Skip to content

Releases

OpenSuperBot v0.6.0: desktop app, first-run wizard, and file attachments

Choose a tag to compare

@github-actions github-actions released this 13 Nov 12:07

A desktop app for macOS and Linux (.dmg, .deb, .AppImage), a first-run setup wizard, a settings page and a refined workbench UI — so trying OpenSuperBot needs neither Node, pnpm, Docker nor a terminal — plus @ mentions in group threads, file and PDF attachments, a startup release check (with no auto-updater while builds are unsigned), and a tool-scoping fix so only turns you start can change rules or routines.

First-open notes for desktop installers (walkthrough and SHA-256 verification in docs/desktop.md):

  • macOS (mac-arm64.dmg / mac-x64.dmg): ad-hoc signed, not notarized with an Apple Developer ID yet. Drag OpenSuperBot into Applications, open it once, click Done (not Move to Trash), then click Open Anyway in System Settings → Privacy & Security.
  • Ubuntu 24.04+ (linux-amd64.deb / linux-arm64.deb): prefer the .deb over the .AppImage so bubblewrap and the AppArmor userns profiles are installed automatically.

Security

  • Only turns you start can change a teammate's standing setup. Saving or forgetting a rule and creating or stopping a routine used to work in any turn, so text in a webhook payload, a page a routine read, or a teammate's handoff could have left a new routine or removed one of your rules behind. These four tools are now offered only in a turn you started, refused if the model names them anyway in any other turn, and the attempt is logged. Handing work to a teammate still works from any turn.

Added

  • A first-run wizard. A fresh install opens on a few short steps instead of an empty workspace: pick the language, pick an AI service from cards (each with a link to where its key is made), paste the key and test it, then pick a model from the ones the service offers. A failed test says what went wrong and what to try next. It then shows what your teammates' computer has to work with (the browser it found, or a download with its size, progress and a mirror in China; Docker; the sandbox for commands) and ends in Chief's thread with three example tasks to try. Every step before the computer check can switch to demo mode instead. Installs set up with environment variables never see it.
  • A settings page, behind the gear in the sidebar: the default model with a way into the AI models dialog (or, in demo mode, a button to connect a real model), the language, which computer teammates get (with why automatic picked the one in use, and a note when a change needs a restart), the browser and where it downloads from, and an About section with the version, the data folder, where the logs go and what it runs on. Settings pinned by environment variables are shown read-only with the variable to change. The setup can be run again from here, keeping the saved key when you pick the same service, and closed half way.
  • Demo mode says so where you work. A thread whose teammate runs the scripted demo model shows a banner under its header (in a group, naming who) with a button to connect a real model, and the model chip reads "Demo mode" instead of the service and stub.
  • Links to a thread. #thread=<id> at the end of the UI's address opens that conversation; a link to one that is gone opens the first. GET /api/summary gives how many teammates are working, how much waits on you and how many routines are on.
  • Ask one teammate in a group. Type @ in a group thread to pick who should answer; only they run a turn, and the chief closes with a dispatch table only when it is named too (named alone, it just answers). For other clients, POST /api/threads/group:…/messages takes an optional mentions list of teammate ids, and an @id written in the text counts the same; display names in the text are not matched. A named id that is not in the group is refused with 400, and so is a non-empty mentions in a direct thread. With no one named, the round is the same as before. Requested in #12.
  • Send a teammate a file (API). A client can upload an image, a text file or a Markdown file (POST /api/uploads, JSON with base64, 8 MB each) and attach up to 8 to a message with attachments. The file is copied into the uploads/ folder of the teammate's workspace, where read_file can open it. A teammate on a model that sees images gets the picture in front of the model on the first turn it meets it, and it stays in view for that turn even when the teammate takes screenshots; later turns carry only its name. Each teammate keeps its own record, so one you ask later in a group still gets the file and the picture. Uploads are served from GET /api/uploads/:id, removed with DELETE, and capped in total by OPENSUPERBOT_UPLOADS_MAX_MB (2048). See docs/attachments.md. Requested in #12.
  • Send files from the app. A paperclip next to the message box, pasting an image, or dragging files onto the conversation attaches them; they upload at once and send with your message, with or without words. In the thread an image is a thumbnail and any other file a strip to download, and a sent file can be removed. The app turns away a file that is too large or of the wrong kind before uploading it, says why an upload failed and lets you try again, and warns when the teammate's model cannot see images. A message that could not be sent goes back into the message box with the reason, where a failed send used to vanish.
  • Send a teammate a PDF. The gateway extracts the text when a PDF is uploaded and puts it beside the document on the teammate's computer (report.pdf and report.pdf.txt, one heading per page), because read_file cannot read a PDF itself. A scan, which has no text, and a document that cannot be parsed are accepted and reported as such, to you in the upload's reply (text: ok, truncated, none or failed) and to the teammate in the conversation. Extraction runs in a separate thread with limits on memory, time (20 seconds), pages (200) and size. CJK radical look-alikes and Latin ligatures in the extracted text are replaced by the real characters, so the text can be searched. This adds one dependency, unpdf, which has no native code.
  • The desktop app (desktop/, Electron 44.9.3). The gateway can run inside the app: it tells the app when it is ready and on which port, shuts down in order when the app asks (routines stop, computers and their Chrome close), downloads the browser with the app's own runtime, and gives commands in the sandbox a node that is the app's runtime. Run from source, nothing changes. The shell starts the gateway, signs its window in without a sign-in link, restarts a crashed gateway, and shows an error page with redacted diagnostics and the log tail when it cannot start. Its window shows only the gateway's UI: other sites open in your browser, the gateway's own pages open in an app window, signed in, other kinds of links are blocked, attachments download to your Downloads folder without overwriting, and pages may only show notifications and write to the clipboard. If port 4747 is taken (a gateway run from source, say) it takes the next free one and keeps it, so webhook addresses stay put; Settings → About shows the port. Closing the window keeps it running, with a menu bar icon (tray on Linux) and a Dock badge for what waits on you; approvals, questions and sign-in requests notify while the window is away; quitting from the app asks first while work or routines are running. It can open when you log in, off unless you tick it on the wizard's last step or in Settings: on macOS as a login item, falling back to a LaunchAgent when the system will not take an unnotarized app (and saying so when you have turned it off in System Settings), on Linux as an autostart entry. When update checks are enabled in Settings (the default), startup makes one GET to GitHub's releases/latest and shows a download link in the menu bar / tray and in Settings → Desktop app when a newer stable version is out; turning it off makes zero requests, and the app never downloads or installs updates automatically. Pages get no Node; the gateway's own page gets the app version, release notice and start-at-login, checked by the app on every call. See docs/desktop.md and desktop/README.md.
  • Mac and Linux installers, release workflow and packaged smoke test. pnpm desktop:package and .github/workflows/release.yml build two macOS .dmgs (Apple silicon and Intel, ad-hoc signed, not notarized) and Linux .deb and .AppImage installers for x64 and arm64, verify packaged builds end to end with scripts/smoke-desktop.ts (pnpm run smoke:desktop: Playwright _electron with chromiumSandbox: true, stepping through demo mode, driving a real browser screenshot and checking that quitting leaves no Chrome or gateway process behind), and publish a GitHub Release with SHA256SUMS.txt. The app carries the gateway with only its production dependencies, outside the asar, and a single SQLite binary for its platform; a dmg is about 135 MB. The AppImage uses the static AppImage runtime, so it opens without FUSE 2 (not installed by default since Ubuntu 22.04), and the gateway strips the AppImage's library, schema and data paths before launching a browser or sandboxed commands. On Ubuntu 24.04+, where AppArmor blocks unprivileged user namespaces by default, the .deb pulls bubblewrap in and installs AppArmor userns profiles for the app, its downloaded browser and bubblewrap; under an AppImage on those systems, the commands switch probes bubblewrap and stays locked with a clear explanation instead of running without a sandbox.

Changed

  • Refined workbench UI and macOS window chrome. The client uses a deeper obsidian zinc surface with hairline borders, consistent inline SVG icons in place of charac...
Read more

OpenSuperBot v0.5.0: each teammate on its own AI

Choose a tag to compare

@wolfqing wolfqing released this 08 Nov 09:14

Each teammate can run on its own AI: Claude, OpenAI, xAI, Gemini, MiniMax and other OpenAI-compatible services, each connection with its own key. Also a Chinese UI, local computers and start at sign-in on Windows, CI on Linux, macOS and Windows, a first slice of browser workflow recording, and several security fixes.

OpenGrokBot is now OpenSuperBot, a name that no longer borrows xAI's "Grok". Existing installs upgrade in place: see Coming from OpenGrokBot.

Security

  • A local computer's screen stream (/screen/:botId/ws) was served without the gateway's token and origin checks; it is now guarded like the noVNC path.
  • Sandboxed commands can no longer read the operator's home directory. Both sandboxes hide it entirely and put back only runtime managers (~/.nvm, ~/.volta, ~/.pyenv, ~/.cargo/bin, …) read-only; before, Seatbelt denied a short list of key folders and bubblewrap denied none of them. The self-check now also fails if a command can list the home directory, and the switch refuses to turn on when no self-check is wired.
  • write_file on a local computer opens the final path component without following symlinks.
  • On Windows the gateway token file is limited to your account: its access list is reset (which also clears entries written on the file itself, such as the creator's default SYSTEM and Administrators grants), inherited access is removed, and only your user (by SID) is granted. Before, 0600 did nothing there and the file kept whatever its folder allowed. Start-up warns if it cannot restrict the file, and an existing token is tightened on the next start.
  • A local computer's file tools refuse Windows device names (CON, NUL, COM1, … — reading CON would wait for input forever) and colons, which would write into an alternate data stream.
  • The gateway runs on hono 4.13.16 and @hono/node-server 2.1.6, which fix several advisories. The one that reached the gateway: a client that dropped its connection while the gateway was still setting up a WebSocket (a screen viewer while the teammate's browser starts, say) raised an unhandled socket error and stopped the process (GHSA-qv5m-9pm6-hqp3). The server also no longer turns a backslash in the path into /, which could carry a request past a reverse proxy's path rules (GHSA-jpvv-r467-5p4w). The rest do not apply here: the origin check reads the Host header rather than the request URL (GHSA-7v5p-43qr-wch5), the token check is a single * middleware, and the gateway uses neither mount() nor the form validator. ws 8.22.1, @types/node 26.6.6 and @vitejs/plugin-react 6.2.1 came in the same update.

Added

  • Each teammate can run on its own AI. AI models (the AI button next to + in the sidebar) lists your connections — a service, its API key and a default model — and adds, tests, edits or deletes them; the top of a teammate's thread shows which connection and model it uses, and a click switches it from its next turn. The connection set up first (the settings and env vars from before) stays the default for every teammate that has not picked one. Schema v7 adds a connections table and two columns on bots; each extra connection's key lives in <data>/llm-keys/<id>, never in the database. Only your account can read it (mode 0600 in a 0700 folder; on Windows, an ACL for your account alone), and the API only ever returns it masked. Deleting a connection moves its teammates back to the default.
  • Claude through its own Messages API (the official @anthropic-ai/sdk), not the OpenAI-compatible shim: tool calls, screenshots, and Claude's thinking blocks handed back untouched within a turn. Requests use automatic prompt caching, and on Anthropic's own API the server-side refusal fallback for the models that have it; a decline that remains becomes a readable error instead of an empty reply. A custom connection can also speak this API, for services that offer an Anthropic-compatible endpoint.
  • The OpenAI Responses API as a third way to talk to a model, with the reasoning items handed back within a turn (store: false, encrypted reasoning).
  • More presets, checked against each provider's docs on 2026-11-07: Claude (claude-opus-5-5, effort medium), Gemini through AI Studio's OpenAI-compatible endpoint (gemini-3.5-flash-lite), MiniMax for the global and China sites (MiniMax-M3.1-Flash, reasoning effort low instead of its default max), and MiniMax M Plan, the one subscription plan whose terms let any compatible tool connect: the form shows its terms and wants a tick before it saves one. The GLM, Kimi and Alibaba coding plans are left out on purpose; their terms forbid use from your own apps, bots or automation. OpenAI and xAI now talk over the Responses API: every GPT-6 model can call tools there (on Chat Completions most cannot), OpenAI's preset asks for reasoning effort low, and xAI marks Chat Completions deprecated. AI models lists the services and which subscriptions other apps may use.
  • A Claude Pro/Max sign-in token (sk-ant-oat…) is refused wherever a key is saved or tested, with a note to create an API key in the Claude Console instead: Anthropic does not allow other apps to use subscription sign-ins.
  • GET/POST /api/connections, PATCH/DELETE /api/connections/:id, POST /api/connections/test, GET /api/connections/:id/models and PUT /api/bots/:id/llm. The default connection keeps the settings rules: fields pinned by env vars stay locked. Changing a connection's address without a new key drops the stored key, a test against a new address never sends the old key, and moving a connection to another service resets the models teammates picked on it. A Claude connection without a key refuses to run instead of letting the SDK pick up Anthropic credentials or ANTHROPIC_CUSTOM_HEADERS from the machine; a reply cut off at the output limit is reported instead of acted on.
  • The gateway keeps settings of its own (schema v6 adds a settings table): provider, API base, model, vision, language, computers, browser mirror and update checks. An env var still wins over a saved value. The model API key can live in <data>/llm-key instead of OPENSUPERBOT_API_KEY. Like the gateway token next to it, only your account can read it (mode 0600; on Windows, an ACL for your account alone), and it is out of reach of sandboxed commands and containers. The AI models dialog writes the model settings; the rest have no UI yet.
  • Model settings apply without a restart: the API base, model, key and vision are read again before each model call, so a change takes effect on the next call while a call already in flight finishes as it started. The choice of computers and the browser mirror still apply at startup.
  • Settings API for the coming wizard and settings page. GET /api/settings returns each value and where it comes from (env, saved, default); the key is only ever returned masked (sk-…3456), with its source (env, file, none); needsRestart lists the saved changes that wait for a restart. PUT /api/settings saves a batch, key included, or nothing at all: values pinned by env vars are refused as locked. POST /api/settings/test tries a provider before you save, first GET /models and then a 16-token chat, and sorts failures into codes (bad_key, quota, rate_limited, model_not_found, unreachable, timeout, provider_error). The saved key is only sent when the API base is unchanged, so one provider's key never goes to another. GET /api/settings/models lists the saved provider's models.
  • Provider presets (GET /api/settings/providers) for DeepSeek, Kimi, Qwen (Model Studio, Beijing), GLM, xAI, OpenAI, OpenRouter, a local Ollama and any OpenAI-compatible endpoint: address, where to get a key, a default model that calls tools, whether it takes images. They were checked against each provider's docs on 2026-11-03. A preset can add request fields the provider needs: DeepSeek and Qwen get thinking turned off; OpenAI's GPT-6 models only call tools on Chat Completions with reasoning_effort: "none". The connection test sends the same fields, and treats a 400 that names the key (xAI's answer to a wrong key) as bad_key.
  • GET /api/setup tells the coming first-run wizard whether it is needed and what the computers have to work with. It is not needed once there is a key, a provider that needs none, the demo model, or any model env var, so existing setups never see it. It reports Docker, the sandbox, the browser found or missing, and how big the download is (about 190 MB, 370 MB on disk). In a mainland China time zone it suggests the npmmirror.com mirror. POST /api/setup/browser downloads the browser ahead of time, and /ws pushes setup.browser progress (which item, percent, size) as it goes.
  • The first local computer and the wizard share one browser download instead of possibly running two into the same folder. The browser mirror is read when the download starts, so changing it no longer needs a restart.
  • With computers on auto, a gateway whose Docker is running but whose bot image is not built now uses local computers, instead of handing out Docker computers that fail on first use. An image built before the rename (opengrokbot/bot:dev) counts as built. The log says why and gives the docker build command. GET /api/setup adds docker.imageReady and the reason the computers were picked (why: chosen, docker_running, docker_not_running, docker_image_missing), and a choice made on the settings page is logged as such rather than as OPENSUPERBOT_COMPUTERS. Settings ask for a restart only when it would change the computers: switching auto to local on a gateway that already runs local computers does not.
  • The UI speaks Chinese. Every label, button and hint in the web UI comes from a dict...
Read more

OpenGrokBot v0.4.0: teammates' computers without Docker

Choose a tag to compare

@wolfqing wolfqing released this 30 Oct 23:17

Teammates' computers without Docker: on macOS and Linux each teammate can get a browser and a folder of its own on your machine, with commands off by default and sandboxed when on — plus @hono/node-server 2 to clear a security advisory, WebSockets on its built-in support, and croner 11.

Security

  • The gateway runs on @hono/node-server 2.1.5, which clears GHSA-rmxm-3fg6-px4f (serveStatic decoding a path twice). The gateway never used serveStatic, so it was not exploitable here, but no 1.x release fixes it.

Added

  • Local computers — no Docker needed. When Docker is not running (or OPENGROKBOT_COMPUTERS=local), each teammate gets a browser of its own on your machine — your Chrome, Edge or Chromium, else a Chrome for Testing downloaded once into the data dir (OPENGROKBOT_BROWSER, OPENGROKBOT_BROWSER_MIRROR) — with its own profile, and a folder of its own that read_file/write_file cannot leave. The panel shows which kind of computer a teammate has and streams a local browser's screen over CDP. ask_for_login reopens the browser as a real window on the page the teammate was stuck on; I'm done signing in (or the next message) puts it back in the background. macOS and Linux; Windows keeps Docker.
  • Commands on a local computer, off by default. Each teammate has a switch in its computer panel. On, commands run under Seatbelt (macOS) or bubblewrap (Linux), confined to that teammate's folder and a private tmp, with the data dir and your key folders unreadable; the switch refuses to turn on until the sandbox passes a self-check. Off, the shell tool is not even offered to the model.
  • pnpm --filter @opengrokbot/gateway run smoke:local exercises a local computer against a real browser and the real sandbox.

Changed

  • WebSockets — the live stream and each bot's screen — use node-server's built-in support instead of @hono/node-ws, which only declares support for node-server 1.x. Upgrades still pass the gateway's token and origin checks first, and screen frames are relayed byte for byte as before.
  • Routines are scheduled with croner 11. Around daylight-saving changes this follows cron's usual behaviour: a routine set for a time that the clocks skip runs at the first valid time after the jump, and one that repeats through the day keeps running through the repeated hour.

Full changelog: v0.3.1...v0.4.0

OpenGrokBot v0.3.1: a second look at webhooks

Choose a tag to compare

@wolfqing wolfqing released this 30 Oct 13:28

Fixes from a second review of 0.3.0 — mostly webhooks: redeliveries that were dropped, signed requests that
could be replayed, and bodies buffered before they were refused — plus a Node 26 bot image and steadier tests.

Security

  • A webhook delivery refused because the last run was still going is no longer remembered as "seen", so
    GitHub's redelivery of it runs instead of being dropped as a duplicate.
  • A captured signed webhook request can no longer be replayed under a different (or no) delivery id: the
    signature itself is remembered per routine.
  • Oversized webhook bodies are refused as they arrive instead of being buffered first, so a caller with a bad key
    cannot make the gateway allocate the whole request.
  • A space-separated list of channels, handles or email addresses ("#ops #dev", "bob@acme.com alice@acme.com")
    is treated as a list and never gets Always allow, as the changelog already promised.
  • The action log keeps query strings and fragments out of page addresses, and masks tokens, passwords and
    authorization headers in logged commands.

Changed

  • The bot image is built on node:26-bookworm-slim (the current LTS line) instead of Node 22.
  • The bot image build gives apt 300 seconds per download and ten retries instead of the defaults, so
    fetching Chromium over a slow mirror no longer fails the build.
  • CI runs on actions/checkout v7, actions/setup-node v7 and pnpm/action-setup v6, all on the Node 24 action runtime.

Fixed

  • A standing rule that differed only by case or trailing punctuation could be saved twice and then removed by
    neither route; saving now dedupes the way removal compares, and the panel reports an ambiguous match as such.
  • Answering a question card twice at the same moment started two turns; the second now gets 409.
  • Web: a reply that arrived while two thread loads overlapped (switching away and straight back) could vanish
    until the next reconnect; a question card whose answer the gateway refused stayed on "Answered"; two timing-
    dependent tests.
  • Web tests that failed now and then on a busy machine. The full-page App tests acted in the gap between the first
    render and the effects that run after it, before the live connection existed or while a "close the panel on a
    thread switch" effect was still queued; they now wait for both. A render test could also run past the 5-second
    limit when the CPU was saturated, and its unfinished typing then leaked into the next test; web tests now get 15
    seconds.

Full changelog: v0.3.0...v0.3.1

OpenGrokBot v0.3.0: teammates you can leave work with

Choose a tag to compare

@wolfqing wolfqing released this 22 Oct 13:27

Teammates you can leave work with: scoped standing approvals, presence on the roster, question cards, webhook and watch routines, batch approvals, shared skills and an action log — most of it borrowed from Grok Bot, Meta Muse and OpenAI dots — plus 40-step turns, Markdown replies and the security review that came with them.

Security

  • GitHub can call a webhook routine without a key in the address. Give GitHub the routine's address without ?key= and its key as the webhook Secret; the gateway checks the X-Hub-Signature-256 signature against the raw body. GitHub's setup ping is answered without waking anyone, a repeated delivery (X-GitHub-Delivery) runs only once, and form-encoded deliveries are read too. For GitHub in the computer panel copies the three settings.
  • A webhook payload can no longer step outside its fence. The lines around it carry the payload's own SHA-256 fingerprint and every payload line starts with │, so a payload cannot write a closing line or pose as your operator.
  • Two long destinations can no longer share a rule. Destinations used to be cut at 200 characters and lower-cased whole, so two webhook URLs with the same first 200 characters, or two URLs whose paths differ only in case, matched each other's rules. Now nothing is cut, URL paths keep their case, and a destination over 200 characters always asks. A rule saved earlier for a URL with capital letters in its path asks once more.
  • Upgrades keep a way back. Before the gateway upgrades an existing database's schema it writes a full copy next to it (opengrokbot.db.v<old>-before-v<new>.<time>.bak, the three most recent are kept). If the copy cannot be written it does not upgrade.

Added

  • Watch routines. "Keep an eye on the status page" creates a routine that only looks: each run gets read-only tools (pages, files, skills, a report, a question, a hold), with no shell, typing, clicking, handoffs or rule changes. A tool it was not given is refused even if the model asks for it. When nothing changed it replies "Nothing new." and the gateway posts nothing. What a watch run holds always waits for you; a standing rule never applies to it.
  • Batch holds. When one action goes to many recipients or records, the teammate holds it once with one line per item. The chip lists them with checkboxes; untick any and approve the rest. The teammate is told exactly which items were approved and which were left out, never to add one, and to stop at the first failure. A batch never becomes a standing rule, and a bare 👍 approves all of it.
  • Shared skills. Markdown playbooks in <data>/skills/*.md are listed, by name and one line, in every teammate's instructions; read_skill fetches one when the work calls for it. Only you write them — they live outside every bot's computer. The computer panel lists them.
  • Action log. Every tool a teammate calls is recorded with what it acted on, whether it worked, how long it took and what started the turn (you, a routine, a watch, a webhook, an approval, a handoff, the group or your answer to a question), plus your approval decisions and every webhook call. Text typed into a page is never logged, only the field. The computer panel shows the latest 30, GET /api/bots/:id/actions serves more, and entries older than 90 days are dropped.
  • Standing approvals. An approval chip for an email, message or post to one destination now has Always allow…: approve it and let that teammate send the same kind of thing to the same destination without asking, for 1 hour, 24 hours or until you revoke it. The next matching hold goes straight through, marked Allowed by your rule. Purchases, bookings and anything the bot does not classify always ask. The rules are listed, with a revoke button, under Allowed without asking in the computer panel. Only your own signed-in browser can create one; a bot cannot.
    • The destination has to be one place: one email address, one channel or handle, one account, page or thread. Wildcards, lists, "everyone" and a whole service such as "twitter" never get Always allow, and the chip says why.
    • A rule only lets a hold through in a turn you started: your message, your decision, your answer or your question to the group, with no webhook payload in view. Work started by a routine, a webhook or a teammate's handoff always asks, and the chip says which rule would have covered it.
    • The teammate is never told which destinations a rule covers, so something it reads cannot steer it to one.
  • Presence in the sidebar. Each teammate's avatar shows whether it is working, waiting on you (an approval, a question or a sign-in), stuck (its last turn errored, ran out of budget or hit a turn limit) or done — it finished something while you were in another thread. A group row only shows that someone is working.
  • Question cards. A teammate can ask you to choose between two to six options (ask_operator); you tap one, or tick several and submit, and it carries on. Open questions count in the sidebar badge and the tab title. Typing a reply instead closes the card.
  • Forgetting a rule. Tell a teammate to drop a standing rule and it removes that line from MEMORY.md (forget_memory), with a red diff chip. The computer panel lists every standing rule with a remove button, which posts the same chip, so every change to a rule is still visible in the thread.
  • Webhook routines. A routine can start when another system calls it, instead of — or as well as — on a schedule. Each one gets its own secret address (copy it from the computer panel); a POST with a body wakes the teammate with that body attached as outside data. The body is posted into the thread as an outside data card when the teammate picks it up, so you can see what set it off. The address works without the gateway token, and the key never enters the thread or the model's context.
  • Per-turn limits that fit real work. A teammate can now take 40 model steps in one turn instead of 12 (OPENGROKBOT_MAX_STEPS), and a turn is paused after 20 minutes (OPENGROKBOT_TURN_MINUTES, 0 turns it off; checked between steps). At either limit the teammate gets one last text-only call to say what it finished, what is left and what it needs — instead of a bare "stopped" line. Say "continue" to pick it up.
  • The computer panel opens at once. While a sleeping computer wakes up (10–20 seconds) the panel shows what it is waiting for and for how long. If you move to another teammate meanwhile, the late answer no longer reopens the old panel.
  • Teammate replies render as Markdown: tables, lists, code, quotes and links (which open in a new tab). Images are shown as a placeholder and never loaded, and raw HTML from the model is dropped. A single newline stays a line break, so a dispatch table keeps one item per line. What you type is still shown as typed.
  • Day dividers and time stamps in every thread: a time on the first message of the day and after any pause longer than five minutes.
  • Versioned database schema. The schema version lives in SQLite's user_version and migrations run on start, one step per transaction. A v0.2 database is adopted in place; one written by a newer gateway is refused with a clear message instead of being half-upgraded.
  • Dependency audit in CI (pnpm audit --audit-level high) and Dependabot for npm, GitHub Actions and the bot image's base.

Changed

  • hold_for_approval takes a kind (email, message, post, purchase, booking, other) and an exact target. Holds from older prompts without them count as "other", so they always ask.
  • The database schema moves to v2 and then v3 on start (approval kinds and targets, a rules table, routine webhook keys; routine modes, batch items, the action log).
  • A teammate is marked done in the sidebar only when it said something while you were elsewhere, so a quiet watch run does not flag it.
  • The live stream's status events carry waiting and blocked as well as thinking and idle.
  • Node 22 is pinned: engines in the manifests and .nvmrc, which CI reads.
  • Dependencies bumped within their existing ranges (hono 4.13.13, vite 8.3.5, vitest 4.1.12, playwright-core 1.64.1, react 19.3.1 and their transitives). The audit against the public registry went from 24 advisories (6 high) to none.
  • pnpm test runs the gateway and web suites one after the other instead of side by side.

Fixed

  • Three timing-dependent tests that failed intermittently on CI: the order of parallel group replies, queued messages being answered in one turn, and the panel closing on a thread switch.

Full changelog: v0.2.1...v0.3.0