Repository navigation
When the lead splits your request into tasks, it now comes back to the same chat once they have all ended, with one combined answer and one report. Questions from the lead are numbered cards you answer in the chat, #general stays short, and "Needs you" becomes a secondary shortcut.
Added
- One combined answer per request. The tasks the lead opens from your message stay tied to it. When all of them have ended, the lead answers once in the same chat, without you asking again, and writes one report: important findings, optional suggestions, overlaps and contradictions, and what wasn't done or couldn't be verified, each point with the tasks it comes from. Below it, a line from the records gives the task outcomes and file changes. "Open the report" opens it; "Write the summary again" appears when the result was written from the records alone.
- No paid answer while something waits for you. A short status line names the waiting tasks instead; their cards are already in the chat. Cancelled or deleted tasks make the answer start with "Partly done." A later change brings one update, never the same answer twice. At most 3 model calls per request; a failed call isn't retried on its own with any provider, and a free summary from the records is written instead. Opening the app starts no model call.
- Numbered questions from the lead. When the lead needs something from you, it opens a card in the chat (Q-1, Q-2…; numbers are never reused). "Answer" starts your message with the number, a message starting with "Q-3: …" goes straight to that question, and "Not needed" closes it. An unnumbered answer is matched only to one clear question; otherwise you are asked for the number. An open question holds the combined answer, and an answer approves nothing.
- A look at the project. With its "Read files" permission, the lead gets short hints from the project folder's top level (files, manifest signs, the start of the README). They are hints, not facts; hidden files aren't read and symbolic links aren't followed. The lead opens only tasks that fit the project, and a worker reports a task that doesn't apply.
Changed
- A quieter #general: how the work was split, the cards that really wait for you, short status lines and the combined answer. The routine steps of a request's tasks (sent to review, approved, sent back, automatic recovery, merged) go to the task thread.
- "Needs you" is a secondary shortcut: a quiet icon when nothing waits, its name and count when something does. It also lists the lead's open questions and takes you to the card in the chat.
- A chat message can be up to 20,000 characters. A longer one isn't sent, the chat says why, and your draft stays.
- The success rate in Reports counts only runs that ended in success or failure.
- A task that was waiting for your approval when the app closed now says it never started. It is still held, and nothing starts on launch.
- Accessibility: task cards open with Tab and Enter or Space, the send button has a name, and notifications are announced to screen readers.
- MCP SDK 1.32.1 (GHSA-6prh-2h8m-c8cw): HTTP MCP connections follow redirects only within the same origin, so extra headers and OAuth token requests no longer go to another host. Electron 44.9.3.
- Database schema v27 (requests, lead questions, request follow-up). A copy of your database is saved in the backups folder before the upgrade.
Fixed
- Every request is followed up, however old: before, only the newest 200 were tracked, so an older request could get no answer.
- A question card further back than the loaded chat is now fetched and focused when you open it from "Needs you". A deleted card says so; your draft stays.
- Small edits to a finished task (title, description, order) no longer bring a paid update or a new message.
- A message no longer comes back as "remembered" in its own reply.
- Text from web pages, apps, project memory and files can no longer close the markers that keep it apart as data. This lowers the risk of prompt injection but doesn't remove it.
Known limitations
- These changes were tested with automated tests (real SQLite, Demo provider and a stubbed Claude Code). They were not tried with a real Claude Code or Codex agent, or by hand in the Windows app.
- An MCP server that redirects to another host (or port) no longer connects. Enter the final address.
Download
| Platform | File |
|---|---|
| Windows 10/11 (x64) | Agentbits-Setup-0.13.0.exe |
| macOS, Apple Silicon | Agentbits-0.13.0-arm64.dmg |
| macOS, Intel | Agentbits-0.13.0-x64.dmg |
| Windows is the main target. The macOS packages open on a real Mac, but not every feature has been tried there yet. They are not signed by Apple: on first launch go to System Settings → Privacy & Security → Open Anyway (how). |
- Node.js is not needed to run the app. Claude Code CLI and Codex CLI are not bundled: install and sign in separately (
claude auth login,codex login). - The installers are not code-signed yet. Windows SmartScreen will warn you: More info → Run anyway.
- Upgrading: close Agentbits and run the new installer over the old one; your data in
%APPDATA%\Agentbitsis kept. The database is upgraded on first launch (a copy is saved in the backups folder first), so make a backup first if you might go back to an older version (Settings → Data management). An older version cannot open the upgraded database.
Verify the download
SHA-256 checksums are in the SHA256SUMS-*.txt files. A checksum only shows the file was not corrupted; it is not a signature.
Get-FileHash Agentbits-Setup-0.13.0.exe -Algorithm SHA256Expected: 4e1ade81db2b6977444f2688f1048508030399ee18ede148bb276ca98c75b09e
Full list of changes: CHANGELOG.md
License
PolyForm Strict 1.0.0: free for personal and other noncommercial use.
Türkçe: Lider isteğini görevlere böldüğünde, hepsi bitince aynı sohbete tek bir toplu cevap ve tek bir raporla döner; her madde dayandığı görevleri gösterir. Seni bekleyen bir şey varken ücretli cevap yazılmaz, kısa bir durum satırı yazılır; istek başına en fazla 3 model çağrısı yapılır ve başarısız çağrı kendiliğinden yeniden denenmez. Liderin soruları sohbette numaralı kartlar (Q-1, Q-2…): "Cevapla", "Q-3: …" ya da "Gerek yok". #general sadeleşti, rutin adımlar görev akışında; "Bana ihtiyaç var" ikincil bir kısayol oldu ve eski soru kartlarını da getiriyor. MCP SDK 1.32.1 (yönlendirme yalnız aynı kökende izlenir; başka host'a yönlendiren MCP sunucusu bağlanmaz, son adresi gir) ve Electron 44.9.3. Bu değişiklikler otomatik testlerle denendi; gerçek Claude Code / Codex ile ve Windows uygulamasında elle denenmedi. Yükseltmek için Agentbits-Setup-0.13.0.exe'yi eskisinin üzerine kur; veritabanı v27'ye yükseltilir ve eski sürüm onu açamaz.