Skip to content

Wisp v2 - #756

Open
M0xoo wants to merge 1222 commits into
mainfrom
wisp/we-are-going-to-work-on-wisp-v2-8fb64e
Open

M0xoo wants to merge 1222 commits into
mainfrom
wisp/we-are-going-to-work-on-wisp-v2-8fb64e

Conversation

@M0xoo

@M0xoo M0xoo commented Nov 8, 2026

Copy link
Copy Markdown
Collaborator

What and why

How I tested it

  • npm test passes
  • UI change checked against design/checklist.md (desktop + phone, light + dark, split + feed), screenshot attached
  • New UI is reachable from a spell (AGENTS.md → Everything is customizable by spells)
  • User-facing change has a line under Unreleased in CHANGELOG.md

@M0xoo M0xoo changed the title Wisp/we are going to work on wisp v2 8fb64e Wisp v2 Nov 9, 2026
M0xoo added a commit that referenced this pull request Nov 12, 2026
…dows)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@M0xoo M0xoo added the ci:full Run the full CI matrix on this PR label Nov 14, 2026
M0xoo and others added 27 commits November 15, 2026 18:41
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…elease.yml npm finding recorded

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…entry opens on Computers, keeps every computer connected, launching one runs the shell there with the platform look and phone navigation

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…er filter, Search, me menu, project chips, New task FAB / iOS compose) and Add computer (QR with typed fallback, Trust the new key, Stay connected ON for SSH without push) (ADR-0065 step 13, part); a list at its top stays at its top; phone plugin styles load after the kit

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ter its first connect showed every task twice)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s chunk (strings-split)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ines in the right column (no overlap at 320-390 px)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… or folder

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ries row, B + Waiting for you, C + reply), Android/iOS, light/dark

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…task's own actions menu, details from the menu), New task (Android full-screen page, iOS sheet) and search (M3 search view, iOS search field) (ADR-0065 step 13)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…m all v2 conversations); AGENTS.md: read §3 before any change; phone project chips logged as a decision breach

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e, desktop New task included (owner)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…pick C): a ring row under the Tasks bar, a full-screen story (the task page's answer card and composer), arrows and × beside every gesture; browser and desktop keep v1's Crew section

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… build in progress): wisp.threads plugin, port order, name clash with ADR-0068's Ledger

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…read' tab; 'Ledger' stays the task history

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… New task's place row, the sidebar head and the phone Tasks bar use the same control; its list is the ⌘P list (name = only, + adds, − removes, No folder and Other folder… at the end); phone chip rows gone

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…folder (owner)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
M0xoo and others added 30 commits November 18, 2026 10:45
…chip does; phone sheet tabs never cut; an answer's long-press menu opens at the finger

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- phone-pages: the Worktree-in-More tests waited on a closed menu that stays in the page, burning 4 of their 5 s;
  they look only in the open menu now (4.4 s → ~100 ms, 60/60 under a throttled CPU).
- serve_budget: Send anyway answers turnId or queued depending on whether the earlier turn has ended (Windows is
  slower); the test reads either and checks it's the client's id (20/20, 20/20 under load).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…puter comes back; crew faces line up with the dots

- CLI: words that start with a built-in command (`computer on --grid …`, any catalogue verb) are never a task: an
  unknown flag or word prints that command's usage and exits 2 (words::VERB_PATHS, checked against main.rs; the 34
  untyped plugin verbs now declare their arguments; serve_cli_verbs covers every verb). `computer on --grid <grid>`
  signs in there first, as the Grid's + Add panel says. A computer removed from the Grid registers anew on `on`.
- A crew job's face sits on its dot's spot (18 px centred on the 7 px dot, v1's crew dot), so its title and details
  line up with every other row (owner 2026-11-18).
- Phone: an answer's long-press and ⋯ hold every answer action; the reply chip folds to its icons instead of cutting.
- docs/owner-list.md: #45 (the project selector "did nothing": the dev app died with the lead's session; it runs in its
  own session now) and #46.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… prompts stay tasks; offline computers wait on phones; one model name everywhere

- CLI: a built-in group (computer, cloud, grid, crew, plugin(s), skill, completions, sdk, recipes, mcp, machines,
  trigger…) followed by a word that isn't its verb lists its verbs and exits 2, never a task; a plugin's name is a
  command only when one of its verbs follows, so `wisp review the auth module` stays a task (decided by the owner's
  delegate). `--` or one quoted argument always makes a task. HELP and the docs CLI page say so.
- `wisp computer on` after a Remove waits for the dial and prints "on, online"; a sign-in with no grid goes to the
  last grid this wisp used.
- Grid: any navigation to an offline computer lands on its waiting page (/_grid/offline), also on phones where the
  app's service worker answered from cache; it reopens the app by itself.
- One name for the Default model (a per-client store over runtime.defaultModel): task facts, Compare, the card title,
  the list row's tooltip.
- New task and the reply box: the chip whose panel is open stays on the bar, the others fold into More, a folded chip
  keeps its clear.
- Phone: the answer's ⋯ ends its action row; its menu shows alone (no selection bubble beside it).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… or a prompt word; Send says when a scheduled follow-up goes

- `wisp version` prints the version; `wisp update` checks and `wisp update apply` installs (one word after it; longer
  lines stay tasks); `wisp machine …` is `machines`; `wisp share …`, `usage`, `spell(s)` and more are strict groups;
  `wisp help <group>` prints its usage. 27 prompt words (`run`, `plan`, `ship`…) still start a task without their
  argument (decided by the owner's delegate). An e2e reads every command word from the source and v1's list.
- `wisp wait <unknown id>` errors at once (it hung 30 minutes); `wisp btw` is always the command.
- The task page's Send says "Send at 15:24" with Start later set.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… no Subagents chip), same sizes, 668 KB

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…redaction keeps ordinary words; init asks outside a project

- A line of one word that isn't a command, or a word and an id (`rm 123`, `kill 123`), never starts a task: it names
  the real command (`wisp login` → `wisp cloud signin`, `doctor`, `config`, `upgrade`…) and exits 2. `--` or a quoted
  prompt still makes a task; `wisp run the tests` prints its title.
- Sharing takes out another task or crew member only where the text refers to it (its mention chip or bare id), never
  its title as a word ("fix the login bug" stays as written; ADR-0103 supersedes that line of ADR-0027).
- `wisp init` outside a git repo or project asks first, or refuses with `--here`; it never edits a parent's AGENTS.md
  silently.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…` is a task; `ls --all` lists everything

- share.preview answers what the dialog reads (shared, work, targets; additive, with bindings); the dialog draws only
  its content in the shell's dialog (no second box) and an older server's answer reads as "sign in first" instead of
  throwing. An e2e golden of the real answer feeds the dialog's tests. The "Its work" line says what goes: its answers
  and the list of files it changed.
- `wisp add` is the command only for a folder that exists, a path or a git link; other words are a task; a missing
  path is refused. `wisp ls --all` lists every task. Slip hints point to read-only or matching commands.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… after the press; Share shows what goes and what's hidden; one stray word is a slip, two or more a prompt

- Fork… opens v1's Fork task dialog (harness on the task's machine, model, thinking, mode, the turn, the draft) from the
  row menu, the head ⋯, ⇧F and phones; nothing is made until its Fork. Rename from the head opens its dialog.
- A pick in a desktop flyout runs after the press ends, so the dialog it opens stays open (Task options' Set goal,
  Add to crew, Save as recipe; Recipes › New).
- Share: "What goes" (the ask, steps, answers and files with its work) and "Hidden before it goes" (one line per kind,
  with counts), from share.preview at the report level; phones get a sheet with ✕ and native choice rows.
- CLI: after a strict group, exactly one unknown word is a slip (error with the `wisp -- "…"` escape), two or more make
  a prompt; id-taking commands are commands only with an id-shaped argument (`wisp show me where the bug is` is a
  task). `wisp attach last` attaches to the newest task.
- Rerun's bar wraps on phones instead of crushing its mode; Merge's confirm reads cleanly.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…in words, whole-path share redaction, CLI slips; Home's top moves the desktop window

- Fork from a row reads the task's turns before deciding (it said "Nothing to fork yet" on a finished task).
- One Rename dialog (the row's: name selected, nothing sent unless it changed) from the head, the row and phones.
- A run is named in words everywhere (one helper, useRunWords): "Runs on Claude Code · Opus · High", never ids;
  Rerun starts on the harness the task ran on.
- Share hides the project and home folders only as whole paths, in every form (~/, /private/tmp, Windows), longest
  first; a sibling folder keeps its name.
- CLI: plugin/mcp/skill with two or more other words are a prompt; one stray word after a command or prompt word is a
  slip with the `--` escape; `ls -a` is --all; `wisp fork` and `wisp help me` point the right way; init takes only
  auto, ask or manual. Six plugins verbs declare their arguments.
- Phones: sheet and dialog ✕ are 44 pt / 48 dp targets.
- Desktop: Home's top padding is a window drag strip, as a task's head is (owner, 2026-11-18).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…, performance round 1, Settings → Agents

Audit A checked ~290 actions on scratch wisps with fake harnesses (docs/qa/audit-A-*.md). Fixed:
- Worktree safety: Restore works only in the task's own folder (it reset the project's main checkout); Merge keeps a
  worktree with uncommitted or unpushed work and asks; Fork, Rerun and Compare get their own worktree and branch;
  Bring in works; Fix conflicts doesn't spend a turn without a conflict.
- Orchestration: attachments and the Delegate team reach the agent; Side discussion and /btw run apart from the main
  thread; a pending question shows in `wisp show`/`wait`; Start later keeps Worktree and shows the scheduled task;
  Review by another harness runs; Done on a running task stays put.
- Automations: recipes and triggers edit, pause and delete as v1; schedule triggers have their time and project;
  trigger test, apply --prune, `wisp run` and Save as recipe work; Drop goal clears its question.
- Setup: projects must be real folders; CLI tasks in a project get its worktree; enum settings from the CLI and
  formulas; the chosen grid survives a declined sign-in; a second account doesn't switch or unlink this computer;
  `wisp mcp` has v1's task tools; import keeps ntfy and lists what it didn't carry; completion, export, plugins
  lists, Trust in Settings → Plugins, settings search over every setting; connectors keep their env and multi-paste.
- Tasks: Delete and archive confirm or Undo; Changes shows what's changed now; file links open the file; list keys on
  the picked row; phone swipes run their action.
- Settings → Agents: account rows carried the wrong id (every account control was dead); sign-in progress is live;
  status checks run in parallel; Add another account as v1; gh setup-git after a GitHub sign-in.
- Performance: the edge's answer cache is bounded (60 list reads reached 1 GB); a CI check that spike locks are
  current (the server bench was blind); worktrees.view, the machine view, startup, the first list, share preview,
  thread health and `wisp ls` stop walking every task; kit style recalcs −67–83 % (open a task 369 → 123 ms, side
  panel 359 → 60 ms, ⌘K 240 → 69 ms at 4× CPU); shell hot paths; the first screen lazy-loads more.
- CLI: `wisp run "the tests"` launches; non-TTY lines follow the slip rules.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e while open, tests for deliberate changes

- Bring in compared git's C:/… worktree paths with wisp's C:\… as strings and skipped every stray worktree on
  Windows; it compares folders now (offer.rs, with a Windows-spelling unit test).
- The kit's menu keyed rows by object: entries that changed while a flyout was open (a plugin's later part landing)
  remounted it closed, so Task options › Add to crew… did nothing in the first second. Rows key by value now.
- Tests: MCP lists v1's task tools; the /btw prompt carries the rules and the task so far (ADR-0050's checks kept)
  and the fake reads its hang cue from the question; scratch repos get a git identity and core.autocrlf=false;
  paths built with Path::join; Task options asserts its dialog opened.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…first-screen budget ratchet

Audit B checked every action again on scratch wisps with fake harnesses (docs/qa/audit-B-*.md): 42 findings, fixed.
- Tasks: a phone long-press no longer runs the item under the lifted finger; Ship only on worktree tasks, with v1's
  reason elsewhere; Compare finds keyed and bare ids; Mark unread and Save for later say so with Undo; prompt bubbles
  keep their turn's harness; the iOS 360 bar title and the live card's status keep their words; the list mounts once
  (WebKit's ResizeObserver loop); every Grid link shows it's going and a sleeping box says it's waking.
- Orchestration: side runs get their launch files ({tmp}); Rerun of a worktree task gets its own worktree; agents keep
  Notes (wisp_notes); MCP lists only what it can run (ADR-0107); agent-launched tasks get worktrees; crew jobs,
  triggers and recipes get worktrees as v1; Start later says when; crew CLI fixes; --compare keeps --agent; help
  never launches.
- Setup: one router refuses a plugin's name plus a stray word; projects need real folders in import and apply;
  harness Update says when nothing changed; no npm registry call at start (accounts.updateCheck, WISP_UPDATE_CHECK);
  Auto stays out of search, help and formulas while off; wisp-grid dev boxes use the fakes; Trust and Remove confirm.
- Server perf: a fresh client's list reads page by page off the actor (ADR-0106; patch p99 after 10 fresh clients
  273 → 105 ms); blame one git log (7×); machine.read from the tick's sample; usage 12×; `wisp ls` the newest 50;
  touching/log/tracked parse only matching files; share scrub with a RegexSet off the request thread.
- UI perf: ⌘K 940 → 61 ms (one bare-id index, a narrowing ranker, dialog first); picker data kept 30 s; the orb on the
  compositor; dialogs don't restyle the page; autosize one read; submenus open on hover at once; Settings, side tabs,
  New task and J/K draw the frame first; a left task's rows leave memory after 30 s.
- The first screen loads 15–27 KB less, and CI holds it with a ratchet (ADR-0105).
- The kit menu fix's TaskView canShip use-before-declare.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ppy -D warnings)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…tart later, one yes starts every held turn, picked rows alike, compare title, WebKit observer notice

- Phone apps: a doc, table or code in My Stuff opened the raw file with window.open; the phone screen hands in
  `show` now, so they open in wisp's reader and viewer (apps still go over the whole screen).
- + › Start later › a time left the caret on `+`, so Enter reopened the menu; a pick in a row's own menu hands the
  caret back to the box (reply and launcher).
- Trust and start on one held card trusted the program but left the other turns the same change held; they start too,
  as Trust does (serve_accounts test).
- A phone keeps :hover on the last tapped row, which beat the picked tint in Select; hover shades rows only for a
  hovering pointer, and never a picked or selected one.
- A compare group's row is laid out as a task row: the title has the first line (55 → 197 px of a 360 px phone).
- WebKit's ResizeObserver loop notice (lazy parts landing between observer callbacks, delivered next frame by spec) is
  handled at the entry; other errors are untouched.
- The harness picker's data module loads with the picker, keeping the desktop first screen under its ratchet.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…omponent (Windows)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…macOS, one panel per account, never-finished accounts signed out

- Antigravity: agy 1.5 refuses a sign-in with no terminal ("authentication required. Run 'agy' to log in"); on
  macOS a tty sign-in now runs under BSD script(1) as it does under util-linux script on Linux, and agy prints
  Google's link and reads the code back.
- The sign-in box is the kit's SettingsSignIn (v1 .sx-login): numbered steps, an "Open the sign-in page" button (the
  OAuth address was spelled out in full), the code in a dashed box with Copy, "Waiting for you to finish there…",
  Continue; v1's provider names (ChatGPT, Google, xAI). Under an account it takes the inset's band.
- A harness with several accounts showed its first account's sign-in twice (the harness row and the account row are
  the same login); only the account's shows.
- The kit spinner is centred over a button; in the box it sits in its line (it had floated by the harness title).
- An account whose first sign-in didn't finish (failed, cancelled, cut by a restart) is signed out, so its row offers
  Sign in; it had no state and no control.

Tested on the dev app with the owner's real accounts: Antigravity, Codex Main, Claude Te1, light and dark.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sign out on an account whose sign-in was running read "Signed out" beside the sign-in's Cancel, and the CLI's login
kept going (it could sign the account back in after the sign-out). Logout cancels a live sign-in first
(serve_accounts sign_out_during_a_sign_in_ends_that_sign_in).

Checked Remove and Sign out end to end: on the dev app with a throwaway Codex account (Remove asks, the row, its folder
and its running `codex login` go), and on a scratch wisp with fake CLIs (account and harness Sign out answer in 3 ms;
a CLI that stays signed in says so).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…r, agent launches in worktrees, recurring runs reuse one worktree, CLI words, project counts

Orchestration (docs/qa/audit-C-orchestration.md):
- Quick follow-ups raced the queue: some were lost with a raw UNIQUE constraint error, the rest ran out of order.
  Every start is decided under the task lock and keeps the stored queue (followup_race test).
- A run's wisp_launch passed only prompt, cwd, agent, model and mode: worktree, project, effort, subagents and the rest
  now reach runtime.create; the stdio `wisp mcp` launch gets a worktree too; wisp_stop answers once stopped.
- Crew check-ins and scheduled triggers reuse one worktree per member or trigger and project, and a clean run leaves
  none (65 worktrees in 40 minutes before; ADR-0108). In place stays sticky as in v1, says so, and the CLI keeps its
  own default. `wisp rerun` runs at its commit and leaves no detached folder. Worktrees' look always ends.
- Copilot's /btw gets its own session; the Rerun dialog says where it runs; Ship › Merge PR… opens its confirm and
  says what happens to the worktree; queue order/edit, crew schedules in words, rerun -d, CLI output without raw
  epochs or JSON, the project column, one defaultModel ask on Home, the stdio MCP lists /api/mcp's tools.
Setup (docs/qa/audit-C-setup.md):
- Manage projects counts tasks by project; removing a project asks first; This computer reads the disk at once;
  CLI nested values as lines; export names an unknown format or kind; `wisp router` is refused while Auto is off;
  ⌘K ranks a word start first; Setup import errors in one sentence with no stale plan; plugin install wording.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
openExternal tried the opener IPC first, which the page is never granted, and only then fell back to window.open:
after the click had ended, where a webview may treat it as an unrequested popup. Links out now open in the click
itself (platform.test.ts, failing first). A routing unit test pins where a link goes in the packaged app and in dev
desktop (wisp's own pages stay; everything else, previews and other localhost ports included, goes to the browser);
the qa harness's mouse op takes modifiers and the middle button.

The owner's "links don't open" was mostly outside wisp: orphaned headless Playwright Chromes from QA runs had become
the front Chrome and swallowed every URL open on the Mac (stopped).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…stration, setup, tasks); Windows stop race

Side panel (owner, 2026-11-19; docs/qa/audit-side-panel.md): Changes, PR, Shell and Preview are always in the strip,
labelled, in v1's order (replaces UX 2026-11-18 #5's "only tabs with content, the rest behind +"); the kit's Tabs fit
turns the last tabs to icons only while they overflow, the open one keeps its label; arrows keep walking the tabs;
Changes diffs a worktree only, else the edits wisp saw; a gone folder is said in words, no raw toasts; Open file… only
on git's diff and never leaks keys; End shell asks first; a reattached shell fits, edge to edge as v1.
Audit D orchestration: a turn held for project trust starts on Start now, with v1's trust card; Skip's words; wisp
show tells a held turn; MCP tools name their fields; a Stop landing while a turn's CLI starts stops the run (the
Windows serve_task failure).
Audit D setup: a formula can't add a second project for another spelling of a folder; YAML on/off for enum settings;
Auto off hides router from export, validate and apply; sign-in Copy turns into a check; a CLI that stays signed in
says so; CLI output and wording fixes.
Audit D tasks: Needs you puts asking tasks first; held rows name their harness; the details strip fits; phone plan
rows line up; Android's full-window reader.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…1280

The kit's Tabs fit measured before its steps were drawn and ran to all-icons; one measuring pass now picks the first
arrangement that fits: Thread, Watch and pinned tabs to icons first, then a tighter row, then Preview, then the rest.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…shes only the built site

Website and docs (owner, 2026-11-19): a new landing page (Manage the code your agents write: Declare, Run, Track)
with real app screenshots, every feature from the app, the TUI and the CLI, the desktop and phone apps as headline
features, and a Get the app page with the phone-app waitlist (the Grid's /api/waitlist, kind phone-app); about 25
guides with App/TUI/CLI tabs, the Linear guides included; SEO tags, sitemaps, llms.txt; only measured numbers;
phone first paint 980 → 308 ms; 0 broken links. Pages now uploads a staged folder of the built site only: the plan,
ADRs and QA notes in docs/ were being published.

Audit E (docs/qa/audit-E-*.md): a CLI task in any git repo gets a worktree (v1); the trust card lays out on phones and
desktop as v1's; Ask to fix on a failing check is back; settings plans show what a new setting changes from; channels
stop retrying a refused token; Esc in a machine's Edit stays in Manage; one harness order everywhere; Home puts the task
asking you first; Open file… hands the keys to the file; phone Changes keep file names whole; Copy path shows a check;
the PR tab says PR #1; and more.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Update in the desktop app failed with "The desktop app updates itself when it restarts." and there was no updater.
DownloadApp now does what v1's fetchApp did: download the release asset for this OS and arch (refused without a
checksum), verify it, stage it (macOS: unzip, check the bundle, swap it in beside the old one and put the old one back
on failure; Windows: the installer, run /S /R after wisp stops; Linux: the AppImage), then relaunch through hand_off,
which now waits for the old shell on Linux too and quotes every path for sh. On Windows the relaunch helper leaves the
server's job so it outlives it. install::detect finds Tauri's resources/app layout. WISP_UPDATE_FEED (loopback only)
lets a packaged build be tried against a local feed. Reviewed by a second agent; F-upd-1 stays partial until the
packaged app is seen updating end to end (owner decision, docs/v2-plan.md §3, 2026-11-19).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…; docs sidebar lists its pages

Every how-to block (landing and 27 guides) has Desktop app · Phone app · TUI · CLI tabs: a computer opens Desktop app
with real desktop screenshots, a phone opens Phone app, picked before paint and remembered across the site and the docs
(JS off: Desktop app). The docs sidebar showed every group empty, so clicking did nothing: Starlight 0.43 matches a
group's pages by a src/content/docs filePath ours (docs/site) never has; the groups now name the real paths, and the
docs build fails on an empty group (check-sidebar). Audit F reports added (docs/qa/audit-F-*.md).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…asks and undoes, the reply box grows

Audit F (docs/qa/audit-F-*.md), each with a test that failed first:
- A switched-off plugin's CLI words are refused (exit 2) with how to switch it on, never a Full-auto task.
- The Shell tab hands the keys to the terminal (typing ran the list's shortcuts: `e` marked the task done).
- Revert a hunk: Reverted · Undo; Revert file asks first as v1 did, then Undo; a new file is never lost.
- The reply box wraps and grows to 260 px as v1's did.
- A new task's page no longer crashes before it has a worktree; committed files stay listed in Changes.
- Manage projects stays in the switcher with four projects; v1 connectors keep their env and headers; `wisp update`
  says when it hasn't checked; update and connector commands take what they need; merge and fork say what happened
  to the branch and folder; phone held-run buttons back to 44 px / 40 dp; the compare card inside the phone margins;
  and the smaller wording fixes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
d4c98da let every detached start leave the desktop shell's kill-on-close job whenever wisp was in one, so closing
the window no longer stopped what wisp had started (the Windows smoke DESK2-05 then timed out). Only the update's
relaunch helper may break away now (SpawnOpts::leave_every_job, pinned by a test that no other site sets it); the
server learns it's in the shell's job from WISP_IN_PARENT_JOB. DESK2-05 fails fast and names what failed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… race

- Hub::sleep with no clock returned at once, so ntfy's poll spun (207k fetches in 2 s) and the Phone row was never
  written: the macOS-only flake B60. It sleeps; the test waits for the status, then the row (450 stress runs green).
- `wisp sandbox claude begin --bad` started a real sign-in: the command is typed and refuses unknown flags.
- serve_cli_verbs runs its independent CLI calls 8 at a time; the fake harnesses build once per run under a lock
  (a copy could find the binary missing while another cargo replaced it: 5 of 50 runs at 8 in parallel failed);
  the fifty-worktrees Bring in test runs its turns alongside. About 2.3 min off the Windows suite.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…one menus drill in a sheet; Goal… scales

- Formulas declare goals by name (`goals:`, `goals.default` for every goal's judges, vote and rounds) and tasks take
  `goal: <name>` or an inline `{ text }`; `tasks.default` gives every declared task its defaults; validate names an
  unknown goal (ADR-0110, owner 2026-11-19). A goal is attached by name with `wisp "…" --goal <name>`, `wisp goal set
  <task> <name>`, MCP launch, and the app's Goal… menu.
- On a touch or narrow browser every kit menu opens as a bottom sheet and a submenu drills in place with a back row;
  desktop keeps its side flyouts, the phone apps their native menus (the Goal… flyout ran off a phone's screen).
- Goal… and Recipes with many items: a filter box, recently used first then A–Z, seven shown, then All goals… (a
  searchable picker); `/goal ` suggests in the prompt and `--goal <TAB>` completes in bash, zsh and fish.
- The Goal… code loads after the first screen: desktop first-screen JS 176,793 B br, under the 177,152 B ratchet.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

ci:full Run the full CI matrix on this PR

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant