@enricai/barnacle 1.12.78 → 1.12.80

package/dist/scripts/recon-browser.js

@@ -48,7 +48,7 @@
  * per act; healing and replans push the upper bound higher on flaky sites).
  */
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.writeFixtureToTempFile = exports.windowHasTransitionBody = exports.windowHasAdvanceTransition = exports.waitForTransitionBody = exports.verifyFillReadback = exports.shouldWarnMissingAdvancePattern = exports.shouldVetoFallbackAdvance = exports.shouldSkipTechnique = exports.selectRadioGroupOption = exports.selectBodyExcerpt = exports.resetBillingErrorFlagForTests = exports.rephraseWithLLM = exports.renderUnfocusedObserve = exports.renderLeafInvalidFields = exports.probeStepBeforeAttempts = exports.probeLeafInvalidContainers = exports.pollEnumerate = exports.parseSelectStep = exports.parseRadioStep = exports.parseCaptureIndex = exports.pairInvalidWithErrors = exports.normalizeDateValue = exports.narrowInvalidFormControl = exports.logBillingErrorIfPresent = exports.latestCaptureIndex = exports.isWizardExitAction = exports.isUploadAffordanceLabel = exports.isSubmitRevealedInvalid = exports.isDomOnlyAdvanceVerified = exports.isAdvanceStep = exports.isAdvanceStalled = exports.hasBillingErrorBeenLogged = exports.formatValidationRejectedReason = exports.findRecentPageTransition = exports.findRecentBackendError = exports.fillTextDatepickerInput = exports.fillHtml5DateTimeInput = exports.extractSubmitFailureEvidence = exports.extractGaEventEvidence = exports.describeAttemptEffectSignals = exports.chooseRequiredSelectOption = exports.capturesAfterIndex = exports.buildRadioIdXPath = exports.RECON_FLOW_FILE_SCHEMA = void 0;
+exports.writeFixtureToTempFile = exports.windowHasTransitionBody = exports.windowHasAdvanceTransition = exports.waitForTransitionBody = exports.verifyFillReadback = exports.shouldWarnMissingAdvancePattern = exports.shouldVetoFallbackAdvance = exports.shouldSkipTechnique = exports.selectRadioGroupOption = exports.selectBodyExcerpt = exports.resetBillingErrorFlagForTests = exports.rephraseWithLLM = exports.renderUnfocusedObserve = exports.renderLeafInvalidFields = exports.probeStepBeforeAttempts = exports.probeLeafInvalidContainers = exports.pollEnumerate = exports.parseSelectStep = exports.parseRadioStep = exports.parseCaptureIndex = exports.pairInvalidWithErrors = exports.normalizeDateValue = exports.narrowInvalidFormControl = exports.logBillingErrorIfPresent = exports.latestCaptureIndex = exports.isWizardExitAction = exports.isUploadAffordanceLabel = exports.isSubmitRevealedInvalid = exports.isDomOnlyAdvanceVerified = exports.isAdvanceStep = exports.isAdvanceStalled = exports.hasBillingErrorBeenLogged = exports.formatValidationRejectedReason = exports.findRecentPageTransition = exports.findRecentBackendError = exports.fillTextDatepickerInput = exports.fillHtml5DateTimeInput = exports.extractSubmitFailureEvidence = exports.extractGaEventEvidence = exports.describeAttemptEffectSignals = exports.chooseRequiredSelectOption = exports.capturesAfterIndex = exports.buildRadioIdXPath = exports.MAX_TRANSIENT_STEP_RETRIES = exports.RECON_FLOW_FILE_SCHEMA = void 0;
 exports.resolveGotoWaitUntil = resolveGotoWaitUntil;
 exports.detectRejectionInResponseBody = detectRejectionInResponseBody;
 exports.isFlowTruncated = isFlowTruncated;
@@ -63,7 +63,9 @@ exports.extractSubmitJudgeRequiredFields = extractSubmitJudgeRequiredFields;
 exports.applyFailedStepFlagsToResumingBridgeStep = applyFailedStepFlagsToResumingBridgeStep;
 exports.filterReplanDuplicatingNextAuthored = filterReplanDuplicatingNextAuthored;
 exports.isReplanReproposingFailedStep = isReplanReproposingFailedStep;
+exports.decideReplanNoProgress = decideReplanNoProgress;
 exports.hasPageAlreadyAdvancedPastStep = hasPageAlreadyAdvancedPastStep;
+exports.hasSubmitStepDomAdvanced = hasSubmitStepDomAdvanced;
 exports.isReplanRegressingAcrossAuthBoundary = isReplanRegressingAcrossAuthBoundary;
 exports.isSubmitShapedInstructionText = isSubmitShapedInstructionText;
 exports.seedSubmitStepFromOwnInstructionText = seedSubmitStepFromOwnInstructionText;
@@ -108,8 +110,10 @@ const session_1 = require("../scraper/session");
 const session_teardown_1 = require("../scraper/session-teardown");
 const spa_readiness_1 = require("../scraper/spa-readiness");
 const stagehand_guard_1 = require("../scraper/stagehand-guard");
+const transient_provider_error_1 = require("../scraper/transient-provider-error");
 const watchdog_1 = require("../scraper/watchdog");
 const judge_llm_batch_1 = require("./judge-llm-batch");
+const recon_capture_rebind_1 = require("./recon-capture-rebind");
 const recon_shared_1 = require("./recon-shared");
 const resume_fixture_1 = require("../testing/resume-fixture");
 const client_1 = require("../testmail/client");
@@ -591,6 +595,11 @@ exports.RECON_FLOW_FILE_SCHEMA = v4_1.z.union([
          * "<tenant>.<ats>.com"].
          */
         ownBackendHostnames: v4_1.z.array(v4_1.z.string().min(1)).optional(),
+        /**
+         * Declares the ATS cannot be replayed as raw HTTP, so the generator emits
+         * a browser-flow-only plugin regardless of HTTP analysis.
+         */
+        browserFlowOnly: v4_1.z.boolean().optional(),
         /**
          * Optional site-specific class-name prefixes that wrap form/error
          * state. The Haiku invalid-fields judge uses these as additional
@@ -1171,6 +1180,28 @@ function isReplanReproposingFailedStep(newSteps, failedStep) {
     const failedSig = stepSignature(failedStep);
     return newSteps.every((s) => stepSignature(s.instruction) === failedSig && !hasCompoundBridgeClause(s.instruction));
 }
+/** Max times one step may be re-run after a provider-only failure cascade. */
+exports.MAX_TRANSIENT_STEP_RETRIES = 2;
+/**
+ * Cause-aware wrapper over {@link isReplanReproposingFailedStep}: a re-proposal
+ * of only the failed step is a guaranteed re-fail for a real selector/page
+ * problem, but a legitimate bounded re-run when every recorded attempt failed
+ * on a transient provider error. Retries are counted per step signature in
+ * `retryCounts` (mutated on "retry"); once exhausted, or when the dump is
+ * missing/empty/mixed, the original abort stands.
+ */
+function decideReplanNoProgress(newSteps, failedStep, attempts, retryCounts) {
+    if (!isReplanReproposingFailedStep(newSteps, failedStep))
+        return "proceed";
+    if (!attempts || !(0, transient_provider_error_1.summarizeAttemptTransience)(attempts).all)
+        return "abort";
+    const sig = stepSignature(failedStep);
+    const used = retryCounts.get(sig) ?? 0;
+    if (used >= exports.MAX_TRANSIENT_STEP_RETRIES)
+        return "abort";
+    retryCounts.set(sig, used + 1);
+    return "retry";
+}
 /**
  * Normalize a URL to origin + pathname (query and hash stripped) for
  * same-page-vs-navigated comparison. Returns null on unparseable input so
@@ -1212,6 +1243,18 @@ function hasPageAlreadyAdvancedPastStep(stepStartUrl, currentUrl, stepInstructio
         return false;
     return (0, phantom_click_1.isPlausibleStepDestination)(stepInstruction, currentUrl);
 }
+/**
+ * DOM-advance counterpart of {@link hasPageAlreadyAdvancedPastStep} for
+ * non-final submit-shaped steps, whose SPA wizards swap content without a URL
+ * change. Reuses the shared {@link hasPageAdvanced} judgement over the
+ * step-start and live visible-text signatures; a final step is never credited
+ * here because its completion is owned by the submit-destination judge.
+ */
+function hasSubmitStepDomAdvanced(isFinalStep, preVisibleTextSignature, liveVisibleTextSignature) {
+    if (isFinalStep)
+        return false;
+    return (0, phantom_click_1.hasPageAdvanced)({ networkCount: 0, url: "", bodyHtmlLength: 0, visibleTextSignature: preVisibleTextSignature }, { networkCount: 0, url: "", bodyHtmlLength: 0, visibleTextSignature: liveVisibleTextSignature });
+}
 /** Word-boundary phrase patterns identifying an account-creation/registration step. */
 const ACCOUNT_CREATION_PATTERNS = [
     /\bcreate\s+(an\s+)?account\b/,
@@ -1234,13 +1277,13 @@ const ACCOUNT_CREATION_PATTERNS = [
  * in". Scopes the Sign-In match to the step's own action clause (via
  * `splitIntoActionClauses`, shared with `isPlausibleStepDestination`) so a
  * step that merely mentions sign-in in descriptive context isn't mistaken
- * for one that is itself about signing in. Pure.
+ * for one that is itself about signing in. The completed-steps side is
+ * clause-scoped the same way, so a step that only mentions account creation
+ * in descriptive text (e.g. a wizard step title) isn't mistaken for a
+ * completed account-creation step. Pure.
  */
 function isReplanRegressingAcrossAuthBoundary(newSteps, completedSteps) {
-    const hasAccountCreation = completedSteps.some((s) => {
-        const norm = normalizeInstruction(s);
-        return ACCOUNT_CREATION_PATTERNS.some((p) => p.test(norm));
-    });
+    const hasAccountCreation = completedSteps.some((s) => (0, phantom_click_1.splitIntoActionClauses)(s).some((clause) => ACCOUNT_CREATION_PATTERNS.some((p) => p.test(clause))));
     if (!hasAccountCreation)
         return false;
     return newSteps.some((s) => (0, phantom_click_1.splitIntoActionClauses)(s.instruction).some((clause) => phantom_click_1.SIGN_IN_PATTERNS.some((p) => p.test(clause))));
@@ -1263,10 +1306,18 @@ const SUBMIT_SHAPED_INSTRUCTION_PATTERNS = [
  * combination that no longer matches what this particular step does).
  * Matches on word boundaries against normalized (lowercased, whitespace-
  * collapsed) text, mirroring {@link SIGN_IN_PATTERNS}/{@link ACCOUNT_CREATION_PATTERNS},
- * so a phrase like "commit the change" does not false-positive on "submit". Pure.
+ * so a phrase like "commit the change" does not false-positive on "submit".
+ * Negated guardrail clauses ("Do NOT click ... proceeds to checkout") are
+ * blanked first: a step that declares itself not a submit is never seeded as
+ * one. Pure.
  */
 function isSubmitShapedInstructionText(instruction) {
-    const norm = normalizeInstruction(instruction);
+    // The loose wording patterns below would seed an explicit `submitStep: true`,
+    // which overrides a declared non-submit intent everywhere downstream; defer
+    // to the same classifier the submit-judge gates read so the two cannot disagree.
+    if ((0, flow_runner_1.classifyStepSubmitIntent)(instruction) === "non-submit")
+        return false;
+    const norm = normalizeInstruction((0, flow_runner_1.stripNegatedSubmitClauses)(instruction));
     return SUBMIT_SHAPED_INSTRUCTION_PATTERNS.some((p) => p.test(norm));
 }
 /**
@@ -2224,24 +2275,27 @@ async function main() {
         // rather than a previous attempt's partially-replanned plan.
         const plan = [];
         const replanEvents = [];
+        const transientStepRetries = new Map();
         try {
             const stagehand = session.stagehand;
-            const page = await stagehand.context.awaitActivePage();
+            let page = await stagehand.context.awaitActivePage();
             // Phase label is mutated between flow steps so the single CDP listener
             // always tags captures with the currently active phase.
             let currentPhase = "home";
-            const stopCapture = wireNetworkCapture(page, counter, signalCounter, recentCaptures, recentCaptureMeta, () => currentPhase, () => {
-                // Live read so SPA history navigation between captures stays accurate.
-                // Fallback to the initial url on parse failure (e.g. about:blank early
-                // in the goto cycle) so we don't accidentally mark a capture as
-                // cross-origin and miss the user-action signal.
+            // Live read so SPA history navigation between captures stays accurate.
+            // Fallback to the initial url on parse failure (e.g. about:blank early
+            // in the goto cycle) so we don't accidentally mark a capture as
+            // cross-origin and miss the user-action signal.
+            const currentPageOrigin = () => {
                 try {
                     return new URL(page.url()).origin;
                 }
                 catch {
                     return new URL(url).origin;
                 }
-            });
+            };
+            const capture = (0, recon_capture_rebind_1.createCaptureRebinder)(page, (target) => wireNetworkCapture(target, counter, signalCounter, recentCaptures, recentCaptureMeta, () => currentPhase, currentPageOrigin));
+            const stopCapture = capture.stopAll;
             logger.info(`navigating to ${url} (waitUntil: ${GOTO_WAIT_UNTIL})`);
             // A navigation wait that never resolves must not discard the run. Ad-heavy
             // commercial sites (analytics/session-replay beacons on timers) never reach
@@ -2316,7 +2370,7 @@ async function main() {
                 const currentOrigin = originOf(readLiveUrl());
                 if (currentOrigin !== "" && currentOrigin !== lastOrigin) {
                     logger.info(`origin changed ${lastOrigin || "(none)"} → ${currentOrigin}; re-gating on SPA hydration`);
-                    await (0, spa_readiness_1.waitForSpaReady)(page, logger);
+                    await (0, spa_readiness_1.regateSpaAfterOriginChange)(page, logger);
                     lastOrigin = currentOrigin;
                 }
                 // Debug: dump the full DOM right before this step's cascade runs. Lets
@@ -2360,6 +2414,10 @@ async function main() {
                 // in-frame navigation would otherwise be invisible to the top-level
                 // page.url() this used to read.
                 const urlAtStepStart = await (0, frame_target_1.readCurrentFrameUrl)(page, frameTarget);
+                const textAtStepStart = step.submitStep
+                    ? (await (0, flow_runner_1.snapshotPage)(frameTarget, signalCounter).catch(() => undefined))
+                        ?.visibleTextSignature
+                    : undefined;
                 // A child frame CDP has just attached to may still be sitting on
                 // about:blank (the moment right after Target.setAutoAttach fires and
                 // before the OOPIF's own navigation lands). waitForSpaReady above only
@@ -2385,6 +2443,13 @@ async function main() {
                         : (0, flow_runner_1.executeStepWithHealing)({
                             stagehand,
                             page,
+                            onPageAdopted: (adopted) => {
+                                if (adopted === page)
+                                    return;
+                                page = adopted;
+                                capture.adopt(adopted);
+                            },
+                            abortSignal: session.deathAbortSignal,
                             frameTarget,
                             step: step.instruction,
                             optional: step.optional,
@@ -2395,7 +2460,10 @@ async function main() {
                             emailStepConfig: step.emailStepConfig,
                             allocatedInbox,
                             flowHasSubmitSemantics: (0, flow_runner_1.flowHasSubmitSemantics)({
-                                steps: plan.map((s) => ({ submitStep: s.submitStep === true })),
+                                steps: plan.map((s) => ({
+                                    submitStep: s.submitStep === true,
+                                    instruction: s.instruction,
+                                })),
                                 submitEndpointPattern,
                                 requireSubmitEndpointMatch,
                             }),
@@ -2513,8 +2581,13 @@ async function main() {
                     // matches, so skip the replan dispatcher entirely and resume with
                     // the remaining tail.
                     const urlAfterFailure = await (0, frame_target_1.readCurrentFrameUrl)(page, frameTarget);
-                    if (!step.submitStep &&
-                        hasPageAlreadyAdvancedPastStep(urlAtStepStart, urlAfterFailure, step.instruction)) {
+                    const urlAdvanced = !step.submitStep &&
+                        hasPageAlreadyAdvancedPastStep(urlAtStepStart, urlAfterFailure, step.instruction);
+                    const domAdvanced = !urlAdvanced &&
+                        step.submitStep === true &&
+                        hasSubmitStepDomAdvanced(i === plan.length - 1, textAtStepStart, (await (0, flow_runner_1.snapshotPage)(frameTarget, signalCounter).catch(() => undefined))
+                            ?.visibleTextSignature);
+                    if (urlAdvanced || domAdvanced) {
                         logger.info(`${(0, flow_runner_1.formatStepPrefix)(i, () => plan.length)} verification failed but the page already advanced past this step (${urlAtStepStart} → ${urlAfterFailure}); treating as completed and resuming remaining tail`);
                         consecutiveStaleSkips = 0;
                         lastSuccessNetworkCount = signalCounter.n;
@@ -2543,7 +2616,11 @@ async function main() {
                     // /tmp/recon/graphql/ during a sweep: some tenants hit this exact
                     // failure — only interruption_check + GA tracking POSTs fired, no
                     // integrated_apply, yet trailing-grace declared success.
-                    if (step.optional && i >= plan.length - flow_runner_1.TRAILING_GRACE_WINDOW) {
+                    const stepDeclaredNonSubmit = (0, flow_runner_1.isStepDeclaredNonSubmit)({
+                        step: step.instruction,
+                        submitStep: step.submitStep ?? false,
+                    });
+                    if (step.optional && !stepDeclaredNonSubmit && i >= plan.length - flow_runner_1.TRAILING_GRACE_WINDOW) {
                         // Trailing-grace check: did the submit actually land somewhere in
                         // the recent capture history? Ask the same Haiku judge — it has
                         // multi-signal reasoning to distinguish real submit POSTs from
@@ -2555,6 +2632,11 @@ async function main() {
                         const trailingGraceVerdict = await (0, verify_submit_1.verifySubmitWithLLM)({
                             client: anthropic,
                             input: {
+                                ...(await (0, flow_runner_1.buildSubmitJudgeStepContext)({
+                                    step: step.instruction,
+                                    isFinalStep: i === plan.length - 1,
+                                    target: (0, frame_target_1.mainFrameTarget)(page),
+                                })),
                                 // A dead session throws synchronously here too — fall back to ""
                                 // rather than let the trailing-grace check itself crash the run.
                                 pageUrl: (() => {
@@ -2634,16 +2716,21 @@ async function main() {
                     // failed step's re-emission. originalRemaining is re-appended below.
                     // The untruncated failure-time DOM re-verifies completed fills so a
                     // silently-reset field isn't trusted as still-completed.
-                    const { bodyHtmlAtFailure, fieldValuesAtFailure } = (() => {
+                    const { bodyHtmlAtFailure, fieldValuesAtFailure, attemptsAtFailure } = (() => {
                         try {
                             const dump = JSON.parse((0, node_fs_1.readFileSync)(dumpPath, "utf8"));
                             return {
+                                attemptsAtFailure: dump.attempts ?? null,
                                 bodyHtmlAtFailure: dump.bodyOuterHtml ?? null,
                                 fieldValuesAtFailure: dump.fieldValuesAtFailure ?? null,
                             };
                         }
                         catch {
-                            return { bodyHtmlAtFailure: null, fieldValuesAtFailure: null };
+                            return {
+                                bodyHtmlAtFailure: null,
+                                fieldValuesAtFailure: null,
+                                attemptsAtFailure: null,
+                            };
                         }
                     })();
                     const newSteps = filterCompletedFromReplan(rawNewSteps, completedSteps, step.instruction, bodyHtmlAtFailure, fieldValuesAtFailure);
@@ -2660,7 +2747,11 @@ async function main() {
                     // cascade on the identical click that just exhausted it. Abort now
                     // instead of waiting REPLAN_CYCLE_THRESHOLD repeats for the cycle
                     // detector — that many dead cascades cost minutes of wall-clock.
-                    if (isReplanReproposingFailedStep(newSteps, step.instruction)) {
+                    const noProgressDecision = decideReplanNoProgress(newSteps, step.instruction, attemptsAtFailure, transientStepRetries);
+                    if (noProgressDecision === "retry") {
+                        logger.warn(`replan #${replanIndex} re-proposed the failed step after provider-transient errors only; allowing bounded re-run`);
+                    }
+                    if (noProgressDecision === "abort") {
                         const noProgressMessage = `replan #${replanIndex} re-proposed only the just-failed step ("${step.instruction.slice(0, 60)}") with no new bridge; resuming would re-fail identically — aborting`;
                         logger.error(noProgressMessage);
                         throw new errors_2.StepVerificationError(noProgressMessage, "replan-cycle-detected");
@@ -2794,8 +2885,7 @@ async function main() {
             // maxTransportRetries budget re-running the same head of the flow.
             const sessionTimeoutHit = session.getSessionTimeoutHit?.();
             if (sessionTimeoutHit) {
-                logger.error(`Browserbase session hit its timeout (${sessionTimeoutHit.configuredTimeoutSeconds}s) after ${completedSteps.length} steps — raise BROWSERBASE_SESSION_TIMEOUT_SECONDS`);
-                process.exit(1);
+                throw new errors_2.SessionLifetimeExceededError(`Browserbase session hit its timeout (${sessionTimeoutHit.configuredTimeoutSeconds}s) after ${completedSteps.length} steps`);
             }
             const cdpTransportClosedError = session.getCdpTransportClosedError?.();
             if (cdpTransportClosedError) {
@@ -2842,6 +2932,13 @@ async function main() {
             await snapshotAndPersistCookieJar(page, jarCounter, "run-complete", currentPhase, plan.length - 1);
             logger.info(`recon complete — ${counter.n} captures written to ${runDir.root}`);
         }
+        catch (err) {
+            // Runs before the finally closes the session so any error that escaped
+            // the step loop (deathSignal rejection, page.url() timeout, uninitialized
+            // Stagehand) is upgraded to the non-retryable lifetime error when the
+            // session died of its own configured lifetime.
+            throw (0, errors_2.isSessionLifetimeExceededError)(err) ? err : (0, session_teardown_1.resolveSessionDeathError)(session, err);
+        }
         finally {
             // Replay-the-discovered-path: if any replan fired and the user provided
             // a flow file, write the improved plan back so the next run starts
@@ -2897,20 +2994,29 @@ async function main() {
     // re-thrown unchanged (same instance, same type) once the retry call
     // settles. This is plain error propagation, not a second retry loop.
     let nonTransportError;
+    let lifetimeError;
     try {
         await (0, retry_1.withScraperRetry)(async () => {
             try {
                 await runFlowAttempt();
             }
             catch (err) {
-                if (err instanceof errors_2.CdpTransportClosedError)
+                if ((0, errors_2.isSessionLifetimeExceededError)(err)) {
+                    lifetimeError = err;
+                    throw err;
+                }
+                if ((0, errors_2.isCdpTransportClosedError)(err))
                     throw err;
                 nonTransportError = err;
             }
         }, { maxAttempts: config_1.config.scraper.maxTransportRetries });
     }
     catch (err) {
-        if (err instanceof errors_2.CdpTransportClosedError) {
+        if (lifetimeError) {
+            logger.error(`${lifetimeError.message} — raise BROWSERBASE_SESSION_TIMEOUT_SECONDS`);
+            process.exit(1);
+        }
+        if ((0, errors_2.isCdpTransportClosedError)(err)) {
             logger.error(`Stagehand tore down the CDP transport mid-flow on every attempt (${config_1.config.scraper.maxTransportRetries} of ${config_1.config.scraper.maxTransportRetries}): ${err.message} — refusing to report success`);
             process.exit(1);
         }