demobites 1.31.0 → 1.33.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "demobites",
-  "version": "1.31.0",
+  "version": "1.33.0",
   "description": "The DemoBites agentic recorder — you prompt, it films a real browser, and DemoBites turns the take into an editable demo bite.",
   "bin": {
     "demobites": "launcher/index.mjs"

package/scripts/record.mjs

@@ -394,7 +394,10 @@ const postFrame = async () => {
     // the JPEG OUTSIDE the page with ffmpeg. No ffmpeg → no thumbnail, never a glitch.
     const { data } = await within(frameCdp.send("Page.captureScreenshot", { format: "jpeg", quality: 60 }), 4000);
     const small = await within(shrinkJpeg(Buffer.from(data, "base64"), 480), 4000);
-    if (small) await within(postEvent("frame", { ...frameStep, jpeg: small.toString("base64") }), 8000);
+    // A BLANK STILL IS NO STILL (founder 2026-10-18: a white box with the red pill while the first page loaded). A
+    // 480 px JPEG of a white page is under 3 KB; a page with anything on it is well over. The app keeps the last real
+    // still, or says the page is loading.
+    if (small && small.length > 3000) await within(postEvent("frame", { ...frameStep, jpeg: small.toString("base64") }), 8000);
   } catch (e) {
     if (!frameWarned) { frameWarned = true; console.error(`frame: dropped (${String(e?.message ?? e).split("\n")[0]})`); }
   } finally { framing = false; }
@@ -1319,11 +1322,28 @@ if (connectedBrowser) { try { await connectedBrowser.close(); } catch {} }
 if (ownedBrowser) { try { await ownedBrowser.close(); } catch {} }
 fs.writeFileSync(path.join(DIR, "manifest.json"), JSON.stringify(manifest, null, 2));
 if (failure) {
-  console.error(`TAKE FAILED at step ${manifest.steps.length + 1}: ${failure.message}`);
+  const n = manifest.steps.length + 1;
+  const N = STORYBOARD.steps.length;
+  const st = STORYBOARD.steps[n - 1] || {};
+  console.error(`TAKE FAILED at step ${n}: ${failure.message}`);
   console.error("Partial raw.webm + manifest.json saved in", DIR);
-  process.exit(1);
+  // THE REASON REACHES THE PERSON (founder 2026-10-18, "Recording did not finish" with nothing else: "please
+  // analyze"): the step, its words, and what the camera could not do, in plain words. Exit 3 tells plan.mjs the
+  // failure is already posted.
+  const m = /no visible target for (.+)$/.exec(failure.message);
+  const what = m ? `the camera could not find ${describeTarget(m[1])} on the page` : failure.message.replace(/\s+/g, " ").slice(0, 160);
+  const label = st.name || st.label || (st.selector ? describeTarget(st.selector) : st.action) || "";
+  await postEvent("failed", { message: `The recording stopped at step ${n} of ${N}${label ? ` (${label})` : ""}: ${what}.`, step: n, N, reason: failure.message.slice(0, 300) });
+  process.exit(3);
 }
 const cursorKinds = [...new Set(mouseEvents.map((e) => e.cursor))].join(", ");
+/** A selector in the person's words: has-text('X') → "X"; aria-label → its words; else the selector itself. */
+function describeTarget(sel) {
+  const t = /has-text\(['"](.+?)['"]\)/.exec(sel)?.[1] || /aria-label=['"](.+?)['"]/.exec(sel)?.[1];
+  if (!t) return `"${sel}"`;
+  const kind = /menuitem/.test(sel) ? "in the menu" : /role=tab/.test(sel) ? "among the tabs" : /button/.test(sel) ? "as a button" : "";
+  return `"${t}"${kind ? ` ${kind}` : ""}`;
+}
 console.log(
   `DONE raw.webm + manifest.json in ${DIR} — ${manifest.duration.toFixed(1)}s, ` +
   `${manifest.steps.length} steps, ${manifest.shots.length} camera shots, ` +

package/skill/scripts/plan.mjs

@@ -22,7 +22,7 @@ import { estimateTake } from "./clock.mjs";
 import path from "node:path";
 import { spawnSync } from "node:child_process";
 import { fileURLToPath } from "node:url";
-import { listening, postEvent } from "./listen-event.mjs";
+import { listening, postEvent, poll } from "./listen-event.mjs";
 
 const here = path.dirname(fileURLToPath(import.meta.url));
 const dir = process.argv[2];
@@ -218,7 +218,7 @@ const firstMessage = async () => {
     // storyboard is the base; the note says what to keep an eye on; the probe finds what moved.
     job.retake_of && job.storyboard
       ? `## This is a retake\nThe person wants the same demo filmed again, the same length and the same shots, because the app changed since. Their note: "${(job.retake_note || "no note").replace(/"/g, "'")}". ${job.storyboard.storyboard ? "Start from the previous storyboard below: keep its steps and its lines; probe it; fix only what no longer resolves or what the note names; then submit_storyboard. Never explore when the probe passes." : "Only the previous storyboard's lines survive (below): rebuild the same steps with the same lines, find the controls again, probe, then submit_storyboard."}\n\n## The previous storyboard\n${JSON.stringify(job.storyboard.storyboard ?? job.storyboard.steps ?? job.storyboard)}`
-      : codeOn ? `Write the storyboard, probe it, fix it, probe again, then submit_storyboard.` : `There is no code: explore first (explore from the starting address, looking_for = the brief's nouns), then write the storyboard, probe it, fix it, probe again, then submit_storyboard.`,
+      : codeOn ? `Write the storyboard from the page's own words, probe it, fix it, probe again, then submit_storyboard. Read a file only when the page's words do not name a control you need; never as a first move. You have three dry runs and two minutes from the first one; when a control of the brief is not where you looked, ask_person at once: the person is there and knows the app.` : `There is no code: explore first (explore from the starting address, looking_for = the brief's nouns), then write the storyboard, probe it, fix it, probe again, then submit_storyboard. You have three dry runs and two minutes from the first one; when a control of the brief is not where you looked, ask_person at once: the person is there and knows the app.`,
   ];
   return parts.join("\n\n");
 };
@@ -226,6 +226,44 @@ const firstMessage = async () => {
 let storyboard = null;
 let creates = "";
 const messages = [{ role: "user", content: [{ type: "text", text: await firstMessage() }] }];
+// THE PLANNING CAP (founder 2026-10-18, a 13-minute storyboard: "inconceivable"): the planner gets FOUR dry runs, or
+// three minutes from the first one, whichever comes first. After that no probe runs; the next move is the storyboard
+// as it stands, with the steps it could not find named in `unsure`, so the person answers the question instead of
+// waiting while it rewrites and retries (that job: 18 dry runs, 13 model turns).
+// THE PERSON LEADS (founder 2026-10-18: "if the planner gets stuck, he should put the customer into discussion…
+// I wouldn't go more than 2 minutes"): three dry runs, or two minutes from the first, and the planner ASKS — one
+// short question through the chat (ask_person); the answer comes back as the tool's result and the counters start
+// again. The person is there and knows their app; a storyboard guessed in minute six is worth less than a question
+// in minute two.
+const CAP_PROBES = 3;
+const CAP_MS = 2 * 60 * 1000;
+let probes = 0;
+let firstProbeAt = 0;
+let asked = 0;
+const capped = () => probes >= CAP_PROBES || (firstProbeAt && Date.now() - firstProbeAt > CAP_MS);
+const CAP_TEXT = () => asked < 2
+  ? `STOP (${probes} dry run${probes === 1 ? "" : "s"}, ${Math.round((Date.now() - (firstProbeAt || Date.now())) / 1000)} s since the first). No more dry runs, no more reading. Call ask_person NOW with ONE short question in plain words: what you were looking for, where you looked, and how they want to play it (where it is, or what to show instead). The person is there and knows the app. Their answer comes back as the result; then fix the storyboard from their words, probe once, and hand it in.`
+  : `STOP. Hand in the storyboard NOW with submit_storyboard: keep every step that passed; a step that failed becomes a hover or a settle on the nearest element the page does have, or goes; name in \`unsure\` what you could not find.`;
+/** The question goes to the chat; the planner waits here for the person's word (up to 30 minutes). */
+const askPerson = async (text) => {
+  const q = String(text || "").trim().slice(0, 600);
+  if (!q) return "ask_person needs a question.";
+  say("Asking you in the chat…");
+  if (!listening) return "(no chat here: decide yourself and hand in the storyboard)";
+  await postEvent("question", { text: q });
+  const started = Date.now();
+  while (Date.now() - started < 30 * 60 * 1000) {
+    const st = await poll();
+    if (st?.status === "cancelled" || st?.status === "failed") { console.log("  The person cancelled."); process.exit(0); }
+    if (st && st.status !== "waiting_answer") {
+      const a = st.answer && typeof st.answer === "object" ? st.answer : null;
+      asked += 1; probes = 0; firstProbeAt = 0;
+      return a?.note ? `The person says: "${String(a.note).slice(0, 2000)}"` : "The person said to go on as you see fit.";
+    }
+    await new Promise((r) => setTimeout(r, 3000));
+  }
+  return "No answer came in 30 minutes: hand in the storyboard as it stands, with what you could not find in unsure.";
+};
 const plan = async () => {
   say("Writing the storyboard…");
   storyboard = null;
@@ -255,12 +293,21 @@ const plan = async () => {
         else if (u.name === "harvest_words") { say(`Reading the app's own words on ${(input.urls || []).length} page${(input.urls || []).length === 1 ? "" : "s"}…`); out = harvest(input.urls); }
         else if (u.name === "explore") { say(`Walking the app from ${where(input.from)}, looking for ${String(input.looking_for || "the brief's words").slice(0, 60)}…`); out = explore(input.from, input.looking_for, input.depth); }
         else if (u.name === "screenshot") { say(`Looking at ${where(input.url)}…`); out = screenshot(input.url); }
-        else if (u.name === "probe") { say(`Dry run of the storyboard, ${(input.storyboard?.steps || []).length} steps…`); out = probe(input.storyboard); }
+        else if (u.name === "ask_person") { out = await askPerson(input.question); }
+        else if (u.name === "probe") {
+          if (capped()) out = CAP_TEXT();
+          else {
+            probes += 1; if (!firstProbeAt) firstProbeAt = Date.now();
+            say(`Dry run of the storyboard, ${(input.storyboard?.steps || []).length} steps… (${probes} of ${CAP_PROBES})`);
+            out = probe(input.storyboard);
+            if (capped()) out += `\n\n${CAP_TEXT()}`;
+          }
+        }
         else if (u.name === "submit_storyboard") {
           say("Handing in the storyboard…");
           const bad = validate(input.storyboard);
           if (bad) out = `Not accepted: ${bad}`;
-          else { storyboard = input.storyboard; creates = typeof input.creates === "string" ? input.creates : ""; out = "Accepted."; }
+          else { storyboard = input.storyboard; creates = typeof input.creates === "string" ? input.creates : ""; if (typeof input.unsure === "string" && input.unsure.trim()) storyboard.unsure = input.unsure.trim().slice(0, 240); out = "Accepted."; }
         } else out = `Unknown tool ${u.name}`;
       } catch (e) { out = `Tool failed: ${e.message}`; }
       results.push({ type: "tool_result", tool_use_id: u.id, content: typeof out === "string" ? [{ type: "text", text: out }] : out });
@@ -311,7 +358,7 @@ const step = async (name, args, what) => {
   console.log(`  ${new Date(t0).toISOString()} ${what}…`);
   const r = spawnSync("node", [script(name), ...args], { cwd: process.cwd(), stdio: "inherit", env: process.env });
   console.log(`  ${new Date().toISOString()} ${what}: ${((Date.now() - t0) / 1000).toFixed(1)}s`);
-  if (r.status !== 0) { console.error(`  ! ${name} failed (exit ${r.status})`); await postEvent("failed", { message: `${what} did not finish` }); process.exit(1); }
+  if (r.status !== 0) { console.error(`  ! ${name} failed (exit ${r.status})`); if (r.status !== 3) await postEvent("failed", { message: `${what} did not finish` }); process.exit(1); }
 };
 const sb = JSON.parse(fs.readFileSync(sbFile, "utf8"));
 if (Array.isArray(sb.prep) && sb.prep.length) await step("cleanup.mjs", [dir, "--prep"], "Preparing the data");