TabHelm 0.31.0

Added

• Self-analysis (Experimental flag): TabHelm looks back at your chats and explains what went wrong and why — a TabHelm bug, site knowledge or a skill the agent was missing, the agent, your setup or the site.
  • Diagnose this chat (＋ menu, /diagnose [what went wrong], or Diagnose when a task fails): the top causes, what is happening right now (a question waiting for you, a stuck agent, the host offline) and one-click fixes — open the right setting, draft a skill or a site note from what failed and what worked, switch agents.
  • Report (History → Report, /report): every chat of the last 7 / 30 / 90 days, findings grouped by who can fix them, "fix these first", tool and agent tables; hidden findings come back only if they grow, ones that stopped happening move to Probably fixed.
  • Copy bug report, Copy as a fix task, Export .md / .json and Ask the agent… — always previewed first, redacted: no messages, typed text, page content, titles, full URLs, file paths or secret values; site and skill names become site-1, skill-1. The issue page opens with an empty URL.
  • The agent can run its own self-check with tabhelm_diagnose (this chat; other chats only after you allow it).
  • Uses a redacted digest of host.log (crashes, agent sign-in problems, relay errors) — the host never sends raw log lines.
• npm run test:selfanalysis (237 unit tests, including canary values that must never appear in any output); npm run test:flags covers Diagnose, /report, the agent's self-check and the bug-report preview.

Fixed

• A reopened chat could stop saving. Reopening a chat that contained an approval or a question card threw while replaying it, and the chat silently stopped being saved; exporting such a chat to Markdown failed too.
• The agent was not told that your skills changed when other start-up notes (secrets, single-tab mode) were also due.
• Scheduled runs waited 15 minutes when the agent asked for permission. They now decline at once and save "needs approval (scheduled runs cannot ask)"; a confirmation with no panel to answer it says so instead of looking like your "no".
• Agents are offered only the tools of features that are on (Developer tools, Page scripts, Take over, Password manager, Memory), instead of tools that answer "this feature is off". Turning one on applies from the next chat.
• The panel closing, New chat or opening another chat mid-task now records how the task ended (cancelled / panel closed); errors in the panel itself are kept with the chat (redacted) for the self-analysis.
• The agent picker no longer jumps to another agent while one is connected (it did when another panel started a new chat), so a model you pick is saved for the agent that is actually answering.
• host.log is rotated at 5 MB (one old copy, host.log.1) and each line carries the host's process id.
• Tests write their own host log instead of ~/.tabhelm/host.log.

Install or update: curl -fsSL https://ivlad003.github.io/tabhelm/install.sh | bash, then press ↻ for TabHelm on chrome://extensions.
Or download tabhelm-extension.zip below and load it unpacked. Full history: CHANGELOG.md.

TabHelm 0.30.0

Added

• Skip waiting in recordings (Experimental flag, with how much of each wait to keep): still stretches — loading, the agent thinking — are cut from the video and marked "⏩ N s of waiting skipped". A 45 s task typically becomes a ~10 s video; step lists still seek to the right moment.

Fixed

• Recordings could not be seeked and did not preview in Slack. MediaRecorder writes a streaming (fragmented) file with a keyframe only every ~20 s. Videos are now encoded with WebCodecs into a regular H.264 MP4 with its index at the start (fast start) and a keyframe every 2 s; the encoder also keeps working while the side panel is hidden. (Vendored mp4-muxer, MIT.)
• Back after the agent opened another page could land on about:blank — Chrome skips pages left without a click — and the agent was then stuck ("Chrome does not allow extensions to control about:blank"). Navigation now works from any page, and Back returns to the page the agent came from.
• Tests no longer read or overwrite the user's ~/.tabhelm/state.json (the e2e test keeps the host's backup in its own profile).
• End-to-end tests no longer fetch the live news feed (a banner appearing mid-test shifted the Settings layout under a click).

Install or update: curl -fsSL https://ivlad003.github.io/tabhelm/install.sh | bash, then press ↻ for TabHelm on chrome://extensions.
Or download tabhelm-extension.zip below and load it unpacked. Full history: CHANGELOG.md.

TabHelm 0.29.1

Changed

• Guide auto-play is a setting: Settings → Page → Auto-play guides. On: every guide plays by itself with no prompt. Off: guides wait for you (the agent is told why if it asked for auto-play) and ▶ Auto plays a single guide. The per-guide "Play this guide automatically?" prompt is gone.

Fixed

• Guide steps for fields inside a <label> show the label text ("Color"), not the field's options.

Install or update: curl -fsSL https://ivlad003.github.io/tabhelm/install.sh | bash, then press ↻ for TabHelm on chrome://extensions.
Or download tabhelm-extension.zip below and load it unpacked. Full history: CHANGELOG.md.

TabHelm 0.29.0

Added

• Work in one tab (Settings → Page, or ＋ → Stay in this tab): the agent stays in the chat's tab — no new tabs or switching; "open in a new tab" and pop-ups load in this tab.

Fixed

• Pressing Send jumped to the last tab of the chat's group when Record video was on: recording started on the agent's last tab and pulled it to the front. It now records the tab you are looking at and never switches on Send; each new message also makes that tab the agent's default.
• Guide auto-play did not start when the agent asked for it (the tool schema dropped auto), and it could only click. Steps now have an action — click, type a value, select an option, check — performed by auto-play; in manual mode typing steps complete when the value changes.
• Messages typed while the agent works were treated as prompt injection (they arrived inside a tool result). Enter now interrupts the current step and sends the message as a normal user message asking the agent to continue.
• Recording across tabs now also follows the agent when it only reads another tab, so the video shows every tab it works with.
• Guide scrolling: each step's element is scrolled into view smoothly whenever it is not comfortably visible (with room for the arrow and the bubble), so recordings show where the guide goes next.
• Voice (Chrome recognition): a result arriving just after Stop no longer throws in the panel (found by the monkey test).

Install or update: curl -fsSL https://ivlad003.github.io/tabhelm/install.sh | bash, then press ↻ for TabHelm on chrome://extensions.
Or download tabhelm-extension.zip below and load it unpacked. Full history: CHANGELOG.md.

TabHelm 0.28.0

A feature release that rounds up everything since 0.19. No code changes since 0.27.3 beyond the panel's news: the update banner and five news items now announce these features in the panel and on the website.

Highlights since 0.19

• Remote host through your own relay (0.25) — agents on a VPS or another PC, paired with a code; relay ready for Dokploy.
• Guide mode (0.21–0.27) — the agent marks every step on the page; Next / Back across steps and tabs; ▶ Auto plays it with pauses; Detailed guide flag.
• Pick element / Select area and Page scripts & styles (0.27) — hand exact elements to the agent; a library of CSS/JS presets run from ＋ or automatically.
• Context meter, messages while the agent works, Copy / Text under replies (0.24–0.26).
• Skills from whole repositories, skills made from a chat (0.21–0.23).
• History survives reinstalling the extension; delete one chat or all (0.19).
• Reliability: Auto-approve covers routine confirmations, recording follows the agent across tabs and never comes out empty, voice keeps the last words, no lost messages after New chat; a monkey test (npm run test:monkey) and a full smoke / e2e pass (0.27.1–0.27.3).

Install or update: curl -fsSL https://ivlad003.github.io/tabhelm/install.sh | bash, then press ↻ for TabHelm on chrome://extensions.
Or download tabhelm-extension.zip below and load it unpacked. Full history: CHANGELOG.md.

TabHelm 0.27.3

Fixed

• Voice: the last words before Stop were lost. The recorder's final chunk arrives after Stop, but the text was delivered before it was transcribed (and a late chunk could throw in the panel). The final chunk is now transcribed and included, exactly once.
• Microphone permission: after you press Allow microphone on the one-time page, it closes, takes you back to your tab and the recording you asked for starts by itself — no second click.
• Video recording on still pages and hidden windows: the screencast only sends frames when the page repaints, so a still page or an occluded window gave a choppy or empty video. Recording now brings the tab forward, captures the first frame right away and fills silences with screenshots (about 2 frames a second); a task that ends instantly gets a clear note instead of an error.
• Tab groups in pop-up / app windows: a clear message (open the page in a regular window) instead of Chrome's "Grouping is not supported".

Install or update: curl -fsSL https://ivlad003.github.io/tabhelm/install.sh | bash, then press ↻ for TabHelm on chrome://extensions.
Or download tabhelm-extension.zip below and load it unpacked. Full history: CHANGELOG.md.

TabHelm 0.27.2

Fixes from a full smoke / end-to-end / monkey test pass (npm run test:monkey is new).

Fixed

• Relay crash: one oversized or malformed WebSocket frame from any client — even without the key — took the whole relay down. Socket errors now close only that socket.
• Relay rate limit used the first X-Forwarded-For entry, which a client can forge; it now uses the address the proxy appends.
• Path traversal in recordings: the recording id was used in a file name unchecked; ids are now validated so nothing can be written outside the recordings folder.
• Lost messages after New chat / switching agents: a message sent while the new agent was still starting could overtake the start and fail with "No agent is connected" or "Invalid params". Messages now wait for the start (and for sign-in) to finish.
• Stopping, New chat or switching agents during a reply no longer shows a red "ACP connection closed" error.
• Changing the model while the agent is starting no longer fails with "No agent session"; the choice is applied once the session is ready.
• Browser tools return clear "Invalid arguments" errors instead of internal TypeErrors (browser_fill_form, browser_highlight and any tool given wrong argument types); the host ignores non-object messages and validates critic / save_recording input.
• Clearer texts: no more "press Connect" (there is no Connect button), and tab-audio capture explains that the TabHelm icon must be clicked on that tab first.
• The end-to-end test runs against built releases (it no longer needs the stripped test hook).

Install or update: curl -fsSL https://ivlad003.github.io/tabhelm/install.sh | bash, then press ↻ for TabHelm on chrome://extensions.
Or download tabhelm-extension.zip below and load it unpacked. Full history: CHANGELOG.md.

TabHelm 0.27.1

Fixed

• Auto-approve now covers TabHelm's own routine confirmations — JavaScript on the page, using another tab, cookies/storage changes, saved scripts, guide auto-play — and logs them as ✓ Auto-approved steps. Uploading files from disk and the opt-in safety gates (sensitive actions, site rules, critic) still ask.
• Recording follows the agent across tabs: when the agent acts on another tab (also via tabId without switching), that tab comes to the front and the video continues there; before, background tabs were not captured.

Install or update: curl -fsSL https://ivlad003.github.io/tabhelm/install.sh | bash, then press ↻ for TabHelm on chrome://extensions.
Or download tabhelm-extension.zip below and load it unpacked. Full history: CHANGELOG.md.

TabHelm 0.27.0

Added

• Guide auto-play: ▶ Auto in a guide lets the agent click each step itself after a pause (Settings → Page); ⏸ Pause takes over again. The agent can offer it with auto: true after you agree.
• Pick element and Select area in the ＋ menu: pick one element by clicking or many by dragging a rectangle; they go into your message with refs, CSS selectors, text and HTML.
• Page scripts & styles (Experimental flag): a library of CSS / JS presets with URL patterns — run from the ＋ menu on any page, or automatically on matching pages; agent tools browser_style, browser_script_save, browser_script_list, browser_script_run.

Fixed

• The context meter no longer shows an empty "ctx" button before the agent reports usage.

Install or update: curl -fsSL https://ivlad003.github.io/tabhelm/install.sh | bash, then press ↻ for TabHelm on chrome://extensions.
Or download tabhelm-extension.zip below and load it unpacked. Full history: CHANGELOG.md.

TabHelm 0.26.1

Added

• Copy / Text buttons under every agent reply: Markdown source, or plain text with links written out as label (url) (and rich text with clickable links for docs and mail).

Install or update: curl -fsSL https://ivlad003.github.io/tabhelm/install.sh | bash, then press ↻ for TabHelm on chrome://extensions.
Or download tabhelm-extension.zip below and load it unpacked. Full history: CHANGELOG.md.