Repository navigation
v0.51.0
Added
-
Sandboxes: a task in a fresh box of its own. Type /sandbox and a task in New task, and it runs in a new box on your wisp Grid instead of on this computer or in your cloud computer. Your project goes along as it is here, uncommitted changes included, so private repos need no login there. Deleting the task deletes its sandbox. Inside, the agent has root (
sudo apt-get, under gVisor), and what it installs comes back when the sandbox wakes. It also has a real browser it drives through the new Browser connector. You watch that browser live, and can click in it, in the task's new Browser tab in the side panel. A project's.wisp/sandbox.shruns once when a sandbox gets it. You sign in to Claude once for all sandboxes in Settings → Cloud → Sandboxes (in a cloud computer: Settings → Connect). wisp runsclaude setup-token, and the token goes straight to your grid, sealed there and never kept on your computer. The same card has Env vars every sandbox starts with, and lists every sandbox with what it's for. Subagents can run in a sandbox of their own: pick Sandbox as a delegate's place. Crew members can run each job in a fresh sandbox, or keep one of their own: Their jobs run on → A sandbox per job / A sandbox of their own. Only Claude Code (or Auto) runs in sandboxes for now. Previews of a box's pages now load inside wisp on a computer too (their cookie works in a frame of another site). Grid: up to 10 sandboxes per user (sandboxes.perUser), 2.5 GB each (hostd.container.sandbox); see grid/README.md → Sandboxes. Spells: thesettings.cloud.sandboxesandpanel.browserparts, thebrowserpanel tab and theboxicon. -
Drafts, and tasks that start after another. Type /draft and a task in New task (or in any task's reply box) to save it without starting it: it waits in the new Drafts section, above For later. Click it to edit the prompt, then Start now, Start later… or Delete. Type /after, pick a task (a menu searches your tasks as you type; pasting a task's ID works too), then the prompt: the new task waits in Scheduled and starts once that task is marked done (at once if it already is). From a task's reply box it runs in that task's project, on its harness. If the task it waits for is deleted or archived first, it moves to Drafts. API:
draft: trueorafter: <task id>instead of a time onPOST /later. Spells: thesection.draftsanddrafts.rowparts. -
Run the task's app and see it beside the task. The side panel has a Preview tab (after Shell) on a task with a worktree: Run starts the app from the worktree, with
PORTset to a free port wisp keeps for the task, and shows it in the panel. The command comes fromrunin.wisp/worktree.json(with an optionalport), else a guess from package.json (npm run dev, elsenpm start, said to be a guess), and you can edit it before you run. While it runs, a slim bar says where (Running · localhost:41664) with Pick element, Logs, Restart, Stop and Open in browser; until the app answers, its output shows in its place, and if it exits it says the exit code and its last lines with Run again. Pick element outlines what's under the pointer in the page; a click adds a line naming it (In the preview,<button class="primary">"Save" (selector:main > form > button.primary):) to the task's reply box. Stop ends everything the command started; deleting or archiving the task, or quitting wisp, stops it too. Also Run app / Stop app in the task's More, its right-click menu and ⌘K. NewGET/POST/DELETE /api/sessions/{id}/runandGET /api/sessions/{id}/run/log; spells reach it through thepanel.preview,panel.preview.bar,panel.preview.run,panel.preview.pickandpanel.preview.logsparts andpanel'stab: "preview". -
Start tasks from Slack, Linear or any webhook. Automations → Triggers has three new kinds: When a webhook is called, When someone mentions wisp in Slack and When a Linear issue is created. Each gets a secret URL (Copy, Regenerate) to paste where the calls come from: anything that can POST (its JSON or form body fills the prompt through
{{body.ref}}and{{text}}, optionally signed with an HMAC secret), your Slack app's Event Subscriptions (the signing secret is checked and old requests are refused; mentions and messages in one channel, optionally only with a keyword;{{text}},{{user}},{{channel}}), or a Linear webhook (its signature checked; issues created, optionally only with a label;{{title}},{{url}},{{description}}). They start in Plan mode, at most one task per trigger every 30 seconds, and the trigger's editor lists its last five calls with what came of them (started, skipped or rejected, and why). Off by default: Settings → Automations → Webhook triggers; off, the URLs answer 404. On wisp Cloud the URL works as it is; on your computer, put a tunnel in front of it. NewPOST /api/hooks/{token}; spells set the switch withwebhookTriggersand reach the rows through thesettings.webhookTriggers,automations.hookUrlandautomations.deliveriesparts. -
Start a task from an issue. New task has an Issue button at the end of its project row (in the phone app, in the bar after Delegate; behind More options until your first task) that lists the picked project's open GitHub issues, newest first: #123 title, its labels and who it's assigned to. Type to filter, or past the 50 newest to search GitHub. Picking one fills the prompt with Fix #123: title, the link and the issue's text, the caret at the end, so you can add to it before you send; a draft it replaces comes back with Undo. With a Linear connector that has an API key, the same picker lists the Linear issues assigned to you; without one it offers Connect Linear… (Settings → Connectors). When gh isn't installed or signed in, or the project has no GitHub remote, the picker says so with the button that fixes it. Also in ⌘K: Start from an issue…. New
GET /api/issues?project=&q=&source=github|linear; spells reach it through thelauncher.issueandissues.pickerparts and theissueicon. -
Stage and commit from Changes. In a worktree task's Changes, each file has a + before Viewed to stage it (− on a tint once it's all staged, a dot on the + when part of it is), and each hunk a Stage / Unstage beside its Revert, so you pick what goes in a commit as in VS Code's Source Control. A commit row under the summary takes a one-line message (the task's title when you focus it empty), says 2 staged, and Commit (Enter) commits those, or everything when nothing is staged; the toast says Committed abc1234. Every flip is instant and comes back with a red toast if git refuses; nothing is staged or committed while the agent works. ⌘K on a task has Commit staged changes…. Spells reach the toggle, the hunk's button and the commit row (
panel.changes.stage,panel.changes.stageHunk,panel.changes.commit); the API isPOST /sessions/:id/changes/stageand/changes/commit. -
Review Changes like a pull request: revert a hunk or a file, tick files Viewed. In the side panel's Changes, hover a hunk of a diff and click Revert on its first line (always shown on a phone) to put just that part back as it was before the task; a file's
⋯has Revert file…, which asks first, then brings a changed or deleted file back from where the branch left the base, or removes a file the task added. Both happen at once with Reverted and Undo on the toast, which puts the file back exactly as it was. wisp won't revert while the agent is working on the task, or a part that changed since you looked (Changes reloads instead). Each file also has a Viewed box, as on GitHub: ticking it folds the file and dims its row, the summary counts 2 of 7 viewed, and a file whose diff changes afterwards is unviewed again (kept per task in this browser). NewPOST /api/sessions/{id}/changes/revert; spells reach it through thepanel.changes.revertHunk,panel.changes.viewedandpanel.changes.viewedCountparts. -
Worktrees bring what git ignores, and set themselves up. A task's new worktree now gets the ignored files the repo's
.worktreeincludelists (gitignore-style lines, as Claude Code reads them; without one, the root's.envfiles), and runs the repo's setup commands from.wisp/worktree.json({ "setup": ["npm ci"] }) or a Cursor.cursor/worktrees.json(setup-worktree), in the worktree, withROOT_WORKTREE_PATHset to your checkout. A step that fails stops the rest and the task's branch line says Worktree setup failed (its tooltip says why); the branch's tooltip lists what was copied and run. Spells reach the warning through thecard.setupWarnpart. -
Your plan's usage before you start. The harness picker in New task (and a reply's) now says how much of each harness's plan is gone, its fullest window (5h 40% used), not only when it's near your line; the tooltip says when it resets.
-
@ a file in a prompt. Typing
@and part of a name in New task (with a project on this computer picked) or in a task's reply box now offers the folder's files too, under Files, after the crew and My Stuff;/in what you type looks in paths. Picking one puts its path in (@src/auth.js), the way the harnesses read a file mention. NewGET /api/files?project=; spells reach it through themention.pickerandmention.fileparts. -
Jira in the issue picker, and Sentry and PagerDuty triggers. New task's Issue picker lists the Jira issues assigned to you when a Jira or Atlassian connector has its site, email and API token (Jira's search with
assignee = currentUser(), still open), beside GitHub and Linear;source=jiraonGET /api/issues. Automations' new trigger offers Sentry alert and PagerDuty incident: a webhook trigger with its service's signature header (Sentry-Hook-Signature,X-PagerDuty-Signature, including PagerDuty'sv1=signatures during a secret rotation) and a prompt that reads the alert or incident's title and link. -
Language support in the editor. With a language server for the file (TypeScript and JavaScript, Python, Go, Rust: found on your PATH, in the repo's
node_modules/.bin, or installed by wisp), the editor asks it: ⌘-click or F12 goes to the exact definition, ⇧F12 lists every reference, resting the pointer on a name shows its type and docs, F2 renames it across the task’s files, Ctrl+Space completes the word you’re typing, and problems show as you type: wavy line numbers in red or amber with the message as their tooltip, and n problems in the editor's head listing them. Without a server, definition and references use the text search, and the editor offers Install TypeScript support (or Python):typescript-language-serverorpyright, into~/.wisp/lsp, once for every task. Renaming writes the files, so it waits for a saved file and an idle agent. NewPOST /api/sessions/{id}/lspandPOST /api/lsp/install; spells reach it through thepanel.editor.problemsandpanel.editor.hoverparts. -
Search the task's files, and go to a definition. The editor's Open file… (and ⌘P in the editor) finds text inside the files too: an In files group with each line, the words marked, opening the file at that line. In the editor, ⌘-click a name (Ctrl-click on Windows and Linux) or press F12 to open where it's defined: the one place at its line, or a list when several define it. New
GET /api/sessions/{id}/grep(q=text,def=a name). -
Find past work by what was said. ⌘K now searches inside every task's prompts and answers, not only titles: type three letters or more and an In prompts and answers group follows the rest with the task, where the words were (You asked or It answered) and the line around them. Every word you type must be in one prompt or one answer. New
GET /api/search?q=. -
Open any file of a task. Changes has Open file… (and ⌘K Open a file… on a task): the task folder's files by name, the ones it changed first, open in the editor. The lines the task added are marked in the editor's gutter, and the marks follow the lines you add or remove. New
GET /api/sessions/{id}/files; spells reach it through thepanel.changes.openpart. -
Edit a changed file in wisp. Each file in the side panel's Changes has a pencil (and Edit here in its ⋯): the file opens over the list, highlighted, with line numbers, Tab and Shift+Tab to indent, and your browser's own undo. Save (⌘S) writes it to the task's folder (its worktree, or the project folder without one) at once, and Changes shows your edit in the diff. It won't save while the agent is working on the task (its next edit would land on yours), and if the file changed since you opened it, it offers Load theirs or Keep mine instead of overwriting. Esc goes back to Changes, asking first if you have unsaved work; switching to another task keeps the editor with its task. New
GETandPUT /api/sessions/{id}/source; spells reach it through thepanel.changes.edit,panel.editor,panel.editor.codeandpanel.editor.saveparts. -
A shorter New task on your first run. Until you've sent a task, the box is the prompt, the mode, the harness and Send, with More options for the rest (another computer, compare, delegate, goal, worktree, start later). Once you open it or send a task, the full bar stays. Spells reach it through the
launcher.morepart. -
Comment on a line of a diff, like a review. In the side panel's Changes, hover a line of a diff and click the + that comes up over its sign (on a phone, tap the line number), say what should change, and the comment waits above the reply box with your other comments on the answer, numbered, with a mark on that line. Shift-click another line (on a phone, tap a second number while the card is open) to comment on the lines between. It goes out at the top of your next follow-up naming the file, the line numbers (as the file is now; a removed line by its old number, and it says so) and the quoted code, and the sent prompt shows it as a card (server/auth.js:42–48). Clicking it in the tray opens Changes on that file and flashes its mark. The tray now has Send comments while the box is empty. Spells reach it through the
panel.changes.comment,panel.changes.commentMarkandcomposer.sendCommentsparts. -
Pull to refresh your computers. On the Android app's first screen, pull the list down from the top and let go to ask your grids again at once, instead of waiting for the next refresh: a computer you just switched on shows Online right away. A circle follows your finger, fills as you pull, and spins until the grid answers.
-
Download on the phone. The Android app now has Download on pictures and videos in answers, in the full-screen viewer, and on files in My Stuff (it was hidden there because the app couldn't save files). Pictures go to Pictures/wisp and videos to Movies/wisp, where the gallery shows them. Other files go to Download/wisp. No storage permission is needed. On Android 8 and 9 the file opens the share sheet instead. Over a cloud computer the download uses the app's sign-in. If the app is older than the computer's wisp, the button stays hidden.
-
Add a goal to a task that's already going. The reply box has a Goal button (the target) beside the team picker. On a task without a goal it opens the same Done when line as New task: press Enter there to set the goal on its own (an idle task is checked at once), or send a message with it and the goal is checked after that follow-up's turn. While a goal is going the button is lit and opens Edit goal…; after one is met it starts a new one. Esc closes the line, and its text survives a reload. API:
goal: { text }onPOST /sessions/:id/reply. Spells: thecomposer.goalandcomposer.goalLineparts. -
Judges for a goal. A goal can have judges: up to five agents, picked like Delegate (any harness with its model and thinking level, or a crew member on theirs), that each check the work on their own after every round. Pick how their votes add up: All must say it's met, a Majority, or Any one. The goal is met only when they agree that way; can't be done takes a majority of them, and a judge that can't answer counts against All. Add them from the Judges button in the launcher's Done when line (remembered for the next goal, like delegates) or in Set goal… / Edit goal…. The goal band shows each judge's face tinted by its last vote and the count (2/3), its tooltip each judge's reason, and a round sent back says 2 of 3 judges say met (all must agree) with what the dissenting judges want next.
goal: { text, judges, vote }onPOST /sessionsandPUT /sessions/{id}/goal(judges left out keep the ones it had;[]clears them). Spells reach it through thegoal.judgesandlauncher.goalJudgesparts and thejudgeicon. -
wisp on a slow connection. The phone app's first open of a computer (and its first open after an update) no longer pays a round trip per file: wisp's ~120 modules come in one request that its service worker unpacks before the page loads them, and the 30 stylesheets are one file. On a 3G link (2 s round trips, 400 kbit/s) the task list shows in 18 s instead of 60 s, and in 6.7 s instead of 18.8 s at 600 ms and 1 Mbit/s. The page's scripts go without their comment lines (78 KB less over the wire, line numbers unchanged). Before the list is in, its rows show as placeholders the shape of the real ones, and a task opened before its copy arrives shows its bar, prompt and answer as placeholders (with its real status pills) instead of Loading…. The phone's saved snapshot keeps the tasks you read lately and the ones fetched ahead, so after Android stops the app they open at once. An open task is fetched once at a time (a running task's updates no longer pile up on a slow link, and an older answer can't paint over a newer one), and a task's load that hangs on a dead connection gives up after 45 s and tries again. Spells reach the placeholders through the
list.placeholderandsplit.placeholderparts. -
Waiting work goes when a plan resets early. OpenAI hands out banked Codex resets that land with no notice, and until now work waiting for a plan (follow-ups held by Automations stop at, tasks waiting at a limit with Wait for the reset, runs held by Ask near the limit) waited for the reset time the CLI last gave, days later for a weekly window. Now, when a run's report or a live read shows a window reset early (much less used, and a new period before the old one ended), that work goes at once. Crew members and triggers skipped for the plan stop saying so, the phone hears how many tasks started, and the window's line in Plan budget reads Reset early 10:12 · resets Sat 10:12. Settings → Usage → Check Codex while work waits (off by default): while work waits on a Codex account, wisp asks Codex's app-server how its windows stand every 10 minutes (
account/rateLimits/read: no model turn, nothing spent; it never redeems a reset credit). Claude Code has no such read, so an early Claude reset is seen with its next run. Codex's session files can end with a line for another limit (premium) with no windows; wisp now skips it instead of reading no windows. Spells set it withliveLimitsand reach the row throughsettings.liveLimits. -
Automations stop at, per harness. Settings → Usage → Plan budget now has a row for Claude Code and one for Codex under Automations stop at: Same (follow the row above), Off, 50%, 75% or 90%. OpenAI is moving ChatGPT and Codex to one shared allowance, so a lower Codex line (say 75% while Claude Code stays at 90%) leaves the rest for your own chats. Each account's bars mark its own harness's line. Spells set them with
autoStopClaudeandautoStopCodex, and reach the rows through thesettings.autoStopBypart. -
Drain watch. Settings → Usage → Drain watch (off by default) warns when an account's plan suddenly moves much faster for the same work: when its runs in the last two hours took twice or more the share of a window per unit of work (Claude Code's reported cost, else tokens) that its runs took over the two weeks before. The account's block in Plan budget gets a card: Using your plan 3.1× faster than usual · Since 14:05, 4 runs took 18% of the weekly window for $2.10 of work. That much work usually takes 6%. 1 of them ended in an error and still took 3%. Copy report puts the numbers in words for a support ticket, and Dismiss hides the card until a later run shows the drain again. With ntfy or Telegram set up, your phone hears of it too, at most once every two hours per window. New
POST /api/agents/{id}/drain/dismiss; spells set it withdrainWatchand reach it through thesettings.drainWatchandbudget.drainparts. -
Addresses on a map. Agents can answer with a
mapblock: one or more places (street addresses, or coordinates) on a Google map in the answer, each with Open in Google Maps (the Maps app on a phone). With several, the list under the map picks which one it shows, numbered so a typed answer can name one, and"route": trueadds Route: directions through them in order. Exports and chat apps get the places as Google Maps links. Spells reach it through themap,map.head,map.stageandmap.listparts. -
Babysitting mode (Terminals). Settings → General → Babysitting mode (Terminals): each task you send opens straight in the agent's own CLI, in a terminal inside wisp, already working on your prompt, instead of wisp's cards. wisp still makes its folder, worktree and title, and keeps the prompt as its turn. A follow-up from wisp's box is typed into the terminal; once the terminal ends, opening the task (or a follow-up) resumes the same conversation (Claude Code's is named by wisp with
--session-id). Schedules, crew jobs, spells, goals, teams and compare runs still run in wisp. Off by default; spells set it withterminalSessions. With it off, start a prompt with/terminal(offered in the new-task box's/menu) and that one task starts and opens in its terminal; the CLI, the API and the phone read it too. The task opens the moment you send it, on its terminal's Starting up screen, rather than after its folder and worktree are ready. -
wisp blocks off. Settings → General → wisp blocks: off, agents answer in plain markdown instead of a
wispblock of cards, and wisp shows it without the answered in plain text note. Crew jobs and spells keep their blocks (wisp acts on them). Spells set it withblocks. -
See what a task changed, beside it. The side panel has a Changes tab: every file the task changed since its branch left the base (committed or not, new files included), with A/M/D/R,
+n −nin fixed columns, and each file's diff opening in place under its row (several at once, Expand all). Lockfiles and build output fold into one row; a file's⋯has Open in editor and Copy path. It follows the agent while it works, patching rows in place. A task without a worktree shows the edits wisp captured instead, and says so. Open it from the task's Changes button,D, More, right-click, ⌘K Show changes, or All changes in this task under a turn's files. Once the panel is open, its head is now tabs (Changes · PR · Shell, and Watch once there's a page), so you move between them without going back to the task; the place line says 3 uncommitted (was 3 changed). NewGET /api/sessions/{id}/changes; spells reach it through thepanel.changes,panel.changes.summary,panel.changes.listandsplit.changesparts,panel.tab: "changes"andpanel.changes.generated. -
Share a task, and import one. Share… (in a task's menu and its More) puts the task on your wisp Grid at a link: anyone with it opens the task in a browser, drawn the way wisp draws a task page (the same reports, delegates and details column), or only people with the passkey you set, each delegated run on a page of its own. Someone signed in to the same grid pastes the link in New task: it shows as a card (who shared it, its turns, delegates, crew, connectors and skills, and what carries it on here), and Enter imports it with its history, work (branch and uncommitted changes, into their checkout of the repo), team and delegates' runs, the crew members it delegated to, its connectors (a command one starts switched off) and skills. A harness or model they don't have is swapped for the nearest one, and the card says so and lets them pick another. Secrets in the conversation are redacted; connectors never take their keys. Spells reach it through the
launcher.shareandshare.dialogparts. -
Phone app: swipe a crew member's thread to the left for its About page. The column a wide pane shows beside the thread (schedule, delegates, notebook, brief, kit) comes in under your finger as a page of its own, the same way a task's side panel does; drag it to the right, tap Back or press the phone's Back to return to the thread. Spells turn it off with
crew.page.swipe: falseand restyle it through thecrew.aboutpart. -
Phone app: swipe a task to the left for its Changes, PR and Shell. On a task's page, drag to the left and the side panel comes in under your finger, on the tab you last looked at; drag it to the right to go back to the task. It follows the finger, the task slides and dims under it, and letting go past 40% or flicking finishes the move at your speed, otherwise it springs back. Diffs and the shell's key row keep their own sideways scroll, and the shell doesn't bring up the keyboard. Spells turn it off with
panel.swipe: false. -
My Stuff in the phone app's top bar. The app's floating New task hides the row that held the My Stuff button, so My Stuff was only reachable through Search. It now has its own button beside Search. Spells reach it through the
stuff.topButtonpart.
Fixed
-
/terminalno longer refuses a task with a goal, a team or a crew member. Typing/terminal …with Done when set, delegates picked, compare runs added or@Adain the prompt used to show an error and keep the prompt. Now it starts the task in your terminal and leaves those out: you drive the agent there yourself. It's still refused where no terminal can open (Windows, or a harness that can't resume its sessions). -
Changes and the editor, said in words. A file's stage toggle in Changes says Stage, Part or Staged in Viewed's style (a bare +/− beside the +3 −2 counts read as one), and on a phone a file's Stage · Viewed · ⋯ sit on a line under its name so the name stays whole. An empty commit message shows the one Commit will use (the task's title). The harness picker's plan use is plain grey, amber only near your line. The editor's language support offer is a quiet Add TypeScript support in its head instead of a note that pushed the code down; hover cards sit above the name's line; references and definitions say 3 references to add. Preview's Pick and the app's Issue have their words; the issue picker keeps one height from loading to loaded; trigger kinds are named like the others (Webhook called, Slack mention, Linear issue created).
-
Changes and the task page, tidied. Each hunk in a diff has a head row like git's
@@line (Line 12, and Revert), so Revert never sits on code; Viewed says its word on a phone too; Revert file… names the file it changes. The overlap notice above the reply box is one line with +3 more. wisp's rounds of a goal say which round they are (Round 3: what's still missing), a step quieter than your prompts. The editor shows Save only when there's something to save. Lists in Settings and dialogs open wisp's own menu on a computer (a phone keeps its own picker). -
The task's bar says its names on a narrow page. Wherever the task is too narrow for its details column (a phone, a window under 900px, the side panel open) its actions wrap under its facts with their names instead of bare icons beside labelled ones, and the branch takes the line above them, whole. On a phone the bar is one row (Changes, Ship it…, Mark done, More), with Transcript, Rerun… and Take over at the top of More (and nothing in More that the page already shows), so the first turn starts higher. Each of these actions now has one name everywhere, the side panel's tab names: Changes, Shell, Take over and Transcript in the bar, More, the right-click menu, ⌘K and the keys sheet (they were also Show changes, Open shell, Take over in a terminal and Show the full transcript). Side by side, the reply box keeps its bar when you click elsewhere, so the notices above it no longer jump 30px each time. The goal's bar never draws full before the goal is met (it showed 100% beside 0/1 and Stalled), and a paused goal's line says why it paused without repeating the check that the round wisp sent back already shows.
-
Labels read whole, one type scale again. A task row's third line no longer cuts to Claude Code · Default mo… · This compu…: it names the model only when you picked one (its tooltip still says Default model), and a crew member's details drop it the same way. On a phone, a task's list of turns shows each prompt whole instead of Make the tests p… beside Fake claude fin…, and the goal's check line wraps to two lines instead of ending in not_…. A crew member's role takes two lines on its Org card. The 70 font sizes that had drifted off the scale (12px, 13.5px, 11px) are back on 11.5 / 12.5 / 13 / 14, so a screen has at most five sizes. ⌘K opens on your tasks first, with the crew under them; a crew member's chores (delegates, share, track record) show once you type. The row's hover Save for later is no longer a button inside the row's button (screen readers met a nested control; L and More still have it), and the time on a round wisp sent back reads at AA contrast. The task's bar says Ship it… and Push to PR #12, the words its menus and ⌘K use (it said Ship it, Push to #12 and Ship again). A turn from today shows its time alone (23:49). On a phone, the overlap notice above the reply box is one row and + n more (two rows covered a quarter of the answer), and in the app every control takes a 44px tap (switches, segments, fields, chips, dialog buttons, the mini bar), without moving anything. Screen readers get a name for every switch, list and field in Settings, every dialog's ×, and the harness picker's lists; Settings' tabs and the Run on menu hold only what those roles allow, and a link in a sentence is underlined. Removed lines in a diff no longer fade (their code fell under AA contrast). A comment's highlight on a diff line lights its number and sign, so the line keeps its added or removed colour, and on a phone the comment sheet sits square at the bottom. The phone app's Settings no longer offers Open files in, which only a desktop can do. Settings' headings are one size, and a bare
<small>no longer drops to 10.4px. -
Delegate faces no longer cover Send on the phone. When the New task bar scrolled under the Send button, the delegates' faces showed on top of it; Send now covers them.
-
Round spinners in the Android app. The spinner on Opening was squeezed into a flat oval (the button was too narrow for it and the word); it stays round now, and every spinner on the computers screen is an arc on a faint ring.
-
Buttons line up, in every state. Lit, hovered and focused buttons had uneven boxes. Goal turned on in the prompt box was a 26×30 sliver on a phone, and 30px tall beside 32px Attach and Start later; Worktree, Delegate and the place were 32×30. In a task's reply box Attach and Start later were taller than Send; the sidebar's search was 32×30 beside 32×32 buttons; Cancel (28px) sat beside Save (30px) in every dialog; a crew member's Pause folded to a 36×30 oval on a narrow window, their Remove was 30×28 and their reply bar's buttons were shorter than its Send; Auto's logo sat off-centre on a phone and when comparing; a turn's tools and thumbs were 24×22; the task's icon actions on a narrow window were 28×24; the project bar's + was 28×26 and Other folder… as its icon an oval; a project's colour and remove buttons were 26 and 32px; My Stuff's view toggles were 28×24. Each is now a square, or one height per bar, and in the phone app the logo-only harness picks, Other folder… and My Stuff's options are full 44px targets again.
node qa/buttons.mjschecks every button at rest, hovered and focused in 36 views, at three widths, light and dark, and in the phone app. -
The Org chart pans on its own on a phone. A chart wider than the screen (cards placed on a wider one) now scrolls sideways by itself; the page, the hint above it and the member's details under it stay put.
-
wisp opens and answers much sooner. With 770 tasks: the desktop app shows your tasks about 0.95 s after you open it (was 3.7 s),
wispserves its page in 0.2 s (was 1.75 s), and the server holds 33 MB instead of 155 MB. A start reads a small light copy of each task (sessions/.light/) and leaves its turns' activity, file changes and raw answers on disk until something reads them (opening the task, a follow-up); the list and the usage stats never need them. The task list keeps each task's summary until the task changes, and an unchanged list is sent from the last one (0.7 ms instead of 15 ms). Saves are written in the background. After a restart wisp no longer checks every worktree's git status at once (thousands ofgitprocesses that held the server for seconds): a few at a time, newest first, once the page has loaded, with 3gitcalls each instead of 5. The desktop app starts on the PATH it read last time and asks your login shell again behind it (half a second or more on every open), and wisp finds git without askingxcrun. -
Launching a task doesn't wait on the network. While you type, wisp fetches the project's repo (
POST /api/launch/warm), so Enter starts the agent in its worktree at once instead of after agit fetch(half a second or more to GitHub). -
Smaller answers, everywhere. Big API answers go brotli-compressed to browsers (a 2 MB task: 134 KB instead of 540 KB gzipped, and faster to make), and the page's files at brotli's best, so the phone app and wisp Grid move a quarter of the bytes. The page asks for its first lists while its scripts load instead of after, asks for the machines once instead of twice, skips laying out the rows of the list that are off screen, and a dialog fading out no longer swallows the click you make under it.
node qa/desktop-perf.mjstimes the page in a desktop browser (load, opening tasks, New task, Settings, search, launch), asqa/mobile-perf.mjsdoes for the phone app. -
Buttons sit where they should on phones. In the phone app, Crew's Org sat 3px below the heading's line (a desktop rule loaded after the touch one), and the top bar's My Stuff icon sat 4px right of its ring. In phone browsers, a task's buttons (Transcript, Rerun, Take over, Changes, Ship, Done, More) sat at the left of their 44px tap areas, and Crew's Org and + sat 8px below the heading; section headings there now take a thumb (44px), as in the app.
-
Phone app: the follow-up box's Delegate button is centred. With a crew member (or one delegate) picked, their face sat at the left of the 44px button with empty space on its right; the computer button had the same offset.
-
The phone app stays quick with thousands of tasks and on a bad connection. Stress-tested with 4,000 tasks, a task of 400 turns, a 300 KB answer and a 130 KB prompt, at 600 ms per request and 1 Mbit/s: opening, Back, search and typing stay under 150 ms. A long task arrives with only its latest 8 turns whole (the older ones fill in when you unfold one: 64 KB instead of 2 MB); a long answer or prompt is laid out only where you read it; rows scrolled past aren't restyled on every page change; coming back, the list brings only what changed (3 KB instead of 680 KB); and when the network returns, the open task and the live updates reconnect at once instead of at the next retry.
-
Buttons in the phone app react as you touch them. The pressed look used to show only once your finger lifted, 130–210 ms after the touch; now a button shows it within a frame and a task row after about 90 ms (a beat later, so starting a scroll doesn't flash it), as native buttons do.
-
Faster page changes in the phone app, most of all away from home. Opening a task, Back and New task take a third of the work they did (measured with
node qa/mobile-perf.mjson a slowed phone CPU: Back about 420 ms → 30 ms, opening a task about 165 ms → 75 ms). Several style rules made any change to a row restyle the whole page, the project chips refit on every Back, and the phone now holds 20 finished tasks at a time instead of 60 (more come as you scroll). Over a slow tunnel: the tasks on screen are fetched ahead while you look at the list, so opening one shows it at once; a task is downloaded once per open instead of twice and without each edit's whole file from before it; and a task or list that hasn't changed comes back as a bodiless 304. -
Terminals inside wisp feel like a native terminal. A big paste (or typing while the agent starts up) no longer hangs the terminal: the pty helper waited on the paste going in while the program waited on its output going out. Keys go out without waiting for the last one to arrive, in order, so on a cloud computer or a slow link each key shows about one round trip later instead of two. Output is read about twice as fast (scrolling logs 2.4×, full-screen TUIs 1.9×), and reopening a terminal with a long history paints in about 100 ms instead of 300 (the older scrollback fills in while idle). Frames from programs that use synchronized output (Codex) are never shown half drawn, programs that ask are told when the terminal gets and loses focus (Claude Code, Codex), and the server no longer copies the whole kept output on every chunk it prints.
-
Babysitting mode's terminal fits its pane. A task's CLI now starts at the size of the pane it opens in, so its first screen (banner, rules, boxes) isn't drawn 100 columns wide and left that way. On a phone, and after the window crossed 700px, the terminal measured its cells in the wrong font size: lines stopped short of the right edge (47 columns where 52 fit) or ran past it; it measures the font it really uses now. A quick burst of resizes could leave the CLI at a size it passed through; the last one wins now. Full-screen CLIs that paint their own background (opencode, Grok) fill the box to its edges instead of sitting in a lighter frame, and the task's header keeps its buttons at the pane's right edge with the list hidden. The measuring, resize and edge fixes also apply to Take over's terminal and the side panel's Shell.
-
A key after a terminal session ends resumes it, and nothing else. As the CLI exited, the task moved from Working to Needs you and the terminal lost the keyboard, so the key meant to resume it ran wisp's shortcut instead (
aarchived the task,emarked it done). Moving a task between sections keeps its focus now. -
Take over keeps its transcript when wisp was started from Claude Code. A wisp started from a Claude Code terminal (
npm run dev:webthere) passed that session's markers to the terminals it opens, so the claude in them kept no transcript and couldn't be resumed (No conversation found). They're dropped now, as they already were for wisp's own runs, and there are more of them. -
Take over without the JSON. A Claude Code session you take over is now told to answer you in plain text, so its replies in the terminal no longer end in a
wispblock meant for wisp's page. -
Collapsing a file's diff in a report's Files. Once a file's diff was open, clicking its chevron (or the row) tried to open the file in the editor instead of closing the diff. On a shared task's page, Open in editor did nothing at all: there's no editor there, so it's no longer shown.
-
Updating a published app on a Windows grid. Publishing a new version right after someone opened the link could fail with Something went wrong on the grid: Windows still held the file the visitor had just read, so the app's folder couldn't be swapped. The grid now waits a moment and tries again.
-
A quick query in an app's data isn't stopped on a busy computer. When wisp (or the grid) was too busy to read the answer in time, a query that took a millisecond could fail with That query took over 5 s and was stopped, and the database restarted. The limit now counts from when the query starts, and is checked only after any answer already waiting is read.
-
Coming up's rows line up. On the home page the first row under Coming up sat a little to the left of the others, and the rows were shorter than Needs you's: they picked up the styles of the open task's "delegated from" trail, which uses the same class name. Those styles now stay in the trail.
Get wisp 0.51.0
| macOS (Apple silicon) | wisp-mac-arm64.dmg |
| macOS (Intel) | wisp-mac-x64.dmg |
| Windows | wisp-windows-x64-setup.exe · Arm |
| Linux | AppImage · .deb · Arm: AppImage, .deb |
| npm (Node 20+) | npx wisp-agents@0.51.0 or npm install -g wisp-agents |
| Homebrew | brew install m0xoo/tap/wisp-agents |
The desktop app updates itself from these releases. Checksums are in SHA256SUMS.txt.