@withbuddi/buddi 0.1.0-pre.18 → 0.1.0-pre.19

package/README.md

@@ -57,8 +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
@@ -78,6 +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).
 
@@ -342,8 +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;