zerocheck 0.1.2 → 0.1.3

package/README.md

@@ -9,7 +9,7 @@ Paste a checklist, review one drafted check and browser verification result per
 Requires Node.js 20.19 or newer and a Zerocheck project/account. Create a project in the web app’s Settings, then authorize a device or create a project token.
 
 ```sh
-npm install --save-dev zerocheck@0.1.2
+npm install --save-dev zerocheck@0.1.3
 npx zerocheck login
 npx zerocheck init --project your-project-id --url http://localhost:3000
 npx zerocheck install
@@ -149,7 +149,7 @@ Use the installed CLI as a stdio MCP server, with an explicit project directory:
   "mcpServers": {
     "zerocheck": {
       "command": "npx",
-      "args": ["--yes", "zerocheck@0.1.2", "mcp", "--project-dir", "/absolute/path/to/project"]
+      "args": ["--yes", "zerocheck@0.1.3", "mcp", "--project-dir", "/absolute/path/to/project"]
     }
   }
 }
@@ -167,6 +167,6 @@ Local execution sends selected check YAML, the page context needed for AI (inclu
 
 ## Repository development
 
-The CLI bundles the shared engine source into `dist/index.js`; it has no unpublished engine-package dependency. From this repository, run `npm run build`, `npm run check`, and `npm test` in `packages/cli`. Tests include a real MCP client communicating with the bundled CLI over stdio. The generated workflow and documentation use version `0.1.2`; update that pin together with the package version for a release.
+The CLI bundles the shared engine source into `dist/index.js`; it has no unpublished engine-package dependency. From this repository, run `npm run build`, `npm run check`, and `npm test` in `packages/cli`. Tests include a real MCP client communicating with the bundled CLI over stdio. The generated workflow and documentation use version `0.1.3`; update that pin together with the package version for a release.
 
 Hosted runs and imports return a queued handle when browsers are occupied; CLI and MCP keep polling until they finish. The queue holds 20 waiting jobs for up to five minutes and never automatically resumes interrupted jobs after restart. Recent imports is available in the web app’s Checks page.

package/dist/index.js

@@ -6979,7 +6979,7 @@ async function loginCommand() {
   let device;
   try {
     device = await apiRequest("/v1/cli/auth/device", {
-      body: { device_name: deviceName, client_version: "0.1.2" },
+      body: { device_name: deviceName, client_version: "0.1.3" },
       authed: false
     });
   } catch (err) {
@@ -7492,16 +7492,47 @@ async function isFrameVisible(frame) {
   return true;
 }
 
+// ../../src/agent/observation-recovery.ts
+function isTransientObservationError(error2) {
+  const message = error2 instanceof Error ? error2.message : String(error2);
+  if (/closed|cancel|abort|policy|ambiguous/i.test(message)) return false;
+  return /Execution context was destroyed|Cannot find context with specified id|Frame was detached|frame has been detached|Stale computer-use observation: frame/i.test(message);
+}
+async function retryObservation(page, read, options = {}) {
+  const deadline = Date.now() + (options.timeoutMs ?? 5e3);
+  for (let attempt = 0; ; attempt++) {
+    options.signal?.throwIfAborted();
+    if (page.isClosed()) throw new Error("Browser page closed during observation.");
+    const remaining = deadline - Date.now();
+    if (remaining <= 0) throw new Error("Browser observation timed out before a complete snapshot was available.");
+    try {
+      const value = await read(remaining);
+      options.signal?.throwIfAborted();
+      if (Date.now() > deadline) throw new Error("Browser observation timed out before a complete snapshot was available.");
+      return value;
+    } catch (error2) {
+      if (attempt >= 2 || page.isClosed() || options.signal?.aborted || !isTransientObservationError(error2) || deadline - Date.now() <= 50) throw error2;
+      options.onRetry?.(error2);
+      await page.waitForTimeout(50);
+    }
+  }
+}
+
 // ../../src/agent/page-state.ts
-async function capturePageState(page, logger, secrets = [], targetPolicy) {
+async function capturePageState(page, logger, secrets = [], targetPolicy, options = {}) {
+  return retryObservation(page, (remaining) => capturePageStateOnce(page, logger, secrets, targetPolicy, remaining), options);
+}
+async function capturePageStateOnce(page, logger, secrets, targetPolicy, timeoutMs) {
   targetPolicy?.assertCurrentUrl(page.url());
   const policy = observationPolicy(page, targetPolicy);
-  const [screenshot, snapshot, url2, title] = await Promise.all([
-    captureScreenshot(page, 5e3, secrets, policy),
+  const observations = await Promise.allSettled([
+    captureScreenshotOnce(page, timeoutMs, secrets, policy),
     getAccessibilitySnapshot(page, logger, secrets),
     page.url(),
     page.title()
   ]);
+  for (const observation of observations) if (observation.status === "rejected") throw observation.reason;
+  const [screenshot, snapshot, url2, title] = observations.map((observation) => observation.value);
   const frameTrees = [];
   const blocked = /* @__PURE__ */ new Set();
   for (const frame of page.frames()) {
@@ -7604,6 +7635,9 @@ function buildTreeFromCDP(nodes) {
   return nodeMap.get(nodes[0].nodeId) ?? null;
 }
 async function captureScreenshot(page, timeoutMs = 5e3, secrets = [], targetPolicy) {
+  return retryObservation(page, (remaining) => captureScreenshotOnce(page, remaining, secrets, targetPolicy), { timeoutMs });
+}
+async function captureScreenshotOnce(page, timeoutMs, secrets, targetPolicy) {
   const policy = observationPolicy(page, targetPolicy);
   const mask = [];
   const owners = [];
@@ -8125,8 +8159,8 @@ async function waitForVisibleText(page, assertion, timeoutMs, signal, targetPoli
     targetPolicy?.assertCurrentUrl(page.url());
     let found = false;
     for (const frame of targetPolicy ? page.frames() : [page.mainFrame()]) {
-      if (targetPolicy && (await blockedFrameOrigins(frame, targetPolicy)).length) continue;
       if (!await isFrameVisible(frame)) continue;
+      if (targetPolicy && (await blockedFrameOrigins(frame, targetPolicy)).length) continue;
       const locator = frame.getByText(assertion.text, { exact: true });
       const count = await locator.count();
       for (let index = 0; index < count; index++) {
@@ -9746,8 +9780,12 @@ var BrowserAgent = class {
     const before = /* @__PURE__ */ new Map();
     for (const assertion of this.config.assertions ?? []) {
       if (step.lineNumber >= 0 ? assertion.lineNumber <= step.lineNumber : assertion.lineNumber < 0 && assertion.lineNumber >= step.lineNumber) continue;
-      const observed = await deterministicPredicate(page, assertion.argument, this.config.targetPolicy);
-      if (observed !== void 0) before.set(assertion.raw, observed);
+      try {
+        const observed = await deterministicPredicate(page, assertion.argument, this.config.targetPolicy);
+        if (observed !== void 0) before.set(assertion.raw, observed);
+      } catch (error2) {
+        if (!isTransientObservationError(error2) || page.isClosed() || this.signal?.aborted) throw error2;
+      }
     }
     const result = await this.withAI(page, step, start, STEP_EXECUTION_SYSTEM, lookup.reason, establishedIdentity);
     this.signal?.throwIfAborted();
@@ -9788,7 +9826,7 @@ var BrowserAgent = class {
     for (let attempt = 0; attempt < 2; attempt++) {
       let executionStarted = false;
       try {
-        const state = await capturePageState(page, this.config.logger, this.config.secrets, this.config.targetPolicy);
+        const state = await this.observe(page);
         const plan = await this.plan(step, state, system, recovery);
         this.signal?.throwIfAborted();
         const groundedTarget = await resolveVisualTarget(page, plan, state, this.config.secrets ?? [], this.config.targetPolicy);
@@ -9798,13 +9836,20 @@ var BrowserAgent = class {
         return pass(step, start, { resolvedVia: "ai", confidence: plan.confidence, _resolvedPlan: plan });
       } catch (error2) {
         const planningRetry = !executionStarted && /temporar|unavailable|rate limit|429|50[234]|ECONNRESET|fetch failed/i.test(String(error2));
-        if (attempt || this.signal?.aborted || planningRetry && this.config.retries === 0 || !planningRetry && !safeToReground(error2, this.signal) && !/Stale computer-use observation/.test(String(error2))) throw error2;
+        if (attempt || this.signal?.aborted || planningRetry && this.config.retries === 0 || !planningRetry && !safeToReground(error2, this.signal) && !(!executionStarted && /Stale computer-use observation/.test(String(error2)))) throw error2;
         recovery = redactText(String(error2), this.config.secrets ?? []);
         this.recoveries.push({ kind: planningRetry ? "grounding" : "interaction", reason: recovery, at: (/* @__PURE__ */ new Date()).toISOString() });
       }
     }
     throw new Error("Computer-use grounding exhausted its recovery budget.");
   }
+  observe(page) {
+    return capturePageState(page, this.config.logger, this.config.secrets, this.config.targetPolicy, {
+      timeoutMs: Math.max(1, this.deadline - Date.now()),
+      signal: this.signal,
+      onRetry: (error2) => this.recoveries.push({ kind: "interaction", reason: redactText(String(error2), this.config.secrets ?? []), at: (/* @__PURE__ */ new Date()).toISOString() })
+    });
+  }
   async plan(step, state, system, recovery) {
     const plan = await groundAction({
       ai: this.config.ai,
@@ -9847,7 +9892,7 @@ var BrowserAgent = class {
     const { deadlineMs, cleanDescription } = parseWaitDescription(step.argument);
     const outcome = await pollForPresence({ description: cleanDescription, deadlineMs, pollIntervalMs: 500, check: async () => {
       this.signal?.throwIfAborted();
-      const state = await capturePageState(page, this.config.logger, this.config.secrets, this.config.targetPolicy);
+      const state = await this.observe(page);
       const text = await this.config.ai.complete({ system: WAIT_FOR_ELEMENT_SYSTEM, screenshot: state.screenshot, text: redactText(buildWaitForElementPrompt(cleanDescription, state.accessibilityTree), this.config.secrets ?? []) });
       const parsed = extractJSON(text);
       return { present: parsed?.present === true, confidence: typeof parsed?.confidence === "number" ? parsed.confidence : 0, reasoning: typeof parsed?.reasoning === "string" ? parsed.reasoning : void 0 };
@@ -9867,7 +9912,7 @@ var BrowserAgent = class {
       if (passed) this.confirmCandidates(step);
       return passed ? pass(step, start, { resolvedVia: "deterministic" }) : fail(step, start, `Expected exact text "${visible.text}" to be ${visible.visible ? "visible" : "absent"} within ${timeoutMs}ms.`, "assertion", { resolvedVia: "deterministic" });
     }
-    const outcome = await evaluateAssertion(this.config.ai, redactText(step.argument, this.config.secrets ?? []), await capturePageState(page, this.config.logger, this.config.secrets, this.config.targetPolicy));
+    const outcome = await evaluateAssertion(this.config.ai, redactText(step.argument, this.config.secrets ?? []), await this.observe(page));
     if (outcome.confidence < 70) throw new Error(`AI could not verify the expected outcome: ${outcome.reasoning}`);
     return outcome.pass ? pass(step, start, { confidence: outcome.confidence, resolvedVia: "ai" }) : fail(step, start, `Assertion failed: ${outcome.reasoning}`, "assertion", { confidence: outcome.confidence, resolvedVia: "ai" });
   }
@@ -9875,7 +9920,7 @@ var BrowserAgent = class {
     let last;
     const performed = /* @__PURE__ */ new Set();
     for (let i = 0; i < 5; i++) {
-      const state = await capturePageState(page, this.config.logger, this.config.secrets, this.config.targetPolicy);
+      const state = await this.observe(page);
       const plan = await this.plan(step, state, ACT_SYSTEM);
       const identity = JSON.stringify([plan.action, plan.selector, plan.value]);
       if (performed.has(identity)) return fail(step, start, "The agent requested the same action again without confirmed completion; stopped to avoid duplicate writes.", "interaction");
@@ -9883,7 +9928,7 @@ var BrowserAgent = class {
       this.signal?.throwIfAborted();
       const actionResult = await executeAction(page, plan, { ...this.policy(step), groundedTarget: await resolveVisualTarget(page, plan, state, this.config.secrets ?? [], this.config.targetPolicy) });
       last = pass(step, start, { resolvedVia: "ai", confidence: plan.confidence, _resolvedPlan: plan });
-      const after = await capturePageState(page, this.config.logger, this.config.secrets, this.config.targetPolicy);
+      const after = await this.observe(page);
       const response = await this.config.ai.complete({ system: ACT_COMPLETION_SYSTEM, screenshot: after.screenshot, text: redactText(buildCompletionCheckPrompt(step.argument, after.accessibilityTree), this.config.secrets ?? []) });
       if (extractJSON(response)?.complete === true && actionResult.execution?.outcomeObserved) return last;
       if (actionResult.execution?.effect !== "observation" && !actionResult.execution?.outcomeObserved) return fail(step, start, "The action has no independently observed completion. Stopped before another action.", "interaction");
@@ -11499,7 +11544,7 @@ jobs:
           node-version: '22'
       - name: Install Zerocheck
         run: |
-          npm install --prefix "$RUNNER_TEMP/zerocheck-cli" --no-save --package-lock=false zerocheck@0.1.2
+          npm install --prefix "$RUNNER_TEMP/zerocheck-cli" --no-save --package-lock=false zerocheck@0.1.3
           echo "$RUNNER_TEMP/zerocheck-cli/node_modules/.bin" >> "$GITHUB_PATH"
       - name: Restore learned targets
         uses: actions/cache@v4
@@ -17532,7 +17577,7 @@ var runner = z2.enum(["local", "hosted"]).default("local").describe("Local brows
 var output = (value) => ({ content: [{ type: "text", text: JSON.stringify(value) }], structuredContent: { result: value } });
 var error = (value) => ({ isError: true, content: [{ type: "text", text: JSON.stringify({ error: "operation_failed", message: value instanceof Error ? value.message : String(value) }) }] });
 function createMcpServer(services) {
-  const server = new McpServer({ name: "zerocheck", version: "0.1.2" });
+  const server = new McpServer({ name: "zerocheck", version: "0.1.3" });
   server.registerTool("list_checks", {
     description: "List the repository YAML checks, IDs, exact contents and revisions. Does not execute them.",
     inputSchema: { paths: z2.array(z2.string()).optional(), environment },
@@ -17672,7 +17717,7 @@ async function mcpCommand(options) {
 }
 
 // src/index.ts
-var program = new Command().name("zerocheck").version("0.1.2").description("Turn your team\u2019s manual release checklist into repeatable browser tests.");
+var program = new Command().name("zerocheck").version("0.1.3").description("Turn your team\u2019s manual release checklist into repeatable browser tests.");
 var directory = (command) => command.option("--project-dir <path>", "Repository project directory", process.cwd());
 var environment2 = (command) => directory(command).addOption(new Option("--env <environment>", "Named project environment").choices(["dev", "staging", "production"]));
 var execution = (command) => environment2(command).addOption(new Option("--runner <runner>", "Browser execution location").choices(["local", "hosted"]).default("local"));

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "zerocheck",
-  "version": "0.1.2",
+  "version": "0.1.3",
   "description": "Turn manual release checklists into repeatable browser tests locally, in CI, or hosted.",
   "type": "module",
   "repository": {