local-kanban 1.15.0 → 2.0.0-beta.1

package/CHANGELOG.md

@@ -8,6 +8,727 @@ version it goes out as and opens an empty one again. Released sections are histo
 
 ## [Unreleased]
 
+## [2.0.0-beta.1]
+
+**2.0 beta.** The board stopped being a list of tasks next to Claude and became the place you work with it: Claude starts
+inside the board, you talk to it in the card, watch every session side by side, and a fresh Claude checks the work. A
+beta goes to everyone, like a release: `npm install -g local-kanban` and the update button both bring it, and your
+data stays. To stay on the last 1.x, run `npm install -g local-kanban@1.16.0`.
+
+### Changed
+
+- **The README shows the board as it is now**, in the dark theme, with Claude moving cards in the first picture.
+- **Update checks understand betas.** A board on a beta is offered the next beta and the release it leads to; before,
+  both looked like the version it already had.
+
+- **The chat says what Claude is doing while nothing lands in the transcript.** A long script takes a minute or two to
+  write, and all that time the chat said only "Claude is working…". The light is green now and tells how long the
+  transcript has been silent and, for a Claude inside the board, whether it is thinking or writing and how many tokens
+  it wrote. A counter that stood still for a minute with nothing to wait on turns the light yellow: "no answer for 1 min".
+- **Steps tell writing from running.** A step and a folded run of steps say how long Claude took to write them and how
+  long they ran: "wrote for 1 min 31 s · ran for 19 s". Before, only the second number was shown, and a session of
+  minutes looked like commands of seconds.
+- **Health lists the rules with `paths:` too.** The rules of a project's `.claude/rules` that Claude Code loads only
+  when it works with their files are listed with the CLAUDE.md files of subfolders, with their size: past the
+  per-file limit Claude Code reads only the start of such a file.
+
+## [1.16.0]
+
+### Added
+
+- **Health: how big the instructions Claude reads are.** Switch it on in Settings → Health, and the sidebar gets a
+  section that measures every project: its own CLAUDE.md in lines and characters, and everything a session started in
+  its folder reads (the CLAUDE.md of the folders above, imports, rules, auto memory), against the norms — under 200
+  lines per CLAUDE.md, as the Claude Code docs ask, and the 40,000 and 120,000 characters from which Claude Code itself
+  warns. The CLAUDE.md files of subfolders are listed apart.
+- **Claude done by time.** A switch over the section shows the runs as before, or every task Claude finished in one
+  list, the latest on top, with its project on the row. The choice is kept in the browser.
+- **Automations: a check of Review that starts by itself.** Switch them on in Settings → Claude Code, and the sidebar
+  gets a section where a rule reads "When [project] finishes its work, wait [3] min and start a check of the tasks in
+  Review". Once every task Claude took in that project has reached Review and nothing waits for you, the board counts
+  the pause down — beside ▶ of Review and over Claude now — and starts the check inside the board, as the button does.
+  A rule runs once; the page may be closed. A question of Claude, a task handed back to To do or a used-up limit holds
+  it, and the pause starts again after.
+
+- **An answer cut off by a lost connection goes on by itself.** When the connection to the API breaks in the middle of
+  Claude's answer, Claude Code ends the turn and waits — nothing told you the work had stopped. Now the card says
+  "connection lost · going on", and the board tells its own Claude to go on from where it stopped (up to three times in
+  a row, after 15 s, 1 min and 3 min). When it cannot — a Claude started outside the board, or it broke again — the card
+  says "answer broke off" and the board notifies you. A second "Claude is waiting for you" (or question) of the same
+  task now shows its banner and sound as well — it used to replace the first one silently.
+- **The board over Wi-Fi.** Settings → Wi-Fi (and the Wi-Fi button beside Settings, orange while it is on) opens the
+  board to people on the same network: a QR code and a link open it on their phone or laptop. The address has no
+  password — anyone on that network who opens it can do on the board what you can, including talking to Claude on your
+  computer — so keep it on only on a network of people you trust. Switching off closes the board to the network.
+- **The first-run wizard says what the board is now** — the card's chat with questions as a form, Claude now with every
+  Claude side by side and Continue — and asks whether to show Claude as a chat or a terminal. The check of the computer no
+  longer reads "Claude's questions: not connected": the board gives its own Claude the hook. Every board sees the wizard
+  once more after the update.
+- **Continue a Claude of the board that has stopped.** A Claude started inside the board and stopped from outside (or
+  ended by itself) keeps its terminal: the card says "Session interrupted" and opens on it, and Continue — in the card's
+  terminal, in its chat, and in Claude now (the list and Terminals) — starts Claude again in that terminal with its
+  conversation, where it waits for your word. Claude now keeps an interrupted run in sight instead of folding it among
+  the ones nobody leads.
+- **Claude as a chat.** A card has a Chat tab in place of "Claude's work": the conversation of its Claude in turns —
+  your words, what Claude wrote, its steps folded under "N steps" with their time and a ✗ where one failed, and how long
+  each turn took. What a command returned is never shown. The Claude the board started can be answered right there:
+  words go at once to a Claude waiting at its prompt, and to a busy one they wait in a queue (✕ takes them back) and
+  go once it has been free for two seconds; a question Claude asks is answered right in the chat, with the keys of its
+  terminal. A Claude in Terminal or the Claude app is read, not spoken to. Whether a card and a column of Claude now →
+  Terminals show Claude as its chat or its terminal is one choice in Settings → Claude Code. The tasks over a column
+  open and close like a shutter, and opened steps fold back. A picture pasted, dropped or picked with the clip goes
+  with the message as a picture — a thumbnail over the field, ✕ takes it out — and Claude gets it as an image, not
+  as the path of a file in your words. Claude's answers read as Markdown: headings, lists, tables, quotes and code.
+  A message reaches Claude for certain: words left in its input field are cleared first, the board sees the message
+  arrive before it presses Enter and leave after it, and when it does not, the chat says why instead of looking sent.
+  A Claude at work can be stopped from the chat (■ next to Send) — as Esc in its terminal, never twice at once.
+  ⌘↵ (Ctrl+↵ off a Mac) sends at once even while Claude works — it takes the words into the turn going on — and a
+  waiting message has its own Now; words sent mid-turn show in that turn.
+  A block of code in Claude's answer has a copy button, and a command (a block of a shell) runs in a new Terminal
+  window in the project's folder once you have seen it whole and said yes — only a command Claude wrote in that chat.
+  ⛶ in a chat of Claude now → Terminals spreads that column over the window, smoothly, and takes it back the same way
+  (Esc too) — the same chat, what you typed and the place you read kept.
+- **⌘F opens the search over projects and tasks** (Ctrl+F off a Mac) — from any screen, from a field too. Pressed
+  again while the search (or any other window) is open, it is the browser's own find. The shortcut can be changed in
+  Settings → Hotkeys.
+- **A task moves to another project.** "Move to project…" in the card's ⋯ menu and in the right-click menu of a card:
+  the task keeps its comments, attachments, checklist, links and issue, takes the next number of that project, and
+  its old key still leads to it (the card's timeline and `kb show` say where it came from). A task Claude is working
+  on is not moved until it is handed in.
+- **A screenshot reaches Claude from its terminal.** Paste a picture (or drop one) into the answer field of the card's
+  Terminal tab or of a column in Claude now → Terminals: it is kept as an attachment of the task, and its path goes
+  into the field, so Claude opens the picture when the answer is sent.
+- **The answer field suggests handing the task in.** While Claude waits for a word on a task in Doing or Deploy, the
+  empty field shows a grey hint — "Move KEY to Review, so I can check it later"; Tab puts it in, typing hides it.
+- **A check of Review can think at its own level.** Settings → Claude Code has a second "How hard Claude thinks", for
+  checks: a check started from Claude review or the Review column takes it, every other task the level of all starts.
+  Left as it is, a check thinks as every task does.
+- **Claude now → Terminals: every run of Claude side by side, each with its terminal.** A switch at the top of the
+  section lays the live runs out as columns, like an agents window: the tasks of the batch above, and — for a Claude
+  the board started — its screen, its keys and a field to answer it, without opening a card. A Claude in the Terminal
+  app or the Claude app has no screen here; its column says where it is.
+  A ✕ beside "Open in the card" closes a column's terminal: a Claude that is done and has nothing to ask goes off
+  the section at once (a Claude still running is asked about first).
+- **Settings → As an app: how to make the board an app of its own.** The steps for the browser the page is open in
+  (Chrome and other Chromium browsers, Edge, Safari on a Mac, iPhone and iPad, Android; Firefox on a computer is told
+  to use another browser), the others folded below, and an *Install* button where the browser offers it. A board
+  opened over plain http at a LAN address is told why Chrome greys the menu item out. The installed app opens light.
+- **The terminal of a card shrinks and grows.** Drag the short bar over its keys (or press ↑/↓ on it) to make the
+  screen lower or taller; a double click gives back the usual height. The height is remembered on this computer.
+- **A new look, in the spirit of the new Claude app.** Near-black (or, with the light theme, cream) monochrome, a
+  white main button, neutral highlights instead of orange, serif titles, icons in the sidebar, thin borders and
+  rounder cards, a *New task* row at the top of the sidebar and soft transitions for screens, the task card, windows
+  and lists. Behind the board, clouds and waves made of grey symbols drift from right to left in three layers of
+  depth — the far ones smaller, dimmer and slower, the near ones hiding them — and a click on empty space (not on a
+  button, card or field) sends a sonar pulse that lights them up in blue as it spreads, the lit side of every form
+  brighter than its shadow. It moves at the display's frame rate, can be turned off in *Settings → Appearance →
+  Symbols in the background*, and stops with reduced motion and on a hidden tab. The theme still picks light or dark.
+- **A welcome wizard that asks how you want to use the board — and it comes back once after an update.** Besides the
+  language, it now asks how ▶ should start Claude (inside the board, in Terminal, in the Claude app or only copy the
+  prompt — the ways this computer lacks are greyed with the reason), what the board shows (Claude's usage, limits,
+  work feed, compact cards) and the GitHub sync, which it shows as it is when it is already on. A board that was in
+  use before goes through it once after updating, without the demo step; its projects, tasks and settings stay as
+  they are, and only what you change is written. *Settings → About* can open it again, and has a line with a
+  Telegram contact for questions.
+- **A card knows which Claude sessions worked on it.** Claude Code gives every command it runs a
+  session id, and `kb` now sends it along. When a session moves a task into work, leaves a note on
+  it or ticks its checklist, the board ties the two together: the card shows how many sessions
+  worked on the task and when the latest one did, and `kb show` prints the same, plus the transcript
+  of an earlier session — so a task sent back for rework can be picked up knowing what was already
+  tried. Grooming a backlog card does not count as work on it. Nothing of this leaves your machine.
+- **Claude's work, live, in the card.** A task some Claude session worked on gets a switch next to
+  its comments: *Claude's work* lists what the session did on this task — the person's request, what
+  Claude wrote, every step (commands, files read and edited, searches, helper agents and how long
+  they ran), which steps failed and what the person answered to its questions. It updates by itself
+  while Claude works. A session that works through several tasks is cut by task, so each card shows
+  only its own part. Read from Claude Code's own transcripts on this machine; nothing is sent anywhere.
+- **A card in work says whether Claude is working on it, waiting for you, or gone.** The badge follows
+  the Claude session live: *working*, *waiting for you* (a question, a permission, or it simply stopped
+  with the task still in work) and *session closed* when nobody is working on it anymore. The
+  dashboard's *Waiting for you* lists these tasks with how long they have waited, and when Claude
+  stops to wait for you the board plays the same sound and shows the same notification as for a task
+  ready for review, under the same settings.
+- **▶ starts the tasks in Claude.** The *To do* column has ▶ next to its copy button, and a card that
+  Claude has not taken yet has ▶ in its header (a manual task, labelled `noclaude`, has none — and loses it
+  the moment it is labelled so): the board opens Claude in the project's folder with the
+  job already written — a new session in the Claude app, or a terminal running `claude`. A card some
+  session already worked on gets ↗, which opens that very session (or resumes it in a terminal) — the
+  quick way to answer a Claude that is *waiting for you*. **Settings → Claude Code → Start tasks** chooses
+  the app, the terminal, or copying the prompt only; *Automatically* takes the app when it is there.
+- **Claude works right inside the board.** With `tmux` on the machine, ▶ opens nothing at all: Claude
+  starts at once in the background, in the project's folder, and the card opens on its new *Terminal*
+  tab — Claude's screen as it is, a field to answer it, the keys its questions need (a digit to pick an
+  option, the arrows, Esc, Shift+Tab) and a button to stop it. Claude keeps working when the board
+  restarts, and a Claude that has finished leaves its last screen. A second ▶ on a task that Claude is
+  already working on shows that Claude rather than starting another one. **Settings → Claude Code** has
+  *Inside the board* as a way to start tasks (and *Automatically* takes it first), and what Claude may
+  do there without asking: decide by itself what is safe (the default), edit files, ask about
+  everything, or only make a plan — skipping the checks altogether is not offered. A project without a
+  folder on this computer gets a one-click way out right there: create a new empty folder for it, or
+  choose the one with its code — and Claude starts as soon as it is set. The `claude` command keeps its
+  own sign-in, apart from the Claude app: when it is not signed in, the tab says so and signs it in
+  with one button, and the web addresses on its screen (the sign-in one included) are links. Claude
+  started there runs the board's own commands (`kb take`, `note`, `review`…) without asking, whatever
+  its mode — but not `kb done`, `cancel` or `rm`, and `kb mv` only to the working columns; and when it asks about every command in the manual
+  mode while the board is set to auto, the tab offers to switch it to auto in one click. It runs the
+  board's own `kb`, whatever version is on your PATH, and its card shows *Claude's work* even when the
+  `kb` it used did not say which session it was. Claude does not quit after handing a task in — it waits
+  for your answer — so the board closes its terminal by itself once every task it worked on is accepted,
+  cancelled or deleted; a task in *Review* keeps it, and a Claude in the middle of a reply finishes first.
+  The tab keeps the card as wide as the other tabs do: Claude's screen is laid out for the card, rather
+  than the card growing for the screen. A comment can go to that Claude too: the **>_** button next to the
+  send arrow hands it to Claude in the terminal and saves it, so Claude carries on with the task or the
+  talk without you switching tabs. It does that only while Claude waits for you at its prompt; while it
+  works or asks a question there, nothing is typed and the comment stays in the field.
+
+- **Claude now: what Claude is working on, top to bottom.** A new item under *Dashboard* in the sidebar
+  lists every task in work across the projects, grouped by the run of Claude that holds it. A run says
+  how much of it is handed in (*handed in 1 of 3*, with a ring); each task shows its step and how far it
+  is — by its checklist when it has one, by its stage otherwise (Prep 25%, Doing 50%, Deploying 75%,
+  Review 100%). The task Claude is on right now is marked, the ones waiting their turn follow it, and
+  runs nobody leads any more — the session closed, or nothing moved for two hours — are folded at the
+  bottom, so stuck work is easy to find. The sidebar item counts the tasks in work and shows a dot while
+  some Claude is working; it all updates live, checklist ticks included.
+- **Claude review: hand the acceptance of a project to Claude with one button.** A new sidebar item
+  under *Claude now* lists the tasks in Review by project, in the order of the sidebar. *Check in Claude* makes a check task in that
+  project — the tasks in Review at that moment are listed in it — and starts a fresh Claude on it: it
+  checks every task with tests of its own, simulates failures, moves what passes to Done and leaves the
+  report on the card of the check. While a check is open the button shows it instead — including when its Claude waits for an answer in the terminal, which opens right there — so the same tasks
+  are never checked twice at once — even two tabs pressing at once make one check. A check carries the
+  `check` label, so it is recognised on a board in any language, and the Claude started on it inside the
+  board moves exactly the tasks it lists to Done without asking; once that Claude has exited, the check
+  says its session is closed rather than that Claude is checking. The `kanban` skill says how such a
+  check is done.
+- **The sidebar shows how much of your Claude limits is used.** The 5-hour and weekly windows of a Pro or
+  Max subscription sit at the foot of the sidebar, with the time the board heard of them. Claude Code hands
+  these figures only to its status line, so `kb statusline` forwards them: put it in front of your own
+  status line (`kb statusline | your-command`) or use `kb statusline --line` on its own. A window whose
+  reset time has passed is not shown, and with nothing current the footer says why: *not connected · connect* when
+  the status line does not go through the board (*not connected · how* when your own status line has to take
+  `kb statusline` by hand), *go to another board* when another board holds it, and *no data*
+  otherwise — which is what a board used only from the Claude desktop app gets, since the app runs no status line.
+  A click on it opens **Settings → Claude Code**. A board that hangs does
+  not slow your status line down: kb sends the figures and hangs up without waiting for an answer. Switch to
+  another Claude account (`/login`) and the footer shows that account's figures from its first reading on.
+- **Answer Claude from the card.** When Claude Code is about to ask you something — a permission, a
+  question, a plan to approve — the PermissionRequest hook `kb hook permission` puts the request on top of
+  the card of the task the session works on, and the card's badge says *needs your answer*. Allow or deny it
+  with a reason, pick the answers (or write your own), approve or reject the plan with what to change.
+  Claude's own dialog stays usable all along, and whichever side answers first wins; a request answered in
+  Claude leaves the card by itself. An answer given the moment Claude has already taken one in its own window
+  does not pass for a given one: the card says it did not reach Claude. A tab whose answer lost to another tab's
+  says so, and a request the board lost when it was restarted is named on the card instead of vanishing. The whole command is shown, never cut short, and every answer is written
+  into the task's history — never sent to GitHub. Answers are taken from the board's page, which keeps other
+  web sites out but not programs on your computer — Claude included, if it may run `curl` without asking.
+- **Connect the board to Claude Code in one command.** `local-kanban hooks install` adds the two entries
+  above to Claude Code's settings.json — the status line through `kb statusline`, in front of your own, and
+  the PermissionRequest hook — after showing what it will change and asking. Your other hooks and settings
+  stay, the previous file is kept in the board's backups, and `local-kanban hooks uninstall` takes off
+  exactly what it added. A settings.json made read-only, or a folder it cannot write, is reported and left as
+  it was. Your status line keeps getting all of its input whatever shape its command has,
+  and keeps working even if kb cannot start; a settings.json that is a link stays a link with the same
+  permissions, a file Claude Code changed while you read the change is not written over, and running the
+  command again after a Node upgrade brings the board's entries up to date. Your settings.json keeps its own
+  indentation and line ends, and the copy of the previous one is a `.json` file in the board's data folder —
+  in the setup wizard, the folder you chose there. A board removed without disconnecting it leaves Claude Code
+  quiet: the hook checks that the board is still there before it runs, and the README says how to take the
+  entries out by hand. The setup wizard offers the same step.
+- **Settings → Claude Code.** One place to see whether the board's work with Claude Code can work on this computer:
+  the `claude` program, `tmux`, the Claude app and a terminal — the ways ▶ starts a task; the Claude folder, the
+  board's skill in it, the sessions and transcripts the cards read; whether the status line and Claude's questions
+  go through the board, and when the limits came last. Every missing piece says what does not work without it and how
+  to fix it, and a piece only one way of working needs — tmux, the Claude app, the connection — is marked as optional
+  (–), not as missing (✕). A copy of the board's skill that is not the version this board ships — one left from an
+  earlier version, or changed by hand — is said to be so, and **Settings → Skills** takes the board's version in one
+  click, keeping the old file among the backups. **Connect** writes the board's two entries into Claude Code's settings.json — the same as `local-kanban
+  hooks install`, after showing exactly what it will write — and **Disconnect** takes them off again. How tasks start
+  and what Claude inside the board may do moved here from General. The two switches at its top, **Show on the
+  board**, hide Claude's usage (the dashboard block and the card's line) and the limits footer — for anyone who
+  would rather not see them, or whose footer only ever says "no data" because they work in the Claude app. With two
+  boards on one computer only one can be connected: the section names the board that holds the connection, and
+  connecting this one asks before it takes the connection over.
+- **What a task cost.** A card some Claude session worked on shows a usage line under its sessions: the
+  tokens first, then roughly what they would cost on the API (≈, at the prices of the date shown — with a
+  subscription the money is notional), and how full the conversation memory was on the last answer; hover it
+  for the breakdown (input, output, cache reads and writes). Helper agents started for the task count as its
+  usage. `kb show` and `kb take` print the same line. The dashboard shows Claude's usage for the chosen
+  period and the tasks that cost the most — click one to open it. It is counted from Claude Code's
+  transcripts on this machine, once and then only what is added; nothing is sent anywhere. Claude Code often
+  leaves out the final count of a helper's answers, so their output is under-recorded — the card says so
+  instead of pretending the number is exact.
+- **Claude done.** A section under *Claude review* lists what Claude has finished in the last seven days: every
+  run — a batch started with ▶ or a single task — once none of its tasks is left in work, the most recently
+  finished on top, with how many of its tasks wait in Review and how many are accepted. Hand-ins made without a
+  session the board knows are put together by project and time and say so.
+- **Check the Review column from the board.** The Review column of a project has the same button as *Claude
+  review*: it makes a check task for the tasks in the column and starts Claude on it; while a check is open, the
+  button opens its card instead.
+
+### Changed
+
+- **A phone opens the board its own way.** Until you pick otherwise, a phone gets the dark theme, Claude as a chat, and
+  Claude now on the chats of its runs (Terminals); the language follows the phone's, as on a computer. A computer —
+  and a tablet — keeps the light theme, the terminal and the list.
+- **Claude reads a whole batch of tasks before starting the first.** Given several keys at once (▶ or 📋 on To do), the
+  board's skill had Claude read and do them one by one — what a later task said (a key, a decision, the same place)
+  reached it only after the first was done. Now it reads them all first and plans the order.
+- **The first run is a whole screen.** The welcome wizard no longer sits in a small window: the words and the
+  choice are on the left, and on the right a live picture of the step — a card going to Claude and back for your
+  check, Claude's terminal on a card, what this computer has for Claude Code (the same check as Settings → Claude
+  Code), the board as the switches leave it, the task and its GitHub issue. Pointing at a way of starting Claude shows
+  what ▶ does that way. Under "reduce motion" the pictures stand still; on a narrow screen the words come alone.
+  A board that went through the earlier wizard shows the new one once after the update; projects, tasks and settings
+  stay as they are.
+- **The project of a run opens its board.** In Claude now, Claude review and Claude done the project name at the top of
+  a run or a group is a link to that project's board.
+- **The terminals of Claude now → Terminals are as wide as their columns.** A column lays its terminal out to its own
+  width again (the card's narrower width left the right part of it empty), and keeps showing the old screen while
+  Claude reprints the conversation for the new width, so the start of the session no longer scrolls by — the card's
+  Terminal tab does the same. When a card and a column (or two tabs) show one terminal, its window takes the narrower
+  width instead of changing back and forth. A screen scrolled up has a ↓ in its lower right corner that takes it back to the end.
+- Enter in the title of a card moves on to its description instead of leaving the field.
+- **The light theme is crisper.** Text is near-black instead of grey, borders and the selected item stand out from the
+  cream, the sidebar has a tone of its own, and the symbols in the background are a quieter haze.
+- **The board opens in the light theme until you pick one.** It used to follow the system, so a computer in dark
+  mode got a dark board without anyone choosing it. *System* is still there in *Settings → Appearance*, and a theme
+  you already picked stays.
+- **Two themes: light and dark.** The theme *Claude* looked the same as *Dark* in the new look, so it is gone; a board
+  set to it opens dark, and the theme cards show the real colours of each theme. *System* still follows the computer.
+- **No note pops up at the bottom right when Claude starts to wait for you.** The board still rings, the system
+  notification still goes out when it is on, and the block on top of the sidebar lists what waits.
+- **A long list inside a window shows its scrollbar** — the skills in *Settings → Skills* and an open dropdown such
+  as the deploy skill of a project — narrow and grey, so it is clear there is more below.
+- **The sections of the sidebar start with a small grey icon** — the first letter of the name ("Pinned", "Other",
+  your own sections) — in line with the icons of Dashboard and Settings above them; the projects carry none.
+- **The title of a screen stands alone.** The line of description right of it (counts, the key and domain of a
+  project, "quick task capture") is gone.
+- **The header of the Terminal tab fits one line**: a few words ("Working", "Waits for you", "Waits for your
+  answer", "Finished · code 3") and a short "Stop" or "Close" button next to them, instead of a sentence that
+  wrapped to two lines and pushed the button down.
+- **"Claude now" puts no dot in front of its tasks** — the run's badge already says what its Claude is doing — only
+  ✓ in front of one handed in, and the task Claude is on is not highlighted — the line under its title says it.
+- **The board no longer zooms by accident.** A pinch on a trackpad or a phone screen, Ctrl with the mouse wheel
+  and a double tap used to zoom the whole page, and the board slid past the edges of the window. These gestures
+  do nothing now; ⌘/Ctrl with + and − and the browser's View menu still zoom for whoever needs larger text.
+- **Claude's work is a tab you switch on.** It is off by default now: *Settings → Claude Code → Show on the
+  board → Claude's work*. Off, the card does not show the tab or read the transcript for it, and a card with no
+  terminal shows no tabs at all — switched off while a card stands on that tab, the card goes back to its comments.
+- **A card in work opens on its Terminal.** A task in Preparation, Doing or Deploy whose Claude runs inside the
+  board opens on the *Terminal* tab — there are no comments yet, and the terminal is what is going on.
+- **The terminal of a card scrolls back.** Claude inside the board starts with Claude Code's classic renderer
+  (`CLAUDE_CODE_DISABLE_ALTERNATE_SCREEN=1`): the fullscreen one draws on the terminal's alternate screen, which keeps
+  no history, so the card showed the 40 rows in sight and nothing above them. The card's screen now reaches 1000
+  lines back, and what you read up there stays put while Claude goes on printing. Terminals started before this keep
+  no history.
+- **Words typed to the terminal of a task back in To do move it to Preparation**, and of a task in Review — back to
+  *Doing*, as the move of that Claude: it keeps the task when it takes it itself. Keys alone (Esc, the arrows, a bare Enter), a digit answering a question of
+  Claude's and a manual task move nothing.
+- **"Claude now" shows a Claude of the board that is still at work** with none of its tasks in work — busy, or
+  with commands running in the background after handing its task in (*working in the background*). Such a run is
+  not in *Claude done* meanwhile.
+- **A project Claude is at work in has a green pulsing dot** in the sidebar, in place of the orange one — also while
+  Claude works with commands in the background, or goes on after handing its tasks in; and a Claude that comes back
+  from the background to its prompt rings *waiting for you*.
+- **A Claude of the board stopped at the usage limit says so.** Its tasks say *limit reached · until 12:50am* (the
+  reset as Claude Code gives it), the outline of a card in work stops turning, *Claude now* and *Waiting for you* say
+  the same, and the board rings and notifies once when Claude stops — before, the cards went on saying *working*. A
+  look at the terminals that tmux did not answer keeps the stop rather than ringing it again, and a dialog that only
+  quotes the limit's line is no stop.
+- **Check in Claude waits for the work of its project.** While Claude is still at work on a task of the project
+  (Preparation, Doing or Deploy), the button in *Claude review* and the ▶ of the Review column are off, and a press
+  on them names those tasks. A blocked task, a task whose Claude session has closed, and one stuck for two hours with
+  no Claude on it do not hold the check. The press reads the project's tasks afresh, so a search narrowing the board
+  neither lets a check past work nor leaves tasks out of it.
+- **The Terminal tab of a card says what its Claude is doing** — working, at its prompt with commands running in
+  the background, waiting for an answer, stopped at the usage limit, or done with its turn and waiting for a word —
+  instead of *Claude works in the background* over every live terminal, also of a Claude that handed its task in
+  hours ago.
+- **"Claude review" no longer lists the report of a check**; a project with only reports in Review has no group, and
+  the count of the section in the sidebar leaves them out too.
+- **The lists of "Waiting for you" on the dashboard fold** by their heading and stay folded on this computer.
+- **The folders not on the board fold** under one row of the sidebar, folded by default.
+- **No footer of Claude's limits while the status line does not go through the board** — it said *not
+  connected · connect* to everyone who works without it. Connecting stays in *Settings → Claude Code*.
+- **Settings → Claude Code says why macOS keeps asking "node would like to access data from other apps".** Claude
+  inside the board runs under the board's `node` when pm2 runs the board (else under the app the board was started
+  from), macOS asks in its name and keeps no answer for a program without an app bundle. The row names which one to
+  give *Full Disk Access* — the path of that `node`, or the app — and opens that list.
+
+- **The board finds Claude Code where it is, not only on its PATH.** Besides the PATH it looks where the installers put
+  `claude` and, on a Mac, at the copy of Claude Code the Claude app keeps for itself — so ▶ keeps starting Claude
+  inside the board when the command-line install is gone. Settings → Claude Code says where it was found; when it is
+  nowhere, the section shows the install command itself.
+- **▶ in the Claude app no longer moves the tasks to Preparation.** The app only gets the prompt — you send it — and the
+  task waited in Preparation, a column only Claude moves, even when nothing was sent. Claude moves its tasks itself.
+- **`kb take` names the cards left for you.** When To do holds only tasks you do by hand, `kb take`
+  and `kb next` answer `empty (2 by hand: CK-12 CK-14)` instead of a bare `empty`, so Claude tells you
+  there is nothing for it rather than that the column is empty.
+- **`kb take` takes the task in one step on the board, a pinned card first.** Two sessions that took a
+  task at the same moment could both get the same one; now the board picks and moves it at once. A card
+  the person pinned (📌) in To do is taken first, and `kb` prints the pin in its lists. `kb lbl` adds and
+  removes labels instead of writing the whole set back, so calls side by side keep each other's labels.
+- **The server runs on Fastify 5.** Fastify 4 went out of support on June 30, 2025, and was never
+  meant for Node 24 and 26, which the board supports. Nothing changes in how the board looks or
+  behaves; the same security headers are sent with every page.
+- **A task you start goes to *Prep* at once.** ▶ on *To do* or on a card used to leave the tasks in *To do* until
+  Claude came up and took them — a good ten seconds, and in the Claude app until you pressed Enter there. Now they
+  move to *Prep* the moment Claude starts, whichever way it is started, and the card says *starting* until Claude
+  is up. The Claude that was started holds them, so another session cannot take them meanwhile — a second ▶ while the
+  first one is still starting Claude is refused, and a second Claude is not given a task another one works on. A start that is
+  refused or fails leaves the tasks where they were.
+- **Claude started inside the board shows no next-prompt suggestions.** In the card's terminal the dim
+  suggestion Claude Code puts in its empty prompt looked like text typed and not sent, and the card's ↵ on an
+  empty field sends a bare Enter. The board now starts its Claude with suggestions off — the variable Claude
+  Code documents for it, `CLAUDE_CODE_ENABLE_PROMPT_SUGGESTION=false`. Claude you start yourself keeps them.
+- **Updating from Settings → About brings the Claude Code skills along.** The skills the setup copied into the Claude
+  folder were left as they were, so Claude kept reading the rules of the version it was installed with until you ran
+  `local-kanban skills`. Now a copy that matches the board's version is brought up to date with the board, and one
+  that already differs from it stays as it is — for the board's own skill **Settings → Claude Code** says so and leads
+  to the button that takes the board's version. `local-kanban skills` and that button never write through a link —
+  a skill file linked to a file of your own is left alone. The README's steps for updating by hand say to restart
+  the board too.
+- **`kb mv` no longer closes a task.** Done and Cancelled are your word, so they are `kb done` and `kb cancel`
+  only: `kb mv <key> done` (or `cancelled`) is refused with exit code 2 and names the command to use. Claude started
+  inside the board runs `kb mv` without asking, and through it could move any task to Done without the question
+  `kb done` gets.
+- **A card in work is easier to see and to read.** Its moving contour is twice as thick and visible all the way
+  round, and the card has a warm tint under its text; the middle is drawn on a layer of its own, so a browser that
+  composites with the graphics card does not let the moving colour through under the text.
+- **The icon left of a section's title** is as tall as the title's capitals and centred on them.
+
+### Fixed
+
+- **Typing the title of a new task on an iPhone no longer throws the card down.** Once the task was created, the board
+  scrolled its new card into sight, and with the keyboard up that moved the whole page: the title went off the top and
+  the comment field took its place. On a narrow screen the board now scrolls only its own columns.
+- **A picture can be added to a task from a phone.** A phone can neither paste nor drop a picture, and a new task had no
+  way to add one at all. The card now has a "picture" chip beside "+ label": it picks pictures from the device, and a new
+  task is created by them as by a pasted one.
+- **No scroll bar in a column on an iPhone.** iOS drew its own bar at the edge of a column scrolled with a finger,
+  whatever the page asked; the column now cuts that strip off.
+- **`kb review` leaves no note when the move is refused.** It wrote the note first and moved the task after, so a task
+  that could not go to review (straight from To do) still got the note, and the retry added a second one. Now the task
+  moves first and the note follows; if the note does not save, kb says the task is in review and how to add it.
+- **Words sent to a Claude at work no longer vanish from the chat.** Claude Code holds what you send while Claude is in
+  the middle of a step (a command of several minutes, say) and takes it only when the step ends; the chat said
+  "sending…" for a minute and then dropped the words, as if they were lost. Now the chat shows them as "Claude reads it
+  when the step it is on ends" for as long as Claude Code holds them — after a reload and on another device too.
+- **A note from another Claude no longer takes the task away from the Claude working on it.** A Claude of one session
+  that left a note (or ticked a checklist item) on a task another Claude was working on became that task's latest
+  session — the card lost "session interrupted" and Continue of the real one. Now a note or a tick ties a new session to
+  a task only when no other session works on it; moving its status still does.
+- **Claude sessions go on after the computer restarts.** When the Mac froze or restarted, every terminal of the board
+  went with it, and its sessions read "session closed" with no way back — the chat even said the Claude was not started
+  inside the board. Now such a session reads "session interrupted" with **Continue** in the card's chat, in Claude now
+  and in its column of Terminals, as long as its conversation is on disk: Claude comes back to the same conversation.
+  Such sessions have their own view, **Broken terminals**, beside Terminals in Claude now (with how many there are), so
+  Terminals keeps only the Claudes that are running.
+- **The board on a phone.** The sidebar stood beside everything and left the board, a card and the columns of Claude
+  now 170 points of 390 — the page ran off the side. On a narrow screen the sidebar now hides: ☰ in the top bar opens
+  it over the page, and a pick in it, a tap beside it or Esc closes it; a card opens over the board, the whole width.
+  Closed, the panel is out of the way of Tab too — the focus no longer goes into it unseen.
+- **A long link or line in a comment wraps** instead of running past the edge of the card, where it was cut off unread.
+- **The session of a chat is chosen in the chat's own style** — a quiet chip with a chevron, like the model in the
+  field, and its list the chat's menu with a check by the session shown.
+- **A session picked in a chat that does not load says so.** The chip named the session picked while the chat, and its
+  field, stayed on the other one, with no word; the list now goes back to the session shown and the chat says why.
+- **A command of the chat runs as its block says.** Typed into the zsh of a Mac's Terminal, a block was read by zsh: a
+  `# comment` ran as a command or became the arguments of the one before it (`rm -f x # old log` removed `old` and
+  `log`), a `!` in a string failed, and a ```bash block split its words the zsh way. The block now goes whole to the
+  shell it names (bash, or zsh for a ```zsh block). A block longer than the board runs has no Run button.
+- **Continue, once.** A Claude of the board killed outright (SIGKILL) read "Finished" and folded away instead of
+  "Session interrupted" with Continue; two Continues of one session from two of its cards at once started it twice
+  (the second failed, or stopped the first); and Continue took a task the session never held along with its own.
+- **The pictures of a message do not weigh on the board.** The ten pictures of one message, asked for together, each
+  read and parsed the whole message — 400 MB of memory for 30 MB of pictures; the message is read once for them now.
+  A WebP or JPEG thumbnail keeps its place before it decodes, as a PNG did (the next words no longer jump), and a wide
+  thumbnail beside a tall one keeps its shape.
+- **A question read off the screen keeps an option's description apart from its label** — at a narrow width a label
+  that ended near the edge took the description in, and the "recommended" mark was lost — and keeps the lines of its
+  text.
+- **A chat spread over the window leaves the chats beside it where they were read.** In Safari the columns beside it
+  jumped by a few lines as it went over the window.
+- **Opened steps come into view.** "N steps" right over the chat's field opened under the edge, and the steps had to be
+  scrolled to; the chat now scrolls along as the steps open, until the last of them is over the field (or the head of a
+  long list is at the top).
+- **Every question of Claude in the chat is a form.** A question whose request never reached the board (the board was
+  restarted while it waited, or it waited past half an hour), and a permission Claude asks in its terminal, were the last
+  rows of its screen — cut off at the top — with the keys 1 2 3. The chat now reads it off the screen and shows the same
+  form as any question: the whole text, a row per option with its description, an answer of your own; a long one fades
+  at its edge while there is more below.
+- **The board away for a moment.** A refresh of the board while it did not answer left "Failed to fetch" uncaught —
+  in the browser's console and, once the board was back, in Settings → Errors — and the list unread. The board now
+  says "tasks did not load" by the title and reads them again once it answers, and a card the board reopens after a
+  move no longer stacks "The board is not responding" over the window: it follows the move when the board is back.
+- **A cleanup of leftover hooks no longer stops Claude's sessions of the board.** The board's hook for Claude's questions
+  went on the command line of every Claude it started, so `pkill -f "kb hook permission"` (meant for leftover hooks)
+  stopped all of them at once. The hook now goes to Claude as a file in the board's data folder. Other words of the
+  command line (the task's key, the allowed `kb` commands) are still there — never stop processes by a pattern.
+- **A command Claude Code moved to the background is background work.** A command that outruns its two minutes goes on
+  in the background by itself, and a skill can run as a helper there; the chat now names such work over its field with
+  the time it started, and a Claude done with its turn while only such commands run is said to wait for you — not
+  «Claude works in the background…» over nothing.
+- **The chat holds up under odd answers and a busy hand.** A long run of spaces in a heading, deep quotes or a line of
+  brackets in Claude's answer no longer hangs the tab or blanks the chat, and a wide table scrolls instead of squeezing
+  its words to a letter a line. Words the board could not deliver come back to the field of their own task, before
+  anything typed since, even after you went to another task; a picture put to a message and sent at once goes with it;
+  the light of a stopped Claude goes out after a failed look too, and a closed card stops asking the board. Words sent
+  "Now" from the queue stay in the chat until Claude takes them, and the queue waits its two seconds after them.
+- **The chat says why a message was not sent.** With too many messages already waiting for a busy Claude, the refusal
+  flashed and went; it now stays under the chat, in words that say what is wrong, until the words are edited. A chat
+  that could not load says so and tries again instead of "Loading…" for good, and background work Claude stopped
+  itself no longer runs in the chat forever.
+- **A chat column on a phone has a height of its own**, as the terminal has: it no longer grew to the whole
+  conversation and opened at its first words.
+- **⌘F in fields.** The search shortcut rebound to a bare letter no longer swallows that letter while typing in a
+  field, and ⌘F opens the search from the chat's field too.
+- **The first-run wizard.** A change of language no longer leaves the page one solid colour when the board does not
+  answer; Back stands with its arrow on the edge of the words above it; and a failed check of the computer in the step
+  Permissions says so with a button to check again, instead of the macOS row silently going.
+- **A check of Review moves what passes to Done without asking again.** Its Claude moved the tasks with a loop, which
+  the permission check of Claude Code sometimes refused, and the person had to say "move them yourself". The check now
+  tells it to move each task with a command of its own — the one the board allows it.
+- **A card left open while its task moves to another project follows it.** Moved from another tab or by `kb`, the
+  task's open card went on showing its old key and no "Moved from"; it now opens again under the new key, with an edit
+  in progress saved and a half-typed comment kept.
+- **The answer field of Claude now → Terminals no longer lags.** With a few columns open, every letter typed laid out
+  their screens of a thousand rows again; a key now costs what it costs with no screens at all.
+- **Words typed to a task in Review that are not about it no longer leave it in Doing.** Typing into the terminal of a
+  handed-in task sends it back to Doing, as before; once Claude's turn is over, a task it did not touch — no move, no
+  note, no commit and no changed file in the project, an edited file git does not track yet included — goes back to
+  Review by itself, with the time it had. So does one whose Claude moved on another of its tasks in that turn: the
+  words were about that one. The card's timeline says which moves were made by typed words and which the board undid.
+- **`kb` stops quietly when the reader of its output goes away** (`kb show KEY | head -1`), instead of printing a
+  stack and exiting with the code that means "the board is not responding".
+- **Claude now shows a Claude that put its task back to To do and waits for your answer.** An `ask` task goes back
+  to To do with Claude's questions; its terminal waited at the prompt, and the section was empty. A task the person
+  moved to To do afterwards does not count.
+- **A terminal in Claude now → Terminals opens at its end, not at the start of the session.** A column widened its
+  terminal's window on every opening of the section, and Claude Code then printed the whole conversation again from
+  its first line. A column now only narrows a window its lines do not fit.
+- **The background of symbols no longer fades out and back when the window is hidden.** The last frame stays, and
+  the clouds go on from where they were once the window is back.
+- The ✕ of a column in Claude now → Terminals is as tall as "Open in the card" beside it.
+- **A task handed in stays in Review when you go on to the next one in the same terminal.** Words typed to a card's
+  terminal send its task from Review back to Doing — and after a hand-in, words like "now take the next one" did that
+  too: the finished task sat in Doing and kept its run in Claude now. When the same Claude takes another task into work
+  without having touched this one since, it goes back to Review with the time it had — and the task Claude went on to
+  is the one its card and Claude now say it works on.
+- **Copying from a short terminal screen keeps the lines Claude ended apart.** On a fresh screen, where no text had
+  reached the edge yet, the longest line was taken for the width of the window, and `git push` under
+  `git commit -m "…"` was copied as part of it.
+- **New task shows every task added last.** A task added from outside right after another (`kb add` in a batch) did not
+  show under the form until something else happened on the board.
+- **Handing in one of two tasks in work leaves Claude on the other.** A Claude with two tasks in Doing that handed one
+  in was on no task at all: the card of the other did not say "working", and Claude now marked none of its tasks.
+- **Text copied from a terminal is the text Claude wrote.** A selection on the screen of a card's Terminal tab or of
+  a Terminals column used to copy the screen's rows: a command Claude laid out over three rows of a narrow window came
+  with line breaks and indents inside it and broke in the shell. Now the rows the width broke go back together, the
+  indent the whole selection shares goes, and lines Claude ended stay ended.
+- **A Claude that waits for a command of its own no longer calls you.** When the turn that just ended started work in
+  the background (a measurement, a build, a background agent) that is still running, the card and Claude now say
+  "working in the background" and nothing rings: the end of that work wakes Claude by itself. A tunnel or a dev server
+  left running does not count — Claude does not wait for those — and after two hours it calls you again.
+- **Claude now no longer says "working" over runs whose tasks are all done.** Newer Claude Code marks a Claude at
+  its prompt with commands still running in the background as `shell`, and the board took the unknown word for a
+  turn in progress. Such a Claude now reads "working in the background" — on cards, in *Claude now* and in the
+  card's *Terminal* tab — and a terminal whose tasks were all accepted closes again. A Claude with no task in work
+  that has sat at its prompt for over two hours, a background command still running, leaves *Claude now* for
+  *Claude done*: a command that long with no turn of Claude is a leftover, not work.
+- **Terminals of the board that only wait no longer pile up.** A terminal whose tasks are all in Review (or put
+  back), with its Claude idle at the prompt for six hours, now closes by itself, as one whose tasks were accepted
+  does; a question on its screen or a command still running in the background keeps it. Forty such terminals,
+  some ten days old, held 12 GB of a 16 GB Mac. The conversation stays: ↗ on the card picks it up again.
+- **The tabs of a card stay on one line.** In a card 360 pixels wide *Claude's work* and its count, and
+  *Terminal* and its dot, broke onto a second line. A count over 999 is written short (*1k*), and only the
+  words of a tab give way, so the count and the dot always show.
+- **A clicked card keeps its ring while it opens.** A board event arriving meanwhile (Claude moving cards) put the
+  ring back on the card open before; and the answer to an earlier, slower click no longer replaces the last one.
+- **Claude's screen on the Terminal tab follows its answer.** Pressing a key or sending an answer lets go of a
+  selection on the screen, which used to hold the screen still; a mouse button pressed on the screen holds the
+  redraw too, so a slow drag selects.
+- **A question of a board's Claude always rings.** One that the page first heard of from its own refresh used to
+  be taken as known and never rang; a line of Claude quoting a question's footer no longer reads as a question.
+- **One run is never in *Claude now* and *Claude done* at once.** A manual task in work stands on its own in
+  *Claude now* and no longer holds Claude's run there.
+- **The Review column's ▶ wakes up after another tab made the check,** instead of staying dead until a refresh.
+- **Settings:** the install command of Claude Code (and every command shown to be run) can be selected and
+  copied; the effort list stays inside the Settings window; a stored effort level the board does not know is not
+  sent; the advice for an API key names both places it comes from; *GitHub sync* says above its fields that the
+  repository is made by you and that `gh` needs the `project` scope.
+- **Settings → Claude Code tells the truth in more corners.** "The status line runs" is shown only for a recent
+  reading while the status line goes through this board; a `KB_CLAUDE_DIR` pointing at a missing folder disables
+  Connect instead of creating it; a `KB_CLAUDE_BIN` pointing nowhere is blamed on the variable; a Claude folder that is
+  not writable no longer offers Connect or keeps a copy of settings.json on every try; a refusal is said once; losing
+  the connection keeps the list on screen; and a double click on Connect no longer closes the confirmation.  Installing or refreshing a skill — from Settings → Skills, by an update, by `local-kanban skills` or the setup — no
+  longer creates a missing `KB_CLAUDE_DIR` folder either. An update of a clone no longer logs that its own skill "did
+  not come along".
+- **Smaller fixes found while accepting this release.** A hook written by an earlier version of the board reads as
+  "written by an earlier version — Connect brings it up to date", not as a board that moved. Duplicate on one of two
+  open checks opens the other one instead of advising ▶, which that check would refuse. A tmux that does not answer in
+  time no longer turns a fresh start of Claude into "session closed". An open card's status log keeps two moves made
+  outside that end where they began. Answering the refusal of Send back keeps the reason you typed. Settings → Skills
+  says a refusal in the board's words, not the server's.
+- **Settings → Skills names the right folder.** It used to say the skills live in `~/.claude/skills` even when the
+  board reads them from `KB_CLAUDE_DIR` or `CLAUDE_CONFIG_DIR`; it now shows the folder it really uses and where that
+  comes from, as Settings → Claude Code does.
+- **`local-kanban hooks` and `local-kanban skills` keep their copies with the right board.** Run from a terminal,
+  they kept the copy of the file they replaced in the data folder of their own process, so a board with its data
+  elsewhere never showed it under Backups. They now ask the running board where its data is; when it does not
+  answer, they write nothing unless you name the folder with `--data <dir>`.
+- **A refusal of the board is said in its words, everywhere.** On a task deleted in another tab, removing a
+  link, adding a checklist item, deleting a comment, Send back, Accept and ▶ now say the task is gone and close
+  its card, instead of failing without a word. An edit of an open task the board fails or refuses is said in the
+  card — a failure waits behind "not saved" and is sent again — instead of being lost. Settings, the project
+  wizard, the project's page and Chaos say why something was refused in the board's language.
+- **An image too large for the board is said to be too large in Safari too.** Pasting an image over 6 MB
+  into a card, a comment or Chaos said "no connection to the board" in Safari about half the time; the board
+  now checks the size before sending and says the image is larger than it takes.
+- **An open card's status log follows a move made elsewhere.** When a task is moved by `kb mv`, in another
+  tab or by starting Claude, the open card's status log now shows the new line at once, not only after the card
+  is opened again.
+- **Uploading a backup can no longer delete files.** A JSON request to the upload route could name any
+  file — the board's own database included — and the board deleted it; the route now takes only the uploaded
+  file itself, and a broken upload no longer blocks every later one. A board opened from a backup no longer
+  writes to the live board's files, takes no snapshots of itself and cannot update the code it shares with
+  the live board, and copying its attachments no longer freezes the live board for half a second.
+- **A snapshot is never rotated away right after it is taken.** From schema version 10 on, the snapshot
+  taken before a migration would have been deleted at once; west of Greenwich, *Back up now* right after the
+  update deleted the snapshot it had just taken; saving the `deploy` skill removed the snapshots
+  of `deploy-tickets` and its own; two saves within a second lost the previous text. Unfinished snapshot
+  files left by a crash are cleaned up at start, and snapshot names and the times under Backups follow the
+  local clock instead of UTC — the snapshot taken before a migration, the one made with *Back up now* and the
+  copies kept when a skill or a project's CLAUDE.md is saved included, with the UTC offset in their names so that older names still sort
+  in their true order.
+- **The card keeps your edits, and only yours.** An edit made elsewhere while the cursor sat in a field is
+  no longer undone by the next click in the card; a card left open on the dashboard or in Chaos follows the
+  task; a closed card is never saved again as a copy (Cmd+Enter with the sidebar pinned, or a retry after
+  the server came back), and an edit made while the server was down reaches it even after the card is
+  closed. Project settings and the CLAUDE.md and skill editors ask before dropping typed text, Save sends
+  only what changed, the board's keys stay out of open windows, and an open tab picks up a new version of
+  the page's markup or styles.
+- **A big board stays light.** The card loads Claude's work only on its tab, and a session with dozens of
+  helper agents no longer re-reads all their transcripts on every update (with 30 helpers that kept the
+  board above its memory limit). A note, a link or an attachment changes one card instead of reloading the
+  whole list in every tab, a burst of events is drawn once, the dashboard stops recounting on events that
+  change none of its numbers, and the endless pulses no longer redraw the page every frame (reduced motion
+  stops them). The list of all projects no longer comes back empty from 16,384 tasks on.
+- **Search shows the answer to what you typed.** A slow answer to an earlier, shorter query could stay on
+  the board as the result; now only the latest request counts, typing waits a little longer before asking,
+  and a task that arrives while a search is on is filtered like the rest.
+- **`kb` and the API refuse malformed input instead of acting on it.** A task key went into the address
+  as is (`kb note '../backups?' …` took a database snapshot), `kb constructor` crashed, a foreign service on
+  `KB_URL` passed for the board, `prefix: null` renamed every task to `NULL-N`, and a slug with `../` wrote
+  outside the data folder. Errors now name the field. The error log no longer loses its history to one
+  failure repeated on many tasks; `local-kanban skills` refreshes the skills in place and keeps a copy of
+  what it replaces; `--data=~/x` lands in the home folder; deleting the demo closes its GitHub issues.
+- **Claude's exit code shows on Linux too.** tmux there sometimes missed the signal of Claude's exit, and
+  the card said only "Claude has finished"; the board now reminds tmux once.
+- **A task you have just made is a whole card at once.** A task made with **+** had no ▶ in its card, no
+  *+ link*, no status log and no *Cancelled* among its statuses until the card was closed and opened again —
+  so it could not be started in Claude right away. Now they are there as soon as the task is made.
+- **Smaller interface fixes.** A right-click change or a comment made while the server is down waits for it
+  instead of failing silently; the Russian board writes sizes and hours with a decimal comma, names the
+  Local and Other sections in Russian and puts the word for "subfolders" in the right form for the number.
+- **Claude inside the board: a language, long replies, an honest start.** A board run by pm2 has no
+  `LANG`, so its Claude worked in the C locale — `grep -i` missed capital Cyrillic, `wc -m` counted bytes;
+  it now gets the person's language or a UTF-8 one. A reply or a comment longer than about 8 000 Cyrillic
+  letters failed with "command too long"; text now reaches Claude whole up to the promised 14 336
+  characters, and a key name in a reply (`C-c`) is typed, not pressed. A prompt longer than tmux takes is
+  refused before anything starts, and *auto* takes it to the app or the Terminal instead. A project whose
+  folder is not on this computer no longer starts Claude in the home folder from the Terminal or the app —
+  the board offers to make the folder, as it already did inside the board. The *working* badge no longer
+  trusts a process id that another program got later. The Terminal tab looks at the screen only after the
+  last look came back, with one call to tmux, and is sent the screen only when it changed.
+- **A retried GitHub issue no longer comes twice.** When the answer to creating an issue was lost, the
+  retry looked for the issue through GitHub's search, which indexes minutes late — exactly when the retry
+  comes — and made a second one. It now looks among the newest issues first, and searches only after that.
+- **▶ right after the last Claude of the board was stopped no longer fails.** tmux was still shutting down
+  its server, and the start got "server exited unexpectedly"; the board now tries such a start once more.
+- **The board always starts on matching page and code.** Safari, and the board installed from it as an app,
+  reopens a window and goes back and forward from its cache without asking the board — telling it to check
+  first does not stop that. After an update this could pair the new code with the old page: the board stayed
+  on its loading placeholder and then came up empty, and *New task* did nothing. The page, its code and its
+  styles are no longer kept in the browser cache, the installed app takes the page from the board every time,
+  and should a script of the board still fail while the board loads, the page reloads once by itself.
+- **Found while these fixes were accepted.** One helper transcript the board cannot read no longer turns the
+  whole *Claude's work* tab into an error; a board whose first load failed (the server answered 500 for a
+  moment) tries again instead of staying empty until the page is reloaded; `kb lbl` against a board that was
+  updated but not restarted says to restart it instead of printing labels it did not set; and `local-kanban
+  skills` names a skill it could not refresh and exits with code 1. Twenty requests for *What's new* or the
+  update check that arrive together share one trip to GitHub instead of twenty (and one run of `gh` when the
+  anonymous request is refused); a `kb take` of an older version can no longer take a task another session
+  already has in *Prep* — it is told the task is taken, and the kanban skill tells Claude to leave such a task
+  and say so; and a page whose code failed to download at start
+  reloads once instead of staying empty.
+- **The Update button on a clone installs the dependencies too.** It ran `git pull` and nothing else, so
+  a release that changed the dependencies restarted the board on new code with the old packages. It
+  now runs `npm install` after the pull, the way `npm run update` always has. A clone that picks up
+  this release with a plain `git pull` (which is all the button did until now) keeps working on its old
+  packages until `npm install` runs.
+- **Claude's work keeps the step that handed a task in.** Status changes are stored to the second, so a `kb
+  review` run within the same second as the move fell out of the task's feed, or showed up in the next
+  task's. That second now belongs to the task handed in, and a step whose answer came after the move shows
+  how long it took and whether it failed.
+- A card that Claude takes while it is open now locks right away (also in Preparation) and hides “+ label”,
+  “ask?”, “+ link” and the link and checklist delete buttons; when the task leaves work the card unlocks
+  with those buttons back and shows what Claude changed meanwhile — no need to reopen it. Before, a label
+  picked on such a card showed up but was never saved, and taking `noclaude` off a task in work was lost. The card's ⋯ menu and the link picker, if open at that moment, close too.
+- When the server does not answer while you switch projects or search, the board keeps the project and the
+  tasks you had, tells you so (“project did not open”, “search did not reach the server” next to the title),
+  and finishes the switch or the search by itself once the connection is back — a click on the notice
+  retries at once. When the server answers with an error instead, the board says so and tries again by itself in a few seconds; the first load of a board that fails shows its columns with a note and a “Try now” button, and a switch that takes a moment shows the loading skeleton instead of the previous project's cards under the new title. Before, the board went blank and lost the selected project without a word.
+- A change made with the right click on a task that was deleted elsewhere now says “This task no longer
+  exists” in the board's language and takes the leftover card off the board, instead of “Not saved: task not
+  found”.
+- The rest of the task card speaks the board's language too. The link picker, typing into Claude's
+  terminal, starting Claude and pasting an image say why the server refused in the board's words instead
+  of its English (“the task to link to was not found”, “limit of attachments per task: 30”); pinning,
+  cancelling or deleting a task that was deleted elsewhere says so and clears the leftover card instead
+  of doing nothing. A server error now says the board answered with an error and that the details are
+  under Settings → Errors, rather than that the board is not responding.
+- A comment sent while the server is down now stays in the card's comments as “not sent — goes out once the
+  connection is back” (with × to drop it) and is sent once when the connection returns; sending the same
+  text again while it waits no longer makes a duplicate.
+- The “working” dot of the card's Terminal tab goes out when Claude exits — no need to reopen the card, and also when the task was still in To do or Backlog.
+- A new task the server does not create says why at the top of its form — no connection, a server error
+  (it is retried from the “not saved” badge or with your next change) or a refusal in the board's words —
+  instead of failing silently, and typing on no longer sends a request for every letter.
+- **Claude is never started on a task you do by hand.** ▶ and the copy button on *To do* put every task of the
+  column into Claude's job, the ones labelled `noclaude` too, and `kb take` could pick such a task. Now both buttons
+  leave manual tasks out (a column of manual tasks only has neither), the board refuses to start Claude on one —
+  also from a tab that has not heard of the label yet, which then says why and loses its ▶ — and `kb take` passes
+  manual tasks by.
+- **The button that archives a project says so on the Russian board.** It read as the name of the settings
+  section: the Russian dictionary held the word twice and the second translation won without a word. Each
+  word is in the dictionary once now (the English name of that settings section is *Archived*), and a test
+  keeps it that way.
+- **The status line's limits reach a board that reads the request a moment late.** `kb statusline` hung up the
+  moment its request was out, and a board with an asynchronous step before reading it lost the reading. kb now
+  waits up to 300 ms for the board's answer — a board that never answers still costs the status line a moment,
+  not a second.
+- **A second open check is no longer made silently.** Putting the check label on a task, or sending the report of
+  an earlier check back to To do, while the project has an open check goes through and says so — on the page and in
+  `kb`. *Duplicate* on the open check itself says it is still open and how to start it again, instead of opening the
+  same card, and no longer lands in Settings → Errors; a check in Backlog can be copied.
+- **A Claude that exited at once, or a start nobody sent in the Claude app, no longer reads as starting or
+  checking.** A Claude started inside the board that exited right away (not signed in, crashed) showed *starting*
+  for two minutes; now its card says *session closed* at once. A task or a check started in the Claude app says
+  *waiting to start in the app* until someone presses Enter there, and *session closed* after two hours — the
+  Claude review section used to say *Claude is checking* for good.
+- **Claude started in the Terminal or the Claude app works with the board that started it.** Its `kb` went to
+  the board on the default port (3100), so on a board on another port Claude moved tasks on a different board.
+  The Terminal command now carries the board's address, and for the Claude app (and a copied prompt) the prompt
+  ends with a line naming it — only when the board is not on the default port.
+
+### Security
+
+- **`@fastify/static` 10.1.3 closes two advisories** — a path-traversal bypass of route guards
+  (GHSA-83w8-p2f5-377r) and a bypass of `allowedPath` (GHSA-8pvw-jcv7-9cmj). The board was not
+  exposed to either, because it protects every request the same way rather than individual static
+  files, but `npm audit` reported them to everyone who installed the package. The version is pinned
+  on purpose: later releases need Node 22.12 or newer, and the board still supports all of Node 22.
+
 ## [1.15.0]
 
 ### Added