stuntdouble 0.1.1 → 0.1.3

package/dist/{mcp-O2LGDSU7.js → mcp-MOXTYXZU.js}

@@ -1,24 +1,25 @@
-import {
-  startUi
-} from "./chunk-VWTYG7G4.js";
 import {
   ActorSchema,
   ChecklistSchema,
   ConfigSchema,
   appendJsonl,
   describeRunScore,
-  isAlive,
+  effectiveStatus,
+  isLive,
   newChecklistScoring,
   projectDir,
   readJson,
   readRuns,
   resolveHaltedChecks,
   resolveSkippedChecks,
+  runningMarker,
   scoreChecklistRun,
   sdDir,
+  startUi,
+  tallyRun,
   writeFileAtomic,
   writeJsonAtomic
-} from "./chunk-JAB46QV7.js";
+} from "./chunk-7IC2G3Z3.js";
 
 // src/mcp.ts
 import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
@@ -1420,17 +1421,19 @@ ${where}` : where,
 };
 
 // src/mcp.ts
+import { dirname } from "path";
 import { z } from "zod";
 
 // src/run.ts
-import { mkdir, mkdtemp, readFile, rm, stat } from "fs/promises";
+import { createHash } from "crypto";
+import { mkdir, mkdtemp, readdir, readFile, rm, stat } from "fs/promises";
 import { homedir, tmpdir } from "os";
 import { join } from "path";
 
 // package.json
 var package_default = {
   name: "stuntdouble",
-  version: "0.1.1",
+  version: "0.1.3",
   description: "Stunt Double Local: run Stunt Double checklists in a local Chrome from your own coding agent session, with the runs kept as files and a local viewer",
   keywords: [
     "stunt-double",
@@ -1503,7 +1506,7 @@ function checkLine(check) {
 function buildBrief(input) {
   const { actor, checklist, url, checks, resumed } = input;
   const persona = [
-    `You are ${actor.name}${actor.description ? `: ${actor.description}` : ""}.`,
+    `You are ${actor.name}${actor.description ? `: ${actor.description.replace(/[.\s]+$/, "")}` : ""}.`,
     actor.system_prompt ?? ""
   ].filter(Boolean).join("\n");
   const knowledge = actor.knowledge.length ? [
@@ -1522,6 +1525,7 @@ ${checks.map(checkLine).join("\n")}`,
       "How to work:",
       "- Stay in persona: act and judge as this person would, not as a tester.",
       "- Read the page with read_page and find. Ask for a screenshot only when how something looks is the question.",
+      '- get_page_text and read_page "all" return up to 20,000 characters: keep them for long text such as a policy.',
       "- Judge each check with evaluate_check as soon as you have observed it, while the evidence is on screen.",
       "- A passed verdict needs a locator: text visible on the page, or a CSS selector, that shows the evidence. A failure may describe something missing.",
       "- Cite only what you saw in this run. Never pass a check on what a page probably does.",
@@ -1698,6 +1702,12 @@ var PlaywrightDriver = class _PlaywrightDriver {
         channel: "chrome"
       });
     } catch (chromeError) {
+      if (/ProcessSingleton|SingletonLock|already in use/i.test(chromeError.message)) {
+        throw new Error(
+          "Another run is already using this actor's kept login (persist_login). Finish that run first, or try again when it ends.",
+          { cause: chromeError }
+        );
+      }
       try {
         context = await chromium.launchPersistentContext(options.profileDir, settings);
       } catch {
@@ -1746,6 +1756,7 @@ var PlaywrightDriver = class _PlaywrightDriver {
 var MAX_LOCATOR_TEXT = 120;
 var MAX_LOCATOR_SELECTOR = 400;
 var SETTLE_MS2 = 350;
+var FIND_TIMEOUT_MS = 5e3;
 function cleanLocator(locator) {
   const clean = (value) => value?.replace(/\s+/g, " ").trim() || null;
   const selector = clean(locator?.selector);
@@ -1754,9 +1765,13 @@ function cleanLocator(locator) {
     text: clean(locator?.text)?.slice(0, MAX_LOCATOR_TEXT) ?? null
   };
 }
-async function focusEvidence(page, query) {
+async function focusEvidence(page, query, timeoutMs = FIND_TIMEOUT_MS) {
   if (!query.selector && !query.text) return null;
-  const via = await page.evaluate(findAndScroll, query).catch(() => null);
+  let timer;
+  const via = await Promise.race([
+    page.evaluate(findAndScroll, query),
+    new Promise((ok) => timer = setTimeout(() => ok(null), timeoutMs))
+  ]).catch(() => null).finally(() => clearTimeout(timer));
   if (via) await page.waitForTimeout(SETTLE_MS2).catch(() => {
   });
   return via;
@@ -1821,6 +1836,7 @@ var PACKAGE = `${package_default.name}@${package_default.version}`;
 var STOPPED = "This run was stopped from the viewer. Do not use the browser again: call finish_run now with a summary of what you found.";
 var NOT_JUDGED = "Not judged before the run finished.";
 var PRIVATE_REFUSED = "Local and private addresses other than the app under test are not reachable.";
+var SETUP = "Run /stuntdouble setup to create one.";
 var LOG_TEXT_CHARS = 500;
 var PRESETS = {
   desktop: { width: 1280, height: 720, touch: false },
@@ -1837,11 +1853,21 @@ function stamp(date = /* @__PURE__ */ new Date()) {
   return `${date.getFullYear()}${p(date.getMonth() + 1)}${p(date.getDate())}-${p(date.getHours())}${p(date.getMinutes())}${p(date.getSeconds())}`;
 }
 var now = () => (/* @__PURE__ */ new Date()).toISOString();
+function ago(iso) {
+  if (!iso) return "not started";
+  const minutes = Math.round((Date.now() - Date.parse(iso)) / 6e4);
+  if (minutes < 1) return "just now";
+  if (minutes < 60) return `${minutes} min ago`;
+  const hours = Math.round(minutes / 60);
+  return hours < 48 ? `${hours} h ago` : `${Math.round(hours / 24)} days ago`;
+}
 var LocalSession = class {
   root;
   launch;
   active = null;
   ui = null;
+  /** The run this session last finished, so a second finish_run is answered, not refused. */
+  finished = null;
   constructor(options) {
     this.root = sdDir(options.project);
     this.launch = options.launch ?? PlaywrightDriver.launch;
@@ -1856,6 +1882,28 @@ var LocalSession = class {
     });
     return this.ui.link(path);
   }
+  /** The viewer's port while it is running in this process. */
+  get viewerPort() {
+    return this.ui?.port ?? null;
+  }
+  /** open_ui: the link, and a line on each of the latest runs. */
+  async overview() {
+    const link = await this.openUi();
+    const runs = (await readRuns(this.root)).slice(0, 5);
+    if (runs.length === 0) return `${link}
+
+No runs yet.`;
+    const lines = runs.map((e) => {
+      if (!("run" in e)) return `- ${e.id}: unreadable (${e.error.split("\n")[0]})`;
+      const { run } = e;
+      const status = effectiveStatus(this.root, run);
+      const t = tallyRun(run);
+      const outcome = run.score ? `${status}, ${run.score.outcome}` : status;
+      return `- ${run.checklist}: ${outcome}, ${t.passed}/${t.total} checks passed, started ${ago(run.started_at)} (${run.id})`;
+    });
+    return [link, `Latest runs, newest first:
+${lines.join("\n")}`].join("\n\n");
+  }
   async startRun(input) {
     if (this.active) return this.resumeActive(input);
     if (!/^[a-z0-9][a-z0-9-]*$/.test(input.checklist)) {
@@ -1865,12 +1913,20 @@ var LocalSession = class {
       join(this.root, "stuntdouble.json"),
       ConfigSchema,
       ".stuntdouble/stuntdouble.json"
-    );
+    ).catch((err) => {
+      throw /file not found$/.test(err.message) ? new Error(`${err.message}. ${SETUP}`) : err;
+    });
     let checklist = await readJson(
       join(this.root, "checklists", `${input.checklist}.json`),
       ChecklistSchema,
       `.stuntdouble/checklists/${input.checklist}.json`
-    );
+    ).catch(async (err) => {
+      if (!/file not found$/.test(err.message)) throw err;
+      const slugs = (await readdir(join(this.root, "checklists")).catch(() => [])).filter((f) => f.endsWith(".json")).map((f) => f.slice(0, -".json".length));
+      throw new Error(
+        `${err.message}. ${slugs.length ? `The checklists here are: ${slugs.join(", ")}.` : SETUP}`
+      );
+    });
     const actor = await readJson(
       join(this.root, "actors", `${checklist.actor}.json`),
       ActorSchema,
@@ -1932,7 +1988,8 @@ var LocalSession = class {
       checkNavigation: refuseNavigation
     };
     const tempProfile = actor.persist_login ? null : await mkdtemp(join(tmpdir(), "stuntdouble-profile-"));
-    const profileDir = tempProfile ?? join(homedir(), ".stuntdouble", "profiles", checklist.actor);
+    const project = createHash("sha256").update(this.root).digest("hex").slice(0, 12);
+    const profileDir = tempProfile ?? join(homedir(), ".stuntdouble", "profiles", project, checklist.actor);
     await mkdir(profileDir, { recursive: true });
     let driver;
     try {
@@ -1977,13 +2034,16 @@ var LocalSession = class {
       everyScreenshot: config.model_screenshots === "every_action",
       stopped: false
     };
+    this.finished = null;
     await this.save();
+    await this.markRunning();
     await this.log("start_run", input, input.resume ? "Resumed" : "Started");
     const opened = await this.browserTool("navigate", { url: run.url });
     const pending = this.pending();
     const viewer = await this.openUi(`/runs/${run.id}`);
     return [
       `Run ${run.id} ${input.resume ? "resumed" : "started"}. Watch it or stop it at ${viewer}`,
+      unweightedNote(checklist),
       buildBrief({
         actor,
         checklist,
@@ -1991,10 +2051,12 @@ var LocalSession = class {
         checks: pending,
         resumed: Boolean(input.resume)
       }),
+      // Not the navigate result: it carries the page title, and this reply is
+      // read by the main session, which never reads page content.
       opened.isError ? `The start page did not load: ${opened.text}
-Try navigate again, or judge the checks you cannot reach as "blocked".` : opened.text,
+Try navigate again, or judge the checks you cannot reach as "blocked".` : `The browser is open at ${run.url}: read the page with read_page before acting.`,
       this.nextStep()
-    ].join("\n\n");
+    ].filter(Boolean).join("\n\n");
   }
   /** One of the browser toolset's direct tools, on the active run's browser. */
   async browserTool(name, input) {
@@ -2002,7 +2064,7 @@ Try navigate again, or judge the checks you cannot reach as "blocked".` : opened
     if (await this.stopRequested(active)) return { text: STOPPED, isError: true };
     const page = await active.driver.context.activePage();
     const host = page ? URL.parse(await page.url())?.hostname ?? "" : "";
-    const tools = active.guarded && host && !active.trustedHosts.some((t) => hostMatchesDomain(host, t)) ? active.guarded : active.trusted;
+    const tools = active.guarded && !isTrustedHost(host, active.trustedHosts) ? active.guarded : active.trusted;
     const result = await tools.run(name, input);
     const refusal = active.driver.takeRefusal();
     if (refusal) {
@@ -2017,6 +2079,7 @@ Try navigate again, or judge the checks you cannot reach as "blocked".` : opened
     return { text: result.text, ...image2 ? { image: image2 } : {}, isError: Boolean(result.error) };
   }
   async evaluateCheck(input) {
+    if (!this.active && this.finished) return this.alreadyFinished();
     const active = this.requireActive();
     if (await this.stopRequested(active)) throw new Error(STOPPED);
     const { run, checklist } = active;
@@ -2095,6 +2158,7 @@ Try navigate again, or judge the checks you cannot reach as "blocked".` : opened
     return [recorded, ...notes, this.nextStep()].join("\n");
   }
   async finishRun(input) {
+    if (!this.active && this.finished) return this.alreadyFinished();
     const active = this.requireActive();
     const stopped = await this.stopRequested(active);
     const { run, checklist } = active;
@@ -2131,8 +2195,11 @@ Try navigate again, or judge the checks you cannot reach as "blocked".` : opened
       passedCount: run.results.filter((r) => r.status === "passed").length,
       totalCount: checklist.checks.length
     });
+    const t = tallyRun(run);
+    const tally = score.weighted ? ` ${t.passed}/${t.total} checks passed.` : "";
+    this.finished = { id: run.id, report: viewer };
     return [
-      `Run ${run.id} ${run.status}: ${score.outcome}. ${counts}. Self-judged: the actor recorded its own verdicts.`,
+      `Run ${run.id} ${run.status}: ${score.outcome}. ${counts}.${tally} ${score.weighted ? "" : `${unweightedNote(checklist, true)} `}Self-judged: the actor recorded its own verdicts.`,
       failed.length ? `Failed:
 ${failed.map(line).join("\n")}` : "",
       blocked.length ? `Not judged:
@@ -2156,6 +2223,7 @@ ${blocked.map(line).join("\n")}` : "",
     run.status = "running";
     await this.clearStop(run.id);
     await this.save();
+    await this.markRunning();
     await this.log("start_run", input, "Resumed");
     const viewer = await this.openUi(`/runs/${run.id}`);
     return [
@@ -2182,6 +2250,10 @@ ${blocked.map(line).join("\n")}` : "",
     await this.ui?.close();
     this.ui = null;
   }
+  alreadyFinished() {
+    const { id, report } = this.finished;
+    return `Run ${id} has already finished: its report is at ${report}. Start another with start_run.`;
+  }
   requireActive() {
     if (!this.active) throw new Error("No run is going: call start_run first.");
     return this.active;
@@ -2224,7 +2296,7 @@ ${blocked.map(line).join("\n")}` : "",
   async markInterrupted() {
     for (const entry of await readRuns(this.root)) {
       if (!("run" in entry) || entry.run.status !== "running") continue;
-      if (entry.run.pid !== process.pid && isAlive(entry.run.pid)) continue;
+      if (entry.run.pid !== process.pid && isLive(this.root, entry.run)) continue;
       await writeJsonAtomic(join(this.root, "runs", entry.id, "run.json"), {
         ...entry.run,
         status: "interrupted"
@@ -2239,6 +2311,11 @@ ${blocked.map(line).join("\n")}` : "",
     });
     if (active.tempProfile) await rm(active.tempProfile, { recursive: true, force: true });
     await this.clearStop(active.run.id);
+    await rm(runningMarker(this.root, active.run.id), { force: true });
+  }
+  async markRunning() {
+    const { run } = this.requireActive();
+    await writeFileAtomic(runningMarker(this.root, run.id), String(run.pid));
   }
   async clearStop(id) {
     const stopFile = join(this.root, ".live", "stop");
@@ -2264,6 +2341,15 @@ ${blocked.map(line).join("\n")}` : "",
     });
   }
 };
+function isTrustedHost(host, trustedHosts) {
+  return host !== "" && trustedHosts.some((t) => hostMatchesDomain(host, t));
+}
+function unweightedNote(checklist, atFinish = false) {
+  if (checklist.checks.some((c) => c.type === "weighted")) return null;
+  const threshold = checklist.pass_threshold == null ? "" : ` Its pass_threshold (${checklist.pass_threshold}) applies to weighted checks only.`;
+  if (!atFinish && !threshold) return null;
+  return `No score: this checklist has no weighted checks, so it passes only when every pass/fail check passes.${threshold}`;
+}
 function scoreLocalRun(checklist, results) {
   const scoring = newChecklistScoring({
     scoreType: checklist.score_type,
@@ -2368,9 +2454,24 @@ function createServer(session) {
   server.registerTool(
     "open_ui",
     {
-      description: "The local viewer for this project: every run with its verdicts, screenshots and activity, and the checklists and actors. Returns a link to open in a browser."
+      description: "The local viewer for this project: every run with its verdicts, screenshots and activity, and the checklists and actors. Returns a link to open in a browser and a line on each of the latest runs."
     },
-    () => attempt(() => session.openUi())
+    () => attempt(() => session.overview())
+  );
+  server.registerTool(
+    "doctor",
+    {
+      description: "Check what local runs need: Node, Chrome, the viewer port, this project's stuntdouble.json and how runs are billed. Read only. Returns one line per check, each starting ok, FAIL, note or --.",
+      annotations: { readOnlyHint: true }
+    },
+    () => attempt(async () => {
+      const { doctor } = await import("./doctor-7PW2QD3M.js");
+      const { lines } = await doctor({
+        project: dirname(session.root),
+        viewerPort: session.viewerPort
+      });
+      return lines.join("\n");
+    })
   );
   return server;
 }
@@ -2388,4 +2489,4 @@ export {
   createServer,
   serveMcp
 };
-//# sourceMappingURL=mcp-O2LGDSU7.js.map
+//# sourceMappingURL=mcp-MOXTYXZU.js.map