If you use the DeepSeek Harness (npm i -g @deepseek-ai/dsh), pick it in the model picker like any other harness. PandaCode launches dsh --profile acp, lists the models dsh offers and its reasoning efforts (Off, Low, High, Max).
• Your key, only where it is needed. An API key is optional: save one in Settings → Providers (it stays in the system keychain), set DEEPSEEK_API_KEY, or let dsh use its own stored credentials. A saved key is handed only to the dsh process.
• Permissions through dsh's sandbox. Standard and Accept edits write in the workspace and ask before going further, Plan mode is read-only, Full access never asks. Details are in [docs/approvals.md](../docs/approvals.md).
• Teams can switch models. A task with a team can change its model and reasoning effort from the composer's model picker; you no longer get "Use the team controls to switch the architect".

## Reading and writing

• A reader in the right sidebar. Markdown and text files open as a rendered page with a Rendered / Raw toggle, selectable text, Copy file, a wrap toggle, word count and reading time. Files from the Files tab, file links in messages, and files an agent opens all land here. Agents are asked to open essays, notes and docs here after writing them.
• Code files read better too. Line numbers stay put but are not copied with the code, and long lines can wrap.

## Generative UI in threads

• Agents can answer with interface, not just text. A reply can include tables you can sort, charts (bar, line, pie, area, sparkline), stat tiles, cards, callouts, tabs, accordions, steps, timelines, checklists, plans, file trees, diffs, code with a filename, terminal output, pros and cons, image galleries, forms, buttons and follow-up suggestions, arranged in stacks and grids.
• Agents can build their own views, safely. Custom views are static HTML and CSS drawn in the thread: no scripts, frames or remote content, with a few built-in behaviours (toggle, tabs, show a value, insert a prompt). Pages that need scripts are opened in PandaCode's browser instead.
• Nothing runs without you. Buttons, chips and forms only put a prompt in the composer for you to read and send.
• Code blocks have a Copy button, and collapsible sections and follow-up chips render in replies.

## Fixes

• Typing after an @ mention of a file whose name has spaces no longer reopens the file picker on every space (macOS smart quotes curled the mention's quotes).
• Clicking into a conversation no longer draws a blue focus outline around it.

## Known limitations

• Mermaid diagrams show as labelled source, not drawings.
• An agent's ``html or ``svg code example now renders as a view; its source is one click away under "Source".
• DeepSeek usage limits are not shown in PandaCode; check the DeepSeek platform.
• PandaCode's DeepSeek tests replay recorded dsh sessions; live runs depend on your dsh setup.
• Six browser tests that already failed before this release still fail; they are unrelated to these changes.

Pi joins PandaCode

## Harnesses

• Pi is a new harness. If you use the Pi coding agent, pick it in the model picker like any other harness. PandaCode uses the accounts, models and extensions you set up in Pi, lists its models by provider, and offers Pi's thinking levels. Install it with npm i -g @earendil-works/pi-coding-agent and sign in from its CLI first; Settings → Providers can point to a different Pi executable. Thanks to @ExcuseMeBro for the integration.
• Everything you expect from a task. Pi replies stream in with their tool activity and context usage, accept screenshots, and resume the same Pi session on your next message. Pi works in tasks, teams and bots.
• Steer a running reply. While Pi is working, ⌘/Ctrl Enter sends your message into the current reply through Pi's own steering.
• Pi's slash commands. Commands, skills and prompt templates from Pi and your project show up in the composer's slash menu.
• Permissions you choose. Standard asks before edits, commands and other tools that change things; Accept edits allows file edits; Plan mode allows only reading; Full access runs everything. PandaCode enforces these through a small Pi extension, and its app tools reach Pi through the same authenticated local connection other harnesses use.
• Pi keeps going when it should. A reply that Pi retries after a temporary error, or continues after compacting its context, is not cut off; you see "Retrying" and "Compacting context" while it happens.

## Known limitations

• Pi's interactive console needs Full access; use the task chat for the other permission modes.
• Pi extensions that ask you to pick from a list, type text or edit something are declined for now. Yes/no confirmations work.
• Pi does not report subscription usage, so its usage card says so.
• Pi runs were tested without paid model calls.
• Eight browser tests that already failed in 0.3.8 still fail; they are unrelated to these changes.

A quieter changed-files card

## Conversation

• The change-tracking note moved behind an ⓘ. When PandaCode could not capture every edit in a turn, the changed-files card used to add a paragraph under the file list. That note is now an ⓘ beside "N changed files" and the +/− counts; hover it to read the details.

## Known limitations

• Eight browser tests that already failed in 0.3.8 still fail; they are unrelated to this change.

The conversation stays put when you send

## Conversation

• No more jump when you send. The conversation stays at the bottom while your message and the reply stream in, through tool rows and formatting. Scroll up to read and it leaves you there; scroll back to the bottom and it follows again.

## Queued messages

• Send now. Every queued message has a Send now button. If the agent is idle or the queue is paused, it goes out right away. If an agent is running, PandaCode stops that run and sends this message first; the rest of the queue keeps its order.
• Steer is easier to find. While a Claude or Codex agent is running, a labelled Steer button sends the queued message into the running reply.

## Known limitations

• Send now is not offered in bot conversations, which send their queue on their own.
• Eight browser tests that already failed in 0.3.8 still fail; they are unrelated to these changes.

A clearer workers panel

## Agent teams

• The workers panel shows each worker at a glance. The header reads "Architect is working…" with a task count, and each worker gets its own row: a status icon with its harness logo, the worker's name, and a coloured status (Working.., Needs input, Queued, Done, Failed) followed by its model. Long lists fade out at the bottom.

## Attachments

• No toast when you paste an image. The image already shows up in the composer, so the "Image attached from the clipboard" toast is gone. Dropping files still confirms what was attached.

## For contributors

• Release builds are about 30 seconds faster: the Rust library now builds only the output the desktop app links.
• npm run package:fast builds an app with the new fast-release profile (no LTO, parallel codegen), about twice as fast clean and much faster after an edit. Shipped builds still use the release profile.
• CI caches Rust builds and no longer runs cargo check before Clippy.
• The browser suite runs 4 workers by default (PW_WORKERS overrides), about 3.5 minutes instead of 8.

## Known limitations

• Eight browser tests fail before and after this release (in ui.spec.ts, jev-assist.spec.ts and jev-live.spec.ts); they are unrelated to these changes and still need fixing.

Agent team threads read in order

## Agent teams

• Worker lines stay where they were spawned. While the architect's turn runs, its worker lines appear right after the call that started them. Everything the architect does after that, including its tool calls once the workers finish, now appears below them instead of above.
• A resumed architect turn stays below earlier workers. When the architect picks up again after its workers finish, the earlier worker lines keep their place above the new turn.

## Known limitations

• If the architect starts workers twice in one turn, the second batch appears with the first.

Safer cloud deploys and no more lost queued messages

## Queue

• Queued messages are never lost. If a queued message can't be sent (an attachment can't be read, or the task restarted), it goes back to the front of the queue and the queue waits.
• Bot conversations keep queued attachments. Images, documents and folders attached to a queued bot message now reach the agent. A failed send puts the message back instead of dropping it.

## Cloud and deploys

• The cloud CLI always asks first. pandacode_cloud_cli asks for approval on every call, even in full-access mode, because it can create or delete paid resources.
• AWS credentials stay out of the agent's reach. aws configure, the sts commands that hand out credentials, and --debug are refused. Access key IDs, secret keys and session tokens are redacted from AWS CLI output.
• Cloudflare deploys no longer hang. Large Wrangler output used to stall a deploy until its 30-minute timeout. A timeout now also stops Wrangler's child processes.

## Local VMs

• Linux desktops finish installing. The idle watcher no longer stops a Linux VM while its desktop installs.

A new panda icon and sturdier agent teams

## Agent teams

• Teams no longer stall. Claude architects can no longer wait on background commands, monitors or scheduled wakeups. These never report back after the turn ends. If an architect still ends a turn without handing out work or finishing, PandaCode asks it once to continue before pausing.
• A stuck turn is stopped after 20 minutes without output. A silent worker fails and goes to the architect. A silent architect turn shows a notice, and its workers keep running. Turns waiting on your approval or answer never count as silent.
• Failed workers are handled right away. When a worker fails while others are still running, the architect gets it at once and can retry or replace it.
• Faster and cheaper turns:
  • Team tools no longer ask for approval.
  • Reviews start warm.
  • Each conversation gets the team rules once. Later turns get a one-line reminder and only new messages.
• Fixes:
  • Asking the team a question no longer starts work you didn't ask for.
  • A reused task name runs again for a new request.
  • The assignment limit counts only the current request.
  • One team's missing folder, or one worker that can't start, no longer stops the rest.
  • Pause, Stop and restarts don't use up a worker's tries.
  • Resume works right after Pause and at the round limit (four more rounds, up to 20).
  • A failed retry no longer shows the previous attempt's result.
• You see more:
  • A spinning ring marks each working worker line.
  • "Needs input" appears on the Architect & workers button when a worker waits on you, even with the panel closed.
  • Pauses and failures you didn't ask for appear as a notice.

## Questions

• Questions from Codex render rich Markdown: links, project images, charts, SVG drawings and static HTML views. Remote images stay links, and no script runs.

## Look

• A new panda app icon.

Faster, snappier, and approvals right above the prompt

## Faster agents, fewer tokens

• Follow-ups start about twice as fast with Claude Code. After a turn, the next turn's process starts in the background, so your follow-up skips the CLI's startup (about 4 s to the first word instead of 8–9 s in our tests). At most two wait at a time, each for up to 10 minutes.
• About 650 fewer tokens on every message. PandaCode's fixed instructions (sidebar titles, the run script, change tracking) now go to Claude Code, GLM and MiMo once as a cached system prompt instead of being added to each message. Other harnesses get them on the first message and every eighth.

## An instant interface

• No transitions. Menus, dialogs, panels and settings appear at once. Loading spinners still run.
• Switching threads is one frame. A thread you return to appears formatted and scrolled to the bottom, with no loading flash or jump.
• The prompt stays put. It sits in the same place in every thread.
• ⌘/Ctrl + and − now scale the whole interface, icons included, not just text.

## Approvals and teams

• A slim approval bar above the prompt. Each request is one line with Deny and Approve; several group into one line with Approve all. Long commands no longer overflow.
• Spawned workers are one line each, in the turn that started them. Click one to open it in the Architect & workers panel, which is now solid, compact and has no border.

## Find, GitHub and Git

• Find in thread. Press ⌘/Ctrl F to search your messages and the agent's, then Enter or ⌘G to step through matches.
• GitHub threads open beside the GitHub view, like the board's side panel, so you can keep chatting without leaving it.
• Merge into main. A worktree task's branch menu merges its committed work into your main branch locally, then offers to remove the worktree. Uncommitted changes and conflicts are refused without changing anything.
• Switch branches from the branch chip, including straight back to main.
• Merge a linked pull request from the task's PR tab.
• In the GitHub view, repositories without a push in 30 days fold under Other.

## Look and feel

• Settings is a full page with grouped sections. Resource usage moved there from the status bar, which is gone.
• Sidebar:
  • Threads in All projects are one line under their project's icon.
  • With one project chosen, they're two lines, and the project's icon leads the filter.
  • The selected thread is a solid fill of your accent color.
• A Bots view in the sidebar lists each bot with its latest message.
• Solid, borderless surfaces: dropdowns, modals and the prompt box drop their glass and borders. Corners are tighter throughout.
• Bot conversations no longer show the bot's brief as "Attached context".

## Fixes

• The bot row's settings gear no longer slips out from under the pointer when pressed.
• The Team status line stays inside the conversation column.
• The moodboard list is one tidy line per board.

Agents can use gh, node and Homebrew tools

• Fixes agents not finding gh, node and other installed tools when PandaCode was opened from Finder or the Dock. Agents could push a branch but not open the pull request. PandaCode now gives every agent the same tool locations your terminal has (Homebrew, /usr/local/bin, ~/.local/bin and more).

GitHub works when opened from Finder

• Fixes the GitHub view and other GitHub features saying the GitHub CLI isn't connected when PandaCode was opened from Finder or the Dock, even though gh was installed and signed in. PandaCode now finds gh where your terminal does (Homebrew, /usr/local/bin, ~/.local/bin and more).
• Clearer connect screen. If GitHub can't be reached, the GitHub view shows the GitHub CLI's own message and what to do about it.

Automatic updates and a GitHub view

• PandaCode updates itself. It checks GitHub Releases at launch and every six hours, downloads a new version in the background and verifies its signature. It restarts to install once no agent is working and you've stepped away, or right away from Restart to update in the status bar. Settings → About shows the status, offers Check now, and can turn automatic updates off.
• GitHub view. A third view beside Conversation and Board lists your repositories. For each one: its threads, open pull requests and open issues. Review a pull request, fix its checks, address review comments or merge it when ready; solve an issue in one click or open it in a thread. Every thread is linked to its pull request or issue, and repositories that are only on GitHub can be cloned from there.

Updating from 0.3.0: install this version once from the Releases page. From then on, PandaCode updates itself.

PandaCode is open source

PandaCode is now free and open source under the PandaCode License: the MIT License, plus a requirement that forks used commercially or released publicly say "This is a fork of PandaCode" and link back. The code lives at github.com/aicodeking/pandacode.

• No license key, no account. Every feature works as soon as you install. There is nothing to activate or renew.
• No telemetry. PandaCode does not phone home. Your projects and conversations stay on your machine and go only to the coding agents and services you choose. The only automatic requests are anonymous checks for new versions of your installed coding CLIs and a daily public model price list.
• Updates on GitHub. New versions are published as GitHub Releases for macOS (Apple silicon), Windows x64 and Linux x64. Download the new build and install it over the old one; your tasks and settings are kept.
• Build it yourself. Clone the repository, run npm ci and npm run tauri dev. CONTRIBUTING.md lists the prerequisites, tests and pull request process.
• Contribute. Bug reports and ideas are welcome as GitHub issues, and security problems can be reported privately through GitHub Security Advisories.

Known limitations: release builds are not code-signed yet, so macOS asks you to choose Open Anyway in Privacy & Security the first time, and Windows SmartScreen may warn about an unknown publisher. The in-app update installer is gone; check the Releases page for new versions.

• Bot pages, redesigned. A bot's page now reads like a task: the same column width and text size, replies as plain formatted text, and your messages in the familiar bubble.
• A proper message box. Bots use the full task composer, with chips for the bot's model, permissions and check-ins that open the matching setting.
• Better settings panel. Bot settings sit in a full-height side panel with a section header for each tab, app-sized fields, tidy lists, clear empty states and a close button.

• VM screens show up. The VM and cloud machine screens now appear in the VM panel in the desktop app. Frames were being blocked by the app's security policy; they are now drawn directly.
• Linux VMs come with a desktop. Creating a VM now starts it straight away, and any Linux VM without a desktop gets a light Xfce desktop with automatic login installed in the background the first time PandaCode starts it (a few minutes). The panel shows the console and the progress meanwhile, and agents can run commands right away. "Add a desktop" only appears if that install fails, to try again.
• Clicks on the VM screen land precisely in the native app's WebKit view, including when the panel is narrow.

• Cloud sandboxes for agents. Connect providers in Settings → Agents → Cloud sandboxes, and their machines join the VM panel and the agent VM tools beside local VMs:
  • Daytona: Linux desktops agents can see, click and type into, and run commands in. They stop after 15 idle minutes.
  • Cloudflare: container sandboxes in your own account for commands. PandaCode deploys a small Worker once (needs Docker running and a Workers Paid plan), protected by a random token kept in your keychain; containers stop after 10 idle minutes.
  • AWS: Ubuntu and Windows Server desktops on EC2 in your account, using your AWS CLI sign-in. They have no open ports: commands and the screen go through Systems Manager (screens need the Session Manager plugin). Also AWS AgentCore code interpreter sessions for quick commands.
• Deploy with an agent. A new Deploy button in the header starts a task that ships the project to Vercel, Cloudflare or AWS end to end: it builds, deploys a preview, checks the live URL, then goes to production and reports what changed. Agents run the Vercel, Wrangler and AWS CLIs through one approval-gated tool that is already signed in, so tokens never reach their prompts and are scrubbed from output.
• Keys for Daytona, Cloudflare and Vercel are stored in the OS keychain; AWS uses your existing CLI login.

Known limitations: cloud machines and deploys cost money in your accounts; agents are told to ask before creating paid resources, and every create or deploy asks for approval. Windows desktops take about 10 minutes to set up on first boot. These integrations were tested against each service's published API with simulated responses; first use with real accounts may surface rough edges, so please report them.

• Virtual machines for agents (Apple silicon Macs). A new VM tab in the right sidebar runs Linux (Ubuntu, Debian, Fedora) and macOS (Tahoe, Sequoia) virtual machines on Apple's Virtualization framework.
  • Agents can use them like a person, seeing the screen, clicking, dragging, scrolling, typing and pressing shortcuts, and can run shell commands inside them for isolated work: installing packages, building and running code, starting servers.
  • You can use the same screen yourself: click, scroll and type into it from the panel.
  • "Set up VMs" downloads Tart from Cirrus Labs and checks Apple's notarization and the developer signature before installing it. Tart is free for personal and small-team use.
  • Built to save memory: VMs run without a window, Linux VMs start with 2 GB, copies share disk space with their image, and a VM nobody has used for 10 minutes is suspended (macOS) or stopped (Linux). VMs started by PandaCode are released when it quits.
  • Console-only Linux VMs can get a light Xfce desktop with one click, so agents can use them with computer use.
  • Share the current project with a VM read-only when it starts; the agent copies what it needs inside the VM.
• This release also brings the VM work to Windows and Linux builds, where the VM tab stays hidden (local VMs need an Apple silicon Mac).

Known limitations: downloads are large (Linux about 5 GB, macOS about 30 GB). Windows VMs and cloud sandboxes (Daytona, AWS, Cloudflare) and deployments are coming in the next releases. A VM started outside PandaCode must be stopped and started here before its screen can be shown.

• Virtual machines for agents (Apple silicon Macs). A new VM tab in the right sidebar runs Linux (Ubuntu, Debian, Fedora) and macOS (Tahoe, Sequoia) virtual machines on Apple's Virtualization framework.
  • Agents can use them like a person, seeing the screen, clicking, dragging, scrolling, typing and pressing shortcuts, and can run shell commands inside them for isolated work: installing packages, building and running code, starting servers.
  • You can use the same screen yourself: click, scroll and type into it from the panel.
  • "Set up VMs" downloads Tart from Cirrus Labs and checks Apple's notarization and the developer signature before installing it. Tart is free for personal and small-team use.
  • Built to save memory: VMs run without a window, Linux VMs start with 2 GB, copies share disk space with their image, and a VM nobody has used for 10 minutes is suspended (macOS) or stopped (Linux). VMs started by PandaCode are released when it quits.
  • Console-only Linux VMs can get a light Xfce desktop with one click, so agents can use them with computer use.
  • Share the current project with a VM read-only when it starts; the agent copies what it needs inside the VM.

Known limitations: downloads are large (Linux about 5 GB, macOS about 30 GB). Windows VMs and cloud sandboxes (Daytona, AWS, Cloudflare) and deployments are coming in the next releases. A VM started outside PandaCode must be stopped and started here before its screen can be shown.

• Agent teams. In Settings → Agents → Agent teams, create a named team: choose the models it may use and which one leads. Your teams appear at the top of the model picker. Pick one and the lead plans the work and runs workers using only that team's models; PandaCode refuses any other model it asks for. Picking an ordinary model switches the task back to a single agent.
• Classic team setup. Prefer choosing every worker yourself? Turn on Settings → Agents → Classic team setup and the Team button opens Architect & workers again, with team profiles, per-worker models and run limits.

• No indent while an agent works. The live "Working…" row, its tool groups and the reply streaming under it now start at the same edge as finished replies, so nothing shifts when a turn ends.

• Your messages sit on the right. Message bubbles now line up with the right edge of the conversation instead of floating near the middle on wide windows, and their text reads left to right inside the bubble.
• Tidier activity rows. "Worked for…", tool activity and changed-file rows start at the same edge as replies, without the old indent.

• iOS and Android simulators in the sidebar. Open the phone tab in the right sidebar to see a live, 3D model of the device that you can turn and zoom.
  • Click the screen to tap, drag to swipe, scroll to scroll and type to enter text. Home, Lock, Back and Recents (Android) sit in a dock below, with dark mode and screenshot copy.
  • Devices look like the real products: iPhone 17 Pro in Cosmic Orange with its camera plateau, iPhone Air, iPhone 17, earlier iPhones, iPads, and Pixel phones and tablets, each with its official finishes.
  • Agents can use them too: they see the screen as an image or a list of elements, then tap, swipe, type, press buttons, install and launch apps and open links.
  • iOS uses Xcode's simulators; touch input needs AXe (brew install cameroncooke/axe/axe). Android uses the Android SDK emulator; if it is missing, "Set up Android emulator…" in the device menu installs it with Homebrew (about 2 GB).
• Pull requests in the sidebar. When an agent opens a pull request in a thread, it is linked automatically. The PR tab shows its state, checks, reviews, description, files with diffs and commits. "Link pull request…" in the branch menu attaches an existing one.
• Token saver. The Saver button in the composer turns on per-thread packs that help agents spend fewer tokens: Caveman (Lite, Full, Ultra), Lean code, Quiet tools and RTK compression. Turning on RTK installs it with Homebrew if needed. Hover a name to see what it does.
• Colors that follow your wallpaper. The accent color now comes from the current wallpaper. Choose Classic yellow, a swatch or any color in Settings → Colors.
• VS Code themes. Import a .vsix or install a theme from Open VSX by its ID, such as dracula-theme.theme-dracula. Only the theme's colors are read. While a theme is active, its colors stay put when the wallpaper changes.
• Teams pick their own models. Auto teams are led by the task's model, which chooses each worker's model from your installed ones or as your prompt asks.
• Bot pages. A setup checklist guides new bots, bots can be paused and resumed, and the header has Edit, Settings and a More menu with Reset and Delete.
• A calmer layout. Divider lines between panels are gone, the right sidebar's header blends in, the top-right icons are grouped, and notices are a slim line.
• Cleaner threads. Your messages are right-aligned bubbles without name rows. Copy and Fork appear when you hover a message and no longer take a line.

Known limitations: Android typing accepts plain ASCII text only. Android emulator setup was verified against Homebrew's package definitions, not yet on a fresh machine. Savings shown for Lean code and Quiet tools are estimates.

• Richer Markdown everywhere. Agent replies, streamed text, subagents, team workers, the quick launcher and bot chats now render through Software Mansion's Enriched Markdown engine.
  • Task lists show real checkboxes.
  • GitHub-style notes and warnings (> [!NOTE], > [!WARNING]) appear as colored callouts.
  • Tables keep their column alignment, and numbered lists keep their starting number.
  • LaTeX math, inline $…$ and display $$…$$, is typeset. Prices like "$5" stay plain text.
  • HTML written inside a reply is shown as text, never run.
• A page for each bot. Click a bot in the sidebar to chat with it on its own page, with streamed replies, suggestions and inline approvals. Its memory, schedules, channels, rules and activity sit in a panel beside the chat, and the pencil edits its name, instructions and model. The panel can be hidden, and the app remembers your choice.
• The bots' web chat page uses the same Markdown engine.

Known limitations: optional link titles ([text](url "title")) are not shown. Math in the bots' web chat page is shown as code.

• Bots in the sidebar. A Bots section at the bottom of the sidebar lists your bots and shows when one is working or paused. Click a bot to chat with it right in the main conversation view, with the full composer. Use + to add a bot, and the gear to open its memory, schedules, channels and rules. Bot workspaces no longer appear among your projects.
• Slimmer modals. Fields, dropdowns, buttons, setting rows and tabs in every modal are more compact, with a smaller title and close button.

Known limitation: messages with images or moodboards sent in a bot's chat go to the task directly, without refreshing the bot's memories.

• Clearer, more modern windows. Settings, Agent usage, Bots, Import conversations, Go to anything and every other modal now use a lighter, more transparent glass. The workspace stays visible behind them, edges are fine hairlines with rounder corners and softer shadows, and the muddy tint is gone.
• Fields, dropdowns, setting groups and buttons inside modals are frosted to match, with a soft yellow focus ring. The close button is now a small round control.
• Go to anything and Import conversations each have a single clean search field.

With Reduce transparency or Increase contrast turned on, modals keep solid backgrounds.

• Bots: always-on assistants on your own computer. Open Bots from the sidebar or Go to anything and start from a template or from scratch. Give a bot a job, instructions and standing goals, and choose its harness, model and permissions.
  • Chat with it right in the Bots window, with streamed replies and inline approvals. Each conversation is also an ordinary task you can open in the workspace.
  • It remembers what matters, and you can see, add and remove its memories. It can schedule work for itself (once, every few minutes or hours, daily or on weekdays). Optional check-ins review its goals and message you only when something needs you.
  • Reach it from a private web chat page, Telegram, Slack, Discord or WhatsApp. Each channel has a guided setup with step-by-step links and copy buttons, a one-click Slack app, a Discord invite link, and a credential check before anything is saved. New senders pair with a four-digit code you approve.
  • In any chat, /stop stops its work, /new starts over, and replying yes, no or always answers an approval. Rules can allow or deny kinds of actions; risky ones such as sudo, force pushes and deleting files always ask.
  • Pause, reset or delete a bot at any time. The Activity tab records everything it did.
• Queue and steer while an agent works. The prompt stays open during a run: Enter queues your message to send when the run finishes, and Command-Enter steers the running turn (Claude-family harnesses and Codex). Queued messages can be edited, reordered and removed.
• Usage limits. When Claude Code or Codex hits a usage limit, the task shows when it resets. Continue now, or let it resume automatically at reset.
• Agents can search tasks, queue and steer messages for other tasks, request a stop, and pin tasks in the same project.

Known limitations: bots run while PandaCode is running. WhatsApp needs a public HTTPS address for Meta's webhook, such as a Cloudflare or Tailscale tunnel, and only allows free-form replies within 24 hours of the person's last message. Chat channels carry text only. On macOS, after an update you may be asked to allow PandaCode to read saved channel tokens from Keychain.

• Import conversations from Claude Code and Codex. Open Import conversations from the sidebar or Go to anything, search your conversations, check the ones you want, and import them.
  • Each one becomes a task in the project for its folder. The project is added if needed.
  • Its history keeps its original times, and your next prompt continues the same session.
  • Conversations already in PandaCode link straight to their task.
• New task opens immediately, but the task is only saved once you send its first prompt. Leaving an untouched new task discards it. One with typed text, attachments or a worktree is kept, and New task brings it back.
• When a single project is selected, its tasks are listed without a repeated folder heading, and up to 20 show before "Show more". "More in other projects" is now a single line at the bottom of the sidebar.

Known limitations: only readable text is imported, and tool calls appear by name. Conversations whose folder no longer exists on this computer cannot be imported.

• A more compact sidebar. New task and Search are simple rows with their shortcuts, and the project filter sits in the "All projects" label above your tasks.
• Tasks are grouped into one folder per project, with the most recently active project first. Each task is a single line showing its agent, title, status and age.
• Folders show six tasks with a "Show more" link, and remember when you collapse them. Opening a task from a collapsed folder opens it.
• Pinned tasks stay at the top and show their project. Tasks without activity in the last 24 hours are dimmed instead of moving to a separate Inactive group.
• Drag or use Alt+Arrow keys to reorder tasks within a project.

• Agent usage now opens on an Overview. It shows the estimated API cost for the past week, month, or year, and the average of each agent's most-used limit window.
• The Overview lists every agent with its color, limit meter, cost share, tokens and sessions. It also has a daily cost chart, token totals with cache savings, and a breakdown by model or by day.
• Each agent has its own tab with its limit windows, its cost and models for the period, and a year activity calendar.
• Estimated API cost is read from Claude Code and Codex session logs on this computer, only while Agent usage is open. Only token counts are kept and nothing is uploaded. Reopening reads only new log data.
• Prices come from LiteLLM's public price list, refreshed at most daily, with a bundled copy for offline use. They include cache reads, cache writes and long-context rates.

Known limitations:

• The estimate shows what the same tokens would cost through the API, not what a subscription charges. Batch, priority and fast-mode pricing are excluded.
• Models missing from the price list are shown as unpriced.
• Cost is not tracked for Kimi, OpenCode, GLM, Grok, Antigravity or MiMo. Their calendars count prompts sent from PandaCode.
• The first scan of large Codex histories can take several seconds.

• Run projects directly from the workspace header or a changed-files summary. Agents can maintain a pandacode-run.sh launch script and open the resulting local preview in PandaCode's browser. Runs use a terminal tab belonging to the task.
• Changed-files summaries capture edits attributed to that task and its workers, instead of including unrelated changes from the entire checkout. Saved diffs remain available after the files change again.
• Resize the right sidebar by dragging its edge. The saved width adapts to the available space, including when the left sidebar is collapsed.
• Imported terminal settings keep fonts, colors, and cursor preferences without importing oversized font sizes.
• Worker scheduling respects the team's parallel limit, starts independent assignments together, and explains when work is queued behind a conflicting file assignment or dependency. Read-only assignments can run alongside other work.
• Existing Codex architect sessions refresh their app-tool catalog while preserving conversation context, so new team tools are available after an update.
• The Codex usage panel includes a keyboard-accessible daily activity calendar with per-day token counts.
• Run Script preview paths now use consistent validation on Windows, macOS, and Linux, with native Windows Bash discovery covered by platform-aware checks.

• Run projects directly from the workspace header or a changed-files summary. Agents can maintain a pandacode-run.sh launch script and open the resulting local preview in PandaCode's browser. Runs use a terminal tab belonging to the task.
• Changed-files summaries now capture edits attributed to that task and its workers, instead of including unrelated changes from the entire checkout. Saved diffs remain available after the files change again.
• Resize the right sidebar by dragging its edge. The saved width adapts to the available space, including when the left sidebar is collapsed.
• Imported terminal settings keep fonts, colors, and cursor preferences without importing oversized font sizes.
• Worker scheduling respects the team's parallel limit, starts independent assignments together, and explains when work is queued behind a conflicting file assignment or dependency. Read-only assignments can run alongside other work.
• Existing Codex architect sessions refresh their app-tool catalog while preserving conversation context, so new team tools are available after an update.
• The Codex usage panel includes a keyboard-accessible daily activity calendar with per-day token counts.

• The Codex usage panel (/usage) now shows daily token use as an activity calendar instead of a collapsed list. Each column is a week, and darker squares mean heavier days. The header totals tokens and active days for the period.
• Hover or click a day to see its exact token count. With the keyboard, the calendar is a single stop: arrow keys move by day and week, and Home and End jump to the ends. /usage daily opens with the calendar focused.
• The calendar fits the panel and scrolls on narrow windows, keeping the newest weeks in view.

Known limitations: the calendar appears only for Codex accounts that report daily usage. The Account limits view for other agents is unchanged.

• Switch saved Team profiles while a task is running from either Team or the model selector. Changes queue until the current architect turn finishes; existing workers keep their models and conversations.
• Replace a queued setup, switch architect harnesses with a saved-history handoff, and change profiles on paused or previously used teams without creating another task. Queued choices survive app restarts.
• Fixed a scheduling race between an architect finishing and its next review starting, so team changes preserve pending work.
• Includes the Claude title-tool continuity and local browser-preview restoration fixes from 0.2.28.

• Fixed Claude architect sessions failing with a missing pandacode_title tool reference after a review, retry, or resume. The title tool stays available across turns and safely does nothing when a title update is not requested. Existing conversation history and ToolSearch remain intact.
• Fixed local HTML and SVG previews showing a blank page after switching threads. Reopening an active PandaCode file preview now restores its private browser access, including directory pages, query parameters, and anchors.

• Pin threads above Active and Inactive, and drag threads into your preferred order within each section. Pins and ordering are saved across restarts. Right-click a thread for Pin, Unpin, Move up, and Move down; focused threads also support Option/Alt + Up/Down.
• Terminals now belong to their thread, with multiple tabs that retain their processes and output when switching threads. Minimize a terminal without stopping it; closing a terminal with running commands offers Minimize instead.
• New terminals automatically import supported font, cursor, spacing, and color settings from local Ghostty, Kitty, Alacritty, WezTerm, or Windows Terminal configuration. The most recently modified supported configuration is used; configuration code is never executed.
• Earlier messages load as you scroll up, preserving your reading position while responses continue streaming. This also works in worker conversations and the quick overlay.
• Fixed browser pages going blank after switching threads. Restoring a page and pressing Reload now explicitly restore its visibility.
• Fixed terminal startup output being lost when another tab is busy, and restarting an exited harness console for a new command.

Terminal processes and open tabs are retained while PandaCode is running. Appearance import supports static settings; dynamic configuration and named external themes keep PandaCode's defaults.

• Each thread now keeps its own right sidebar state while PandaCode is open: the selected tab, folder or file, browser URL, and whether the sidebar is open. Switching threads restores that thread's panel; new threads start with it closed.
• Get notified when an installed coding agent has an update. Choose Update from the popup, or check in Settings → Agents. Supported installations use their official updater, show progress, and verify the resulting version. Other installations show manual guidance.
• Agent update checks are cached and updates start only when requested, after running work is stopped. Shared Claude installations used by GLM and MiMo are checked together.
• Delete a thread and its worker history from the task menu, with confirmation. Project files and worktrees are preserved, and running tasks cannot be deleted.
• Fixed browser and file-preview races when switching threads quickly, including keeping an in-flight screenshot attached to its original thread.

Browser pages reopen at their saved URL when switching threads. Agent update availability depends on the installed harness and installation method.

• See account usage limits and reset times for supported, configured agents from the new Usage panel beside Settings and Help.
• Thread titles can now follow the conversation as it develops. Titles you set yourself stay unchanged.
• Tool calls appear in compact, expandable groups of up to five, with agent messages kept in order between groups in conversations, board tasks, and the quick overlay.
• Expanded tool details stay open as responses stream, and the generating status follows the current activity.
• Fixed the browser preview going blank when a slash-command menu opens beside it.

Usage availability depends on the installed harness and account. Unsupported providers show an explicit unavailable state.

• Switch harnesses in an idle task while keeping the conversation and project. A concise handoff explains what carries over, and earlier replies keep their original provider labels.
• Open HTML, SVG, images, and other previewable project files in the built-in browser. File references in agent responses can link directly to a preview or source view.
• Use local and cached slash commands immediately while provider commands load in the background, with clearer loading feedback and retry.
• Improve Windows Codex npm-launcher and Antigravity path compatibility.
• Give Antigravity more time to discover models on a cold start, share successful metadata, and reliably refresh stale model lists.
• Check Antigravity's connection separately from its installation status, and see clearer startup progress and useful provider failure details.
• See Claude's authentication source in Settings and optionally prefer its saved login over inherited environment credentials.
• Reduce snapshot-related startup delays and remove misleading change warnings after replies that did not produce a usable file diff.

Alpha preview. Windows launch behavior and the reported Antigravity cold-start latency still require verification on affected machines. Conversation transfers include bounded text history; private reasoning, tool execution state, and earlier image bytes do not transfer between harnesses.

Unlimited agents and immediate task activity

• Remove the four-agent limit so more tasks can run at the same time.
• Add unlimited parallel workers for architect teams.
• Move tasks into Active as soon as a prompt is sent, and remember that activity after reopening the app.

• Hide empty file-change cards when an agent finishes without changing files.
• Remove the blank space left by those cards while keeping worker progress visible.
• Keep explanations visible when a change snapshot fails or is incomplete.

Grok Build and MiMo

• Add Grok Build with model discovery, reasoning effort, context usage, streaming responses, approval controls, and saved-session resume.
• Add Xiaomi MiMo through Claude Code, with Token Plan region selection, a separate API endpoint option, and API key settings.
• Make both providers available in architect and worker teams, saved profiles, the quick overlay, terminal sessions, conversation forks, and slash commands.
• Keep unsupported image attachments and mockup sketches out of Grok tasks while its CLI does not accept images.

Compact menus and clearer settings

• Keep model, project, attachment, and command menus compact with single-line rows.
• Move longer explanations into information tooltips while keeping help text available to assistive technology.
• Simplify settings, dialog headings, and quick-launcher shortcut controls.
• Keep live GitHub status and worktree previews visible.
• Avoid replaying dialog entrance motion when menus or tooltips open inside it.

Attachments, mockups, and clearer project navigation

• Attach images, documents, and folders from the message box, with image paste and drag-and-drop support.
• Create and edit mockup sketches from the attachment menu and send them to your coding agent.
• Mention folders with @ alongside project files.
• Make hidden tasks easier to find when the sidebar is filtered to one project.
• Keep the Team button's appearance consistent before and after opening its selector.

PDF and Office document text extraction is available on macOS. Images, plain-text documents, folders, and sketches are supported across the desktop platforms.

Fix loading existing workspaces after updating

• Fix the database upgrade that could prevent existing workspaces and threads from loading after version 0.2.18, with a missing worktree_branch column error.
• Apply the latest schema migration to previous installations while keeping existing projects, threads, and messages.
• Add regression coverage for upgrading an existing database and keeping the migration guard in sync with the schema version.

A smoother workspace, connected to GitHub

• A consistent warm glass material and spring motion across menus, dialogs, panels, board cards, toasts, and segmented controls, with reduced-motion and contrast fallbacks.
• A project, checkout, and branch bar above every task's prompt. Move an unstarted task into a worktree while keeping its draft, or create a sibling worktree task from an existing conversation.
• Choose a starting branch, reuse an existing branch, recover a missing worktree, and clean up finished checkouts with checks for uncommitted work and unmerged branches.
• Optional GitHub integration through your signed-in GitHub CLI. Import issues as pending board tasks and push a branch to open a linked pull request.
• Linked issue and pull-request status appears on tasks and the board. Turn GitHub integration off in Settings to hide its controls and stop its CLI requests.

Known limitation: GitHub mention searches and issue-context resolution can fail because the issue lookup requests a pull-request-only CLI field. Board issue imports and pull-request creation use separate commands.

Help, right where you need it

• Open PandaCode support from the new question-mark button on the left. It stays available in board view and with the sidebar hidden.
• Find Help & support in Settings. Copy the app version, platform, and architecture when you want to include them in a support request.

• New tasks start with Full access across supported coding harnesses, using each harness's native permission policy.
• Keep saved task permissions when switching harnesses or using a team profile. Workers and forks keep their parent task's permission choice.
• Install the macOS app with a disk image: open the download and drag PandaCode into Applications.
• Existing installations continue receiving cryptographically verified updates.

• Give worker and subagent popovers the same translucent, blurred material as the sidebar, with clear text and accessible solid-surface alternatives.
• Fix the conversation collapsing into a narrow strip when opening Browser with the left sidebar hidden.
• Preserve resized sidebar widths and responsive layouts when opening the browser, switching views, or increasing text size.

• Fix Test with Jev overflowing at larger text sizes and with long local-server names.
• Use a compact translucent popover with background blur, quieter controls, and explanations behind information buttons.
• Keep Run test visible while options scroll. Expand local servers when needed; empty test history stays hidden.
• Preserve keyboard access, form validation, and reduced-transparency preferences.

• Watch Jev test your app directly in the sidebar browser, with a live view of its isolated Chrome session and a compact action history.
• Open Test with Jev in a small browser popover. Choose a goal or preset, then start watching without leaving your task.
• Stop a test, return to normal browsing while it continues, or reopen a saved test from its report.
• Keep live previews private and lightweight: bounded local frames, capture only while watched, and automatic cleanup when the viewer closes or disconnects.

Jev remains an optional add-on. Browser testing requires its runtime and your provider keys; watching a test adds no model calls.

PandaCode 0.2.12 fixes OpenCode model selection.

• Fixes valid models being rejected with “The harness does not advertise the selected setting” when OpenCode reports a large model catalog. Models later in the list, including DeepSeek Flash, can now start normally.
• Keeps model discovery and launch validation on the same bounded catalog size, including grouped model lists. Unsupported models and reasoning efforts are still checked against the harness's current options.

Your saved models and task settings are preserved. This update is for Apple silicon Macs.

PandaCode 0.2.11 adds optional Jev assistance for agents and teams.

• Set up Jev in Settings → Agents → Jev · Optional add-on. Decision assistance and browser testing are separate switches, both off by default. PandaCode continues to work without a Jev key.
• Use Jev to suggest a configured worker, find relevant recent conversation excerpts, review saved work, and check changed files against project rules. Preview the evidence first and keep reports locally. Suggestions do not automatically change models, assign work, or approve completion.
• Let agents test local app previews with Jev in a separate browser session. Follow live actions, inspect screenshots, and read independent final checks. Browser testing uses a separately installed runtime and your own TypeSafe and text-model credentials; provider usage is billed separately.
• Mention project files with @ in the prompt. Copy messages and responses, or fork a conversation from a selected saved message, including from the board and quick overlay.
• Includes desktop compatibility and process-management improvements. This release is for Apple silicon Macs; other platform downloads are not being published in this update.

Jev is a preview integration. Model judgments are advisory, and browser checks verify only the supplied assertions. Your existing coding providers, approvals, conversation history, and normal browser workflow remain available.

Long saved replies now use less memory when opening a conversation, without changing the stored response or agent context. The quick overlay avoids redrawing unchanged conversations, while continuing to show new text, status changes, and recovered connections immediately.

Automatic team updates now appear as compact Team status rows instead of large user messages containing internal prompts and JSON. The agents retain their full coordination context. Existing team conversations and forks use the same concise presentation, while JSON you write yourself remains unchanged. This applies to conversation, board, worker threads, and the quick overlay.

Team conversations now show a compact live worker summary instead of empty architect-turn file summaries. See each worker's assignment, role, harness/model, and status, and click a worker to open its live thread and saved changes. The summary follows queued, working, approval, paused, and completed states in conversation and board views. Actual file changes and capture errors remain visible.

The model selector now shows the architect model followed by “+1 model” or “+2 models” when a team is enabled, across conversations, the board, and the quick overlay. The count reflects saved worker model selections and stays visible beside long model names. Switching back to a single agent removes the count.

Team setup now opens as a compact popover beside the composer. Choose a saved profile, architect model, and worker models without opening a full modal. Run limits stay tucked away until needed.

Architects and workers can exchange direct messages across harnesses. Open Workers → Team messages to see questions and handoffs. Workers can also save progress checkpoints with completed work, next steps, and blockers; the latest checkpoint appears in each worker’s details and is restored when interrupted work resumes.

New messages and blockers can wake an idle architect within the configured round budget. Coordination is event-driven, with bounded messages and summaries. Existing concurrency, file-scope scheduling, approval, cancellation, and retry limits remain in place.

Saved teams and profiles are retained. Existing Codex teams transfer their saved conversation once on the next explicit run to enable the new tools. Updating never starts an agent automatically.

PandaCode no longer requests macOS Keychain access at startup, including after app updates.

Your tasks, provider settings and local history are preserved.

Fixes PandaCode waiting to open while macOS requests Keychain access. The app now remains responsive during permission checks and explains which password macOS is asking for.

Ad-hoc preview updates can still trigger a macOS Keychain permission prompt because they do not yet use a stable Apple Developer ID signature.

Choose which OpenCode models appear in PandaCode from Settings → Agents → OpenCode models. Search your installed OpenCode catalog, save your selections, and use them across conversations, boards, the quick overlay, and team setup.

Fixes model discovery falling back to Default when OpenCode has a large catalog. Existing task models are preserved, and saved selections remain available after restarting PandaCode.

Fixes a crash on a fresh launch.

Maintenance update with improved release verification and automated macOS preview delivery.

• Get an update notification when a new version is available, then install and restart when your work is idle.
• Keep task menus stable while agent responses finish streaming.