Skip to content

Releases

v2.1.35-fork.13

Choose a tag to compare

@github-actions github-actions released this 12 Nov 05:48

Built on MyGo 0.9.0, MyGo's latest release, past the commit Kero 2.1.35
requires, with the fork's MyGo changes on top. It brings MyGo's changes
since then, among them highlighting of Svelte, HCL/Terraform, ERB,
Jinja and Nix files and of scripts by their #! line; the update window
offering a release that came out while it was open; a transparent
WebView background and the shadow of borderless windows fixed on
Windows; Ctrl+I, Ctrl+M and Ctrl+[ sending Tab, Enter and Esc in the
terminal; and a program killed in the terminal no longer leaving mouse
and key reporting on.

  • Diffs that aren't edited end where git's do: a staged change's
    diff, a commit's and History's no longer show the empty line after a
    file's last newline. A last line without a newline is marked No
    newline at end of file
    , and a line that gained or lost its newline
    shows as that line changed, not as an empty line added or deleted
  • Unstaging chosen lines of a staged change keeps each line's own
    ending in these diffs, and a file's last line its missing one

v2.1.35-fork.12

Choose a tag to compare

@github-actions github-actions released this 11 Nov 07:12
  • A pull request's review session is named after the pull request's
    number and title, #42 Fix the login redirect, not its number alone,
    and its Review mark sits at the right of its row, under the time
    in two-line rows

v2.1.35-fork.11

Choose a tag to compare

@github-actions github-actions released this 10 Nov 12:03
  • Claude Code acts on your page answers again: an answer you send
    from a page, Send to Session… and a review's Send to Agent
    reached Claude Code as a background event, which Claude would not take
    as your words, and it refused to act on your "accept". What you send
    yourself is now typed into Claude's prompt as yours. Orchestrators'
    reports and PR watch messages still arrive as events
  • Orchestrators work in a new way: an orchestrator gives each piece of
    work to a subagent of its own or to a child session, whichever fits.
    It tells each child to show its decisions to you as a page in the
    child's own session, and to report when a decision waits. The
    orchestrator then tells you only which session waits, and does not
    copy the decisions into its own chat
  • Agents can change any setting, the prompts too: kero config get --default <key> prints Kero's own prompt, --text reads a prompt as
    its lines, and kero config set --file <path> <key> sets it from a
    file. --repo owner/name reads or sets one repository's value, such
    as its review prompt, and keeps the other repositories' values
  • Ama answers in a cheerful, quirky way

v2.1.35-fork.10

Choose a tag to compare

@github-actions github-actions released this 09 Nov 11:31
  • Kero's messages no longer type over what you are writing: an
    orchestrator's reports, PR watch messages and a page's answers reach
    the agent through its own way in, as a turn of its own, and your
    half-written prompt stays as it is. Claude Code gets them through
    Kero's plugin, Codex through codex queue, OpenCode through its
    service, and Pi and OMP through Kero's extension
  • Press Update in Settings, or run kero skill setup, once to add
    Kero's plugin to Claude Code. kero skill status shows it as missing
    until then, and messages are typed in as before
  • A long message reaches Claude Code as its start and a file with the
    whole text, which Claude reads
  • Where a message is still typed in, as for an agent with no way in,
    Kero stops pressing Enter for it once you type, so it never sends your
    own text with it
  • New Session with no worktree starts in the project's checkout:
    from a session in a worktree, the dialog with the worktree switch off
    started the new session in that worktree. It now starts it in the same
    folder of the main checkout

v2.1.35-fork.9

Choose a tag to compare

@github-actions github-actions released this 08 Nov 20:39
  • An orchestrator hears of a child's turn when it ends: Kero's hook
    in Claude Code and Codex now runs at the end of each turn too, and
    tells Kero the turn's end, its reply and what still runs in the
    background, so that a report no longer waits on the child's screen and
    transcript, nor reports a compaction as a turn. The Pi, OMP and OpenCode
    plugins tell it as well
  • A turn that fails, as on an API error, is reported with its error, and
    a Codex or Pi turn you interrupt as interrupted
  • An interrupted Claude Code turn, or an agent with no hook, is reported
    from its screen, as before
  • Run kero skill setup once to add the new hooks: kero skill status
    shows a Claude Code or Codex hook of an older Kero as outdated. The hooks
    run in the background and print nothing, so they never hold an agent up

v2.1.35-fork.8

Choose a tag to compare

@github-actions github-actions released this 08 Nov 20:07
  • An orchestrator hears what its child said: when a child of Claude
    Code, Codex or Pi finishes its turn, its report carries the reply it
    ended the turn with, so that the orchestrator answers a question at
    once, without reading the child's screen. A long reply comes as its
    first and last thousand characters, and kero read <child> --reply
    prints it whole
  • The report says what the child left running in the background, as "1
    shell, 1 monitor still running", so that the orchestrator knows it isn't
    done
  • No report for a turn with no new reply, as one that only compacted the
    conversation, nor for one the child ended with its own
    kero report --to-parent
  • Children are told that their last reply goes to their orchestrator, so
    they end a turn with their result or their question
  • kero prompt sends a long text whole: it pastes the text, so that
    its lines reach an agent as one prompt, instead of typing them one by
    one, each line's end an Enter that sent only the last line. It presses
    Enter once the agent has the text, and its timeout says that the prompt
    went and the agent is still working
  • kero read --reply prints an agent's last reply, whole, from its
    transcript

v2.1.35-fork.7

Choose a tag to compare

@github-actions github-actions released this 08 Nov 13:06
  • Deleting Ama's chats: History in Ama's window has Delete Chat, to
    delete one past chat, and Delete All Chats…, to delete all but the one
    running, each after a confirmation. A deleted chat leaves Ama's
    History; its agent keeps its own record of it
  • Ama's window no longer has Clear, which cleared the screen of a
    full-screen agent and so did nothing
  • Ama's header has room between the frog, Ama's name and its status
  • The frog's dot for an answer you haven't read is the theme's green, not
    grey on the grey frog
  • A confirmation in a small window, as Ama's can be, keeps a margin
    around it

v2.1.35-fork.6

Choose a tag to compare

@github-actions github-actions released this 08 Nov 12:13
  • Ama knows Kero's docs: Kero writes its docs into Ama's folder, in
    docs/, with an index of them in its CLAUDE.md and AGENTS.md, so that
    Ama reads the page it needs before it answers, as how setup scripts in
    kero.json work, instead of searching
  • New pages in the docs: Worktrees (with setup scripts), Files &
    Search, Pull Requests, Orchestrators, Ama, The Window and
    Troubleshooting, in English, Japanese and Chinese, and fixes to the
    others: kero daemon stop saves the sessions, kero do runs actions,
    the keybind syntax, the environment that Kero gives agents
  • Ama's icon is the frog of Kero's logo, solid with its face, belly
    and feet, as large as the icons beside it

v2.1.35-fork.5

Choose a tag to compare

@github-actions github-actions released this 08 Nov 11:17
  • Ama's role prompt is a setting: Settings › Agents › Ama › Role
    prompt (ama-prompt) shows Kero's, which you can change, and Reset to
    Default brings back. Kero writes it into the CLAUDE.md and AGENTS.md of
    Ama's folder, which its agent reads from its next chat
  • Arguments for Ama and orchestrators, whichever agent they run:
    Settings › Agents › Ama › Arguments (ama-agent-args) and Orchestrators
    › Arguments (orchestrator-agent-args), as --model opus, go after the
    agent's command as it starts and as it resumes, so that one agent serves
    all three
  • The picker's screenshots box an element at the page's top or side, as
    example.com's <body> at 0, with every line: a side at the picture's
    edge has its line inside the element, and its number, in the page's
    screenshot, inside too
  • Importing sign-ins says that the browser may ask in any of its windows,
    an installed app's (PWA) too, often the one used last. Kero connects to
    the browser, whose installed apps share its profile, and imports the
    same cookies whichever window asks

v2.1.35-fork.4

Choose a tag to compare

@github-actions github-actions released this 08 Nov 10:41
  • Ama's window: Ama opens in a window of its own, above Kero's, which
    you move by its header, to another display too, and resize by its edges.
    Kero remembers its size and its place by the nearest corner of Kero's
    window, in DIPs, so that it keeps its size on displays of other scales:
    it goes with Kero's window as that moves or resizes, and stays inside a
    display as Kero's window nears a screen's edge or a display goes away.
    Right-click the frog, or Settings › Agents › Ama › Window, to put it back
    above the frog
  • Ama's text has a size of its own, a point under the terminals' unless
    Settings › Agents › Ama › Text size (ama-font-size) says otherwise
  • Claude Code's status line is off in Ama's window
    (ama-hide-status-line, on): Kero puts an empty one in the
    .claude/settings.json of Ama's folder, beside what you put there, and
    leaves your own Claude Code settings alone. Claude Code has no way to
    remove it but turning off every hook, Kero's among them, so an empty row
    stays where it was
  • With detailed sidebar rows, an orchestrator's mark and the chevron that
    folds its children sit at the right end of its second line, under the
    time and status, so its name keeps the first line; compact rows keep
    them by the name
  • An orchestrator's role prompt, and every report Kero sends it, tell it
    to answer a report that needs nothing with a plain OK, not a summary
  • A child closed while its orchestrator was busy no longer reaches it as
    "finished its turn: it is idle" afterwards. Kero handed a report to the
    agent's mailbox as soon as it was due, where it waited, already
    written, for the orchestrator's turn to end, and a close never reached
    it. Reports now wait in Kero until the orchestrator is at its prompt,
    and a child that closes meanwhile takes its state lines (idle, waiting,
    failed, exited) out of them: they would have the orchestrator act on a
    session that is gone, and "closed" says all that is left. What it
    wrote with kero report --to-parent still goes, as that is its result
  • The picker's HTML of an element, its parent and its siblings shows a
    <script>, <style>, inline <svg>, <template> or <noscript> as
    its tag and attributes with its contents' size, such as
    <style data-vite-dev-id="…">/* 1.2 KB omitted */</style>, rather
    than the code or path data
  • A message to an agent names the pane's title, the session and the page
    as they are written, titled "C:\Program Files\PowerShell\7\pwsh.exe"
    rather than with every backslash doubled, and on one line: a line break
    in a title, or in a label a page sends with, reads as a space, and a "
    in it as '