@withbuddi/buddi 0.1.0-pre.17 → 0.1.0-pre.19

package/README.md

@@ -1,6 +1,8 @@
 # buddi
 
-<img src="packages/web/public/mascot/core.png" alt="The Buddi Blob, buddi's mascot" width="160" align="right">
+<img src="https://raw.githubusercontent.com/withbuddi/buddi/main/packages/web/public/mascot/core.png" alt="The Buddi Blob, buddi's mascot" width="160" align="right">
+
+[withbuddi.com](https://withbuddi.com) · [npm](https://www.npmjs.com/package/@withbuddi/buddi) · [docs](docs/README.md)
 
 buddi is a personal agent platform you run on your own machine. An agent is a
 markdown file: a front matter that names the tools it may call, and a body that
@@ -55,8 +57,11 @@ wizard offers to fetch Chromium (about 150 MB), or you can run
 `buddi browser install` at any time.
 
 **Telegram.** The wizard can pair your phone: you ask @BotFather for a bot,
-paste the token, and scan the QR code it draws. `buddi telegram pair` does the
-same from a terminal.
+paste the token, and scan the QR code it draws. Skipped it, or want a second
+phone? Settings → Notifications has the same steps, and lists the phones you
+paired. `buddi telegram pair` does the same from a terminal. What arrives on
+the phone, what you can send and the commands are in
+[docs/telegram.md](docs/telegram.md).
 
 **On Linux.** There is no OS keychain, so secrets go in an encrypted file vault
 in the data directory, opened by a key stored beside it (`vault-key`, mode
@@ -76,6 +81,8 @@ tag:
 npm install -g @withbuddi/buddi@next
 ```
 
+To remove it: `buddi uninstall`, then `npm uninstall -g @withbuddi/buddi`.
+
 The full install story, including what is not built yet, is
 [docs/install.md](docs/install.md).
 
@@ -110,6 +117,8 @@ script.
 - "Take a screenshot of example.com." It opens its own browser, asks you
   first, and shows the page on the canvas beside the chat.
 - Drop a PDF or a CSV on the chat and ask about it.
+- Click the camera in the composer, pick another tab of your browser, and ask
+  about what it shows.
 - "Every Friday at 6, send me a recap of the week." A mission, on a schedule.
 
 **A mailbox.** In Settings → Email, add an account with its address and an app
@@ -160,7 +169,7 @@ runs on a timer, and the hosts it talks to. Nothing happens until you approve.
 own, such as sending mail, running a command or acting in a browser, stops the
 run and shows a card with exactly what will happen. You approve or reject it
 on the dashboard or on Telegram. Unknown tools, bad arguments and missing
-configuration never run. [ARCHITECTURE.md](ARCHITECTURE.md) has the model.
+configuration never run. [docs/architecture.md](docs/architecture.md) has the model.
 
 **Missions and watchers.** A mission is scheduled work. A watcher checks
 something on an interval and produces findings; core decides whether a finding
@@ -246,22 +255,14 @@ Most of this is on the dashboard. From a terminal:
 | Command | What it does |
 | --- | --- |
 | `buddi` | open the dashboard (the first run sets everything up) |
-| `buddi doctor` | check every moving part and say what is wrong |
-| `buddi service status` | is the background service running? `restart`, `logs`, `stop` too |
-| `buddi dashboard` | a five-minute sign-in link to the dashboard |
-| `buddi upgrade` | back up, install the new version, migrate, restart |
-| `buddi backup create --encrypt` | one archive now, sealed with your passphrase |
-| `buddi browser install` | download Chromium for the agents' browser |
-| `buddi agents` | every agent, its engine, and whether it can run |
+| `buddi status` | one screen: service, database, agents, what needs you |
+| `buddi ask "…"` | one question, one answer, then exit; `--json` for scripts |
 | `buddi chat` | talk to the default agent in the terminal |
-| `buddi ask "…"` | one question, one answer, then exit |
-| `buddi telegram pair` | a QR code that pairs a phone |
-| `buddi reminders` | what the agents have put on the clock |
-| `buddi plugins list` | what is installed |
-| `buddi mcp` | buddi as an MCP server, for Claude Code |
+| `buddi upgrade` | back up, install the new version, migrate, restart |
+| `buddi backup create` | one archive of everything, now |
 
-`buddi help` prints them all. `buddi status` is `buddi doctor` under another
-name.
+`buddi help` lists every command, and `buddi help <command>` explains one.
+The whole tree is in [docs/cli.md](docs/cli.md).
 
 ---
 
@@ -327,7 +328,13 @@ A tag `v<version>` runs the gate, builds the tarball, publishes it to npm
 (pre-release versions under `next`, others under `latest`) and creates the
 GitHub release.
 
-Read next: [ARCHITECTURE.md](ARCHITECTURE.md) for the design,
+**Changelog.** Every change an owner or plugin author could notice gets a line
+under Unreleased in `CHANGELOG.md`, in the same commit; CI refuses a change to
+`packages/*/src` without one unless the commit message says `[no changelog]`.
+Before a tag, `node scripts/release/changelog.mjs cut <version>` turns
+Unreleased into that version's section, which the release lifts into its notes.
+
+Read next: [docs/architecture.md](docs/architecture.md) for the design,
 [docs/plugins.md](docs/plugins.md) to write a plugin, and
 [docs/README.md](docs/README.md) for the index of everything else. The roadmap
 is kept outside this repository.
@@ -340,8 +347,9 @@ buddi is a 0.1 pre-release. macOS is the reference platform. Linux works and
 is in trial: the file vault, the bundled Postgres and the systemd user unit
 are built, and fixes land as the trial finds them. Windows is not supported
 yet. Native computer control (operating your own apps) is macOS-only. Signing
-in with a Claude subscription and ChatGPT accounts through Codex are
-experiments, off unless enabled
+in with a Claude subscription (on the plan's monthly Agent SDK credits) or a
+ChatGPT subscription through Codex is offered by default;
+`BUDDI_SUBSCRIPTION_SIGNINS=off` hides both
 ([docs/anthropic-oauth.md](docs/anthropic-oauth.md),
 [docs/codex-accounts.md](docs/codex-accounts.md)). Host commands are approved,
 not sandboxed ([docs/host-execution.md](docs/host-execution.md)). Nothing an

package/extension/background.js

@@ -240,6 +240,7 @@ var MAX_TARGETS = 240;
 var LOAD_TIMEOUT = 3e4;
 var SETTLE_MS = 600;
 var SETTLE_STEP_MS = 50;
+var OBSERVE_RETRY_MS = [500, 1e3, 2e3];
 var MIN_FRAME_MS = 100;
 var MAX_CAST_SIDE = 4096;
 var MAX_COORDINATE = 2e4;
@@ -427,12 +428,14 @@ var BrowserCommands = class {
   #contentFile;
   #onFrame;
   #now;
+  #wait;
   /** How many things want the debugger on this tab. A screencast is one of them, and it outlives a command. */
   #attached = /* @__PURE__ */ new Map();
   #casts = /* @__PURE__ */ new Map();
   #castsByTab = /* @__PURE__ */ new Map();
   constructor(chrome2, options = {}) {
     this.#chrome = chrome2;
+    this.#wait = options.wait ?? ((ms) => new Promise((resolve) => setTimeout(resolve, ms)));
     this.#onFrame = options.onFrame ?? (() => void 0);
     this.#now = options.now ?? (() => Date.now());
     chrome2.debugger.onEvent.addListener((source, method, params) => this.#debuggerEvent(source, method, params));
@@ -673,6 +676,37 @@ var BrowserCommands = class {
   async #inject(tabId) {
     await this.#chrome.scripting.executeScript({ target: { tabId, allFrames: true }, files: [this.#contentFile] }).catch(() => void 0);
   }
+  /*
+   * Frame 0 is the main frame by frame id, not by arrival order: Chrome
+   * returns the frames in whatever order they answered, and reading a subframe
+   * as the page would put the model on the wrong document.
+   *
+   * A main frame with no answer is usually a page that is not ready yet: the
+   * content script did not land in a tab a navigation opened, or a heavy page
+   * kept its main thread busy. So inject again and read again, a few times
+   * with growing waits, before telling the model the page is not answering.
+   */
+  async #readFrames(tabId, cancel) {
+    for (let attempt2 = 0; ; attempt2 += 1) {
+      await this.#inject(tabId);
+      const frames = await this.#chrome.scripting.executeScript({
+        target: { tabId, allFrames: true },
+        func: readObservation
+      }).catch(() => []);
+      const top = frames.find((frame) => frame.frameId === 0);
+      if (top?.result) return { frames, top: { frameId: top.frameId, result: top.result } };
+      const wait = OBSERVE_RETRY_MS[attempt2];
+      if (wait === void 0) break;
+      await this.#wait(wait);
+      cancel.check();
+    }
+    const ready = await this.#chrome.scripting.executeScript({
+      target: { tabId, frameIds: [0] },
+      func: readReadyState
+    }).then((results) => results[0]?.result, () => void 0);
+    const state = typeof ready === "string" && ready !== "complete" ? ` The page is still ${ready}.` : "";
+    throw new PreconditionError(`The page has not answered after three tries.${state} Wait a few seconds and observe again.`);
+  }
   async #observe(sessionId, cancel) {
     const session = await this.#session(sessionId);
     const live = await this.#liveTabs(session);
@@ -680,13 +714,7 @@ var BrowserCommands = class {
     const tab = await this.#ownTab(session, key);
     const tabId = tab.id;
     cancel.check();
-    await this.#inject(tabId);
-    const frames = await this.#chrome.scripting.executeScript({
-      target: { tabId, allFrames: true },
-      func: readObservation
-    }).catch(() => []);
-    const top = frames.find((frame) => frame.frameId === 0);
-    if (!top?.result) throw new PreconditionError("The page did not answer. Wait for it to finish loading and observe again.");
+    const { frames, top } = await this.#readFrames(tabId, cancel);
     const ordered = [top, ...frames.filter((frame) => frame.frameId !== 0)];
     const generation = session.generation + 1;
     session.generation = generation;
@@ -1142,6 +1170,9 @@ Scroll: ${scroll.x}, ${scroll.y}`.slice(0, MAX_TREE);
     }).catch(() => null);
   }
 };
+function readReadyState() {
+  return document.readyState;
+}
 function readObservation() {
   const api = globalThis["__buddiBrowser"];
   return api ? api.observe() : null;