@enricai/barnacle 1.6.12 → 1.6.14

package/dist/scraper/flow-runner.js

@@ -29,6 +29,7 @@ exports.windowHasTransitionBody = windowHasTransitionBody;
 exports.windowHasAdvanceTransition = windowHasAdvanceTransition;
 exports.shouldVetoFallbackAdvance = shouldVetoFallbackAdvance;
 exports.isDomOnlyAdvanceVerified = isDomOnlyAdvanceVerified;
+exports.isClickViewSwapVerified = isClickViewSwapVerified;
 exports.countNgInvalidContainers = countNgInvalidContainers;
 exports.describeAttemptEffectSignals = describeAttemptEffectSignals;
 exports.shouldSkipTechnique = shouldSkipTechnique;
@@ -56,6 +57,7 @@ exports.writeFixtureToTempFile = writeFixtureToTempFile;
 exports.attachToSurfacedInput = attachToSurfacedInput;
 exports.surfaceAndUpload = surfaceAndUpload;
 exports.parseSelectStep = parseSelectStep;
+exports.parseFillStep = parseFillStep;
 exports.parseRadioStep = parseRadioStep;
 exports.pollEnumerate = pollEnumerate;
 exports.waitForTransitionBody = waitForTransitionBody;
@@ -87,7 +89,10 @@ const schemas_1 = require("../lib/llm/schemas");
 const logging_1 = require("../lib/logging");
 const call_capture_1 = require("../lib/telemetry/call-capture");
 const call_types_1 = require("../lib/telemetry/call-types");
+const deep_locator_actuate_1 = require("../scraper/deep-locator-actuate");
 const deep_locator_candidates_1 = require("../scraper/deep-locator-candidates");
+const deep_locator_click_1 = require("../scraper/deep-locator-click");
+const deep_locator_scan_1 = require("../scraper/deep-locator-scan");
 const errors_2 = require("../scraper/errors");
 const frame_target_1 = require("../scraper/frame-target");
 const phantom_click_1 = require("../scraper/phantom-click");
@@ -476,7 +481,14 @@ exports.TRAILING_GRACE_WINDOW = 2;
  * shape via Runtime.callFunctionOn.
  */
 const DOM_SNAPSHOT_EXPR = `(() => { const b = document.body; if (!b) return { html: 0, text: "" }; const t = b.innerText || ""; return { html: (b.outerHTML || "").length, text: t.length + ":" + t.slice(0, 200) }; })()`;
-async function snapshotPage(target, signalCounter) {
+/**
+ * Captures the pre/post signal triple the submit-verify cascade diffs.
+ * Accepts the optional `page` so a resolved child `FrameTarget` whose
+ * `url()` rejects (OOPIF detached by the submit it just fired) can still
+ * report the main frame's post-navigation URL instead of throwing the
+ * whole attempt out of `executeStepWithHealing`.
+ */
+async function snapshotPage(target, signalCounter, page) {
     let bodyHtmlLength = 0;
     let visibleTextSignature = "";
     try {
@@ -495,9 +507,16 @@ async function snapshotPage(target, signalCounter) {
         // Snapshot is observational; on failure, defaults to 0/"" so the verifier
         // sees no delta. Real state-class checks already cover the verified path.
     }
+    // A resolved child FrameTarget's url() reads location.href off the CDP
+    // frame session (frame-target.ts's childFrameTarget), which rejects (or
+    // trips its watchdog) once the OOPIF detaches — most commonly right after
+    // a submit click tears down the wizard iframe. Falling back to page.url()
+    // lets the post-submit navigation still register as a urlChanged signal
+    // instead of throwing the whole step out of the cascade.
+    const url = page ? await target.url().catch(() => page.url()) : await target.url();
     return {
         networkCount: signalCounter.n,
-        url: await target.url(),
+        url,
         bodyHtmlLength,
         visibleTextSignature,
     };
@@ -874,6 +893,43 @@ function isDomOnlyAdvanceVerified(params) {
     // non-advancing POST) does not verify it.
     return networkIsRealAdvance || urlChanged;
 }
+/**
+ * Client-side view-swap verification gate. Credits a click that produces
+ * substantial DOM growth with zero network as a verified step outcome when
+ * the step is a plain client-side reveal (not a submit/final step, not an
+ * advance/'Next' step on a site with advanceTransitionBodyPattern configured).
+ *
+ * **Why this exists:** SPA/iframe wizards often swap visible screens
+ * (e.g. method-chooser → basic-info form) via client-side DOM manipulation
+ * with zero network requests. The existing verifyDomEffect (line 4866-4891)
+ * explicitly returns false for non-radio/checkbox clicks, routing those
+ * to the network/URL signal — which correctly verifies navigation but
+ * false-negatives client-side view swaps. This gate fills that gap.
+ *
+ * **Threshold rationale:** 5000B is 10× the trivial boundary (500B) and
+ * safely distinguishes real view swaps (measured case: +49518B) from
+ * validation-error reflows (typically < 2KB per codebase evidence at
+ * line 2040-2044: error text capped at 200 chars, container markup adds
+ * 150-350B per error, worst-case ~10 errors = ~3500B max).
+ *
+ * **Scope guards:** Excludes final/submit steps (those require real network
+ * per isSubmitRevealedInvalid + LLM submit judge) and advance-pattern steps
+ * (those require real transition per isDomOnlyAdvanceVerified/isAdvanceStalled
+ * to avoid wizard-ATS autosave-vs-transition ambiguity).
+ */
+function isClickViewSwapVerified(params) {
+    const VIEW_SWAP_MIN_BYTES = 5000;
+    const { resolvedAction, isFinalStep, submitStep, isAdvanceWithPattern, networkDelta, bytesDelta, } = params;
+    if (resolvedAction?.method !== "click")
+        return false;
+    if (isFinalStep || submitStep)
+        return false;
+    if (isAdvanceWithPattern)
+        return false;
+    if (networkDelta !== 0)
+        return false;
+    return bytesDelta >= VIEW_SWAP_MIN_BYTES;
+}
 /**
  * Read-only count of ng-invalid form controls on the resolved frame. Side-effect-free
  * counterpart to `probeFormValidityBeforeSubmit` (which also auto-fills
@@ -2277,7 +2333,7 @@ function isUploadAffordanceLabel(label) {
     return /\b(upload|browse|select file|choose file|attach|add (a )?(resume|cv|file)|resume\/cv)\b/.test(norm);
 }
 /**
- * Materialize the in-memory resume fixture to a temp file so CDP
+ * Materialize the in-memory upload fixture to a temp file so CDP
  * `DOM.setFileInputFiles` (which requires a filesystem path, unlike
  * Playwright's `locator.setInputFiles`) can reference it. Only used as a
  * fallback when the on-disk fixture path is unavailable. Process-scoped — the
@@ -2687,6 +2743,40 @@ function parseSelectStep(instruction) {
     const questionLabel = quoted.find((q) => q.trim() !== option)?.trim() ?? null;
     return { option, questionLabel };
 }
+/**
+ * Parse a declarative TEXT-FILL flow step into the field label to target and
+ * the value to type into it.
+ *
+ * Why this exists (sibling of `parseSelectStep`/`parseRadioStep`): the deep-
+ * locator cascade's candidate ranking (`resolveDeepLocatorCandidates`'s
+ * `scoreCandidate`) ranks by the instruction's quoted phrases, which for a
+ * fill step is only the VALUE ("Fill in the First Name field with
+ * 'Reginald'" quotes just `'Reginald'`) — no candidate's accessible name
+ * (e.g. "First Name") ever matches a person's name, so every candidate ties
+ * at score 0 and DOM order decides, clicking whatever happens to be first
+ * rather than filling the named field. This parser recovers the FIELD LABEL
+ * ("First Name") so the caller can match it directly against a candidate's
+ * accessible name instead of ranking by the value.
+ *
+ * Recognizes the flow's conventional phrasing (confirmed against
+ * `src/recon/fixtures/shipped-ats-flow-steps.json`):
+ *   "Fill in the First Name field with 'Reginald'",
+ *   "Fill the signature field with 'Name'",
+ *   "Fill in the Acme Non-Employee ID textbox with 'NA'".
+ * Returns null when the step doesn't match this shape (no "fill" verb, no
+ * field/input/textbox noun, or no quoted value) so the caller falls through
+ * to the normal cascade unchanged.
+ */
+function parseFillStep(instruction) {
+    const match = instruction.match(/\bfill(?:\s+in)?\s+(?:the\s+)?(.+?)\s+(?:field|input|textbox)\s+with\s+'([^']+)'/i);
+    if (!match)
+        return null;
+    const fieldLabel = match[1]?.trim();
+    const value = match[2]?.trim();
+    if (!fieldLabel || !value)
+        return null;
+    return { fieldLabel, value };
+}
 /**
  * Parse a single-choice RADIO flow step into the option to click and (when
  * present) the question label that scopes which radio group it targets.
@@ -2740,6 +2830,44 @@ function parseRadioStep(instruction) {
     const questionLabel = quoted.find((q) => q.trim() !== option)?.trim() ?? null;
     return { option, questionLabel };
 }
+function resolveDeepLocatorActuation(step) {
+    const selectParsed = parseSelectStep(step);
+    if (selectParsed)
+        return { kind: "select", value: selectParsed.option };
+    const fillParsed = parseFillStep(step);
+    if (fillParsed)
+        return { kind: "fill", value: fillParsed.value };
+    return { kind: "click" };
+}
+/** Whitespace-collapsed, lowercased comparison key for {@link findDeepLocatorCandidateByFieldLabel}. */
+function normalizeFieldLabel(text) {
+    return text.replace(/\s+/g, " ").trim().toLowerCase();
+}
+/**
+ * Finds the candidate whose accessible name identifies the named field —
+ * `fieldLabel` (from {@link parseFillStep}/{@link parseSelectStep}), NOT the
+ * value being written — among a frame's deepLocator candidates. Exact match
+ * wins first (a bare "First Name" input against a `fieldLabel` of "First
+ * Name"); otherwise a substring match either direction (a `fieldLabel` of
+ * "Acme Non-Employee ID" against an accessible name of "Non-Employee ID", or
+ * the reverse) so minor phrasing drift between the flow's field noun and the
+ * control's own label still resolves. Returns `null` — never a guess — when
+ * no candidate's accessible name relates to `fieldLabel` at all, so the
+ * caller can refuse to act rather than fill/click the wrong control.
+ */
+function findDeepLocatorCandidateByFieldLabel(candidates, fieldLabel) {
+    const normalizedLabel = normalizeFieldLabel(fieldLabel);
+    if (!normalizedLabel)
+        return null;
+    const named = candidates
+        .map((candidate) => ({ candidate, text: normalizeFieldLabel(candidate.accessibleText) }))
+        .filter((entry) => entry.text.length > 0);
+    const exact = named.find((entry) => entry.text === normalizedLabel);
+    if (exact)
+        return exact.candidate;
+    const partial = named.find((entry) => entry.text.includes(normalizedLabel) || normalizedLabel.includes(entry.text));
+    return partial?.candidate ?? null;
+}
 /** Max settle-retry attempts for a primitive's DOM enumerate (see `pollEnumerate`). */
 const PRIMITIVE_ENUMERATE_ATTEMPTS = 5;
 /** Delay between settle-retry attempts. Total cap ≈ ATTEMPTS × this. */
@@ -4414,25 +4542,63 @@ async function probeFormValidityBeforeSubmit(params) {
     }
 }
 /**
- * Adapts `resolveDeepLocatorCandidates` results into `Action`-shaped
- * evidence for `rephraseWithLLM`, which only knows about Stagehand's
- * `Action` type. Never throws: `resolveDeepLocatorCandidates` itself
- * degrades to `[]` on a resolver failure, so this only feeds the rephrase
- * prompt richer evidence when available — a frame-scoped step whose
- * observe AND deepLocator both come back empty just gets the same `[]`
- * evidence it would have gotten before this fix.
+ * Resolves deepLocator candidates scoped to `INTERACTIVE_CANDIDATE_SELECTOR`,
+ * widening once to the unscoped `"*"` hop when the scoped pass finds
+ * nothing. `INTERACTIVE_CANDIDATE_SELECTOR` is what makes a dense OOPIF
+ * form's candidate set tractable (371 nodes down to a handful — see
+ * `deep-locator-candidates.ts`'s module docblock), but it over-narrows a
+ * `div`/`span` tile that carries only a click handler — no `role`/
+ * `tabindex` — so a scoped-empty result would otherwise strand an attempt
+ * the pre-1.6.13 unscoped `"*"` hop still resolved. The widened pass only
+ * runs when the scoped one returns zero, so the dense-form fast path
+ * (`flow-runner.oopif-dense-form-budget.test.ts`) still costs a single
+ * resolve. Returns the `innerSelector` actually used alongside the
+ * candidates so a caller driving `clickDeepLocatorCandidate` re-derives the
+ * SAME hop the candidates were ranked/indexed against.
+ */
+async function resolveDeepLocatorCandidatesWithWidening(page, frameSelector, instruction, timeoutOptions) {
+    const scoped = await (0, deep_locator_candidates_1.resolveDeepLocatorCandidates)(page, frameSelector, deep_locator_scan_1.INTERACTIVE_CANDIDATE_SELECTOR, instruction, timeoutOptions);
+    if (scoped.length > 0) {
+        return { candidates: scoped, innerSelector: deep_locator_scan_1.INTERACTIVE_CANDIDATE_SELECTOR };
+    }
+    const widened = await (0, deep_locator_candidates_1.resolveDeepLocatorCandidates)(page, frameSelector, "*", instruction, timeoutOptions);
+    return { candidates: widened, innerSelector: "*" };
+}
+/**
+ * Adapts `resolveDeepLocatorCandidatesWithWidening` results into
+ * `Action`-shaped evidence for `rephraseWithLLM`, which only knows about
+ * Stagehand's `Action` type. Never throws: `resolveDeepLocatorCandidates`
+ * itself degrades to `[]` on a resolver failure, so this only feeds the
+ * rephrase prompt richer evidence when available — a frame-scoped step
+ * whose observe AND deepLocator (scoped, then widened) both come back
+ * empty just gets the same `[]` evidence it would have gotten before this
+ * fix.
+ *
+ * `instruction` is forwarded so the evidence list is ranked by relevance to
+ * the step, same as the act path — an unranked `[]`-then-DOM-order list
+ * would feed the rephrase LLM its worst evidence first instead of its best.
+ * It's also what {@link resolveDeepLocatorActuation} reads to pick the
+ * evidence's `method`/`arguments` — a rephrase evaluating a "Fill in the
+ * First Name field with 'Reginald'" step should see fill/"Reginald"
+ * evidence, not a hard-coded click, now that the cascade itself can
+ * actuate fill/select. Omitted (the unfocused-observe evidence list) falls
+ * back to `click` with no value, same as before this method inference
+ * existed.
  *
- * `instruction` is forwarded to `resolveDeepLocatorCandidates` so the
- * evidence list is ranked by relevance to the step, same as the act path —
- * an unranked `[]`-then-DOM-order list would feed the rephrase LLM its
- * worst evidence first instead of its best.
+ * Takes the caller's already-resolved `frameTarget` so
+ * `resolveDeepLocatorCandidates` uses the batched frame-scoped evaluate
+ * instead of re-resolving it internally.
  */
-async function deepLocatorCandidatesAsActions(page, frameSelector, instruction) {
-    const candidates = await (0, deep_locator_candidates_1.resolveDeepLocatorCandidates)(page, frameSelector, "*", instruction);
+async function deepLocatorCandidatesAsActions(page, frameTarget, instruction) {
+    const { candidates } = await resolveDeepLocatorCandidatesWithWidening(page, frameTarget.frameSelector, instruction, { frameTarget });
+    const actuation = instruction
+        ? resolveDeepLocatorActuation(instruction)
+        : { kind: "click" };
     return candidates.map((c) => ({
         selector: c.selector,
         description: c.accessibleText || "(no accessible text)",
-        method: "click",
+        method: actuation.kind === "select" ? "selectOption" : actuation.kind === "fill" ? "fill" : "click",
+        ...(actuation.kind === "click" ? {} : { arguments: [actuation.value] }),
     }));
 }
 /**
@@ -4479,7 +4645,16 @@ async function probeStepBeforeAttempts(params) {
                 ? await reresolveFrameTarget()
                 : frameTarget;
             if (effectiveFrameTarget?.frame) {
-                const deepLocatorCandidates = await (0, deep_locator_candidates_1.resolveDeepLocatorCandidates)(page, effectiveFrameTarget.frameSelector, "*");
+                // Pass the already-resolved `effectiveFrameTarget` so the batched
+                // evaluate reuses it instead of `resolveDeepLocatorCandidates`
+                // re-resolving via its own internal `resolveFrameTarget` fallback —
+                // same "reuse the ambient target" rule as every other frame-scoped
+                // call in this function. Innerselector stays "*": this probe only
+                // answers "does this frame have any content", and batching already
+                // makes that reachability check cheap (one evaluate over every
+                // node) — scoping to interactive elements would false-negative a
+                // frame with rendered content but no controls yet.
+                const deepLocatorCandidates = await (0, deep_locator_candidates_1.resolveDeepLocatorCandidates)(page, effectiveFrameTarget.frameSelector, "*", undefined, { frameTarget: effectiveFrameTarget });
                 if (deepLocatorCandidates.length > 0) {
                     logger.info(`${formatStepPrefix(stepIndex, totalSteps)}: observe found 0 candidates (focused and unfocused) but deepLocator found ${deepLocatorCandidates.length} — treating as present (let cascade resolve)`);
                     return "present";
@@ -4499,7 +4674,7 @@ async function probeStepBeforeAttempts(params) {
     }
 }
 async function executeStepWithHealing(params) {
-    const { stagehand, page, step, optional, upload, submitStep, stepIndex, totalSteps, phase, signalCounter, recentCaptures, recentCaptureMeta, anthropic, logger, captureFn, resumeFixture, isFinalStep, submitEndpointPattern, submittedStateSelectors, requireSubmitEndpointMatch, advanceTransitionBodyPattern, successUrlFragments, successPageTitleHints, ownBackendHostnames, knownErrorClassPrefixes, wizardExitButtonLabels, getSuppressedAisdkElementIdErrorCount, trajectory, onStepFailure, } = params;
+    const { stagehand, page, step, optional, upload, submitStep, stepIndex, totalSteps, phase, signalCounter, recentCaptures, recentCaptureMeta, anthropic, logger, captureFn, uploadFixture, isFinalStep, submitEndpointPattern, submittedStateSelectors, requireSubmitEndpointMatch, advanceTransitionBodyPattern, successUrlFragments, successPageTitleHints, ownBackendHostnames, knownErrorClassPrefixes, wizardExitButtonLabels, getSuppressedAisdkElementIdErrorCount, trajectory, onStepFailure, } = params;
     // Mutable (not the destructured const above) so a lost frame-attach race
     // can be upgraded in place once the OOPIF attaches later in the cascade —
     // see reresolveFrameTargetIfLost below. Every existing reference in this
@@ -4573,7 +4748,7 @@ async function executeStepWithHealing(params) {
         page,
         target: frameTarget ?? (0, frame_target_1.mainFrameTarget)(page),
         isUploadStep: upload,
-        fixture: resumeFixture,
+        fixture: uploadFixture,
         logger,
         signalCounter,
         recentCaptureMeta,
@@ -4744,7 +4919,7 @@ async function executeStepWithHealing(params) {
         // replan budget on every frame-scoped probe-absent failure.
         await reresolveFrameTargetIfLost();
         const unfocusedObserve = probeAbsentObservedUnfocused.length === 0 && frameTarget?.frame
-            ? await deepLocatorCandidatesAsActions(page, frameTarget.frameSelector)
+            ? await deepLocatorCandidatesAsActions(page, frameTarget)
             : probeAbsentObservedUnfocused;
         const dumpPath = onStepFailure?.({
             stepIndex,
@@ -4884,7 +5059,7 @@ async function executeStepWithHealing(params) {
         if (attempt > 1) {
             await page.waitForTimeout(attempt * ATTEMPT_BACKOFF_MS);
         }
-        const pre = await snapshotPage(frameTarget ?? (0, frame_target_1.mainFrameTarget)(page), signalCounter);
+        const pre = await snapshotPage(frameTarget ?? (0, frame_target_1.mainFrameTarget)(page), signalCounter, page);
         // Snapshot the meta-tail length so the final-step pattern gate can scope
         // its URL scan to captures added DURING this attempt (not historical
         // tail from earlier steps).
@@ -5017,7 +5192,7 @@ async function executeStepWithHealing(params) {
                         // burn the step budget probing all of them.
                         const runnerUp = ranked[1];
                         if (runnerUp) {
-                            const midPost = await snapshotPage(frameTarget ?? (0, frame_target_1.mainFrameTarget)(page), signalCounter);
+                            const midPost = await snapshotPage(frameTarget ?? (0, frame_target_1.mainFrameTarget)(page), signalCounter, page);
                             const topVerdict = (0, phantom_click_1.classifyPhantomClick)({
                                 actResultSuccess: true,
                                 pre,
@@ -5079,46 +5254,207 @@ async function executeStepWithHealing(params) {
                 // exclusion is applied here by filtering resolved (already-ranked)
                 // candidates against triedSelectors instead — otherwise attempt 4
                 // would re-pick the same failed element and burn the attempt.
+                // `resolveDeepLocatorCandidatesWithWidening` widens the hop to `"*"`
+                // once when the interactive-scoped pass finds nothing — a
+                // `div`/`span` tile with only a click handler (no role/tabindex)
+                // matches nothing scoped, so the widened pass is what still
+                // surfaces it — and reports which innerSelector actually resolved
+                // so the click below re-derives the same hop, not a stale one.
                 await reresolveFrameTargetIfLost();
-                const deepLocatorCandidates = candidates.length === 0 && frameTarget?.frame
-                    ? (await (0, deep_locator_candidates_1.resolveDeepLocatorCandidates)(page, frameTarget.frameSelector, "*", step)).filter((c) => !triedSelectors.includes(c.selector))
-                    : [];
+                const deepLocatorResolution = candidates.length === 0 && frameTarget?.frame
+                    ? await resolveDeepLocatorCandidatesWithWidening(page, frameTarget.frameSelector, step, { frameTarget })
+                    : null;
+                const deepLocatorInnerSelector = deepLocatorResolution?.innerSelector ?? deep_locator_scan_1.INTERACTIVE_CANDIDATE_SELECTOR;
+                const deepLocatorCandidates = (deepLocatorResolution?.candidates ?? []).filter((c) => !triedSelectors.includes(c.selector));
                 if (candidates.length === 0 && deepLocatorCandidates.length > 0) {
-                    const top = deepLocatorCandidates[0];
-                    if (top) {
-                        record.instruction = `deepLocator: ${top.accessibleText || "(no accessible text)"}`;
-                        // Deny-list guard: mirrors the observe branch's refusal below —
-                        // never act on a wizard-exit control regardless of which
-                        // candidate source (observe vs. deepLocator) surfaced it.
-                        if (isWizardExitAction(top.accessibleText, wizardExitButtonLabels)) {
-                            record.errorMessage = `refused wizard-exit control: "${top.accessibleText.slice(0, 60)}"`;
-                            triedSelectors.push(top.selector);
-                            record.triedSelectors = [top.selector];
+                    // A fill/select step must actuate the NAMED FIELD through the
+                    // fill/select seam, never the click-only walk below: candidate
+                    // ranking scores by the instruction's quoted VALUE (see
+                    // parseFillStep's docblock) — for "Fill in the First Name field
+                    // with 'Reginald'" no control's accessible name is ever 'Reginald',
+                    // so every candidate ties at score 0 and the click walk would fire
+                    // on whatever sits first in DOM order (the bug report's wizard
+                    // 'Close' mis-click). Detect that shape first and route
+                    // deterministically to the candidate matching the field's own
+                    // label, refusing to click when no candidate names that field.
+                    const fillStep = parseFillStep(step);
+                    const selectStep = fillStep ? null : parseSelectStep(step);
+                    const fieldTarget = fillStep
+                        ? { kind: "fill", fieldLabel: fillStep.fieldLabel, value: fillStep.value }
+                        : selectStep?.questionLabel
+                            ? { kind: "select", fieldLabel: selectStep.questionLabel, value: selectStep.option }
+                            : null;
+                    if (fieldTarget) {
+                        const matched = findDeepLocatorCandidateByFieldLabel(deepLocatorCandidates, fieldTarget.fieldLabel);
+                        if (!matched) {
+                            const failureMessage = `deepLocator: no candidate matched field "${fieldTarget.fieldLabel}" (refusing to click an unrelated control)`;
+                            record.actResultSuccess = false;
+                            record.errorMessage = failureMessage;
                             attempts.push(record);
-                            failureReasons.push(record.errorMessage);
-                            logger.info(`${formatStepPrefix(stepIndex, totalSteps)} attempt ${attempt}: ${record.errorMessage}`);
+                            failureReasons.push(failureMessage);
+                            logger.info(`${formatStepPrefix(stepIndex, totalSteps)} attempt ${attempt}: ${failureMessage}`);
                             continue;
                         }
-                        triedSelectors.push(top.selector);
-                        record.triedSelectors = [top.selector];
-                        try {
-                            await (0, deep_locator_candidates_1.clickDeepLocatorCandidate)(page, frameTarget?.frameSelector, "*", top.index);
+                        triedSelectors.push(matched.selector);
+                        record.triedSelectors = [matched.selector];
+                        const actuated = fieldTarget.kind === "fill"
+                            ? await (0, deep_locator_actuate_1.fillDeepLocatorCandidate)(page, frameTarget?.frameSelector, deepLocatorInnerSelector, matched.index, fieldTarget.value, { frameTarget })
+                            : await (0, deep_locator_actuate_1.selectDeepLocatorCandidateOption)(page, frameTarget?.frameSelector, deepLocatorInnerSelector, matched.index, fieldTarget.value, { frameTarget });
+                        if (actuated) {
+                            record.instruction = `deepLocator: ${matched.accessibleText || "(no accessible text)"}`;
                             record.actResultSuccess = true;
-                            record.actResultDescription = `deepLocator clicked "${top.accessibleText || top.selector}"`;
-                            // Synthesize a click action so downstream verification (network
-                            // / url / dom) treats this exactly like any other resolved
-                            // click — same idiom as deep-submit-locator/structured-click.
-                            resolvedAction = {
-                                selector: top.selector,
-                                description: record.actResultDescription,
-                                method: "click",
-                            };
+                            record.actResultDescription = `deepLocator ${fieldTarget.kind === "fill" ? "filled" : "selected"} "${matched.accessibleText || matched.selector}" with "${fieldTarget.value}"`;
+                            // verifyDomEffect can't resolve a `deeplocator=` selector (see
+                            // deep-locator-actuate.ts's module docblock: target.locator()
+                            // has no meaning for cross-origin OOPIF hop notation) — the
+                            // write + read-back fillDeepLocatorCandidate/
+                            // selectDeepLocatorCandidateOption already performed IS the
+                            // verification signal, so record it directly instead of
+                            // synthesizing a resolvedAction that a verifier can only ever
+                            // score false for.
+                            record.verifiedBy = "dom";
+                            attempts.push(record);
+                            logger.info(`${formatStepPrefix(stepIndex, totalSteps)} attempt ${attempt}: ${record.actResultDescription}`);
+                            trajectory?.push({ stepIndex, verifiedBy: "dom" });
+                            return "completed";
                         }
-                        catch (err) {
+                        const failureMessage = `deepLocator: ${fieldTarget.kind} on "${matched.accessibleText || matched.selector}" did not verify (read-back mismatch or rejected write)`;
+                        record.actResultSuccess = false;
+                        record.errorMessage = failureMessage;
+                        attempts.push(record);
+                        failureReasons.push(failureMessage);
+                        logger.info(`${formatStepPrefix(stepIndex, totalSteps)} attempt ${attempt}: ${failureMessage}`);
+                        continue;
+                    }
+                    // A select step whose question is phrased un-quoted (parseSelectStep
+                    // returns `questionLabel: null`) has no fieldLabel to route through
+                    // findDeepLocatorCandidateByFieldLabel above, so ranking below falls
+                    // back to the OPTION value alone. Scoring the full interactive-scoped
+                    // `deepLocatorCandidates` here (buttons, inputs, links alongside any
+                    // <select>) would refuse on virtually every dense form: an unrelated
+                    // 'First Name' input or 'Submit Application' button never mentions
+                    // the quoted option any more than the genuine <select> target does,
+                    // so every one of them ties at score 0 right alongside it — not just
+                    // the two genuinely ambiguous <select>s the guard exists to catch.
+                    // A fresh resolve scoped to "select" narrows the tie check to only
+                    // the candidates that could actually satisfy a select-write (a
+                    // non-<select> candidate already reports "not-actionable" and is
+                    // skipped by the walk below regardless — see
+                    // buildSelectFrameCandidateExpr's `el.options || []` lookup), so
+                    // this only refuses when two (or more) real <select>s tie for
+                    // relevance, e.g. two 'Yes'/'No' controls whose own accessible names
+                    // never mention the option text. A unique top-ranked <select>
+                    // (including the common single-<select>-in-frame case, where a tie
+                    // is impossible) still reaches the walk.
+                    if (selectStep && !selectStep.questionLabel) {
+                        const selectCandidates = await (0, deep_locator_candidates_1.resolveDeepLocatorCandidates)(page, frameTarget?.frameSelector, "select", null, { frameTarget });
+                        const optionPhrases = (0, deep_locator_candidates_1.extractTaggedPhrases)(step);
+                        const optionScores = selectCandidates.map((candidate) => (0, deep_locator_candidates_1.scoreCandidate)(candidate.accessibleText, optionPhrases));
+                        const topScore = optionScores.length > 0 ? Math.max(...optionScores) : 0;
+                        const tiedForTop = optionScores.filter((score) => score === topScore).length;
+                        if (tiedForTop > 1) {
+                            const failureMessage = `deepLocator: select step's question is un-quoted and ${tiedForTop} candidates tie for relevance to "${selectStep.option}" (refusing to guess which control to select)`;
                             record.actResultSuccess = false;
-                            record.errorMessage = `deepLocator: click threw ${(0, errors_1.toErrorMessage)(err)}`;
+                            record.errorMessage = failureMessage;
+                            attempts.push(record);
+                            failureReasons.push(failureMessage);
+                            logger.info(`${formatStepPrefix(stepIndex, totalSteps)} attempt ${attempt}: ${failureMessage}`);
+                            continue;
                         }
                     }
+                    // Actionable-candidate walk: a top pick that rejects with the CDP
+                    // `-32000 Node does not have a layout object` error (an unrendered
+                    // node) or is refused by the wizard-exit deny-list costs only that
+                    // one candidate, not the whole attempt — the next ranked candidate
+                    // is tried instead. `attemptTriedSelectors` mirrors the walk's own
+                    // click attempts as they happen (not just on a successful return)
+                    // so a click that throws a REAL error (e.g. a wedged
+                    // `WatchdogTimeoutError`) still feeds attempt 4's exclusion filter.
+                    const attemptTriedSelectors = [];
+                    const denyCandidate = (candidate) => {
+                        const denied = isWizardExitAction(candidate.accessibleText, wizardExitButtonLabels);
+                        if (denied) {
+                            logger.info(`${formatStepPrefix(stepIndex, totalSteps)} attempt ${attempt}: refused wizard-exit control: "${candidate.accessibleText.slice(0, 60)}"`);
+                        }
+                        return denied;
+                    };
+                    // Intent discrimination: observe()'s Stagehand-resolved action
+                    // carries its own method + fill/select arguments (see attempt 1's
+                    // `target.method === "fill"` handling above); the deepLocator
+                    // fallback has no such resolved action, so the step prose is the
+                    // only place the fill/select value can come from. Derived once per
+                    // attempt (not per candidate) — every candidate in this walk is a
+                    // guess at the SAME step's target, so they all actuate the same way.
+                    const actuation = resolveDeepLocatorActuation(step);
+                    const cascadeResult = await (0, deep_locator_click_1.clickFirstActionableCandidate)(deepLocatorCandidates, async (candidate) => {
+                        attemptTriedSelectors.push(candidate.selector);
+                        if (actuation.kind === "select") {
+                            const verified = await (0, deep_locator_actuate_1.selectDeepLocatorCandidateOption)(page, frameTarget?.frameSelector, deepLocatorInnerSelector, candidate.index, actuation.value, { frameTarget });
+                            // selectDeepLocatorCandidateOption never throws on an ordinary
+                            // failed write (see deep-locator-actuate.ts's writeAndVerify) —
+                            // it resolves `false` for both a rejected selectOption() and a
+                            // read-back mismatch. clickFirstActionableCandidate infers
+                            // success from "didn't throw", so a `false` here must become a
+                            // throw or the walk would wrongly report this candidate as
+                            // actuated.
+                            if (!verified)
+                                throw new Error("-32000 Node does not have a layout object");
+                        }
+                        else if (actuation.kind === "fill") {
+                            const verified = await (0, deep_locator_actuate_1.fillDeepLocatorCandidate)(page, frameTarget?.frameSelector, deepLocatorInnerSelector, candidate.index, actuation.value, { frameTarget });
+                            if (!verified)
+                                throw new Error("-32000 Node does not have a layout object");
+                        }
+                        else {
+                            await (0, deep_locator_candidates_1.clickDeepLocatorCandidate)(page, frameTarget?.frameSelector, deepLocatorInnerSelector, candidate.index, { frameTarget });
+                        }
+                    }, { denyCandidate })
+                        .then((outcome) => ({ outcome, error: null }))
+                        .catch((error) => ({ outcome: null, error }));
+                    triedSelectors.push(...attemptTriedSelectors);
+                    record.triedSelectors = [...attemptTriedSelectors];
+                    if (cascadeResult.outcome?.clicked && cascadeResult.outcome.candidate) {
+                        const acted = cascadeResult.outcome.candidate;
+                        const verb = actuation.kind === "select"
+                            ? "selected"
+                            : actuation.kind === "fill"
+                                ? "filled"
+                                : "clicked";
+                        record.instruction = `deepLocator: ${acted.accessibleText || "(no accessible text)"}`;
+                        record.actResultSuccess = true;
+                        record.actResultDescription = `deepLocator ${verb} "${acted.accessibleText || acted.selector}"`;
+                        // Synthesize an action matching the actuation kind so downstream
+                        // verification (network/url/dom — STATE_CLASS_METHODS treats
+                        // fill/selectOption as DOM-verifiable, not click) treats this
+                        // exactly like any other resolved action — same idiom as
+                        // deep-submit-locator/structured-click.
+                        resolvedAction =
+                            actuation.kind === "click"
+                                ? {
+                                    selector: acted.selector,
+                                    description: record.actResultDescription,
+                                    method: "click",
+                                }
+                                : {
+                                    selector: acted.selector,
+                                    description: record.actResultDescription,
+                                    method: actuation.kind === "select" ? "selectOption" : "fill",
+                                    arguments: [actuation.value],
+                                };
+                    }
+                    else {
+                        const failureMessage = cascadeResult.error
+                            ? `deepLocator: ${actuation.kind} threw ${(0, errors_1.toErrorMessage)(cascadeResult.error)}`
+                            : attemptTriedSelectors.length > 0
+                                ? `deepLocator: no actionable candidate (${attemptTriedSelectors.length} not-actionable)`
+                                : "deepLocator: no actionable candidate (every candidate refused by the wizard-exit deny-list)";
+                        record.actResultSuccess = false;
+                        record.errorMessage = failureMessage;
+                        attempts.push(record);
+                        failureReasons.push(failureMessage);
+                        logger.info(`${formatStepPrefix(stepIndex, totalSteps)} attempt ${attempt}: ${failureMessage}`);
+                        continue;
+                    }
                 }
                 else if (candidates.length === 0) {
                     record.errorMessage = "observe returned no candidates";
@@ -5395,7 +5731,7 @@ async function executeStepWithHealing(params) {
                     // prompt, so a resolver error/empty result is fine to swallow.
                     await reresolveFrameTargetIfLost();
                     const candidates = observedCandidates.length === 0 && frameTarget?.frame
-                        ? await deepLocatorCandidatesAsActions(page, frameTarget.frameSelector, step)
+                        ? await deepLocatorCandidatesAsActions(page, frameTarget, step)
                         : observedCandidates;
                     // Fetch live-page evidence so the rephrase prompt can reason about
                     // form state, not just observe candidates. Mirrors the same
@@ -5413,7 +5749,7 @@ async function executeStepWithHealing(params) {
                     // alongside the failed step, not in the top document.
                     const observedUnfocused = await (0, stagehand_guard_1.guardedObserve)(stagehand, undefined, { timeout: exports.STEP_WATCHDOG_MS }, captureFn, frameTarget).catch(() => []);
                     const unfocused = observedUnfocused.length === 0 && frameTarget?.frame
-                        ? await deepLocatorCandidatesAsActions(page, frameTarget.frameSelector)
+                        ? await deepLocatorCandidatesAsActions(page, frameTarget)
                         : observedUnfocused;
                     const submitFailureList = extractSubmitFailureEvidence(recentCaptures, ownBackendHostnames);
                     const gaEventList = extractGaEventEvidence(recentCaptures);
@@ -5466,7 +5802,7 @@ async function executeStepWithHealing(params) {
             continue;
         }
         await page.waitForTimeout(STEP_PAUSE_MS);
-        const post = await snapshotPage(frameTarget ?? (0, frame_target_1.mainFrameTarget)(page), signalCounter);
+        const post = await snapshotPage(frameTarget ?? (0, frame_target_1.mainFrameTarget)(page), signalCounter, page);
         record.post = post;
         if (resolvedAction) {
             record.resolvedMethod = resolvedAction.method ?? null;
@@ -5541,7 +5877,19 @@ async function executeStepWithHealing(params) {
         if (domVerified && !domVerifiedForStep) {
             logger.info(`${formatStepPrefix(stepIndex, totalSteps)} advance step succeeded only via DOM state change (field toggle / non-advancing POST), not a real transition; not treating as verified`);
         }
-        let verified = networkIsRealAdvance || urlChanged || domVerifiedForStep;
+        // Client-side view-swap gate: credit a click that produces substantial
+        // DOM growth (≥5KB) with zero network when it's NOT a submit/final step
+        // and NOT an advance-pattern step. Fixes the UCHealth "Manual Application"
+        // case where a +49KB DOM-only view swap was scored as "no observable effect".
+        const clickViewSwapVerified = isClickViewSwapVerified({
+            resolvedAction,
+            isFinalStep,
+            submitStep,
+            isAdvanceWithPattern: isAdvanceStep(step) && advanceTransitionBodyPattern !== null,
+            networkDelta: post.networkCount - pre.networkCount,
+            bytesDelta: post.bodyHtmlLength - pre.bodyHtmlLength,
+        });
+        let verified = networkIsRealAdvance || urlChanged || domVerifiedForStep || clickViewSwapVerified;
         // Final-step submit-verification gate. Replaces the deterministic
         // submitEndpointPattern regex with a Haiku 4.5 LLM judgment over
         // multi-signal evidence (network captures, page URL/title, DOM
@@ -5654,7 +6002,13 @@ async function executeStepWithHealing(params) {
         if (verified && record.verifiedBy === null) {
             // Preserve a richer verdict set earlier in this attempt (e.g.
             // "submitted-state-dom" from the final-step DOM fallback).
-            record.verifiedBy = urlChanged ? "url" : networkFired ? "network" : "dom";
+            record.verifiedBy = clickViewSwapVerified
+                ? "view-swap"
+                : urlChanged
+                    ? "url"
+                    : networkFired
+                        ? "network"
+                        : "dom";
         }
         // N+16 probe: Stagehand's CDP click sometimes lands on the button without
         // triggering React's SyntheticEvent layer (or jQuery delegated handlers).
@@ -5736,7 +6090,7 @@ async function executeStepWithHealing(params) {
                         probeResult.checked === true &&
                         !ancestorStillInvalid;
                     await page.waitForTimeout(STEP_PAUSE_MS);
-                    const retryPost = await snapshotPage(frameTarget ?? (0, frame_target_1.mainFrameTarget)(page), signalCounter);
+                    const retryPost = await snapshotPage(frameTarget ?? (0, frame_target_1.mainFrameTarget)(page), signalCounter, page);
                     const retryNetworkFired = retryPost.networkCount > pre.networkCount;
                     const retryUrlChanged = retryPost.url !== pre.url;
                     const retryHtmlDelta = retryPost.bodyHtmlLength - pre.bodyHtmlLength;
@@ -6035,7 +6389,7 @@ async function executeStepWithHealing(params) {
     // replan budget on every frame-scoped cascade-exhaust failure.
     await reresolveFrameTargetIfLost();
     const finalObserve = cascadeExhaustObservedFinal.length === 0 && frameTarget?.frame
-        ? await deepLocatorCandidatesAsActions(page, frameTarget.frameSelector, step)
+        ? await deepLocatorCandidatesAsActions(page, frameTarget, step)
         : cascadeExhaustObservedFinal;
     const { pageTitle, pageUrl } = await resolveDumpPageIdentity(page, frameTarget);
     // Discriminator data for "Stagehand sees nothing" failures: capture the raw
@@ -6047,7 +6401,7 @@ async function executeStepWithHealing(params) {
     const bodyOuterHtml = typeof bodyOuterHtmlRaw === "string" ? bodyOuterHtmlRaw.slice(0, 100_000) : null;
     const cascadeExhaustObservedUnfocused = await (0, stagehand_guard_1.guardedObserve)(stagehand, undefined, { timeout: exports.STEP_WATCHDOG_MS }, captureFn, frameTarget).catch(() => []);
     const unfocusedObserve = cascadeExhaustObservedUnfocused.length === 0 && frameTarget?.frame
-        ? await deepLocatorCandidatesAsActions(page, frameTarget.frameSelector)
+        ? await deepLocatorCandidatesAsActions(page, frameTarget)
         : cascadeExhaustObservedUnfocused;
     const dumpPath = onStepFailure?.({
         stepIndex,
@@ -6134,7 +6488,7 @@ async function waitForSpaReady(page, logger, opts = {}) {
  * than let the flow report phantom success.
  */
 async function runHealingFlow(deps) {
-    const { stagehand, page, steps, logger, anthropic, resumeFixture, maxFlowMs } = deps;
+    const { stagehand, page, steps, logger, anthropic, uploadFixture, maxFlowMs } = deps;
     const counter = { n: 0 };
     const signalCounter = { n: 0 };
     const recentCaptures = [];
@@ -6187,7 +6541,7 @@ async function runHealingFlow(deps) {
                 recentCaptureMeta,
                 anthropic,
                 logger,
-                resumeFixture,
+                uploadFixture,
                 frameTarget,
                 isFinalStep: i === steps.length - 1,
                 submitEndpointPattern: deps.submitEndpointPattern ?? null,