agent-fleet 0.25.0

Branch names for a session started from a work item now follow the repository's own rules — its
git-flow settings, a .agent-fleet/branches file or Bitbucket's branching model. Open pull requests in
Work items show their CI status and conflicts. af-aws-exec runs your assume-role and
credential_process profiles, and Settings › AWS profiles/SSM logs in from the profile's row. A
Managed session that fails to start now shows what its CLI printed.

New

• Branch names follow the repository's rules. Starting a session from a work item fills in the
  branch name and its base from the repository's .agent-fleet/branches, its git-flow settings
  (.gitflow or the clone's gitflow.*) or Bitbucket's branching model, says which rule chose the
  base, and notes a name that does not match. It never refuses a name. The rename dialog offers the
  repository's branch kinds, and the branch template gains {ref}, {num}, {project}, {type}
  and {prefix}. Initializing git-flow from the Console is still to come
  (#1126).
  (#1120,
  #1124,
  #1125)
• Open pull requests in Work items show their CI status and merge conflicts — failing, running
  or passing, and a mark when the PR conflicts with its base. Nothing is drawn when GitHub has no
  answer yet.
  (#1113)
• The work item's "already started" list shows session names, and a session that has since been
  archived can be restored from there.
  (#1108)
• af-aws-exec runs your own assume-role (role_arn + source_profile) and credential_process
  profiles when you name the account with --account. It checks the whole chain first and passes
  only short-lived credentials. Assuming a role from a Settings SSO profile is still to come
  (#1109).
  (#1107)
• Settings › AWS profiles/SSM (was "AWS SSM") has Log in and a login-state badge on each
  profile row, so you can sign in before a command asks.
• A Managed session that fails to start shows the last lines its CLI printed, with secrets
  removed, instead of only "did not start in time".
  (#1076)
• create_session can set the child's reasoning effort (it needs model too).
  (#1068)

Improved

• The image studio on a phone: the draft bar stays in view while you type, the composer sits on
  the pane's bottom edge, toasts no longer cover the composer, and a tap outside dismisses them.
  "N pictures are ready" stays until you look, and a studio started in a new worktree takes that
  worktree's name.
  (#1089,
  #1090)
• AI-suggested titles for a session started from a work item keep the item's key (#1146,
  PROJ-123).
• The Work items filter has the same box and × as the other filters, and Escape clears it.
• A translated question card asks again for the fields the first translation dropped.
  (#1114)
• The engine catalogue sizes the memory of gemma-4, LFM2.5 and GPT-OSS models correctly, so they
  can be enabled without declaring their memory by hand.
  (#975)
• llama.cpp sessions with long conversations read faster in the mirror.

Fixed

• codex sessions on Terminal (CLI) ran in the home directory instead of their working copy, could
  not be resumed, and Agent Fleet's own tools could not tell which session called them. Switching a
  Managed codex session to Terminal no longer stops at "This conversation is open in another app".
  (#1131)
• On AWS deployments, an AI suggestion (title, branch name, reply) could fail when the model took
  close to a minute to answer.

Upgrade notes

Nothing to do. The default branch name for a work item changes from feature/issue-1113 to
feature/1113-<title words>, with fix/ for a bug; a branch template you set yourself is kept. On ECS, update.sh now also moves the image engine's ComfyUI tag when it is still on
an earlier release's default (copying the image into ECR first); a tag you chose yourself is left
alone.
(#965)

Agent CLI pins: Claude Code 2.1.284, Codex 0.158.0, OpenCode 1.18.33, Copilot 1.0.89, Antigravity
1.2.12, Cursor 2026.09.28-64d2043 and Muse Code 1.4.0-R4302.1. Kiro is unchanged from 0.24.0.

---

日本語

課題管理から始めたセッションのブランチ名が、リポジトリ自身の規則（git-flow の設定、
.agent-fleet/branches、Bitbucket のブランチモデル）に従うようになりました。課題管理の開いている
プルリクエストには CI の状態と衝突が出ます。af-aws-exec が AssumeRole と credential_process の
プロファイルを扱えるようになり、設定 › AWS プロファイル/SSM ではプロファイルの行からログインできます。
起動に失敗したマネージドのセッションは、CLI が出力した内容を示します。

New

• ブランチ名がリポジトリの規則に従います。 課題管理からセッションを始めると、リポジトリの
  .agent-fleet/branches、git-flow の設定（.gitflow またはクローンの gitflow.*）、Bitbucket の
  ブランチモデルから、ブランチ名と起点を埋めます。どの規則が起点を決めたかを示し、規則に合わない名前には
  注記を出します。名前を断ることはありません。名前を変えるダイアログにはリポジトリのブランチ種別が並び、
  ブランチ名のテンプレートで {ref}・{num}・{project}・{type}・{prefix} が使えるように
  なりました。Console から git-flow を初期化する機能はまだです
  （#1126）。
  (#1120,
  #1124,
  #1125)
• 課題管理の開いているプルリクエストに CI の状態とマージの衝突が出ます。 失敗・実行中・成功と、
  ベースと衝突しているときの印です。GitHub がまだ答えていないときは何も描きません。
  (#1113)
• 課題の「開始済み」の一覧にセッションの名前が出ます。 その後アーカイブされたセッションは、
  そこから元に戻せます。
  (#1108)
• af-aws-exec が、自分で定義した AssumeRole（role_arn ＋ source_profile）と
  credential_process のプロファイルを扱えます。 --account でアカウントを名指ししたときだけで、
  先に経路全体を確かめ、短命の資格情報だけを渡します。設定の SSO プロファイルからロールを引き受ける
  形はまだです（#1109）。
  (#1107)
• 設定 › AWS プロファイル/SSM（旧「AWS SSM」）のプロファイルの行にログインとログイン状態の
  バッジが付きました。コマンドに求められる前にサインインしておけます。
• 起動に失敗したマネージドのセッションは、CLI が最後に出力した数行を示します。 秘密情報は
  取り除いたうえでの表示で、「時間内に起動しませんでした」だけで終わらなくなりました。
  (#1076)
• create_session で子セッションの推論の強さ（effort）を指定できます（model も必要です）。
  (#1068)

Improved

• スマホでの画像のスタジオ: 入力中も下書きの帯が見えたままになり、入力欄がペインの下端に
  付きます。トーストが入力欄を覆わなくなり、外側をタップすると消えます。「N 枚できました」は
  見るまで残り、新しい worktree で始めたスタジオはその worktree の名前になります。
  (#1089,
  #1090)
• 課題管理から始めたセッションの AI 提案のタイトルが、課題のキー（#1146・PROJ-123）
  を残します。
• 課題管理の絞り込みが、ほかの絞り込みと同じ入力欄と × になり、Escape で消せます。
• 質問カードの翻訳で最初に抜けた欄を、もう一度訳しに行きます。
  (#1114)
• エンジンのカタログが gemma-4・LFM2.5・GPT-OSS のモデルのメモリを正しく見積もります。
  メモリを手で宣言しなくても有効にできます。
  (#975)
• 長い会話の llama.cpp のセッションが、ミラーで速く読めるようになりました。

Fixed

• ターミナル（CLI）の codex のセッションが作業コピーではなくホームで動いていました。 再開も
  できず、Agent Fleet のツールは呼び出し元のセッションを見分けられませんでした。マネージドの codex の
  セッションをターミナルに切り替えたときに「This conversation is open in another app」で止まる
  こともなくなりました。
  (#1131)
• AWS の配備で、AI の提案（タイトル・ブランチ名・返信）が失敗することがありました。
  モデルの返答に 1 分近くかかったときです。

Upgrade notes

作業は不要です。課題管理から始めるときの既定のブランチ名が変わります。 feature/issue-1113
から feature/1113-<題名の英単語> になり、バグには fix/ が付きます。自分で設定したブランチ名の
テンプレートはそのまま使われます。ECS では、画像エンジンの ComfyUI のタグが以前の版の既定のままなら、
update.sh がそれも今の既定へ動かします（先にイメージを ECR へ写します）。自分で選んだタグには
触れません。
(#965)

エージェント CLI のピン: Claude Code 2.1.284・Codex 0.158.0・OpenCode 1.18.33・Copilot 1.0.89・
Antigravity 1.2.12・Cursor 2026.09.28-64d2043・Muse Code 1.4.0-R4302.1。Kiro は 0.24.0 から
変わりません。

---

Install (native) — curl -fsSL https://raw.githubusercontent.com/k-k1/agent-fleet-dist/main/install.sh | bash then af start

Assets — agent-fleet-0.25.0.tar.gz (Compose bundle) · agent-fleet-native-0.25.0-linux-amd64.tar.gz (native) · SHA256SUMS

Container images — ghcr.io/k-k1/agent-fleet/control-plane:0.25.0 and ghcr.io/k-k1/agent-fleet/workspace:0.25.0 (pulled by docker compose; the bundle's .env.example already points at them)

Workspace rootfs — rootfs-49fbd15fc22d, fetched on first start and verified against the rootfs.json sha256 inside the native tar. Verify every download against SHA256SUMS.

rootfs 49fbd15fc22d (linux-amd64)

workspace rootfs (content hash 49fbd15fc22d). Referenced by the app release's native tar; not for standalone use.

agent-fleet 0.24.0

The AWS profiles in Settings now work with the AWS CLI, SDKs and build tools in your workspace, and
the Console asks you to log in when one needs it. Deleted worktrees go to the trash and come back
at the same path with their uncommitted changes. Sessions that need you stand out, and Alt+U
jumps to the next one. The image studio is usable on a phone, starts from a repository row and runs
several studios at once. The model pickers show API list price, context window and retirement.

From this release, bugs and requests are tracked as GitHub issues at
https://github.com/k-k1/agent-fleet/issues — please report there rather than on the distribution
repository. Items below end with the issue they came from.

New

• Your Settings › AWS SSM profiles work with the AWS CLI, SDKs and build tools. They are
  written into ~/.aws/config, and af-aws-exec --profile <name> -- <command> runs a command as
  you instead of as the workspace's own AWS identity. When the login has expired, a toast in the
  Console asks you to log in; the device code starts only when you press it, and the agent waits
  about 1.5 minutes for you.
  (#998,
  #1026,
  #1036)
• Deleting a worktree can be undone. The delete goes to the trash with the branch and any
  uncommitted changes; restoring it recreates the worktree at the same path, lays the changes back
  unstaged, and brings its sessions back from the archive. For a worktree deleted before this
  release, the archive offers Recreate working copy from its branch.
  (#1040,
  #1042)
• Sessions that need you stand out, and Alt+U jumps to the next. A question, plan approval or
  permission request puts an amber outline on the row and a ring on the pane; unread rows get a
  bold title. Alt+U walks the waiting sessions first, then the unread ones. A notification is
  now marked read when what it points to is on screen, so a report stays unread until its
  conversation is shown.
  (#1057)
• The model pickers show API list price, context window and retirement, from models.dev and
  from codex itself; a retiring model is badged with its date and successor. This is the API list
  price, not what a subscription is charged. Not shown for cursor, Muse Code, llama.cpp or claude's
  tier names.
  (#1021)
• Desktop notifications that programs raise in a Terminal (CLI) session (OSC 9, 99 or 777)
  now reach the notification center and the OS notification. claude, codex and opencode already
  report through their own hooks.
  (#1058)
• Settings › Machine › Leftover files removes files nothing reads again: stale headless
  Chromium profiles, orphaned ~/.af-work folders, superseded Node patch versions and old Kiro
  versions. The Agent also removes them when it starts, together with old Copilot and Cursor
  versions.
  (#1038,
  #981)
• Settings › Agents › Session › Archive stopped sessions after sets your own period (1 to 30
  days, or Off) instead of the deployment's.
  (#982)
• Clip skip, FLUX guidance and Anima shift can be set from the studio, a catalogue row and
  generate_image. On SDXL, ComfyUI already stops at the second-to-last layer, so only 3 and up
  change the picture.
  (#1035)

Improved

• The image studio works on a phone, starts from a repository or worktree row (Start an
  image studio), and runs several studios at once, each with its own results. A narrow pane
  keeps a draft bar above the composer, the sampler settings fold under Details, and the
  lightbox walks the results with Use as reference and Back to these settings. Every pane
  now edits a studio; a draft from an older pane moves into a new studio once.
  (#1074)
• Work items show every label in its tracker colour, priority first, and say when a query
  matched more than was fetched. GitHub queries fetch 100 rows, and Search the tracker finds
  items past the cut.
  (#993,
  #994,
  #1095)
• The admin and tenant settings reopen where you left them.
  (#1100)
• Settings says when your settings failed to sync, with a Retry, and a copy another account
  left in the same browser no longer mixes into yours.
  (#1023)
• llama.cpp sessions stop a model that keeps calling one tool with changing arguments: it is
  warned at turn 12 and stopped at turn 20.
  (#953)
• Smaller messages: an engine that does not come up in time names the reason ECS gives
  (#967); a refused assistant setting says what
  was wrong, in your language (#971);
  right-clicking a session in the fleet graph opens its menu.

Fixed

• A deletion lock could be undone by a session action running at the same moment, letting the
  next delete through.
  (#950)
• Resuming a Terminal (CLI) session went back to its launch model, effort and mode after you
  had switched them inside the terminal.
  (#987)
• In Managed sessions, Agent Fleet's own tools (handoff, image generation, the studio, browser
  handoff) refused when two sessions shared a folder, and on cursor every call failed.
  (#978,
  #951,
  #989)
• A workspace could stop for idleness under a running or queued image generation, losing the
  queue. Paused jobs still do not keep it running.
  (#949)
• AI assist with Muse Code chosen ran on claude.
  (#1020)
• Resume on a stopped SSM pane skipped the SSO login dialog, and on a narrow pane the sign-in
  link was cut short.
  (#1025)
• Resuming a codex or opencode Managed session whose folder was deleted did not refuse;
  opencode started a new, empty conversation in its place.
  (#1039)
• flux1 image models could not be completed from the catalogue.
  (#1053)
• The image studio: a message sent from it stayed "Pending", every tab read "Image
  generation", the tab had no right-click menu, switching back flashed an empty studio, the phone
  swipe did not leave it, finished trials piled up in the queue, and the agent's change outlines
  spread to every field.
  (#1089)
• Starting a session from a work item could send the first prompt twice.
• A prompt starting with - (a --flag or a - bullet) failed with "invalid flag" and
  never reached the session.
• The composer's history (↑/↓, Ctrl+R) offered text you never typed — operator, schedule and
  peer messages. It now recalls only your own input.
• claude's "Could not refresh your login because another Claude Code process is refreshing it"
  was retried every minute; it now offers signing in again.
• agy sessions showed "waiting for input" while working, and the mirror showed no model on
  agy, llama.cpp and Muse Code replies.
• The CP's MCP list_models refused llama.cpp.
  (#1075)

Upgrade notes

Nothing to do. If you keep your own ~/.aws/config, the Agent now appends a block of its own at
the end; a profile name you already use is never overwritten, and af-aws-exec --list says which
profiles were skipped and why.

Agent CLI pins: Claude Code 2.1.283, Codex 0.157.1, Antigravity 1.2.11, Cursor 2026.09.26-dd393fe,
Kiro 2.24.1 and Muse Code 1.4.0-R4161.1. Copilot and OpenCode are unchanged from 0.23.2.

---

日本語

設定の AWS のプロファイルがワークスペースの AWS CLI・SDK・ビルドツールから使えるようになり、
ログインが要るときは Console が知らせます。消した worktree はごみ箱に入り、未コミットの変更ごと
同じパスに戻せます。対応が要るセッションが目立つようになり、Alt+U で次へ移れます。画像の
スタジオはスマホで使え、リポジトリの行から始められ、いくつも並べて使えます。モデルの選択肢には
API 定価・コンテキスト長・廃止予定が出ます。

この版から、不具合と要望は GitHub の Issue（https://github.com/k-k1/agent-fleet/issues）で
管理しています。報告は配布用リポジトリではなくこちらへお願いします。以下の各項目の末尾には、
元になった Issue の番号を付けています。

New

• 設定 › AWS SSM のプロファイルが AWS CLI・SDK・ビルドツールから使えます。 ~/.aws/config
  に書き出され、af-aws-exec --profile <名前> -- <コマンド> でワークスペース自身の AWS の身元では
  なく、あなたとしてコマンドを実行します。ログインが切れていると Console のトーストがログインを
  求めます。デバイスコードは押したときにだけ発行され、エージェントは 1 分半ほど待ちます。
  (#998,
  #1026,
  #1036)
• worktree の削除を元に戻せます。 削除はブランチと未コミットの変更ごとごみ箱に入り、戻すと
  同じパスに worktree を作り直し、変更をステージしない状態で戻し、中のセッションをアーカイブから
  戻します。この版より前に消した worktree は、アーカイブの作業コピーを作り直すでブランチから
  作り直せます。
  (#1040,
  #1042)
• 対応が要るセッションが目立ち、Alt+U で次へ移れます。 質問・計画の承認・許可の要求がある
  行は琥珀色の枠で、ペインは琥珀色の輪で囲まれます。未読の行はタイトルが太字...

rootfs 7d9fd5aea9a0 (linux-amd64)

workspace rootfs (content hash 7d9fd5aea9a0). Referenced by the app release's native tar; not for standalone use.

agent-fleet 0.23.2

Each terminal session gets its own scratch folder, removed along with the session, and
Settings › Machine can empty the caches go, npm, uv and pip leave in your home. Around them,
a child you asked not to be notified about no longer leaves a red dot behind, the working-copy
delete shows which sessions run in each copy, and the Files pane no longer shows Agent Fleet's own
data folder.

New

• Each Terminal (CLI) session has its own scratch folder, $AF_WORK_DIR. It is
  ~/.af-work/<session>/, created when the session starts and deleted with the session, so an
  agent's probes, review checkouts and screenshots no longer pile up loose in your home. Managed
  sessions do not get it yet.
• Settings › Machine › Tool caches lists the caches go, npm, uv and pip keep in your home.
  They grow without limit and nothing empties them. Measure counts them only when you press
  it; Empty deletes one after a confirmation. A cache that a running build or install is using
  shows In use and cannot be emptied until it finishes.

Improved

• The working-copy delete shows the sessions in each copy — kind icon and display name, with
  a state chip on running ones — so a random folder name no longer hides what the copy was for.
• The user guide's top page puts the release history in its own paragraph right after the
  reference.

Fixed

• With "Notify when a child session is waiting for input" off, the child's row, its pane tab
  and the notification center still showed it as unread. It is now marked read as it arrives.
  Questions and approvals from a child, and other sessions waiting for input, stay unread as before.
• The Files pane could open ~/.local/share/agent-fleet, which holds terminal scrollback,
  the cleanup archive of deleted sessions' transcripts and the usage ledger. It is now blocked
  like the other Agent Fleet folders.
• ecs-ec2: pause.sh left engine GPU instances running, billed until someone terminated
  them, because it stopped the CP — the only thing that terminates them — without looking at
  them. It now waits for engines to go idle (or terminates them with --fast), sweeps any left
  after the CP stops, and also stops the database (--keep-db keeps it). --up starts the
  database before the CP, and --status warns about GPU instances with no CP and a database
  running under a paused CP.

Upgrade notes

Nothing to do. On ecs-ec2, pause.sh now stops the RDS database too; pass --keep-db to keep
it running.

Agent CLI pins: Claude Code 2.1.282, Antigravity 1.2.10 and Kiro 2.24.0. Codex, Copilot,
OpenCode, Cursor and Muse Code are unchanged from 0.23.1.

---

日本語

ターミナルのセッションごとに作業用のフォルダが付き、セッションと一緒に消えるようになりました。
設定 › マシンでは、go・npm・uv・pip がホームに溜めるキャッシュを空にできます。ほかに、通知
しないと決めた子の入力待ちが赤いドットを残さなくなり、作業コピーの削除でそこで動くセッションが
分かるようになり、ファイルペインから Agent Fleet 自身のデータのフォルダが見えなくなりました。

New

• ターミナル（CLI）のセッションごとに作業用のフォルダ $AF_WORK_DIR が付きます。
  ~/.af-work/<セッション>/ で、セッションの起動時に作られ、セッションの削除で消えます。
  エージェントの試し書き・レビュー用のチェックアウト・スクリーンショットがホームに散らばらなく
  なります。マネージドのセッションにはまだ付きません。
• 設定 › マシン › ツールのキャッシュに、go・npm・uv・pip がホームに持つキャッシュが並びます。
  上限なく育ち、ひとりでに空になることはありません。測るは押したときだけ数え、空にするは
  確認のうえ 1 つずつ消します。実行中のビルドやインストールが使っているキャッシュは使用中と
  出て、それが終わるまで空にできません。

Improved

• 作業コピーの削除の各行に、そこで動くセッションが出ます。 種別のアイコンと表示名、稼働中の
  ものには状態のチップが付きます。フォルダ名が乱数でも、何の作業のコピーか分かります。
• 利用ガイドのトップで、更新履歴がリファレンスの直後の独立した段落になりました。

Fixed

• 「子セッションの入力待ちを通知」をオフにしても、子の行・ペインのタブ・通知センターに未読が
  残っていました。 いまは届いた時点で既読になります。子からの質問と許可の要求、ほかのセッションの
  入力待ちはこれまでどおり未読に残ります。
• ファイルペインで ~/.local/share/agent-fleet を開けていました。 ターミナルのスクロール
  バック・削除したセッションの転写を含む掃除のアーカイブ・使用量の台帳が入っているフォルダです。
  いまはほかの Agent Fleet のフォルダと同じく開けません。
• ecs-ec2: pause.sh がエンジンの GPU の箱を止めずに残し、誰かが終わらせるまで課金が続いて
  いました。 箱を終わらせられる唯一の CP を、箱を見ないまま止めていたためです。いまはエンジンが
  アイドルで止まるのを待ち（--fast なら箱を終わらせ）、CP が止まった後に残った箱を掃除し、
  データベースも止めます（--keep-db で残せます）。--up は CP より先にデータベースを起こし、
  --status は CP の居ない GPU の箱と、CP の休止中に動いているデータベースを警告します。

Upgrade notes

することはありません。ecs-ec2 の pause.sh は RDS のデータベースも止めるようになりました。
動かしたままにするなら --keep-db を付けてください。

エージェント CLI のピン: Claude Code 2.1.282 / Antigravity 1.2.10 / Kiro 2.24.0。
Codex・Copilot・OpenCode・Cursor・Muse Code は 0.23.1 から変わっていません。

---

Install (native) — curl -fsSL https://raw.githubusercontent.com/k-k1/agent-fleet-dist/main/install.sh | bash then af start

Assets — agent-fleet-0.23.2.tar.gz (Compose bundle) · agent-fleet-native-0.23.2-linux-amd64.tar.gz (native) · SHA256SUMS

Container images — ghcr.io/k-k1/agent-fleet/control-plane:0.23.2 and ghcr.io/k-k1/agent-fleet/workspace:0.23.2 (pulled by docker compose; the bundle's .env.example already points at them)

Workspace rootfs — rootfs-e94268b418de, fetched on first start and verified against the rootfs.json sha256 inside the native tar. Verify every download against SHA256SUMS.

rootfs e94268b418de (linux-amd64)

workspace rootfs (content hash e94268b418de). Referenced by the app release's native tar; not for standalone use.

agent-fleet 0.23.1

An image studio, where a session's agent rewrites the generation draft as you talk and a person
presses generate, and session deletion that always goes through the trash and never takes a
worktree with it. Around them, a cleanup for the cache deleted sessions leave behind, a view of
the disk Agent Fleet uses, translation for questions and peer messages, and selectable text in PDFs.

New

• Image studio. The image generation pane's Attach an agent binds a session to the draft:
  ask it to read your material and "put her in a harbour at dusk", then "darker", and the agent
  rewrites the prompt, negative, steps, cfg, size, LoRAs and references. A person always presses
  generate. Fields can be locked, every edit is in a history you can rewind to, model, seed and
  mask stay yours alone, and Switch agent rebinds the studio without losing the draft. claude
  and copilot, cursor and kiro attach as Terminal (CLI) only, opencode not as Managed, and Muse
  Code not yet.
• Deleting a session always goes to the trash, and can be undone. Shell and SSM sessions
  deleted from a row, a bulk tidy-up or a working-copy delete now land in the trash as AI sessions
  do. Deleting a session no longer deletes its worktree; deleting a worktree moves its
  stopped sessions to the archive. The trash tab gains Delete permanently: older than 30 days
  — the trash never empties on its own.
• The cleanup dialog clears the cache of deleted sessions — pasted images and images codex
  read — once nothing refers to them any more. Settings › Machine shows the disk Agent Fleet
  uses, with each folder's path; a path opens in the file tree, and an image folder in the
  gallery.
• Questions, answered questions and messages from other sessions have a Translate button
  when they came in another language. Only the display is translated; answers are matched
  against the original.
• Text in a PDF can be selected and copied.
• Release history in the user guide. The guide carries every release's notes and this
  index, and the builtin assistants can answer "what changed in which version" from them (in a
  new conversation).

Improved

• A session can run up to 10 children at once (was 6); the default stays 3.
• Settings › Notifications › Notify when a child session is waiting for input can be turned
  off. Questions and approvals from a child still notify.
• Thinking blocks take less room: a collapsed one is a single line with a preview, and the
  close bar appears only on long ones.
• A pasted image chip opens in the lightbox in the mirror's composer and the launch dialog.
• Long file names in the left pane are shortened in the middle, keeping the extension and
  the end of the name, and the tooltip shows the full path.

Fixed

• The fleet operator and SRE assistant on Antigravity answered "no response from agy"
  whenever the model tried a command or read its reference material. They now answer, a
  refused command is declined in words, and when agy does fail its reason is shown.
• Opening the thumbnail cache folder in the gallery kept making thumbnails of thumbnails.

Upgrade notes

Nothing to do.

Agent CLI pins: Claude Code 2.1.281, Codex 0.156.1, Antigravity 1.2.9 and Kiro 2.23.1. Copilot,
OpenCode, Cursor and Muse Code are unchanged from 0.23.0.

---

日本語

画像生成のスタジオが入りました。セッションのエージェントが会話に合わせて生成の下書きを書き直し、
生成のボタンは人が押します。あわせて、セッションの削除は必ずごみ箱を通り、worktree を道連れに
しなくなりました。ほかに、削除済みセッションが残したキャッシュの掃除、Agent Fleet が使っている
ディスクの内訳、質問と peer メッセージの翻訳、PDF の文字の選択が入りました。

New

• 画像生成のスタジオ。 画像生成ペインのエージェントを付けるで、セッションを下書きに
  結びます。資料を読ませて「夕暮れの港に立たせて」、続けて「もっと暗く」と頼むと、エージェントが
  プロンプト・negative・steps・cfg・寸法・LoRA・参照画像を書き直します。生成のボタンは必ず人が
  押します。欄ごとに錠を掛けられ、編集はすべて履歴に残ってその時点へ巻き戻せます。モデル・seed・
  マスクは人だけのもので、エージェントを替えるで下書きを保ったまま別のセッションに結び直せます。
  claude と copilot・cursor・kiro はターミナル（CLI）でだけ、opencode はマネージド以外で付けられ、
  Muse Code はまだ付けられません。
• セッションの削除は必ずごみ箱を通り、元に戻せます。 行・一括整理・作業コピーの削除から消した
  shell / ssm のセッションも、AI のセッションと同じくごみ箱へ入ります。セッションを消しても
  worktree は消えなくなりました。 worktree を消すと、中の停止中のセッションはアーカイブへ移ります。
  ごみ箱のタブに 30 日より古いものを完全に削除が付きました。ごみ箱がひとりでに空になることは
  ありません。
• 掃除ダイアログで削除済みセッションのキャッシュを片付けられます。 貼り付けた画像と codex が
  読んだ画像のうち、もうどこからも参照されないものです。設定 › マシンには Agent Fleet が
  使っているディスクの内訳がフォルダのパス付きで出ます。パスを押すと左ペインのファイルで開き、
  画像のフォルダはギャラリーでも開けます。
• 質問・回答済みの質問・他のセッションからのメッセージに翻訳ボタンが付きました。別の言語で
  届いたときに出ます。訳すのは表示だけで、回答は原文と照合されます。
• PDF の文字を選択してコピーできます。
• 利用ガイドに更新履歴が入りました。 全リリースのノートとこの索引がガイドに載り、組み込みの
  アシスタントが「どの版で何が変わったか」に答えられます（新しい会話から）。

Improved

• 1 つのセッションが同時に持てる子セッションが 10 本までになりました（これまでは 6 本）。
  既定の 3 本は変わりません。
• 設定 › 通知 › 子セッションの入力待ちを通知をオフにできます。子からの質問と許可の要求は
  オフでも通知します。
• 思考ブロックが場所を取らなくなりました。 畳むと 1 行のプレビューになり、下の「閉じる」は
  長いときだけ出ます。
• 貼り付けた画像のチップを押すとライトボックスで開きます。 ミラーのコンポーザと起動ダイアログの
  両方です。
• 左ペインの長いファイル名は中央で省略されます。 拡張子と名前の末尾が残り、ツールチップに
  フルパスが出ます。

Fixed

• Antigravity のフリート・オペレーターと SRE アシスタントが「no response from agy」を返していました。
  モデルがコマンドを試したり参考資料を読んだりしたときです。いまは答えが返り、断られたコマンドは
  文章で断ります。agy が失敗したときは理由が出ます。
• サムネイルのキャッシュのフォルダをギャラリーで開くと、サムネイルのサムネイルが作られ続けていました。

Upgrade notes

することはありません。

エージェント CLI のピン: Claude Code 2.1.281 / Codex 0.156.1 / Antigravity 1.2.9 / Kiro 2.23.1。
Copilot・OpenCode・Cursor・Muse Code は 0.23.0 から変わっていません。

---

Install (native) — curl -fsSL https://raw.githubusercontent.com/k-k1/agent-fleet-dist/main/install.sh | bash then af start

Assets — agent-fleet-0.23.1.tar.gz (Compose bundle) · agent-fleet-native-0.23.1-linux-amd64.tar.gz (native) · SHA256SUMS

Container images — ghcr.io/k-k1/agent-fleet/control-plane:0.23.1 and ghcr.io/k-k1/agent-fleet/workspace:0.23.1 (pulled by docker compose; the bundle's .env.example already points at them)

Workspace rootfs — rootfs-238b70ece173, fetched on first start and verified against the rootfs.json sha256 inside the native tar. Verify every download against SHA256SUMS.

rootfs 238b70ece173 (linux-amd64)

workspace rootfs (content hash 238b70ece173). Referenced by the app release's native tar; not for standalone use.

agent-fleet 0.23.0

Two new kinds of session: llama.cpp, driven by the fleet's own harness on the deployment's
chat engine or on a llama-server of your own, and Muse Code, Meta's coding agent. Around them,
a fleet graph that draws every session as a lane on a time axis, picture editing by instruction
with masks and reference pictures, AI assistance that can be pinned per feature to an agent and a
model, and stopped sessions that move to the archive instead of being deleted.

New

• llama.cpp sessions. A session kind with no CLI behind it: the fleet's own harness talks to
  the deployment's chat engine, with tools, MCP servers, approvals and exact usage in the mirror,
  and no terminal pane. It is on from the start; Use llama.cpp in Settings › Agents is where
  you switch it off. Your own connection on the same card points it at a llama-server on your LAN instead; that connection
  takes priority, and the card and the engine pill say whether it is reachable before you launch.
• Muse Code sessions. Meta's Muse Code as a session kind. It is not in the image, so the
  Settings › Agents card installs it once into your home, then you sign in with your Meta account
  from the browser or with an API key and launch it like any other agent. Its questions, usage
  and context gauge, skills, MCP servers, fork and handoff all go through the mirror. Muse's own
  OS sandbox cannot run in a workspace, and with it off Muse never asks before running a tool: a
  Muse session reaches the container the way a shell session does.
• A fleet graph. The sessions overview has a Graph switch: one lane per session on a
  time axis, showing when each started and stopped, where it started a child, sent a message or
  reported back, and what state it is in now. Families fold, archived lanes can be shown or
  hidden, and the window pans and zooms from the last 24 hours.
• Editing pictures by instruction. The Qwen-Image-Edit families (2509 and 2511) take a
  sentence describing the change plus up to three reference pictures, and a mask — white is
  repainted — for inpainting. The whole picture is edited — nothing is cropped away — a phone
  photo's orientation is honoured, and a mask of any size marks the same region of the picture.
  Waking the engine and generating are timed separately, and the reproduction info of an edited
  picture carries its prompt, negative and size. Qwen-Image 2.1 joins the families as well,
  generating and editing from one model, with sizes up to 2K.
• Stopped sessions are archived, not deleted. When the stopped-session window runs out the
  session moves to the archive with its conversation, and it can be restored from there. Nothing
  is deleted without a person asking, so a worktree no longer disappears with its session — the
  cleanup dialog reclaims it.
• AI assistance per feature. Settings › AI assistance gives each feature — the title, branch
  name and reply suggestions, file edit suggestions, the work-plan update and answer translation —
  its own on/off, agent and model. A feature that is off hides its
  button, and the card says which agent and model each one currently uses.
• Have another session review a plan. A plan in the mirror gains Review in another
  session: it rejects the plan and starts a review session with the agent you pick, which
  checks the plan against the code without changing anything and answers in a fixed format.
• GPU idle stop is a setting. The engines panel has Stop after per role, in minutes,
  instead of a fixed window.
• Spot needs consent. A Spot offer is bought only for a role where a super admin ticked
  Accept interruption. A pinned Spot offer without it is skipped and the role chooses
  automatically; un-ticking applies to the next instance, not the running one.
• A registered model's card offers "Other versions…" or "Other sizes…" — the same
  publisher's other versions or quantisations, with size and fit — and taking one in creates a
  new row.
• Stop every session under a repository or worktree from its right-click menu, now or after
  the turn, with the reason listed for any session that cannot be stopped.
• The opencode card separates the switch from the billing choice. Use opencode (Off in a
  new workspace) governs the whole kind — off, it leaves the launch menus and a direct API create
  is refused — and the billing choice gains None (my own keys), on which the opencode.ai key is
  never injected.
• An unread notification shows as a red dot on the session, with Mark all as read in the
  notification centre.

Improved

• Launching says what the model list is doing. It shows "Loading models…" and, when the list
  is empty, why: every model excluded in settings, the billing choice leaving none, or a workspace
  that has only just started.
• The WS bar keeps the usage chips of the two agents you used most recently and folds the
  rest into "+N". Each chip can be kept on the bar or always folded.
• The engine pill names the model that is loaded, says In use while a request is running,
  and when it was last used.
• Gallery cards have a right-click menu — copy the path or name, rename, delete, open the
  session that generated the picture — and the lightbox's information panel follows the theme.
  Pictures pasted into the chat open in the same lightbox.
• The changed-files band appears for llama.cpp, Kiro, Muse Code and Antigravity sessions as
  it does for the others.
• The size warning on an edit names the size the picture will actually come out at, for every
  family, rather than the size of the input.

Fixed

• A borrowed engine whose model was still loading answered "did not come up in time" to a
  chat instead of waiting; the wait is now streamed, and a refusal from the far deployment is
  relayed as it came.
• Context usage of a session on the fleet's own chat engine was a guess — it now reads the
  window the engine actually loaded.
• The fit verdict of a chat model's quantisation could be wrong; it is now read from the file
  itself and stays with the row for its lifetime.
• Taking a model in: an undeclared llm file offers Register as well as Delete; editing the
  id on the plan card no longer fails as a stale plan; the dialog no longer plans twice; a split
  family's part landing on another family's file is caught while planning; and moving a
  misplaced file works once the stack below is updated.
• A claude session's mirror went silent during a long run of tools while the terminal kept
  narrating. The narration is claude's thinking — with Fable 5.1 and Opus 5.5 the summarised
  kind — and the mirror was dropping it. It now shows as a collapsible Thinking block, with
  Show thinking expanded on the Claude card in Settings › Agents, and usage is counted once
  per response rather than once per part.
• A long generation on the fleet's own image engine could be reported as failed while the
  engine was still working on it. It now keeps waiting while the job is queued.
• Clicking outside a popover or right-click menu also pressed whatever was underneath — a
  button, a session row, the composer. The click now only closes the menu.
• The Discord bridge could send the same "answer ready" notification twice.
• Copying a quote from the mirror lost its paragraph breaks, and swiping a wide table
  sideways on a phone opened the left pane.

Upgrade notes

• ecs-ec2 deployments that run engines: update the 60-engines stack, and move the ComfyUI
  tag yourself. The template drops the task-level memory limit that killed ComfyUI while it
  loaded larger models, and gives the ingest role the S3 tagging permission the move repair needs;
  update.sh carries both. The ComfyUI image moves to v0.37.0, which the Qwen-Image 2.1
  family requires, but a live stack keeps the tag it has: copy the image to ECR as
  deploy/aws/ecs/README.md describes, then name the parameter once —

  aws cloudformation deploy --stack-name af-ecs-engines \
    --template-file deploy/aws/ecs/cfn/60-engines.yaml \
    --parameter-overrides ImageComfyImageTag=v0.37.0 \
    --capabilities CAPABILITY_NAMED_IAM --no-fail-on-empty-changeset

• A deployment that already lists Spot offers stops buying them until a super admin ticks
  Accept interruption for that role on the engines panel; listing a Spot row is no longer
  consent by itself.

• The llama.cpp engine image can be pinned to a digest with standup.sh --llm-digest sha256:…; left unset, the copy is what it always was. LlmImageTag is unchanged.

• Muse Code is not in the image. It is downloaded from Meta's release channel into a user's
  home only when that user asks for it (about 299 MB); NOTICE says so.

• Disk: worktrees of stopped sessions are no longer removed automatically. The cleanup
  dialog is where they go now.

Agent CLI pins: Claude Code 2.1.280, Codex 0.156.0, Copilot 1.0.88, OpenCode 1.18.32,
Antigravity 1.2.8, and Muse Code 1.3.0-R3401.1 (new, installed on demand). Cursor and Kiro are
unchanged from 0.22.0.

---

日本語

セッションの種類が 2 つ増えました。llama.cpp は、配備のチャットエンジンか自分の
llama-server の上でフリート自身のハーネスが動くもの、Muse Code は Meta のコーディング
エージェントです。あわせて、全セッションを時間軸のレーンとして描く俯瞰図、マスクと参照画像を
使った指示による画像編集、機能ごとにエージェントとモデルを固定できる AI補助、そして期限が来ても
消えずにアーカイブへ移る停止セッションが入りました。

New

• llama.cpp のセッション。 背後に CLI を持たないセッション種別です。フリート自身のハーネスが
  配備のチャットエンジンと話し、ツール・MCP サーバー・承認・正確な使用量はミラーに出ます。
  端末ペインはありません。最初から有効で、切るのは設定 › エージェントの llama.cpp を使うです。
  同じカードの自分の接続先に LAN の llama-server を入れるとそちらが優先され、起動する前からカードと
  エンジンのピルが到達できるかを言います。
• Muse Code のセッション。 Meta の Muse Code をセッション種別として使えます。イメージには
  入っていないので、設定 › エージェントのカードから 1 回だけホームに導入し、ブラウザで Meta
  アカウントにサインインするか API キーを入れて、他のエージェントと同じように起動します。
  質問・使用量とコンテキストゲージ・スキル・MCP サーバー・フォーク・引き継ぎはミラーを通ります。
  Muse 自身の OS サンドボックスはワークスペースでは動かず、それを切った Muse はツールを動かす前に
  尋ねません。Muse のセッションは shell のセッシ...

rootfs e9aed3a35da5 (linux-amd64)

workspace rootfs (content hash e9aed3a35da5). Referenced by the app release's native tar; not for standalone use.