Skip to content

Releases

Plexiform 1.0.2-beta.108

Pre-release

Choose a tag to compare

@github-actions github-actions released this 30 Oct 00:02

Plexiform app update — 29 October 2026 (1.0.2-beta.107)

For macOS (Apple silicon and Intel) and Linux (x64).

What changed

  • Your AI sessions on this computer now know about each other, on by default: each Claude Code and Codex session is told which others are running here (tool, folder name, branch and whether it is working, idle or waiting on you), can leave short notes for them, and gets one line at its start about Plexiform's tools. On macOS, Plexiform does this by adding a tool server named plexiform-bus to Claude Code (one entry in ~/.claude.json) and to Codex (through Codex's own codex mcp add). The first time it adds them, Home shows a notice with Undo, which removes them and keeps them off. Preferences → Advanced → AI sessions on this computer has a switch for each: "Let my AI sessions see each other" and, on macOS, "Add Plexiform's bus tools to Claude Code and Codex".
  • Answering permission prompts from the widget is now on for new installs. Existing installs keep the setting they had. Turn it off with "Answer permission prompts from the widget" in Preferences → Widget, under Permission prompts.
  • In Setups, pressing Cancel on the Apply confirmation now says "Cancelled. Nothing was changed." instead of an error, and you can press Apply again without building the review again (the review still expires when it did). If another window is asking you something, Setups says so.
  • Overview no longer sticks on "Plexiform session state is unavailable right now" after a Settings save: it appears only after two failed reads in a row and goes away on the first good one.
  • Your teammates no longer see you as viewing a session after you minimise Plexiform, leave Team sessions or sign out.
  • Clicking a session in the widget or Overview now brings its terminal forward, including when its window is minimised or the terminal app is hidden. When Plexiform can't, it says why instead of implying it did.
  • A session's card on the widget has its own buttons: open its terminal, show its folder in Finder, open it in your editor, or copy the command that resumes it. The card also shows the folder and how long the session has waited on you.
  • The widget remembers where you put it on each display. Overview's My sessions shows how long each session has been waiting on you.
  • Find a session in Overview's My sessions by folder, task or AI tool. Show in Overview from a notification now lands on that session.
  • A key to bring forward the next session waiting on you, oldest first. It is off until you set a key on the Widget page.
  • Snooze one session's alerts for up to an hour, from its widget card or its notification. Clicking a Plexiform notification now always takes you somewhere useful, and notifications have buttons (Show in Overview and Snooze 1 hour; Stop session for one that is running away).
  • On macOS, Preferences → Advanced now has a Terminal access section listing Terminal, iTerm2, Ghostty and Ghostty Dock and whether Plexiform may switch to their tabs. Test checks one terminal, only when you press it and only if it is already open, so macOS can ask for permission then but never by itself. Open settings takes you to the right place in System Settings.
  • When a teammate stops a turn in a session you shared, Overview and the Team directory now say who (for example "Alex stopped this turn.").
  • Sessions you start in a team workspace are now shared with that team to watch only. To let teammates send to one, open its Share… in Overview and choose Allow sending (it keeps the time the share had left). Teammates can't propose messages to a watch-only session. Terminal Claude channels you attach on purpose still let teammates send. Sessions already shared keep their current access until Plexiform restarts or the session ends.
  • Find agents: search a catalogue of ready-made agents, read what each does, and add one to Claude Code, Codex and other AI tools with one click, with Undo. Plexiform writes a file only when you click +, never replaces one that is already there, and backs up your own agents before taking them over.
  • The plexiform command line tool (Preferences → Advanced → Command line) can now hand over a task: plexiform run "‹task›" picks Claude or Codex for it and runs it in Tasks. Runs that do real work get a board card on their own.
  • Runs Plexiform starts (Tasks, chat, takeover, Explain and the other helpers) now use your own Claude Code and Codex setup: your gateway or cloud provider, model and key. A key never reaches a shell the AI can use: runs with tools get it through a private file the AI can't read, deleted when the run ends.
  • Usage now takes your plan from your own provider instead of guessing Pro, shows turns it can't price as activity rather than spend, and explains its numbers in one place (About these numbers).
  • Sessions and This Mac show which board each project's new work goes to. When a project could go to a team board, Plexiform asks once, with Keep on My board as a choice.
  • Sessions and My sessions are grouped by project, then by status, and the app uses the same words for session states everywhere.
  • The widget's bubble is smaller, stays clear of the main window, and can still be answered on short displays.
  • A new graphite accent colour replaces the old terracotta throughout the app.
  • Linux: clicking a session raises that session's own terminal window, found by its process, never a different one. Your keyring choice is remembered, and the agent bus tools are not added (they need a Mac).
  • The agent bus is now automatic, and you can opt out. AI tools on this Mac can hand work to each other (for example Claude Code to Codex) in any git repository a Claude Code or Codex session here is working in, with no need to add each project first. Your home folder, hidden folders in it (such as ~/.claude and ~/.codex) and ~/Library are never used. To opt out, untick "Automatic" or list "Only these repositories" in Preferences → Advanced → Agent bus, add repositories to "Never these repositories", or switch the bus off there. If you had set the bus up before through the Tasks policy file with an empty list in Preferences, it stays limited to those repositories. Results now say which AI did the work (for example "Codex running in Plexiform").
  • Ghostty Dock, the third-party terminal by Andrew Baker, can be installed from Preferences or Health on Apple silicon Macs. Plexiform opens Terminal with Andrew's own installer command and nothing runs until you press Return there; Plexiform tells you when it is ready.
  • Clicking a session that needs you goes to its exact Ghostty Dock tab even when the session was first seen as plain Ghostty or started before Ghostty Dock was installed, and a session Ghostty Dock shows as waiting is included. Sessions that are blocked on you now come before finished ones, longest-waiting first. When only the terminal app could be brought forward, the widget now says it couldn't pick the tab instead of implying it did.

What has been verified

  • Every change in this release was reviewed before it merged, and the full unit, remote and website test suites ran on a Mac with no new failures.
  • The unit tests also passed on macOS and Linux in a release dry run on GitHub Actions. The release build runs them again on both, and smoke-tests each packaged app, before anything is published.
  • All 62 screenshot test files ran on a recent build of this release: 59 passed. The other three (animated character poses, a page that navigates mid-test and a timing check) failed from the tests' timing, not the app; a test-only fix in this build makes them pass, and none is a visible change.

Current limits

  • Mac builds remain ad-hoc signed and are not Apple-notarized.
  • On Linux, adding or adopting an agent in Find agents needs a working system keyring (libsecret, such as GNOME Keyring, or KWallet). With a weak or unknown keyring, Plexiform refuses rather than store the key unprotected.
  • Builds are for macOS and Linux only; there is no Windows build.
  • The limits listed for earlier betas below still apply.
Read more

Plexiform 1.0.2-beta.101

Pre-release

Choose a tag to compare

@github-actions github-actions released this 22 Oct 09:17
fb640f0

Plexiform app update — 20 October 2026 (1.0.2-beta.101)

For macOS (Apple silicon and Intel) and Linux (x64).

What changed

  • Your Claude Code sessions are back on the board. Since beta.98, sessions started in Claude Code got no card, so "In progress" stayed empty while you worked. They show up again and move between columns as they work, wait and finish.
  • Less clutter from scheduled Hermes jobs. Each scheduled (cron) Hermes job now keeps one card that every run updates, instead of a new card per run. Older cards from past runs can be cleared with "Clear all" on the Idle column.

What has been verified

Desktop (5,586 tests) and privacy test suites pass on this build, with new tests for both fixes.

Current limits

  • The limits listed for earlier betas below still apply.

Plexiform app update — 20 October 2026 (1.0.2-beta.100)

For macOS (Apple silicon and Intel) and Linux (x64).

What changed

  • Project safety. Plexiform checks what your AI changed after every turn. It warns when a turn adds what looks like a password or API key, when a secret would end up in the browser (for example behind NEXT_PUBLIC_ or VITE_), and when a .env file is not ignored by git. You see a blanked-out preview, a link to make a new key, and a one-click undo of that turn. Secrets are never saved or shown.
  • Careful with destructive commands. When an AI asks to run something like rm -rf, a force push, a database reset or a production deploy, the light turns red and Home shows an "Approve with care" card with what it would affect. With restore points on, Plexiform saves one first.
  • Database rules that only look locked. Turn it on per project and Plexiform flags Supabase and Firebase rules that let anyone read or write, and tables without row-level security, with a plain-English fix.
  • Spend safety. "This session burned $X in Y minutes" alerts, daily caps per project, a projected total for today, and a warning when an AI keeps retrying the same failing step. A new rule template alerts when a session crosses an amount.
  • Your Claude plan, not API dollars. Pro and Max users see estimated bars for how close they are to their 5-hour and weekly Claude limits. API-price equivalents are behind "Show what this would cost on the API".
  • Explain mode (off until you turn it on). "Explain this" on permission questions gives one plain sentence, a safe / be careful / risky rating and a suggested answer. Common commands are explained on this Mac; others go to your own claude login with a 2-cent cap. Home shows "What did my AI just do?" after each turn, with "Test it now" when a local dev server is running, and one-click prompts to copy when a session is stuck.
  • Deploys. Plexiform finds Vercel, Netlify and Cloudflare projects and shows building, live with a link, or failed next to the session, using the command-line tools you are already signed in to or a token you paste. A failed build gets a plain-English reason and "Send this error to Claude", which copies a ready prompt. Claude celebrates when a deploy goes live.
  • Easier for people new to coding. Setup asks "Just me, building an app" or "Developer / team", explains in plain words what it adds to Claude Code, and gives a test prompt to copy with help if nothing shows up. Simple mode hides team and developer settings and keeps Home to what matters. Each light colour now means one thing, with a key in Help, Preferences and the More menu. Rule templates read as wishes, such as "Chime when it's done".
  • Build logs and starter kits. Turn a session into a build log, review every line before anything leaves your Mac, then save it as a file or publish a link with "Made with Plexiform — remix this". Anyone can "Start from this" to install a starter kit (repo, CLAUDE.md, skills, MCP servers) after a consent sheet that lists what it may do. Discover now has one catalogue for starters, skills and MCP servers.
  • Smaller follow-ups. When a session crosses your spend line, the "Session crossed my spend line" rule now holds its other nudges until you press "Resume nudges" on its widget card. The widget card shows the latest "What did my AI just do?" recap, a help line and the deploy status. Plexiform keeps the last 50 recaps, scrubbed of secrets, so they survive a restart.
  • Fixes. Light-mode text is readable again in the widget bubble, the Waiting page, Settings and Lights. A saved sign-in is restored when you enter a hub address.

What has been verified

Desktop, remote, site and board test suites pass on this build (desktop 5,585 tests, board 3,282, 0 failures), with the privacy census and the hub tenancy matrix. The team hubs already run this build's hub code.

Current limits

  • Claude limit bars are estimates from this Mac's usage. Anthropic does not publish the limits, and use on other devices is not counted.
  • Deploy status has been tested with simulated Vercel, Netlify and Cloudflare responses, not live accounts.
  • Mac builds remain ad-hoc signed and are not Apple-notarized.
  • The limits listed for earlier betas below still apply.

Plexiform app update — 19 October 2026 (1.0.2-beta.97)

For macOS (Apple silicon and Intel) and Linux (x64).

What changed

  • Search past sessions. Find what you asked an AI about, across your past sessions on this Mac, from the command palette or the Memory page, with project and date filters. Results open the session's timeline. The index stays on this Mac and is scrubbed of secrets.
  • Where the money went. The Usage page breaks spend down by project, branch and pull request, compares it with the period before, and suggests one saving each week: the share of Opus turns that looked routine and what Sonnet would have cost. The weekly digest carries the top projects and the same suggestion.
  • Rules made easy. Start a rule from a template ("Tell me when a session is stuck", "Red banner when CI fails" and more), read each rule as a plain sentence, and have a rule send a desktop notification. Notifications respect quiet hours and meetings.
  • A first-run tour. Five short steps on the demo board within half a minute of setup: the widget, answering a question, rules, spend and team. Skip it any time; take it again from Help.
  • Local models in two clicks. Plexiform finds Ollama and LM Studio running on this Mac, lists their models, starts a session with the one you pick, and says what to do when none is pulled. Sessions show which model they use. You can add your own OpenAI-compatible endpoint; its key is kept in secure storage.
  • Phone. A waiting permission request on your phone's lock screen now has a Review button that opens it straight away; approving still needs your passkey. You can dictate a reply with the microphone and press Send yourself.
  • Team. A heat map of the files your team is working in, with files two people touch flagged; an automatic standup per teammate you can copy; and handoffs that can be declined with a note, with the outcome shown to the sender.
  • Create a Linear or Jira issue from a session. Plexiform drafts it from the session's handover, you edit it, and nothing is sent until you press Create.
  • Ask about your screen. Turn on "Use the screen when I ask" in Voice settings, hold your talk key or long-press Claude, and ask out loud ("where is the Wi-Fi setting?"). Plexiform captures the display under your mouse, asks your own Claude through the claude command (no API key), speaks the answer, and Claude flies to the thing it means on screen. The screenshot is deleted as soon as the answer comes back. macOS asks for Screen Recording permission the first time.
  • Say "Plexiform, do …" to offer a new Task. It starts only when you press Start.
  • The widget no longer fails to appear after an update when macOS asks for permission to a folder your sessions use, such as Desktop or Documents. Plexiform keeps running while that prompt waits for your answer.
  • Fewer interruptions in meetings. Nudges and rule notifications wait while Plexiform sees a meeting or screen share.

What has been verified

Desktop, remote, site and board test suites pass on this build (desktop 5,499 tests, board 3,278, 0 failures), with the privacy census and the hub tenancy matrix.

Current limits

  • Setting your Slack status from Plexiform is built but not available yet: it needs a Slack user permission the team's Slack app does not ask for today. The Slack runbook has the owner step.
  • Phone notification and voice changes reach phones after the team hub is updated.
  • Mac builds remain ad-hoc signed and are not Apple-notarized.
  • The limits listed for earlier betas below still apply.

Plexiform widget update — 19 October 2026 (1.0.2-beta.92)

For macOS (Apple silicon and Intel) and Linux (x64).

What changed

  • Session cards on the widget. Hovering a session on the widget's strip opens a card with the project, what it is doing and for how long, how long it has been waiting on you, this session's spend when known, and a teammate's initials for shared sessions. The card's buttons open the session, answer its waiting input, or pin it.
  • A smarter strip. A pinned session stays first, idle sessions fold into one "+N idle" chip, and teammates' sessions sit on their own row.
  • What just happened, at a glance. A small lane beside Claude shows the last three events (needs you, finished, stuck, teammate collision, CI failed) and fades each one after a few seconds. Click one to jump to it.
  • Two rings round Claude: how long the current turn has run against your usual turn, and today's spend against your budget. They hide when there is nothing to show and respect reduced motion and low power.
  • Widget settings: snap to any screen edge, not just corners; Normal, Compact or Dot size (Dot is the lamp alone); a global shortcut to summon or hide the widget; an accent colour per project; and "Quiet for 30 minutes" from the widget menu, shown by a small moon.
  • More character content: costumes witch, viking, tricorn, beret, antlers, astronaut and heartband; pets penguin, ghost, chick and owl; effects fire...
Read more

Plexiform 1.0.2-beta.100

Pre-release

Choose a tag to compare

@github-actions github-actions released this 20 Oct 09:44
d1b194b

Plexiform app update — 20 October 2026 (1.0.2-beta.100)

For macOS (Apple silicon and Intel) and Linux (x64).

What changed

  • Project safety. Plexiform checks what your AI changed after every turn. It warns when a turn adds what looks like a password or API key, when a secret would end up in the browser (for example behind NEXT_PUBLIC_ or VITE_), and when a .env file is not ignored by git. You see a blanked-out preview, a link to make a new key, and a one-click undo of that turn. Secrets are never saved or shown.
  • Careful with destructive commands. When an AI asks to run something like rm -rf, a force push, a database reset or a production deploy, the light turns red and Home shows an "Approve with care" card with what it would affect. With restore points on, Plexiform saves one first.
  • Database rules that only look locked. Turn it on per project and Plexiform flags Supabase and Firebase rules that let anyone read or write, and tables without row-level security, with a plain-English fix.
  • Spend safety. "This session burned $X in Y minutes" alerts, daily caps per project, a projected total for today, and a warning when an AI keeps retrying the same failing step. A new rule template alerts when a session crosses an amount.
  • Your Claude plan, not API dollars. Pro and Max users see estimated bars for how close they are to their 5-hour and weekly Claude limits. API-price equivalents are behind "Show what this would cost on the API".
  • Explain mode (off until you turn it on). "Explain this" on permission questions gives one plain sentence, a safe / be careful / risky rating and a suggested answer. Common commands are explained on this Mac; others go to your own claude login with a 2-cent cap. Home shows "What did my AI just do?" after each turn, with "Test it now" when a local dev server is running, and one-click prompts to copy when a session is stuck.
  • Deploys. Plexiform finds Vercel, Netlify and Cloudflare projects and shows building, live with a link, or failed next to the session, using the command-line tools you are already signed in to or a token you paste. A failed build gets a plain-English reason and "Send this error to Claude", which copies a ready prompt. Claude celebrates when a deploy goes live.
  • Easier for people new to coding. Setup asks "Just me, building an app" or "Developer / team", explains in plain words what it adds to Claude Code, and gives a test prompt to copy with help if nothing shows up. Simple mode hides team and developer settings and keeps Home to what matters. Each light colour now means one thing, with a key in Help, Preferences and the More menu. Rule templates read as wishes, such as "Chime when it's done".
  • Build logs and starter kits. Turn a session into a build log, review every line before anything leaves your Mac, then save it as a file or publish a link with "Made with Plexiform — remix this". Anyone can "Start from this" to install a starter kit (repo, CLAUDE.md, skills, MCP servers) after a consent sheet that lists what it may do. Discover now has one catalogue for starters, skills and MCP servers.
  • Smaller follow-ups. When a session crosses your spend line, the "Session crossed my spend line" rule now holds its other nudges until you press "Resume nudges" on its widget card. The widget card shows the latest "What did my AI just do?" recap, a help line and the deploy status. Plexiform keeps the last 50 recaps, scrubbed of secrets, so they survive a restart.
  • Fixes. Light-mode text is readable again in the widget bubble, the Waiting page, Settings and Lights. A saved sign-in is restored when you enter a hub address.

What has been verified

Desktop, remote, site and board test suites pass on this build (desktop 5,585 tests, board 3,282, 0 failures), with the privacy census and the hub tenancy matrix. The team hubs already run this build's hub code.

Current limits

  • Claude limit bars are estimates from this Mac's usage. Anthropic does not publish the limits, and use on other devices is not counted.
  • Deploy status has been tested with simulated Vercel, Netlify and Cloudflare responses, not live accounts.
  • Mac builds remain ad-hoc signed and are not Apple-notarized.
  • The limits listed for earlier betas below still apply.

Plexiform app update — 19 October 2026 (1.0.2-beta.97)

For macOS (Apple silicon and Intel) and Linux (x64).

What changed

  • Search past sessions. Find what you asked an AI about, across your past sessions on this Mac, from the command palette or the Memory page, with project and date filters. Results open the session's timeline. The index stays on this Mac and is scrubbed of secrets.
  • Where the money went. The Usage page breaks spend down by project, branch and pull request, compares it with the period before, and suggests one saving each week: the share of Opus turns that looked routine and what Sonnet would have cost. The weekly digest carries the top projects and the same suggestion.
  • Rules made easy. Start a rule from a template ("Tell me when a session is stuck", "Red banner when CI fails" and more), read each rule as a plain sentence, and have a rule send a desktop notification. Notifications respect quiet hours and meetings.
  • A first-run tour. Five short steps on the demo board within half a minute of setup: the widget, answering a question, rules, spend and team. Skip it any time; take it again from Help.
  • Local models in two clicks. Plexiform finds Ollama and LM Studio running on this Mac, lists their models, starts a session with the one you pick, and says what to do when none is pulled. Sessions show which model they use. You can add your own OpenAI-compatible endpoint; its key is kept in secure storage.
  • Phone. A waiting permission request on your phone's lock screen now has a Review button that opens it straight away; approving still needs your passkey. You can dictate a reply with the microphone and press Send yourself.
  • Team. A heat map of the files your team is working in, with files two people touch flagged; an automatic standup per teammate you can copy; and handoffs that can be declined with a note, with the outcome shown to the sender.
  • Create a Linear or Jira issue from a session. Plexiform drafts it from the session's handover, you edit it, and nothing is sent until you press Create.
  • Ask about your screen. Turn on "Use the screen when I ask" in Voice settings, hold your talk key or long-press Claude, and ask out loud ("where is the Wi-Fi setting?"). Plexiform captures the display under your mouse, asks your own Claude through the claude command (no API key), speaks the answer, and Claude flies to the thing it means on screen. The screenshot is deleted as soon as the answer comes back. macOS asks for Screen Recording permission the first time.
  • Say "Plexiform, do …" to offer a new Task. It starts only when you press Start.
  • The widget no longer fails to appear after an update when macOS asks for permission to a folder your sessions use, such as Desktop or Documents. Plexiform keeps running while that prompt waits for your answer.
  • Fewer interruptions in meetings. Nudges and rule notifications wait while Plexiform sees a meeting or screen share.

What has been verified

Desktop, remote, site and board test suites pass on this build (desktop 5,499 tests, board 3,278, 0 failures), with the privacy census and the hub tenancy matrix.

Current limits

  • Setting your Slack status from Plexiform is built but not available yet: it needs a Slack user permission the team's Slack app does not ask for today. The Slack runbook has the owner step.
  • Phone notification and voice changes reach phones after the team hub is updated.
  • Mac builds remain ad-hoc signed and are not Apple-notarized.
  • The limits listed for earlier betas below still apply.

Plexiform widget update — 19 October 2026 (1.0.2-beta.92)

For macOS (Apple silicon and Intel) and Linux (x64).

What changed

  • Session cards on the widget. Hovering a session on the widget's strip opens a card with the project, what it is doing and for how long, how long it has been waiting on you, this session's spend when known, and a teammate's initials for shared sessions. The card's buttons open the session, answer its waiting input, or pin it.
  • A smarter strip. A pinned session stays first, idle sessions fold into one "+N idle" chip, and teammates' sessions sit on their own row.
  • What just happened, at a glance. A small lane beside Claude shows the last three events (needs you, finished, stuck, teammate collision, CI failed) and fades each one after a few seconds. Click one to jump to it.
  • Two rings round Claude: how long the current turn has run against your usual turn, and today's spend against your budget. They hide when there is nothing to show and respect reduced motion and low power.
  • Widget settings: snap to any screen edge, not just corners; Normal, Compact or Dot size (Dot is the lamp alone); a global shortcut to summon or hide the widget; an accent colour per project; and "Quiet for 30 minutes" from the widget menu, shown by a small moon.
  • More character content: costumes witch, viking, tricorn, beret, antlers, astronaut and heartband; pets penguin, ghost, chick and owl; effects fireworks, bats, lightning, clouds, aurora and embers; poses dance, shrug and yawn. All appear in the Lights rule editor.
  • Seasonal pack: around New Year, Valentine's Day, spring, Halloween and winter the character picks up a matching costume, pet and effect, but only on channels no rule sets. Turn it off with "seasonalPack": false in config.json (the older Seasonal looks setting still works).
  • Sprite packs: put a folder with pack.json and PNG/SVG frames in ~/.plexiform/sprites/<pack>/ and set "spritePack": "<pack>" in config.json to show your own frames instead of the drawn character. Packs are checked strictly (plain file names, no links, 512 KB a frame, 64 frames a state, 8 MB a pack; SVGs with scripts or external references are refused) and never run; any problem f...
Read more

Plexiform 1.0.2-beta.97

Pre-release

Choose a tag to compare

@github-actions github-actions released this 19 Oct 17:33

Plexiform app update — 19 October 2026 (1.0.2-beta.97)

For macOS (Apple silicon and Intel) and Linux (x64).

What changed

  • Search past sessions. Find what you asked an AI about, across your past sessions on this Mac, from the command palette or the Memory page, with project and date filters. Results open the session's timeline. The index stays on this Mac and is scrubbed of secrets.
  • Where the money went. The Usage page breaks spend down by project, branch and pull request, compares it with the period before, and suggests one saving each week: the share of Opus turns that looked routine and what Sonnet would have cost. The weekly digest carries the top projects and the same suggestion.
  • Rules made easy. Start a rule from a template ("Tell me when a session is stuck", "Red banner when CI fails" and more), read each rule as a plain sentence, and have a rule send a desktop notification. Notifications respect quiet hours and meetings.
  • A first-run tour. Five short steps on the demo board within half a minute of setup: the widget, answering a question, rules, spend and team. Skip it any time; take it again from Help.
  • Local models in two clicks. Plexiform finds Ollama and LM Studio running on this Mac, lists their models, starts a session with the one you pick, and says what to do when none is pulled. Sessions show which model they use. You can add your own OpenAI-compatible endpoint; its key is kept in secure storage.
  • Phone. A waiting permission request on your phone's lock screen now has a Review button that opens it straight away; approving still needs your passkey. You can dictate a reply with the microphone and press Send yourself.
  • Team. A heat map of the files your team is working in, with files two people touch flagged; an automatic standup per teammate you can copy; and handoffs that can be declined with a note, with the outcome shown to the sender.
  • Create a Linear or Jira issue from a session. Plexiform drafts it from the session's handover, you edit it, and nothing is sent until you press Create.
  • Ask about your screen. Turn on "Use the screen when I ask" in Voice settings, hold your talk key or long-press Claude, and ask out loud ("where is the Wi-Fi setting?"). Plexiform captures the display under your mouse, asks your own Claude through the claude command (no API key), speaks the answer, and Claude flies to the thing it means on screen. The screenshot is deleted as soon as the answer comes back. macOS asks for Screen Recording permission the first time.
  • Say "Plexiform, do …" to offer a new Task. It starts only when you press Start.
  • The widget no longer fails to appear after an update when macOS asks for permission to a folder your sessions use, such as Desktop or Documents. Plexiform keeps running while that prompt waits for your answer.
  • Fewer interruptions in meetings. Nudges and rule notifications wait while Plexiform sees a meeting or screen share.

What has been verified

Desktop, remote, site and board test suites pass on this build (desktop 5,499 tests, board 3,278, 0 failures), with the privacy census and the hub tenancy matrix.

Current limits

  • Setting your Slack status from Plexiform is built but not available yet: it needs a Slack user permission the team's Slack app does not ask for today. The Slack runbook has the owner step.
  • Phone notification and voice changes reach phones after the team hub is updated.
  • Mac builds remain ad-hoc signed and are not Apple-notarized.
  • The limits listed for earlier betas below still apply.

Plexiform widget update — 19 October 2026 (1.0.2-beta.92)

For macOS (Apple silicon and Intel) and Linux (x64).

What changed

  • Session cards on the widget. Hovering a session on the widget's strip opens a card with the project, what it is doing and for how long, how long it has been waiting on you, this session's spend when known, and a teammate's initials for shared sessions. The card's buttons open the session, answer its waiting input, or pin it.
  • A smarter strip. A pinned session stays first, idle sessions fold into one "+N idle" chip, and teammates' sessions sit on their own row.
  • What just happened, at a glance. A small lane beside Claude shows the last three events (needs you, finished, stuck, teammate collision, CI failed) and fades each one after a few seconds. Click one to jump to it.
  • Two rings round Claude: how long the current turn has run against your usual turn, and today's spend against your budget. They hide when there is nothing to show and respect reduced motion and low power.
  • Widget settings: snap to any screen edge, not just corners; Normal, Compact or Dot size (Dot is the lamp alone); a global shortcut to summon or hide the widget; an accent colour per project; and "Quiet for 30 minutes" from the widget menu, shown by a small moon.
  • More character content: costumes witch, viking, tricorn, beret, antlers, astronaut and heartband; pets penguin, ghost, chick and owl; effects fireworks, bats, lightning, clouds, aurora and embers; poses dance, shrug and yawn. All appear in the Lights rule editor.
  • Seasonal pack: around New Year, Valentine's Day, spring, Halloween and winter the character picks up a matching costume, pet and effect, but only on channels no rule sets. Turn it off with "seasonalPack": false in config.json (the older Seasonal looks setting still works).
  • Sprite packs: put a folder with pack.json and PNG/SVG frames in ~/.plexiform/sprites/<pack>/ and set "spritePack": "<pack>" in config.json to show your own frames instead of the drawn character. Packs are checked strictly (plain file names, no links, 512 KB a frame, 64 frames a state, 8 MB a pack; SVGs with scripts or external references are refused) and never run; any problem falls back to the drawn character.
  • Keep-awake on the Burst card now says "Saved, not applied yet" while Burst's machine-wide half waits for the password step in Terminal, instead of "On".

What has been verified

Desktop, remote and site test suites pass on this build (desktop 5,430 tests, 0 failures). The visual regression suite was run on this build and on beta.91's source: the only difference is the character baselines, which grew to include the new costumes and poses. The suite's other failing screenshots fail identically on beta.91's source and are tracked separately.

Current limits

  • Mac builds remain ad-hoc signed and are not Apple-notarized. A changed build may require fresh authorization for Plexiform's own permissions.
  • Sprite packs and project colours are set per Mac; they do not sync.
  • The limits listed for beta.91 below still apply.

Plexiform Claude Burst preview — 19 October 2026

For macOS (Apple silicon and Intel) and Linux (x64). Claude Burst itself is macOS only.

What changed

  • Claude Burst inside Plexiform, verified against Burst v0.20.36. Plexiform now reads Burst's newer state (primary health, Remote Control expectation, hosts entry, certificate trust) and gates every feature on a capability flag, so a 0.19 install keeps working with the newer parts hidden rather than broken.
  • Usage & Burst in the sidebar. One entry opens the Usage optimiser: Burst's status, its dashboard in a locked-down view, and the tools it adds.
  • Guided setup. The optimiser's empty state walks through Install, Trust, Enable and First check, says what each step changes on your Mac (from Burst's own README), and reads progress back from Burst while a step runs instead of sending you to a Terminal and asking you to come back.
  • One cause, one action. Every Burst state names what is wrong and offers one next step: not installed → Install; too old → Update; gateway not answering → Repair, with the last known state shown dimmed and its age; certificate or hosts entry not trusted → Fix; console down → Start console; usage panel not installed → Install. An unknown listener on Burst's port is named and nothing is read from it. Every Burst call from a page answers {ok, reason} and never throws.
  • Home tells you when to look. A single Burst row appears in Home's attention list only when something needs you: broken, too old, not trusted, not responding after several checks, an update available, a limit near, or loose ends. A healthy Burst, an intentionally off Burst or no Burst at all adds nothing.
  • Loose ends on Overview. Each repository row carries Burst's counts of uncommitted files, unpushed commits and stashes for repositories your sessions use; file names only, never contents, diffs or stash names.
  • Keep-awake, Remote Control and mods on the Burst card and the Tools tab. The card shows whether this Mac is being kept awake and whether Remote Control is expected, and lists Burst's session mod and the usage panel (claude-code-cost-sidebar) with Install or Update per mod. Every change to the Mac goes through the same native consent sheet as before. Burst's hotspot settings are never read or written by Plexiform.

What has been verified

Desktop, remote and site test suites pass on this build (desktop 5,340 tests, 0 failures) including a contract test that drives every Burst channel with a failing client, a 0.19 and a 0.20.36 fixture captured from a real install, and the privacy flow checks. On a Mac running the real Claude Burst v0.20.36: detection, all fourteen capability flags, the new state keys, the console, mods, keep-awake, Remote Control, the sidebar entry, the optimiser's Primary chip and embedded dashboard, the This Mac card and the healthy Home page were exercised live and screenshotted. Not exercised live: the setup flow's Install and Update steps and every Burst mutation (keep-awake, mod and panel install, Finder shortcuts, console repair), which rest on unit tests and the consent dialog; the non-empty loose-ends path, which rests on the captured fixture because the live scan found no loose ends.

Current limits

  • Claude Burst is Andrew Baker's project; its install, update and repair run its own scripts in a Terminal after you confirm. Plexiform only reads its loo...
Read more

Plexiform 1.0.2-beta.95

Pre-release

Choose a tag to compare

@github-actions github-actions released this 19 Oct 15:52

Plexiform app update — 19 October 2026 (1.0.2-beta.95)

For macOS (Apple silicon and Intel) and Linux (x64).

What changed

  • Search past sessions. Find what you asked an AI about, across your past sessions on this Mac, from the command palette or the Memory page, with project and date filters. Results open the session's timeline. The index stays on this Mac and is scrubbed of secrets.
  • Where the money went. The Usage page breaks spend down by project, branch and pull request, compares it with the period before, and suggests one saving each week: the share of Opus turns that looked routine and what Sonnet would have cost. The weekly digest carries the top projects and the same suggestion.
  • Rules made easy. Start a rule from a template ("Tell me when a session is stuck", "Red banner when CI fails" and more), read each rule as a plain sentence, and have a rule send a desktop notification. Notifications respect quiet hours and meetings.
  • A first-run tour. Five short steps on the demo board within half a minute of setup: the widget, answering a question, rules, spend and team. Skip it any time; take it again from Help.
  • Local models in two clicks. Plexiform finds Ollama and LM Studio running on this Mac, lists their models, starts a session with the one you pick, and says what to do when none is pulled. Sessions show which model they use. You can add your own OpenAI-compatible endpoint; its key is kept in secure storage.
  • Phone. A waiting permission request on your phone's lock screen now has a Review button that opens it straight away; approving still needs your passkey. You can dictate a reply with the microphone and press Send yourself.
  • Team. A heat map of the files your team is working in, with files two people touch flagged; an automatic standup per teammate you can copy; and handoffs that can be declined with a note, with the outcome shown to the sender.
  • Create a Linear or Jira issue from a session. Plexiform drafts it from the session's handover, you edit it, and nothing is sent until you press Create.
  • Fewer interruptions in meetings. Nudges and rule notifications wait while Plexiform sees a meeting or screen share.

What has been verified

Desktop, remote, site and board test suites pass on this build (desktop 5,478 tests, board 3,278, 0 failures), with the privacy census and the hub tenancy matrix.

Current limits

  • Setting your Slack status from Plexiform is built but not available yet: it needs a Slack user permission the team's Slack app does not ask for today. The Slack runbook has the owner step.
  • Phone notification and voice changes reach phones after the team hub is updated.
  • Mac builds remain ad-hoc signed and are not Apple-notarized.
  • The limits listed for earlier betas below still apply.

Plexiform widget update — 19 October 2026 (1.0.2-beta.92)

For macOS (Apple silicon and Intel) and Linux (x64).

What changed

  • Session cards on the widget. Hovering a session on the widget's strip opens a card with the project, what it is doing and for how long, how long it has been waiting on you, this session's spend when known, and a teammate's initials for shared sessions. The card's buttons open the session, answer its waiting input, or pin it.
  • A smarter strip. A pinned session stays first, idle sessions fold into one "+N idle" chip, and teammates' sessions sit on their own row.
  • What just happened, at a glance. A small lane beside Claude shows the last three events (needs you, finished, stuck, teammate collision, CI failed) and fades each one after a few seconds. Click one to jump to it.
  • Two rings round Claude: how long the current turn has run against your usual turn, and today's spend against your budget. They hide when there is nothing to show and respect reduced motion and low power.
  • Widget settings: snap to any screen edge, not just corners; Normal, Compact or Dot size (Dot is the lamp alone); a global shortcut to summon or hide the widget; an accent colour per project; and "Quiet for 30 minutes" from the widget menu, shown by a small moon.
  • More character content: costumes witch, viking, tricorn, beret, antlers, astronaut and heartband; pets penguin, ghost, chick and owl; effects fireworks, bats, lightning, clouds, aurora and embers; poses dance, shrug and yawn. All appear in the Lights rule editor.
  • Seasonal pack: around New Year, Valentine's Day, spring, Halloween and winter the character picks up a matching costume, pet and effect, but only on channels no rule sets. Turn it off with "seasonalPack": false in config.json (the older Seasonal looks setting still works).
  • Sprite packs: put a folder with pack.json and PNG/SVG frames in ~/.plexiform/sprites/<pack>/ and set "spritePack": "<pack>" in config.json to show your own frames instead of the drawn character. Packs are checked strictly (plain file names, no links, 512 KB a frame, 64 frames a state, 8 MB a pack; SVGs with scripts or external references are refused) and never run; any problem falls back to the drawn character.
  • Keep-awake on the Burst card now says "Saved, not applied yet" while Burst's machine-wide half waits for the password step in Terminal, instead of "On".

What has been verified

Desktop, remote and site test suites pass on this build (desktop 5,430 tests, 0 failures). The visual regression suite was run on this build and on beta.91's source: the only difference is the character baselines, which grew to include the new costumes and poses. The suite's other failing screenshots fail identically on beta.91's source and are tracked separately.

Current limits

  • Mac builds remain ad-hoc signed and are not Apple-notarized. A changed build may require fresh authorization for Plexiform's own permissions.
  • Sprite packs and project colours are set per Mac; they do not sync.
  • The limits listed for beta.91 below still apply.

Plexiform Claude Burst preview — 19 October 2026

For macOS (Apple silicon and Intel) and Linux (x64). Claude Burst itself is macOS only.

What changed

  • Claude Burst inside Plexiform, verified against Burst v0.20.36. Plexiform now reads Burst's newer state (primary health, Remote Control expectation, hosts entry, certificate trust) and gates every feature on a capability flag, so a 0.19 install keeps working with the newer parts hidden rather than broken.
  • Usage & Burst in the sidebar. One entry opens the Usage optimiser: Burst's status, its dashboard in a locked-down view, and the tools it adds.
  • Guided setup. The optimiser's empty state walks through Install, Trust, Enable and First check, says what each step changes on your Mac (from Burst's own README), and reads progress back from Burst while a step runs instead of sending you to a Terminal and asking you to come back.
  • One cause, one action. Every Burst state names what is wrong and offers one next step: not installed → Install; too old → Update; gateway not answering → Repair, with the last known state shown dimmed and its age; certificate or hosts entry not trusted → Fix; console down → Start console; usage panel not installed → Install. An unknown listener on Burst's port is named and nothing is read from it. Every Burst call from a page answers {ok, reason} and never throws.
  • Home tells you when to look. A single Burst row appears in Home's attention list only when something needs you: broken, too old, not trusted, not responding after several checks, an update available, a limit near, or loose ends. A healthy Burst, an intentionally off Burst or no Burst at all adds nothing.
  • Loose ends on Overview. Each repository row carries Burst's counts of uncommitted files, unpushed commits and stashes for repositories your sessions use; file names only, never contents, diffs or stash names.
  • Keep-awake, Remote Control and mods on the Burst card and the Tools tab. The card shows whether this Mac is being kept awake and whether Remote Control is expected, and lists Burst's session mod and the usage panel (claude-code-cost-sidebar) with Install or Update per mod. Every change to the Mac goes through the same native consent sheet as before. Burst's hotspot settings are never read or written by Plexiform.

What has been verified

Desktop, remote and site test suites pass on this build (desktop 5,340 tests, 0 failures) including a contract test that drives every Burst channel with a failing client, a 0.19 and a 0.20.36 fixture captured from a real install, and the privacy flow checks. On a Mac running the real Claude Burst v0.20.36: detection, all fourteen capability flags, the new state keys, the console, mods, keep-awake, Remote Control, the sidebar entry, the optimiser's Primary chip and embedded dashboard, the This Mac card and the healthy Home page were exercised live and screenshotted. Not exercised live: the setup flow's Install and Update steps and every Burst mutation (keep-awake, mod and panel install, Finder shortcuts, console repair), which rest on unit tests and the consent dialog; the non-empty loose-ends path, which rests on the captured fixture because the live scan found no loose ends.

Current limits

  • Claude Burst is Andrew Baker's project; its install, update and repair run its own scripts in a Terminal after you confirm. Plexiform only reads its loopback admin API through a field-by-field whitelist.
  • Mac builds remain ad-hoc signed and are not Apple-notarized; Apple Developer ID signing is still open. A changed build may require fresh authorization for Plexiform's own permissions.
  • Public registration remains closed; invited teams use the sign-in methods offered by their hub.
  • Real second-person acceptance of beta.90's team features, external teammate/device, real-device updater and rollback acceptance remain open, as recorded for beta.90.

Install

Use https://plexiform.dev/download for the published beta. Release assets include SHA256SUMS.txt. A signed update manifest verifies release bytes; it does not establish platform code signing or a completed device update check.

Plexiform 1.0.2-beta.92

Pre-release

Choose a tag to compare

@github-actions github-actions released this 19 Oct 13:57

Plexiform widget update — 19 October 2026 (1.0.2-beta.92)

For macOS (Apple silicon and Intel) and Linux (x64).

What changed

  • Session cards on the widget. Hovering a session on the widget's strip opens a card with the project, what it is doing and for how long, how long it has been waiting on you, this session's spend when known, and a teammate's initials for shared sessions. The card's buttons open the session, answer its waiting input, or pin it.
  • A smarter strip. A pinned session stays first, idle sessions fold into one "+N idle" chip, and teammates' sessions sit on their own row.
  • What just happened, at a glance. A small lane beside Claude shows the last three events (needs you, finished, stuck, teammate collision, CI failed) and fades each one after a few seconds. Click one to jump to it.
  • Two rings round Claude: how long the current turn has run against your usual turn, and today's spend against your budget. They hide when there is nothing to show and respect reduced motion and low power.
  • Widget settings: snap to any screen edge, not just corners; Normal, Compact or Dot size (Dot is the lamp alone); a global shortcut to summon or hide the widget; an accent colour per project; and "Quiet for 30 minutes" from the widget menu, shown by a small moon.
  • More character content: costumes witch, viking, tricorn, beret, antlers, astronaut and heartband; pets penguin, ghost, chick and owl; effects fireworks, bats, lightning, clouds, aurora and embers; poses dance, shrug and yawn. All appear in the Lights rule editor.
  • Seasonal pack: around New Year, Valentine's Day, spring, Halloween and winter the character picks up a matching costume, pet and effect, but only on channels no rule sets. Turn it off with "seasonalPack": false in config.json (the older Seasonal looks setting still works).
  • Sprite packs: put a folder with pack.json and PNG/SVG frames in ~/.plexiform/sprites/<pack>/ and set "spritePack": "<pack>" in config.json to show your own frames instead of the drawn character. Packs are checked strictly (plain file names, no links, 512 KB a frame, 64 frames a state, 8 MB a pack; SVGs with scripts or external references are refused) and never run; any problem falls back to the drawn character.
  • Keep-awake on the Burst card now says "Saved, not applied yet" while Burst's machine-wide half waits for the password step in Terminal, instead of "On".

What has been verified

Desktop, remote and site test suites pass on this build (desktop 5,430 tests, 0 failures). The visual regression suite was run on this build and on beta.91's source: the only difference is the character baselines, which grew to include the new costumes and poses. The suite's other failing screenshots fail identically on beta.91's source and are tracked separately.

Current limits

  • Mac builds remain ad-hoc signed and are not Apple-notarized. A changed build may require fresh authorization for Plexiform's own permissions.
  • Sprite packs and project colours are set per Mac; they do not sync.
  • The limits listed for beta.91 below still apply.

Plexiform Claude Burst preview — 19 October 2026

For macOS (Apple silicon and Intel) and Linux (x64). Claude Burst itself is macOS only.

What changed

  • Claude Burst inside Plexiform, verified against Burst v0.20.36. Plexiform now reads Burst's newer state (primary health, Remote Control expectation, hosts entry, certificate trust) and gates every feature on a capability flag, so a 0.19 install keeps working with the newer parts hidden rather than broken.
  • Usage & Burst in the sidebar. One entry opens the Usage optimiser: Burst's status, its dashboard in a locked-down view, and the tools it adds.
  • Guided setup. The optimiser's empty state walks through Install, Trust, Enable and First check, says what each step changes on your Mac (from Burst's own README), and reads progress back from Burst while a step runs instead of sending you to a Terminal and asking you to come back.
  • One cause, one action. Every Burst state names what is wrong and offers one next step: not installed → Install; too old → Update; gateway not answering → Repair, with the last known state shown dimmed and its age; certificate or hosts entry not trusted → Fix; console down → Start console; usage panel not installed → Install. An unknown listener on Burst's port is named and nothing is read from it. Every Burst call from a page answers {ok, reason} and never throws.
  • Home tells you when to look. A single Burst row appears in Home's attention list only when something needs you: broken, too old, not trusted, not responding after several checks, an update available, a limit near, or loose ends. A healthy Burst, an intentionally off Burst or no Burst at all adds nothing.
  • Loose ends on Overview. Each repository row carries Burst's counts of uncommitted files, unpushed commits and stashes for repositories your sessions use; file names only, never contents, diffs or stash names.
  • Keep-awake, Remote Control and mods on the Burst card and the Tools tab. The card shows whether this Mac is being kept awake and whether Remote Control is expected, and lists Burst's session mod and the usage panel (claude-code-cost-sidebar) with Install or Update per mod. Every change to the Mac goes through the same native consent sheet as before. Burst's hotspot settings are never read or written by Plexiform.

What has been verified

Desktop, remote and site test suites pass on this build (desktop 5,340 tests, 0 failures) including a contract test that drives every Burst channel with a failing client, a 0.19 and a 0.20.36 fixture captured from a real install, and the privacy flow checks. On a Mac running the real Claude Burst v0.20.36: detection, all fourteen capability flags, the new state keys, the console, mods, keep-awake, Remote Control, the sidebar entry, the optimiser's Primary chip and embedded dashboard, the This Mac card and the healthy Home page were exercised live and screenshotted. Not exercised live: the setup flow's Install and Update steps and every Burst mutation (keep-awake, mod and panel install, Finder shortcuts, console repair), which rest on unit tests and the consent dialog; the non-empty loose-ends path, which rests on the captured fixture because the live scan found no loose ends.

Current limits

  • Claude Burst is Andrew Baker's project; its install, update and repair run its own scripts in a Terminal after you confirm. Plexiform only reads its loopback admin API through a field-by-field whitelist.
  • Mac builds remain ad-hoc signed and are not Apple-notarized; Apple Developer ID signing is still open. A changed build may require fresh authorization for Plexiform's own permissions.
  • Public registration remains closed; invited teams use the sign-in methods offered by their hub.
  • Real second-person acceptance of beta.90's team features, external teammate/device, real-device updater and rollback acceptance remain open, as recorded for beta.90.

Install

Use https://plexiform.dev/download for the published beta. Release assets include SHA256SUMS.txt. A signed update manifest verifies release bytes; it does not establish platform code signing or a completed device update check.

Plexiform 1.0.2-beta.91

Pre-release

Choose a tag to compare

@github-actions github-actions released this 19 Oct 11:04

Plexiform Claude Burst preview — 19 October 2026

For macOS (Apple silicon and Intel) and Linux (x64). Claude Burst itself is macOS only.

What changed

  • Claude Burst inside Plexiform, verified against Burst v0.20.36. Plexiform now reads Burst's newer state (primary health, Remote Control expectation, hosts entry, certificate trust) and gates every feature on a capability flag, so a 0.19 install keeps working with the newer parts hidden rather than broken.
  • Usage & Burst in the sidebar. One entry opens the Usage optimiser: Burst's status, its dashboard in a locked-down view, and the tools it adds.
  • Guided setup. The optimiser's empty state walks through Install, Trust, Enable and First check, says what each step changes on your Mac (from Burst's own README), and reads progress back from Burst while a step runs instead of sending you to a Terminal and asking you to come back.
  • One cause, one action. Every Burst state names what is wrong and offers one next step: not installed → Install; too old → Update; gateway not answering → Repair, with the last known state shown dimmed and its age; certificate or hosts entry not trusted → Fix; console down → Start console; usage panel not installed → Install. An unknown listener on Burst's port is named and nothing is read from it. Every Burst call from a page answers {ok, reason} and never throws.
  • Home tells you when to look. A single Burst row appears in Home's attention list only when something needs you: broken, too old, not trusted, not responding after several checks, an update available, a limit near, or loose ends. A healthy Burst, an intentionally off Burst or no Burst at all adds nothing.
  • Loose ends on Overview. Each repository row carries Burst's counts of uncommitted files, unpushed commits and stashes for repositories your sessions use; file names only, never contents, diffs or stash names.
  • Keep-awake, Remote Control and mods on the Burst card and the Tools tab. The card shows whether this Mac is being kept awake and whether Remote Control is expected, and lists Burst's session mod and the usage panel (claude-code-cost-sidebar) with Install or Update per mod. Every change to the Mac goes through the same native consent sheet as before. Burst's hotspot settings are never read or written by Plexiform.

What has been verified

Desktop, remote and site test suites pass on this build (desktop 5,340 tests, 0 failures) including a contract test that drives every Burst channel with a failing client, a 0.19 and a 0.20.36 fixture captured from a real install, and the privacy flow checks. On a Mac running the real Claude Burst v0.20.36: detection, all fourteen capability flags, the new state keys, the console, mods, keep-awake, Remote Control, the sidebar entry, the optimiser's Primary chip and embedded dashboard, the This Mac card and the healthy Home page were exercised live and screenshotted. Not exercised live: the setup flow's Install and Update steps and every Burst mutation (keep-awake, mod and panel install, Finder shortcuts, console repair), which rest on unit tests and the consent dialog; the non-empty loose-ends path, which rests on the captured fixture because the live scan found no loose ends.

Current limits

  • Claude Burst is Andrew Baker's project; its install, update and repair run its own scripts in a Terminal after you confirm. Plexiform only reads its loopback admin API through a field-by-field whitelist.
  • Mac builds remain ad-hoc signed and are not Apple-notarized; Apple Developer ID signing is still open. A changed build may require fresh authorization for Plexiform's own permissions.
  • Public registration remains closed; invited teams use the sign-in methods offered by their hub.
  • Real second-person acceptance of beta.90's team features, external teammate/device, real-device updater and rollback acceptance remain open, as recorded for beta.90.

Install

Use https://plexiform.dev/download for the published beta. Release assets include SHA256SUMS.txt. A signed update manifest verifies release bytes; it does not establish platform code signing or a completed device update check.

Plexiform 1.0.2-beta.90

Pre-release

Choose a tag to compare

@github-actions github-actions released this 19 Oct 08:24

Plexiform 1.0.2-beta.89 — 18 October 2026

For macOS (Apple silicon and Intel) and Linux (x64).

What changed

  • A calmer charcoal-and-copper interface, clearer navigation, and a new woven app icon across the desktop, installer, website and phone assets. The optional desktop characters and status lamps remain available.
  • Find shared sessions by work, teammate or AI. Focus on sessions that are working, need input or have stale reports. The selected session shows the exact identity behind its available actions.
  • Useful task overviews bring the request, reported progress, changed files, reported checks, decisions and next action together. Open that task's conversation or saved context directly.
  • Save a report or note with a task and keep file or commit references alongside it. Reports are restored with the task and tied to checkpoint versions; file references point to current files, not frozen copies.
  • Save interactive checklists, metrics and small tables in a task overview. People can edit cards and tick checklist items; managed AIs can save reports and cards to their own task. Cards keep their source and saved checkpoint version across restart, and conflicting edits retain your draft for review. Metrics and tables are saved data, not live calculations or executable actions.
  • Take over an eligible managed Claude Code or Codex task in Terminal, continue its exact session, then exit and hand it back to Tasks. Plexiform stops the background run first, shows the existing permission context, and requests a single Terminal launch. A launch request is not an AI response; new tasks still start in the background.
  • Before starting local work, inspect other known Tasks in the project, review possible overlap, narrow the request or keep the draft while waiting. These local checks are advisory; waiting does not start work automatically.
  • Managed Claude Code and Codex tasks can request bounded help from each other, exchange messages and read reported results. Helpers use separate working copies of the same repository, committed starting context and the same or narrower permissions. They cannot discover or control arbitrary private sessions.
  • Collaboration is limited to four helpers per task family and two levels. Durable request records prevent automatic duplicate effects after uncertain outcomes. Stored, delivered and answered remain distinct states; messages do not grant permissions or approvals. Task families with spend caps cannot delegate until aggregate budget enforcement exists.
  • Reviewed integration checks current task ownership and saved context before applying changes. Reviewed provider handovers, scheduled follow-ups and draft pull-request workflows preserve their existing permission and confirmation boundaries.
  • Each shared session card now says what its agent is doing right now, in plain words, with a four-step freshness (fresh, idle, stale, offline). Requests are held back from sessions that have stopped reporting, with the reason shown.
  • Pin the teammates and sessions you care about to the top, and save your filter combinations as named views. Pins and saved views are yours alone and never change what is shared.
  • Follow a teammate's shared session and be told once when it needs input, finishes or fails. Followed events appear in your Home Inbox until you mark them seen.
  • Ask a shared session for a status update, a review of your change, its current plan, or to pause at a safe point. Each template drafts the exact text into the existing reviewed composer; nothing is sent until you press Send, and the card shows sent, acknowledged, answered or expired.
  • Press ⌘K to find any teammate, project or shared session and open the exact one. Only work already shared with you is ever listed.
  • Home gains an Inbox: approvals, questions, blockers, failed tasks, teammates' handoffs, requests you are waiting on and followed sessions, ordered by urgency with one action each. New teammates get a four-step tour, shown once per account.
  • "Can you look at this?" sends a teammate a review request tied to your session. They see it in Agent mail with Seen, I'll handle this, reply, Done and Decline; you see its state under Waiting on, on the Sessions page and on Home.
  • A project pinboard on each team board keeps the staging link, designs, useful commands, conventions and current priorities together. Any writer can pin or remove; pin straight from one of your sessions in Plexiform.
  • Session Timeline can share a scrubbed summary of a session with its saved result cards and provenance to the team card, for repositories you opted in to handover sharing. The control appears only when a team hub supports it.

What has been verified

Build 1.0.2-beta.89 (source ff20fc7) is the first candidate carrying the shared-workspace team features. Their checks so far: unit and page tests for the doing-now line, pins and saved views, follow events, Ask templates, the ⌘K palette, Home Inbox and tour; a full "Can you look at this?" exchange (sent, seen, I'll handle this, reply, done) against an in-process team hub; pinboard add, replace, remove and viewer refusal against the hub test tenancy; the privacy census for the new flows. Installed-app and two-person acceptance of these features are recorded separately when they happen. A real local Codex task requested a Claude review, received an acknowledgement and report, and exchanged a follow-up through the task connection. That exchange used normal task permissions and an explicit read-only review request. Claude's native Plan mode blocked the messaging tools; successful messaging in that mode is not claimed. This is local provider evidence, not installed-app, two-person or cross-device acceptance. A real local Codex task also saved a result card through the task connection; the card and its agent provenance survived an engine restart. A separate unpackaged Electron check verified all three card types, saving and restoration, and preservation of a conflicting draft with a fake provider. A controlled native check opened Apple Terminal once and preserved the selected session and working folder. This used a fixture provider; a launch request does not confirm a real model response. Installed acceptance of this build on the owner's Mac: 1.0.2-beta.89 installed from the staged, signature-verified ARM64 archive; the team account and board restored on launch; a real Codex Plan-only task saved a checklist card, the card was edited (version 7) and survived a full app restart; on a second real Codex task, Take over in terminal and Open in Terminal each required a native confirmation, Apple Terminal resumed the exact Codex session in the task worktree, the model answered, the session exited normally and Hand back returned the task to the background run. The staging and production team hubs run this source commit. One known rough edge: the Open in Terminal confirmation sheet can grow taller than the screen because it repeats the full command; the fix is in the next build. Two-person and cross-device acceptance of the team features, and publication, are recorded separately.

Current limits

  • Viewer presence on a session is shown only when the team hub publishes it; this build's hub does not yet. The Timeline summary share appears only when a hub connection for it is configured.
  • Existing sessions need a supported connection and sharing permission. Activity visibility alone is not a messaging channel. The Codex CLI connection does not control Codex desktop conversations.
  • Separate working copies and checked integration protect work managed by Plexiform. Independent agents can still conflict when editing the same shared checkout outside those boundaries; overlap warnings are not universal file locks.
  • Undo preserves later edits and refuses deleted-file or mode reconstruction when original metadata is unavailable.
  • External teammate/device, real-device updater and rollback acceptance remain open. Windows downloads, public phone access and Linear/Jira signed delivery stay gated.
  • Mac builds remain ad-hoc signed and are not Apple-notarized. A changed build may require fresh authorization for Plexiform / Plexiform Safe Storage in the system Keychain dialog.
  • Public registration remains closed; invited teams use the sign-in methods offered by their hub. Provider installation, sign-in, subscriptions and usage charges remain separate. Agent Mail remains free; inbox polling is off by default and sending requires confirmation.
  • Privacy/legal review and an off-device gallery-key backup remain outstanding. No universal cross-platform completion claim is made by this candidate.

Install

Use https://plexiform.dev/download for the published beta. Release assets include SHA256SUMS.txt. A signed update manifest verifies release bytes; it does not establish platform code signing or a completed device update check.

Plexiform 1.0.2-beta.89

Pre-release

Choose a tag to compare

@github-actions github-actions released this 18 Oct 19:59

Plexiform shared workspace preview — 18 October 2026

For macOS (Apple silicon and Intel) and Linux (x64).

What changed

  • A calmer charcoal-and-copper interface, clearer navigation, and a new woven app icon across the desktop, installer, website and phone assets. The optional desktop characters and status lamps remain available.
  • Find shared sessions by work, teammate or AI. Focus on sessions that are working, need input or have stale reports. The selected session shows the exact identity behind its available actions.
  • Useful task overviews bring the request, reported progress, changed files, reported checks, decisions and next action together. Open that task's conversation or saved context directly.
  • Save a report or note with a task and keep file or commit references alongside it. Reports are restored with the task and tied to checkpoint versions; file references point to current files, not frozen copies.
  • Save interactive checklists, metrics and small tables in a task overview. People can edit cards and tick checklist items; managed AIs can save reports and cards to their own task. Cards keep their source and saved checkpoint version across restart, and conflicting edits retain your draft for review. Metrics and tables are saved data, not live calculations or executable actions.
  • Take over an eligible managed Claude Code or Codex task in Terminal, continue its exact session, then exit and hand it back to Tasks. Plexiform stops the background run first, shows the existing permission context, and requests a single Terminal launch. A launch request is not an AI response; new tasks still start in the background.
  • Before starting local work, inspect other known Tasks in the project, review possible overlap, narrow the request or keep the draft while waiting. These local checks are advisory; waiting does not start work automatically.
  • Managed Claude Code and Codex tasks can request bounded help from each other, exchange messages and read reported results. Helpers use separate working copies of the same repository, committed starting context and the same or narrower permissions. They cannot discover or control arbitrary private sessions.
  • Collaboration is limited to four helpers per task family and two levels. Durable request records prevent automatic duplicate effects after uncertain outcomes. Stored, delivered and answered remain distinct states; messages do not grant permissions or approvals. Task families with spend caps cannot delegate until aggregate budget enforcement exists.
  • Reviewed integration checks current task ownership and saved context before applying changes. Reviewed provider handovers, scheduled follow-ups and draft pull-request workflows preserve their existing permission and confirmation boundaries.

What has been verified

A real local Codex task requested a Claude review, received an acknowledgement and report, and exchanged a follow-up through the task connection. That exchange used normal task permissions and an explicit read-only review request. Claude's native Plan mode blocked the messaging tools; successful messaging in that mode is not claimed. This is local provider evidence, not installed-app, two-person or cross-device acceptance. A real local Codex task also saved a result card through the task connection; the card and its agent provenance survived an engine restart. A separate unpackaged Electron check verified all three card types, saving and restoration, and preservation of a conflicting draft with a fake provider. A controlled native check opened Apple Terminal once and preserved the selected session and working folder. This used a fixture provider; a launch request does not confirm a real model response. Final packaging, installed acceptance and publication are recorded separately.

Current limits

  • Existing sessions need a supported connection and sharing permission. Activity visibility alone is not a messaging channel. The Codex CLI connection does not control Codex desktop conversations.
  • Separate working copies and checked integration protect work managed by Plexiform. Independent agents can still conflict when editing the same shared checkout outside those boundaries; overlap warnings are not universal file locks.
  • Undo preserves later edits and refuses deleted-file or mode reconstruction when original metadata is unavailable.
  • External teammate/device, real-device updater and rollback acceptance remain open. Windows downloads, public phone access and Linear/Jira signed delivery stay gated.
  • Mac builds remain ad-hoc signed and are not Apple-notarized. A changed build may require fresh authorization for Plexiform / Plexiform Safe Storage in the system Keychain dialog.
  • Public registration remains closed; invited teams use the sign-in methods offered by their hub. Provider installation, sign-in, subscriptions and usage charges remain separate. Agent Mail remains free; inbox polling is off by default and sending requires confirmation.
  • Privacy/legal review and an off-device gallery-key backup remain outstanding. No universal cross-platform completion claim is made by this candidate.

Install

Use https://plexiform.dev/download for the published beta. Release assets include SHA256SUMS.txt. A signed update manifest verifies release bytes; it does not establish platform code signing or a completed device update check.

Plexiform 1.0.2-beta.82

Pre-release

Choose a tag to compare

@github-actions github-actions released this 15 Oct 19:49

Plexiform 1.0.2-beta.82 public beta — 15 October 2026

This preview is available for macOS (Apple silicon and Intel) and Linux (x64).

What changed

  • Restoring a saved team account while Home, Overview or another local page is open refreshes the cached board target without changing the page you are using. Opening Board afterward shows the restored team's board; an explicit choice to use the local board stays local.
  • Team-account restore and save use asynchronous secure storage on macOS and Windows. Saved sign-ins restore after app readiness while local navigation stays usable; a sign-in succeeds only after its credential is saved securely.
  • On macOS, team-account cryptography runs in a disposable helper so a pending Keychain authorization cannot hold normal Quit open. The helper verifies its running Plexiform parent before accessing secure storage. Quitting retires the helper, and late completions cannot restore an account or resume a runner.
  • The sidebar explains when a saved sign-in is unlocking or being saved. If secure storage is locked, it shows recovery guidance and an explicit Retry saved sign-in action. Ordinary reads do not repeatedly retry a denied authorization. Restoring an account preserves the selected page and remembered workspace, including a local-board choice made while waiting.
  • Cancelling a browser sign-in while secure saving is pending discards its result and requests revocation of the newly issued device token. Signing out prevents a late restore or save from bringing the account back; a cancelled attempt cannot clear a newer sign-in.
  • Home's optional starter-pack detection reads recent project folders in a separate helper process with a deadline. An unavailable folder cannot freeze navigation or leave shutdown waiting for a blocked thread. Dismissed or unnecessary suggestions do not scan those folders.
  • Open a shared Team session to see its task, current reported activity, messages, exact-text receipts and streamed output. Send to the selected session, steer its current turn or request interruption where the provider and explicit share allow it. Every final action is bound to the reviewed session generation and turn.
  • Codex and Claude report fixed activity categories such as thinking, reading, searching and editing. Tool arguments and raw commands are not included. A steering receipt is separate from evidence that the requested correction was applied; its output view starts after that message.
  • Busy sessions without steering let you keep a local draft until sending becomes available. Drafts are not queued or sent automatically. Refused and uncertain actions preserve drafts and are never retried automatically.
  • Watch-only shares remain read-only. Changed accounts, sessions, turns, team views and revoked shares retire pending controls. Sharing excludes earlier history, including same-millisecond history and immediate steering receipts. Passive directory caching keeps larger teams within the existing relay request limits; the selected conversation reads fresh state.
  • Captured AI-session cards now show recorded Activity, editable Task context, exact-session controls, reported files, and relevant Handover actions. Context can be imported into a draft, reviewed, saved, and refreshed without losing edits.
  • Handovers support refresh, copy, native Markdown export, continuation in another supported tool, and timeline navigation. Historical handover text is readable. Comments offer agent delivery only for a current managed run.
  • Session actions recheck the originating window, document, team, workspace, account, and sharing scope after asynchronous work. Native export opens from the correct board window.
  • Optional Burst configuration reads are asynchronous and bounded, reducing a source of desktop startup stalls. Terminal memory-worker shutdown cannot reopen after an explicit stop.
  • Agent Mail is free, including Ask my AI. Inbox polling is off by default, and sending still requires confirmation.
  • Claude Tasks is available on macOS and Linux; Codex remains the default provider. AI tools require their own installation, sign-in, permissions, and provider subscriptions or usage credits.

Current limits

  • Existing sessions need a supported, explicitly enabled connection. The Codex shared daemon supports explicitly attached CLI sessions; it does not reach Codex desktop conversations. Observed activity hooks alone do not provide a messaging channel. Claude/Gemini/local adapters do not all support current-turn steering; only available actions are offered.
  • Real Codex and Claude disposable-session checks passed. Desktop shared-session send, steer and interrupt passed through the real transport with local test accounts and real Codex. Two independent external teammates/devices and every provider/device combination have not been accepted yet. Durable next-turn executable queues, owner-reviewed proposals and granular new delegation records are not included in this preview.
  • Plexiform provides session visibility, shared context, handovers, and isolation for managed Tasks. It does not yet universally prevent independent agents from overwriting work in a shared checkout. Ownership declarations and captured-session collision alerts are advisory.
  • Windows downloads remain unavailable pending runtime and privacy acceptance. Public phone access and Linear/Jira delivery remain gated.
  • Public account registration remains closed. Existing invited teams use Google or GitHub sign-in; production email sign-in is not enabled.
  • Mac builds use ad-hoc signing and are not Apple-notarized. Follow the installation guidance at https://plexiform.dev/download; verify the download checksum.
  • A newly installed ad-hoc Mac build can require a fresh owner approval in the macOS Keychain dialog. Verify that it names Plexiform and Plexiform Safe Storage, and complete that approval directly on your Mac.
  • The asynchronous account change does not make every credential operation nonblocking. Runner tokens and local-board credentials retain their existing storage paths, and Linux retains its guarded desktop password-store backend. Windows runtime acceptance remains open.
  • macOS folder permissions can affect filesystem features. Normal installed beta.67 Home navigation stalled on synchronous directory scanning; the follow-up isolates the optional starter scan with a deadline and process cleanup. Other earlier file-access stalls do not all have a proven cause. Linux device acceptance beyond CI continues.
  • A signed update feed verifies release bytes; it does not establish Apple/Windows platform signing or a completed real-device update acceptance check.
  • Privacy/legal review and the external teammate, device and provider acceptance gates remain open.

Install
Download the installer for your platform at https://plexiform.dev/download. SHA-256 checksums are included in SHA256SUMS.txt. The one-line installer selects the newest published beta by default; PLEXIFORM_CHANNEL=stable selects the stable channel.