Skip to content

Releases

v0.1.129

Choose a tag to compare

@Arttron Arttron released this 17 Nov 12:36

Aigentron v0.1.129

For people who are not programmers: every message in a chat has a link you can copy and hand to an agent, a lead agent can read a subtask's long answer in full, and the technical lines no longer clutter the chat.

New

  • Every message in a chat has a "copy link" button (next to copy and quote). Opening the link scrolls the chat to that message and highlights it until you click in the chat or reload the page. Paste the link to an agent and it reads the message from the server itself (new tool read_message; only messages of the same project).
  • When a subtask's answer is longer than 4000 characters, its lead agent now gets a link to the whole answer in the report and reads the rest itself, instead of being left with a cut text.

Changed

  • The chat view "messages only" no longer shows the technical lines (a tool server that is not connected, the list of tools in a run); they are in "all steps".
  • A message opened from a search result stays highlighted until you click in the chat (before: a short flash).

v0.1.128

Choose a tag to compare

@Arttron Arttron released this 10 Nov 20:11

Aigentron v0.1.128

For people who are not programmers: aigentron addon enable works again on a server without Docker, and an agent can show you a screenshot from the server's Playwright browser in the chat.

Fixed

  • sudo aigentron addon enable <name> failed on every server installed without Docker with «DATA_DIR=/opt/aigentron/data: not found», so no add-on (Browser, Playwright, Voice) could be enabled or re-enabled from the command line. Fixed. (On 0.1.127 you can still run the installer directly: sudo env DATA_DIR=/opt/aigentron/data bash /opt/aigentron/current/infra/addons/playwright/bare.sh install.)
  • A screenshot taken with the server's Playwright browser could not be put into the chat: the browser saves it in its own folder, out of the agent's reach, and the agent said "sent" anyway. The Playwright add-on now saves screenshots in /var/lib/aigentron-playwright/output, the agent can show them with show_images, and its instructions say to report "sent" only after the picture was really shown.

Update note: after updating, run sudo aigentron addon enable playwright once on a server without Docker — it rewrites the Playwright service (the screenshot folder, and the proxy gateway if the 0.1.127 step was not done yet) and restarts it. A Docker install is unchanged: there the agent can still only look at a screenshot, not show it.

v0.1.127

Choose a tag to compare

@Arttron Arttron released this 10 Nov 19:16

Aigentron v0.1.127

For people who are not programmers: the proxy chosen for the agents' browser is now really used on a server where the Playwright add-on is installed without Docker.

Fixed

  • On a bare-metal server the Playwright add-on's service started the browser without any proxy option, so the proxy chosen in Add-ons → Playwright → Browser proxy (or by an agent) was ignored: pages showed the server's own address. The add-on's installer now points the browser at the server's proxy gateway (port 8940, or PROXY_GATEWAY_PORT of the server; none when the gateway is switched off).

Update note: after updating, run sudo aigentron addon enable playwright once on such a server — it rewrites the service and restarts it. (Without waiting for the update you can do it by hand: docs/browser-proxy.md, "bare-metal Playwright add-on".) Then ask an agent to open a page that shows the IP address: with a proxy chosen it must show the proxy's address, not the server's.

v0.1.126

Choose a tag to compare

@Arttron Arttron released this 10 Nov 18:45

Aigentron v0.1.126

For people who are not programmers: agents on the Codex runtime get the Playwright browser again, agents can pick the browser's proxy, add-ons can be stopped and restarted from the dashboard, and a subtask sees the files attached to its main task.

Fixed

  • An agent running on Codex had no Playwright tools ("only browser_ext is available"): Codex does not speak the older SSE protocol that the Playwright server was registered with, so the server was skipped. Codex now gets the server's streamable HTTP address instead. Other providers were never affected.
  • A subtask could not find a file attached to its main task (each task has its own attachments folder). The subtask now gets the lead task's files, read-only, listed in its first message.
  • A tool (MCP) server that did not connect at the start of a run, and the tools the model really received from each server, are now written into the conversation («⚠ MCP server not connected…», «ℹ MCP tools in this run — playwright: 23»), so "the agent says a tool is missing" can be explained afterwards.
  • A model that takes over a conversation without its memory is told that a later message overrides the original request (and, in a scheduled task, that the person's own messages adjust all later runs).

New

  • Agents can choose the proxy of the server's browser: two tools, browser_proxy_list and browser_proxy_use, work with the proxies saved in Add-ons → Playwright → Browser proxy (names only, never a login). The choice is shared by the whole browser; the page shows who made it, and a checkbox forbids it. Every switch drops the open connections so the browser really goes through the new proxy.
  • Add-ons → Playwright and Voice: Restart and Stop / Start buttons (operators; the server must run with root rights and systemd — otherwise the page says why).

Update note: a bare-metal Playwright add-on is not routed through the proxy gateway until its systemd unit gets --proxy-server http://127.0.0.1:8940 --proxy-bypass localhost,127.0.0.1 (docs/browser-proxy.md); the proxy choice in the dashboard has no effect on it before that.

v0.1.125

Choose a tag to compare

@Arttron Arttron released this 10 Nov 12:21

Aigentron v0.1.125

For people who are not programmers: you can search all your conversations, the browser's proxy is chosen in the dashboard, and scheduled tasks remember what you asked for in their chat.

New

  • Search in chats (main menu → Search, or Ctrl/Cmd+K from anywhere): finds words in the conversations of all tasks, in any language («Литве» finds «Литва»). Several words must all be present, "quotes" make an exact phrase, -word leaves a word out. Filters: whose messages, project, agent, status, date; best match or newest. A result opens the task at that message, which flashes once. Inside a task, "🔎 find" searches that conversation. The task list's search box shows "Also in conversations". The admin assistant can search too ("find the chat where we talked about X"). A ".* Pattern" mode takes a regular expression (dates, links, any form of a word); it reads the latest 20 000 messages and is stopped if it is too slow. Old messages are indexed in the background after the update (a bar shows progress, a few minutes, once).
  • Every search box has a clear (×) button.
  • Proxy connections for the agents' browser (Add-ons → Playwright → Browser proxy): paste the line your proxy seller gave (res.proxy-seller.com;10000;user;pass, user:pass@host:port, …), name it, Test, and choose it — or Direct. It applies at once, no restart; the login stays on the server and is never shown again or given to agents. The browser goes through a small gateway in the server; the project's own preview pages stay direct. HTTP / HTTPS proxies only. A proxy can also be set in .env as PLAYWRIGHT_PROXY / PLAYWRIGHT_BROWSER_PROXY in the same formats.
  • The admin assistant can propose a schedule that runs a script (the whole script is on the approval card in the dashboard).

Fixed

  • In a scheduled task, what you asked in its chat for the next runs («add this parameter») is repeated with every run, so a switch of the model or a long history no longer brings the old parameters back. (To make a change permanent, edit the schedule's text.)
  • A model that takes over a conversation is told that a later message overrides the first request.
  • Select-all + Backspace in the message box, the message box on a phone, the floating assistant stuck orange — fixed in earlier releases; included for completeness of the update.

Update note: the database gets a new column and a search index (added automatically on start). On a Docker install, recreate the browser so that it goes through the gateway: docker compose --profile mcp up -d --force-recreate playwright-mcp. Update the extension too only if the icon asks you to.

v0.1.124

Choose a tag to compare

@Arttron Arttron released this 09 Nov 12:35

Aigentron v0.1.124

For people who are not programmers: this is 0.1.123 again, but installable. Version 0.1.123 could not be built (an update to it stops by itself and the previous version keeps running) — use this one.

Fixed

  • The update to 0.1.123 failed while building a package; fixed. Everything announced for 0.1.123 is here:
    • The admin assistant can propose a schedule that runs a shell script on the server (no model, free). Nothing runs until you approve: the card in the dashboard shows the whole script.
    • Message box on a phone: full width, the caret no longer shows a line below the text, the floating assistant no longer covers Send.
    • The floating assistant no longer stays orange / "thinking" after everything is finished.

v0.1.123

Choose a tag to compare

@Arttron Arttron released this 09 Nov 12:29

Aigentron v0.1.123

For people who are not programmers: the admin assistant can set up scripts that run on a schedule, and the message box and the floating assistant behave better on a phone.

New

  • The admin assistant can propose a schedule that runs a shell script on the server (no model, free). Nothing runs until you approve: the card in the dashboard shows the whole script (the chat only shows its start and says that the dashboard has the rest). Other agents and packs still cannot create such schedules.

Fixed

  • Message box on a phone: it takes the whole width (the buttons are a row under it) instead of a third of it; the caret no longer shows a line below the text (the formatted copy of the text could wrap a line differently from the box at the edge); the floating assistant no longer covers the Send button.
  • The floating assistant stayed orange ("waiting for approval") or "thinking" after everything was finished, because it counted events and missed one. It now reads the real numbers from the server (on start, on reconnect, when you come back to the tab, every minute).

v0.1.122

Choose a tag to compare

@Arttron Arttron released this 06 Nov 10:32

Aigentron v0.1.122

For people who are not programmers: two fixes to everyday chat use.

Fixed

  • In the message box, selecting text and pressing Backspace or Delete (or typing over the selection) did nothing — "select all, Backspace" could not clear the box. It works again in every message box (task follow-up, new task, admin chat).
  • In a task's conversation, an agent that only answered a question was recorded as "blocked" ("I need your answers") after every reply, so the task looked failed. A reply that is just conversation is now recorded as done; "blocked" stays for work that really cannot go on without your decision. (This relies on the agent following the instruction, so a model may still occasionally say "blocked".)

v0.1.121

Choose a tag to compare

@Arttron Arttron released this 05 Nov 11:56

Aigentron v0.1.121

For people who are not programmers: the browser extension's chat now works like the main one, the extension stops refusing to work when you are not really using the page, and an old open tab tells you to reload.

New

  • The extension's side-panel chat shows what you send at once ("sending…"), lists the server's queue (with edit and remove), sends on Enter, and formats and quotes like the dashboard.
  • The dashboard notices when the server was updated while the tab was open and shows "Aigentron was updated — reload this page" with a Reload button. (A tab left open across an update keeps running the old code, so a fixed bug can look not fixed.)

Fixed

  • The extension refused to act "because you are using this tab" when you had only scrolled. Now scrolling does not count, only a click or a key press for 1.5 s, and the extension waits and retries by itself before the agent hears of it.
  • Pressing Approve on the in-page card counted as "using the tab", so the agent's next click was refused and the task became blocked. Clicks on the extension's own cards no longer count, and an action you have just approved is never refused for that reason.
  • Editing a queued message in the extension no longer sends it twice when its removal failed; a slow send no longer writes into another conversation; a message is shown once; long unfinished links no longer slow the panel.
  • In the dashboard, a message taken out of the queue no longer comes back as "sending…".

Update note: update the extension too (its icon → "Update the extension").

v0.1.120

Choose a tag to compare

@Arttron Arttron released this 05 Nov 10:30

Aigentron v0.1.120

For people who are not programmers: finished work no longer piles up on the disk, task titles are generated again, and active tasks come first in the list.

New

  • Clearing finished work. Every task works in its own copy of the repository, and these copies are large. Now a copy is cleared when its work has been merged into the main branch (including rebase merges and merged pull requests), or three days after the task is done if its commits are saved on the remote. It is NEVER cleared while it holds uncommitted files, ignored files that are not build output (like a .env), commits that exist nowhere else (including on a detached HEAD), or while the task is running. A merged pull request only counts for the commits it contained. A message to a cleared task brings its copy back from its branch. Settings → General → "Finished work — task checkouts" lets you choose the mode (merged or done / only merged / never) and the days, shows what would be cleared and why, clears on demand, and keeps a log. It is checked every 6 hours.
  • The task list shows tasks that are being worked on first (running, queued, waiting for approval, and leads whose subtasks are working), newest first within each group.

Fixed

  • Task titles were not generated when the default model was unavailable (a local model not running, a revoked key) or when it was a "reasoning" model that answered with nothing. Now every available model is tried in turn, a broken one is skipped for a few minutes, and the log tells you when no model could name a task (Settings → Task titles → Try it shows the details).
  • A task whose working folder had gone missing was started in that missing folder; it is now restored from its branch first.

Update note: the database gets two new columns (added automatically on start). New installations and updates have the cleanup ON with the safe rules above; set "Never" in Settings to keep every copy.