Navigate to a URL and get the semantic DIGEST (~2-3KB): landmark regions with ids, headings with their section, and the top-15 RANKED actionables with complete absolute hrefs in structuredContent (only prose abbreviates them; privacy rules still apply). PDF links include a document hint with URL, title, evidence and source observation/id for an external PDF reader. A successful application/pdf response returns document {type, mediaType, url, title, source?, reader, textExtracted: false}, instead of pretending its text was observed; source retains the prior observation/id/title/href when an exact current link supplied the destination. Navigation metadata includes requestedUrl, finalUrl, navigationUrlsSanitized: true (navigation query values and opaque payloads remain hidden), and redirectChainAvailable. When available, redirectChain lists observed HTTP response URLs/statuses, with redirectChainScope: http, redirectChainTotal and redirectChainTruncated; client-side navigations are reflected by finalUrl, never invented as HTTP redirects. Ids (n_xxx) expire on every new observation. A top entry with placeholder: true is an EMPTY form field whose name is its placeholder — a prompt, never data from the site. Every observation reports authState and cookiesForOrigin: this tool drives ITS OWN isolated per-session BrowserContext, cookie jar and storage. authState is conservatively unknown; a cookie count is evidence, not proof of identity, because authentication can also live in storage, bearer state or the URL. For a task that needs the real signed-in session from another browser, this is the wrong instrument. If the site answered with a bot-mitigation interstitial, structuredContent carries blocked: true and challenge {vendor, reason, status, signal, and vendors when more than one is detected — vendors chain, and a confidently wrong name is worse than unknown for per-vendor retry routing}: the content was WITHHELD, which is a different answer from a page that has little on it — fall back to another fetcher rather than recording an empty result. A request that never reached an HTTP response returns failure {layer: dns|tls|transport|http, code, hostUp} instead of a thrown string — a DNS or certificate failure is neither a block nor an empty page. The open waits (bounded) for window.onload AND briefly watches the fresh document for timer-delayed first paints (entry ads armed via setTimeout at parse time), so late overlays/modals enter the FIRST digest; if the document is STILL not complete, structuredContent carries loading {readyState, waitedMs} and the prose says so — treat the digest as a truthful walk of an UNFINISHED page and re-observe before trusting completeness. Returns observationId (opaque identity of this observation) in structuredContent, and digest (marks/heads/top) in both structuredContent and prose — read the fields, do not parse the text. After a SAME-ORIGIN navigation, structuredContent may also carry carried: which strong-identity elements (data-testid / authored accessible names) persisted from the previous page and how their state/content moved (the cart badge "1"→"2"), plus only-before/only-after COUNTS of page-specific content — those counts are "different page", never removals/additions. Optional redact: session privacy rules — any name/label/text/state string containing a listed term leaves every observation as [redacted], and each observation carries an attestation that the policy ran (policyRevision, rulesActive) — never hit counts, which would tell you whether and how often the hidden term occurs. Raw form values are never returned; sensitive categories use coarse change signals and declare same-bucket uncertainty.

Search text across the WHOLE page (not just the visible part) and get RANKED matches in structuredContent: id, role, name, text (same string, honest label), href (mailto:/tel: pass through intact) and truncated when a value was cut. Optional contextChars allocates a TOTAL text budget across ranked matches: each included match gets context with text, totalChars, observationId and a browser_text continuation if truncated. Context uses the same frozen first-read text as browser_text; contextMatchesOmitted explicitly counts matches without context. The right tool to locate something specific on long pages — do not ask for the full outline.

Climb from a find/digest match to the CARD around it (the nearest container with ≥2 actionables) and observe just that subtree: the way from "found the price text" to "here is the clickable title next to it". Returns the card with fresh ids; the global look baseline stays untouched. When the card carries no prose beyond its actionables and the page splits the logical card across sibling rows (HN-style tables), structuredContent also carries siblingRowText — the metadata row BESIDE the card, as declared text, never merged into the card ids. The right follow-up when browser_find located an inner node and you need its actionable context — never infer the card by id arithmetic.

Act on the page: click (by id from the digest/find, or "x,y"), type (into the focused element — click it first), select a native single-select option by target id and exactly one of value or label (exact, unique match), or enter. Click and select auto-scroll and CONFIRM role/name of the resolved element: read that echo before continuing. Select refuses non-native, multiple, disabled, hidden or covered controls and disabled/ambiguous options. It dispatches input/change when the option changes, does not echo the supplied choice, omits choice arguments from audit logs, and preserves the observation baseline. Privacy rules and readonly policy apply. After EVERY action, call browser_verify, even when select reports selectionChanged:false.

WHAT CHANGED since the last observation — the verification of your action. Returns changed (a faithful negative: if your click did nothing it says so instead of letting you believe you acted), the list of changes with kind (added/removed/state/style/moved) role and name, and what became covered or visible. Possible replacements also carry beforeName, so the prior and current identities are both explicit. Call it after EVERY action instead of comparing screenshots, then pass its diffId to browser_assert to check THIS exact transition. When a baseline exists, a full observation returns diffId, beforeObservationId, afterObservationId, observationId (the after observation), and baselineAdvanced: true. A stored diff covers the FULL evidence, including changes beyond the presentation cap and folded wrappers. No baseline means no diffId. Retention is bounded per session (32 records, 8 MiB total, 10 minutes); an individually oversized record returns diffAvailable: false and diffError: {code: DIFF_TOO_LARGE, message}, without a diffId. structuredContent carries changed, changes (list of {kind, role, name, beforeName?, id, from?, to?}), changesTotal, changesShown, changesOmitted, and changesOmittedByKind — read those rather than parsing the prose. State changes include their before/after state in from/to. Reading aids: the capped changes summary prioritizes state, content, and actionability-related changes before other semantic changes and geometry; omissions, including folded wrappers, are explicitly counted by kind. It omits folded wrapper nodes of an ADDED subtree (identity-free generic wrappers only — authored names never fold; foldedWrappers counts them and changesTotal is the full diff count), and geometryOnly: true flags a diff that is ONLY moved/resized AND changed no actionability — a scope reflow (scrollbar, container resize) you can skim past. After a same-origin navigation, carried reports the strong-identity elements that persisted across pages and their state/content transitions (see browser_open).

Save the currently observed state as a NAMED baseline (before a risky action). NOT undo: a comparison point for browser_diff.

Diff the current state against a checkpoint saved with browser_checkpoint: everything that changed since that known point. Note: the next browser_verify baseline becomes the current state.

Assert the EXACT transition already read from browser_verify by passing its diffId: immutable, full diff evidence; no new observation and no baseline or id-epoch advance. Stored results return diffId, beforeObservationId, afterObservationId, observationId (the historical after observation), baselineAdvanced: false and evidenceSource: stored. Historical evidence survives later navigation; its element ids are historical, not actionable — find again before acting. Stored diffs support ONLY changed, mustInclude, mustNotInclude, only, maxChanges, becameVisible and becameCovered. Mixing diffId with exists, notCovered, url/urlIncludes, ignore, settleMs, retry or keepBaseline (even false) is pass:false; use a separate live assertion for current page predicates. Unknown, expired, evicted, foreign-session or privacy-invalidated ids return pass:false with error {code: DIFF_UNAVAILABLE, message}, never a live fallback. Without diffId, the existing LIVE assertion mode checks the diff since the last observation and consumes its baseline at the END unless keepBaseline:true; calling it after verify therefore checks a NEW interval. Checks any combination of: url (substring of the current URL), changed (expect the diff since the last observation to be true/false — the faithful negative makes "my action did nothing" ASSERTABLE), mustInclude ([{kind, role, name}] entries that must appear in the diff; kind ∈ added/removed/content/state/style/moved/resized — a framework re-render that REPLACES a node reports kind possible-replacement, and added/removed matchers accept it with STRICT side reading: an added matcher matches the after-side name/role, a removed matcher matches ONLY the before-side name/role (never the after side; selector specs never match through the alias), and the check result says "found (via possible-replacement — identity ambiguous)" instead of a plain green), exists (text findable anywhere on the page), notCovered (text whose best match must not be occluded). FAIL-LOUD CONTRACT: unknown spec keys, empty specs and missing baselines are hard pass:false with a reason — confusion never looks green. Returns structured {pass, hasBaseline, attempts, checks[], changes[]}; the diff evidence (with state from/to) travels with every result. Also: mustNotInclude (assert side-effect ABSENCE), maxChanges, becameVisible/becameCovered (actionability deltas), mustInclude entries accept selector and to:{state:value} (directional state — assert the menu IS open), settleMs and retry:{budgetMs} re-walk against the SAME baseline until pass or budget (CSS transitions land mid-flight). exists searches accessible names AND page text. SPA soft navs: results with a baseline include navigated:true + baselineUrl when the URL moved since the baseline was taken — that diff spans two pages of one document; re-observe on settled content (non-zero, stable actionables) before trusting change-based checks.

Scroll WITHOUT acting: by element id (to center), to "top"/"bottom", or to an absolute y in pixels. The one legitimate reason: dense listings hydrate their content lazily on scroll and the semantic walk honestly sees only the DOM that exists — scroll, then browser_verify to see what appeared. Ids from the current observation remain valid (scrolling does not re-observe). Includes a bounded settle for the lazy loaders.

Read visible text of ONE node or section by id, with a bounded character budget (maxChars defaults to 600, maximum 12000). Returns text, textSource, capturedAt (first-read timestamp), totalChars, returnedChars, offset, nextOffset, truncated, observationId and continuation in structuredContent. Pass a non-null continuation directly to browser_text for the next slice; it includes sessionId. offset > 0 requires observationId from the initial read. The first read freezes redacted text at capturedAt (not at observation time) so later slices cannot splice a changing document; re-observe and re-find to refresh it. Snapshot storage is bounded per observation (1000000 UTF-16 units / 128 nodes); exceeding it returns an explicit error. Continuations fail on expired observations, detached ids, navigation or privacy-policy changes. Offsets count JavaScript UTF-16 code units. A cut value must be continued, not recorded as complete; the prose also marks truncation.

Expanded views when the digest is not enough: outline (full structure trimmed to 12KB), map with offset (pages actionables beyond the top), or zoom with id (observes ONLY that subtree — the detail of a region/card; renews ids, global baseline untouched). Explicit escalation — digest first. Returns outline in structuredContent for view:"outline", with truncated when trimmed.

Open an independent browsing session and get its sessionId and environment. Optional viewport, colorScheme and reducedMotion configure responsive/media QA before navigation; defaults are 1280×800, light, no-preference. Each session owns a private BrowserContext, cookie/storage jar, popup tree, observation counter and ids, so several sweeps run AT THE SAME TIME without invalidating or authenticating each other. Pass the returned sessionId on every call belonging to that sweep. Close it with browser_session_close when done.

Get or update this session’s viewport, colorScheme and reducedMotion for responsive, light/dark and motion QA. Omitted settings are preserved; with no settings this only reads the current environment. Returns environment. Updates apply to every page in this session and future popups, preserving cookies, storage, privacy rules and other sessions. The observation baseline remains unchanged (baselineAdvanced:false); call browser_verify after every update to inspect resulting page changes. Changing settings is refused under readonly policy because resize/media handlers may trigger page actions.

Close a session opened with browser_session_open and free its complete context, including every popup it created. Sessions also close themselves after 10 minutes idle, so a crashed run does not leak pages.

List the live sessions with their current URL, observation number and idle time.

Pixels as ESCALATION, not default: snapdom render of the viewport, or of one element (scrolled to center) when you pass an id. Only when the doubt is genuinely visual (layout, color, overlap) — for what changed there is browser_verify.