thursday-agent 0.2.0 → 0.4.0

package/lib/realtime/realtime.tap.ts

@@ -131,3 +131,125 @@ export function createAudioTap(): RealtimeAudio & {
     },
   };
 }
+
+/** What a face moves with on one frame (createVoiceFollower). */
+export type VoiceFrame = {
+  /** How far into its own range the voice is now, 0..1; 0 in silence. */
+  level: number;
+  /** `level` over about half a second: the phrase rather than the syllable. */
+  phrase: number;
+  /** Strength of a syllable that started on this frame; 0 when none did. */
+  onset: number;
+  /** Each band inside its own range, low to high; one the voice never reaches stays 0. */
+  bands: number[];
+};
+
+/** Seconds, except where a line says otherwise. */
+const FOLLOW = {
+  /** A band's ceiling jumps to a new peak and sinks back over `ceilRelease`. */
+  ceilAttack: 0.03,
+  ceilRelease: 2.5,
+  /** Its floor drops to the quiet between syllables and creeps back up over `floorRise`. */
+  floorFall: 0.12,
+  floorRise: 3,
+  /** A band whose range is narrower than this (0..1) carries no shape. */
+  minSpan: 0.12,
+  /** Mean band level (0..1) across which silence turns into voice. */
+  quietBelow: 0.03,
+  voicedAbove: 0.12,
+  fast: 0.05,
+  mid: 0.22,
+  phrase: 0.45,
+  /** A syllable: the fast level rises this far over the mid one, past `onsetMin`, at most once per `onsetGap`. */
+  onsetRise: 0.1,
+  onsetMin: 0.3,
+  onsetGap: 0.12,
+};
+
+/**
+ * Reads the voice's bands the way a face needs them. A voice is loud and narrow:
+ * its bands sit near the top and lean the same way for a whole sentence, so a
+ * face driven by them directly stays swollen and still. Here each band is
+ * followed inside its own recent range — the quiet between syllables is its
+ * floor, the syllables its ceiling — so the face moves with the words whatever
+ * the voice or its volume. One per face; it keeps state between frames.
+ */
+export function createVoiceFollower() {
+  const ceil = new Array<number>(SPECTRUM_BANDS).fill(0);
+  const floor = new Array<number>(SPECTRUM_BANDS).fill(0);
+  const frame: VoiceFrame = {
+    level: 0,
+    phrase: 0,
+    onset: 0,
+    bands: new Array<number>(SPECTRUM_BANDS).fill(0),
+  };
+  let fast = 0;
+  let mid = 0;
+  let sinceOnset = Number.POSITIVE_INFINITY;
+
+  const toward = (from: number, to: number, seconds: number, dt: number) =>
+    from + (to - from) * Math.min(1, dt / seconds);
+
+  return {
+    /** Call once per frame with the seconds since the last call. The frame returned is reused. */
+    read(input: ArrayLike<number>, dt: number): VoiceFrame {
+      let total = 0;
+      let within = 0;
+      let ranged = 0;
+      for (let k = 0; k < SPECTRUM_BANDS; k++) {
+        const x = input[k] ?? 0;
+        total += x;
+        ceil[k] = toward(
+          ceil[k],
+          x,
+          x > ceil[k] ? FOLLOW.ceilAttack : FOLLOW.ceilRelease,
+          dt,
+        );
+        floor[k] = toward(
+          floor[k],
+          x,
+          x < floor[k] ? FOLLOW.floorFall : FOLLOW.floorRise,
+          dt,
+        );
+        const span = ceil[k] - floor[k];
+        const at =
+          span > FOLLOW.minSpan
+            ? Math.min(1, Math.max(0, (x - floor[k]) / span))
+            : 0;
+        frame.bands[k] = at;
+        if (span > FOLLOW.minSpan) {
+          within += at;
+          ranged++;
+        }
+      }
+
+      const mean = total / SPECTRUM_BANDS;
+      const edge = Math.min(
+        1,
+        Math.max(
+          0,
+          (mean - FOLLOW.quietBelow) / (FOLLOW.voicedAbove - FOLLOW.quietBelow),
+        ),
+      );
+      const voiced = edge * edge * (3 - 2 * edge);
+      for (let k = 0; k < SPECTRUM_BANDS; k++) frame.bands[k] *= voiced;
+
+      frame.level = (ranged ? within / ranged : 0) * voiced;
+      fast = toward(fast, frame.level, FOLLOW.fast, dt);
+      mid = toward(mid, frame.level, FOLLOW.mid, dt);
+      frame.phrase = toward(frame.phrase, frame.level, FOLLOW.phrase, dt);
+
+      sinceOnset += dt;
+      frame.onset = 0;
+      if (
+        fast > mid + FOLLOW.onsetRise &&
+        fast > FOLLOW.onsetMin &&
+        sinceOnset > FOLLOW.onsetGap
+      ) {
+        sinceOnset = 0;
+        frame.onset = Math.min(1, fast);
+      }
+      return frame;
+    },
+  };
+}

package/lib/sandbox.ts

@@ -1,17 +1,10 @@
 import { spawn } from "node:child_process";
 import { mkdir, readdir, readFile, writeFile } from "node:fs/promises";
 import { join, resolve as pathResolve, relative, sep } from "node:path";
+import { EXEC_KILL_GRACE_MS, EXEC_TIMEOUT_MS } from "@/config";
 
 export type ExecResult = { stdout: string; stderr: string; exitCode: number };
 
-/**
- * How long one command may run before its process group is killed. Nobody is
- * watching the shell, so a command waiting on a prompt would otherwise hang
- * the job; anything that legitimately runs longer belongs in the background.
- * Named because the shell guide states it to the model (ai/tools/workspace.tool).
- */
-export const EXEC_TIMEOUT_MS = 180_000;
-
 export interface Sandbox {
   /** Absolute. Relative paths resolve from here */
   readonly cwd: string;
@@ -61,6 +54,25 @@ const IGNORE = new Set([
   "coverage",
 ]);
 
+/**
+ * Every file under `dir`, skipping dot entries and the folders a build or a
+ * package manager fills (IGNORE). A folder that cannot be read yields nothing.
+ */
+export async function* walkFiles(dir: string): AsyncGenerator<string> {
+  let entries;
+  try {
+    entries = await readdir(dir, { withFileTypes: true });
+  } catch {
+    return;
+  }
+  for (const e of entries) {
+    if (IGNORE.has(e.name) || e.name.startsWith(".")) continue;
+    const full = join(dir, e.name);
+    if (e.isDirectory()) yield* walkFiles(full);
+    else if (e.isFile()) yield full;
+  }
+}
+
 /**
  * Environment for every shell the agent runs: nobody watches it, so no pager
  * may stall it, and secrets do not travel to children. playwright-cli runs
@@ -108,21 +120,6 @@ export const createSandBox = ({
   const res = (p: string) =>
     p.startsWith("/") || /^[A-Za-z]:/.test(p) ? p : pathResolve(cwd, p);
 
-  async function* walk(dir: string): AsyncGenerator<string> {
-    let entries;
-    try {
-      entries = await readdir(dir, { withFileTypes: true });
-    } catch {
-      return;
-    }
-    for (const e of entries) {
-      if (IGNORE.has(e.name) || e.name.startsWith(".")) continue;
-      const full = join(dir, e.name);
-      if (e.isDirectory()) yield* walk(full);
-      else if (e.isFile()) yield full;
-    }
-  }
-
   // A very simple glob → regex. Only **/, * and ? are supported
   const globToRe = (g: string) =>
     new RegExp(
@@ -155,7 +152,7 @@ export const createSandBox = ({
       const root = res(path);
       const re = globToRe(pattern);
       const out: string[] = [];
-      for await (const f of walk(root)) {
+      for await (const f of walkFiles(root)) {
         const rel = relative(root, f).split(sep).join("/");
         if (re.test(rel) || re.test(f)) out.push(f);
         if (out.length >= limit) break;
@@ -180,28 +177,26 @@ export const createSandBox = ({
         child.stdout?.on("data", (d) => (stdout += d));
         child.stderr?.on("data", (d) => (stderr += d));
 
-        const stop = (why: string) => {
-          stderr += `\n[${why}]`;
+        const group = (name: NodeJS.Signals) => {
           try {
             // A negative pid means the group: the shell and everything under it
-            if (child.pid) process.kill(-child.pid, "SIGTERM");
+            if (child.pid) process.kill(-child.pid, name);
           } catch {
             // Already gone, or no process groups on this platform
-            child.kill("SIGTERM");
+            child.kill(name);
           }
         };
 
-        const timer = setTimeout(
-          () => stop(`Timed out after ${timeoutMs}ms`),
-          timeoutMs,
-        );
-
-        const onAbort = () => stop("Aborted");
-        if (signal?.aborted) onAbort();
-        else signal?.addEventListener("abort", onAbort, { once: true });
+        let settled = false;
+        let killing: ReturnType<typeof setTimeout> | undefined;
+        let letGo: ReturnType<typeof setTimeout> | undefined;
 
         const done = async (result: ExecResult) => {
+          if (settled) return;
+          settled = true;
           clearTimeout(timer);
+          clearTimeout(killing);
+          clearTimeout(letGo);
           signal?.removeEventListener("abort", onAbort);
           resolve({
             ...result,
@@ -210,6 +205,34 @@ export const createSandBox = ({
           });
         };
 
+        const stop = (why: string) => {
+          if (killing || settled) return;
+          stderr += `\n[${why}]`;
+          group("SIGTERM");
+          // A command that ignores SIGTERM would hold its step, and everything
+          // waiting on the job, for good (config EXEC_KILL_GRACE_MS)
+          killing = setTimeout(() => {
+            stderr += "\n[Killed: it did not exit on SIGTERM]";
+            group("SIGKILL");
+            // Output held open by a process that left the group never closes;
+            // what arrived by now is the result
+            letGo = setTimeout(() => {
+              child.stdout?.destroy();
+              child.stderr?.destroy();
+              void done({ stdout, stderr, exitCode: -1 });
+            }, 1_000);
+          }, EXEC_KILL_GRACE_MS);
+        };
+
+        const timer = setTimeout(
+          () => stop(`Timed out after ${timeoutMs}ms`),
+          timeoutMs,
+        );
+
+        const onAbort = () => stop("Aborted");
+        if (signal?.aborted) onAbort();
+        else signal?.addEventListener("abort", onAbort, { once: true });
+
         child.on("close", (code) =>
           done({ stdout, stderr, exitCode: code ?? -1 }),
         );

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "thursday-agent",
-  "version": "0.2.0",
+  "version": "0.4.0",
   "description": "A voice agent that runs on your own machine — talk to her, she hands the slow work to bots with a shell, a browser and skills.",
   "license": "MIT",
   "keywords": [

package/skills/browser/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: browser
-description: "Any browser and anything on the web — a page, a click, a form, a sign-in, a tab. Open a page and read it, fill a form, sign in, download, screenshot, print a page or an HTML file to PDF. This opens a real window on their screen, and it is how a browser is reached even when one is already running on their Mac (`attach`) — a browser is never driven by clicking it through the machine. A page that wants a sign-in is still this skill's job: the user signs in themselves, in the window this opens."
+description: "Any browser and anything on the web — a page, a click, a form, a sign-in, a tab. Open a page and read it, fill a form, sign in, download, screenshot, print a page or an HTML file to PDF. It can put a real window on their screen to show them the thing itself — an order at checkout, a map, a page — and it is how a browser is reached even when one is already running on their Mac (`attach`) — a browser is never driven by clicking it through the machine. A page that wants a sign-in is still this skill's job: a kept session, the browser they are already in, or the window this opens in front of them."
 allowed-tools: Bash(playwright-cli:*) Bash(npx:*) Bash(npm:*)
 ---
 
@@ -15,8 +15,8 @@ Everything the app has to say about the browser is here; the prompts do not
 repeat it.
 
 **Every browser goes through here.** A job that names a browser, a site, a URL,
-a sign-in or a tab is this one's. The window `playwright-cli` opens is a real
-one on their screen — they can watch it and type into it. Never drive Chrome by
+a sign-in or a tab is this one's. With `--headed` the window `playwright-cli`
+opens is a real one on their screen — they can watch it and type into it. Never drive Chrome by
 clicking it through the machine (`osascript`, `peekaboo`, System Events): that
 window cannot be snapshotted or acted on by ref. For the browser they already
 have open — a tab they left, a profile they are signed into — attach instead:
@@ -24,13 +24,27 @@ have open — a tab they left, a profile they are signed into — attach instead
 
 **Your browser is this job's own.** Its session is already in your shell
 (`PLAYWRIGHT_CLI_SESSION`): never pass `-s=`, never `close-all` or `kill-all` —
-other jobs are running theirs. The server closes yours when the job ends.
-
-**Headless unless they should see it.** Add `--headed` when the window is the
-point — a sign-in, something they asked to watch — and say so in your
-`description`. A site that will want a session before it shows you anything — a
-shop, a console, a dashboard, an account page — starts `--headed --persistent`
-rather than finding out headless and starting over.
+other jobs are running theirs.
+
+**Headless is yours; headed is theirs.** A browser opens headless: nobody sees
+it, and the app closes it when the job ends. `--headed` puts a real window on
+their screen, and that window outlives the job — the app never closes it, they
+do. So you can show them the thing itself instead of describing it: the order
+sitting at checkout for them to confirm, the map with the pin dropped, the page
+they asked to watch, a sign-in for them to finish. Open it `--headed`, leave it
+open, and say in your answer that it is on their screen. A window you opened
+headed only to get past a wall and are done with, `close` yourself. Only a job
+they cancel or delete takes its windows with it.
+
+**A payment is theirs to press.** Take a purchase, a top-up or a subscription
+as far as the last screen before money moves, open it `--headed`, and answer
+with what it buys, for how much, and that the pay button is on their screen.
+That is the job done, not stopped — never press it yourself, whatever the
+request said.
+
+A site that will want a session before it shows you anything — a shop, a
+console, a dashboard, an account page — starts `--headed --persistent` rather
+than finding out headless and starting over.
 
 **Blocked, not broken.** A 403, an "access denied" page, a wall about automated
 traffic is refusing the headless browser, not the job. `close`, then
@@ -39,29 +53,36 @@ most of it, and is worth trying before you conclude anything about the site.
 Still refused after that, the wall is real: say so in your report, with the url
 and whatever you did bring back.
 
-**A sign-in is not a wall.** The user is at this screen and signs in
-themselves, in a window you put in front of them. Never type a password, a
-one-time code, or answer a passkey prompt — not with `fill`, `type` or `eval`.
-An email or username they gave you is not a secret: fill it, press next, stop
-at the password. When you do not have it, look before you ask — `memory_recall`
-the note for that site or account carries the address they use; only when it is
-nowhere does the sign-in go back to them. When a page wants the secret half:
-
-```bash
-playwright-cli close
-playwright-cli open <the login url> --headed --persistent
-```
-
-Open the login page itself, not the front door. Then `ask_thursday` in one
-line — what to sign into, that the window is open — with options
-`Signed in` / `Not now`, and stop. The window stays open while the job waits.
-When the answer comes back, continue from a fresh `snapshot`. A captcha is the
-same move. Never `show` — it blocks waiting for annotations nobody will send.
-
-**Keep a sign-in for the next job.** After one succeeds:
-`state-save .auth/<site>.json`. Before a site that will want one, `ls .auth/`
-and `state-load .auth/<site>.json` if its file is there. These files never go
-in a report.
+**A sign-in is a fork, not a wall.** Four ways through, and which one fits is
+yours to read off the job:
+
+- **A session you already kept.** `ls <your folder>/.auth/`, then `state-load`
+  that file and `goto` the site — cookies without a reload leave the signed-out
+  page that was already drawn, which reads as an expired session when it is not.
+  Cheapest when it works, so look before anything else.
+- **The browser they are already in.** `attach --cdp=chrome` carries their own
+  profile and whatever it is signed into.
+- **They sign in themselves.** `open <the login url> --headed --persistent` —
+  the login page, not the front door — then `ask_thursday` in one line saying
+  what to sign into and that the window is open, options `Signed in` / `Not
+  now`, and stop. The window stays open while the job waits; continue from a
+  fresh `snapshot` when the answer comes. A captcha is the same move. Not
+  `show` — it blocks waiting for annotations nobody will send.
+- **You have the credentials.** Ask before you use them: `ask_thursday` in one
+  line naming the account, options `Sign in with it` / `I'll do it`, and on the
+  second take the fork above. An answer to your own sign-in question is the
+  exception — credentials given there are already the yes, so type them.
+
+The line is where the secret comes from, not what kind it is: never guess,
+invent, or go looking for one somewhere they did not point you at. An email or
+username they gave you is not a secret — fill it and press next. `memory_recall`
+the note for that site when you are missing the address they use.
+
+**Keep a sign-in for the next job.** `mkdir -p <your folder>/.auth` and
+`state-save` into it after one succeeds — your own folder, the one your
+instructions name, because a session is yours across jobs and the workspace root
+is nobody's. Say that you kept it: it signs every later job in as them, which is
+theirs to want or not. These files never go in a report.
 
 **Reading a page.** Every command writes a snapshot file and prints its path.
 `find "Create Key"` returns only the matching nodes with a few lines around
@@ -87,7 +108,7 @@ A click that seems to do nothing may have opened a tab: `tab-list`,
 `goto http://127.0.0.1:48800/<file>.html`; `pdf --filename=out.pdf` prints
 it. Kill the server when done.
 
-Snapshots land in `.playwright-cli/` and are cleared after an hour. A file you
+Snapshots land in `.playwright-cli/` and are cleared after a few days. A file you
 name — `--filename=`, a `pdf`, a download — goes under `scratch/`, or
 `artifacts/` when it is the result.
 

package/.next/server/app/api/bot/notes/route/app-paths-manifest.json

@@ -1,3 +0,0 @@
-{
-  "/api/bot/notes/route": "app/api/bot/notes/route.js"
-}

package/.next/server/app/api/bot/notes/route.js

@@ -1,14 +0,0 @@
-var R=require("../../../../chunks/[turbopack]_runtime.js")("server/app/api/bot/notes/route.js")
-R.c("server/chunks/node_modules_consola_dist_chunks_prompt_mjs_02s3hjt._.js")
-R.c("server/chunks/_106co2o._.js")
-R.c("server/chunks/_1php4eq._.js")
-R.c("server/chunks/[root-of-the-server]__005var5._.js")
-R.c("server/chunks/[root-of-the-server]__0-npkh_._.js")
-R.c("server/chunks/[root-of-the-server]__02_2sm-._.js")
-R.c("server/chunks/lib_logger_ts_1exwjut._.js")
-R.c("server/chunks/_1_5mnq4._.js")
-R.c("server/chunks/node_modules_next_dist_0r-70fd._.js")
-R.c("server/chunks/node_modules_zod_v4_1cha_bw._.js")
-R.c("server/chunks/_next-internal_server_app_api_bot_notes_route_actions_0qcsyu7.js")
-R.m(956206)
-module.exports=R.m(956206).exports

package/.next/server/app/api/bot/notes/route_client-reference-manifest.js

@@ -1,3 +0,0 @@
-globalThis.__RSC_MANIFEST = globalThis.__RSC_MANIFEST || {};
-globalThis.__RSC_MANIFEST["/api/bot/notes/route"] = {"moduleLoading":{"prefix":""},"clientModules":{},"ssrModuleMapping":{},"edgeSSRModuleMapping":{},"rscModuleMapping":{},"edgeRscModuleMapping":{},"entryCSSFiles":{},"entryJSFiles":{}};
-

package/.next/server/app/api/memory/tidy/route/app-paths-manifest.json

@@ -1,3 +0,0 @@
-{
-  "/api/memory/tidy/route": "app/api/memory/tidy/route.js"
-}

package/.next/server/app/api/memory/tidy/route.js

@@ -1,22 +0,0 @@
-var R=require("../../../../chunks/[turbopack]_runtime.js")("server/app/api/memory/tidy/route.js")
-R.c("server/chunks/[root-of-the-server]__0emr6go._.js")
-R.c("server/chunks/[root-of-the-server]__0bed6hu._.js")
-R.c("server/chunks/_0a351g7._.js")
-R.c("server/chunks/[root-of-the-server]__0-npkh_._.js")
-R.c("server/chunks/_1d5zmy8._.js")
-R.c("server/chunks/_1php4eq._.js")
-R.c("server/chunks/[root-of-the-server]__005var5._.js")
-R.c("server/chunks/node_modules_next_dist_0r-70fd._.js")
-R.c("server/chunks/[root-of-the-server]__1untqv8._.js")
-R.c("server/chunks/_1_5mnq4._.js")
-R.c("server/chunks/_0zuzmt_._.js")
-R.c("server/chunks/lib_logger_ts_1exwjut._.js")
-R.c("server/chunks/node_modules_zod_v4_0ubdflj._.js")
-R.c("server/chunks/[root-of-the-server]__02_2sm-._.js")
-R.c("server/chunks/_07rtlxv._.js")
-R.c("server/chunks/_1drv2et._.js")
-R.c("server/chunks/node_modules_@ai-sdk_1o5ydf6._.js")
-R.c("server/chunks/_1v9r2ke._.js")
-R.c("server/chunks/_next-internal_server_app_api_memory_tidy_route_actions_0_cfuom.js")
-R.m(245204)
-module.exports=R.m(245204).exports