Generated on 2026-06-05.

The comparison harness runs ax-grep through Puppeteer and compares named role output against agent-browser snapshot.

Operational rule: run comparison suites sequentially. agent-browser and Chromium can exhaust the host when several compare:static* runs overlap. Check for existing browser processes before a run and confirm cleanup afterward.

The historical named role overlap score is intentionally strict: it is an exact match over normalized role:name pairs, with no fuzzy or containment matching. It is useful for tracking whether ax-grep is reproducing agent-browser snapshot output, but it can understate agent usefulness when the missing items are mostly static text.

The harness also reports agentReadiness scores. These are still based on the same exact normalized role:name matches, but split by agent-facing use:

• referenceRecall: how much of the agent-browser named output appears in ax-grep.
• candidatePrecision: how much of the ax-grep named output appears in agent-browser.
• actionableRecall: exact recall for links, buttons, fields, tabs, and other operation targets.
• navigationRecall: exact recall for links, headings, landmarks, and search.
• contentRecall: exact recall for headings, images, table/list structure, and static text.
• structuralContentRecall: exact recall for headings, images, table/list structure, and other non-StaticText content roles.
• textRecall: exact recall for StaticText roles only.
• score: weighted summary for agent parsing: actionable 40%, navigation 25%, structural content 20%, precision 15%. Strict contentRecall and textRecall remain visible because text-heavy pages can expose useful gaps, but raw StaticText volume should not dominate the agent-usefulness score.

The static harness also emits cliAgentSummary, which scores the actual agent-facing --agent compact JSON envelope rather than raw tree overlap. It uses top-level agent routing status, pageCheck, structured content evidence, source links and source quality, readability, requested verification status, follow-up nextSteps, searchResults, and suggestedActions to estimate how directly an agent can decide whether to read, open, or retry a page. averageCliAgentScore in gateSummary tracks that higher-level usefulness separately from agentReadiness, which remains an exact agent-browser snapshot overlap metric. minCliAgentScore enforces the same readiness floor per included target, so a weak search, page-check, or browser-retry case cannot be hidden by strong average results. When a per-target floor fails, weakAgentTargets lists the affected category, URL, scores, status, and primary action for quick follow-up. averageAgentExecutorScore is the executor-focused aggregate. It combines the schema, routing, next, expected-outcome, signal, read-target, command, browser-retry, continuation, response, diagnostic, and verification fields that subagents need to run a search/page-check loop without reconstructing intent from the raw tree. minAgentExecutorScore applies the executor floor per included target. The score includes action-schema completeness, so run-command actions need both a human-readable command and raw commandArgs, read-current actions need readFrom, and interact-browser actions need an explicit browser-interaction signal. averageActionSchemaScore tracks that schema completeness directly across the gate-included targets. averageSearchResultActionScore tracks whether compact search results include rank-specific openResult, command, and raw commandArgs, so search agents can open alternate results without reconstructing commands. averageAgentRoutingIntentScore tracks whether agent.routingIntent correctly summarizes the primary action as reading current payload, opening a URL, searching, retrying with browser HTML, requiring browser interaction, or stopping. averageAgentContinuationModeScore tracks whether agent.continuationMode maps that intent to the executor-facing mode: read, command, browser, capture-html, inspect, or stop. averageAgentNextScore tracks whether agent.next is a canonical executor payload that agrees with continuationMode and mirrors the primary action's command, read pointer, URL, browser interaction, and terminal fields. averageAgentNextShortcutScore tracks whether top-level agent.next* shortcuts mirror that canonical agent.next payload. averageAgentRunbookShortcutScore tracks whether top-level agent.runbook* shortcuts mirror the nested loop runbook contract. averageAgentExpectedOutcomeScore tracks whether agent.expectedOutcome describes the success condition for the next step, including read pointers when the next step is evidence reading. averageAgentPlanShortcutScore tracks whether top-level agent.expectedOutcome* and agent.executionPlan* shortcuts mirror the nested next-step contract. averageAgentSignalScore tracks whether agent.signals exposes structured content, verification, search result, source link, browser, response, and diagnostic signals needed for fast agent routing. averageContentEvidenceMetadataScore tracks whether pageCheck.contentEvidence items include source and bounded score metadata, so agents can prioritize semantic evidence over fallback excerpts. averageReadabilityReasonScore tracks whether compact page checks preserve concise readability reasons, so agents can understand why a page is readable, thin, blocked, or worth retrying. averageAgentReadabilityReasonScore tracks whether the compact top-level agent summary repeats concise readability reasons, so agents can route from the first object before drilling into pageCheck. averageAgentPageMetadataShortcutScore tracks whether agent.page* mirrors root page metadata such as canonical URL, language, author, dates, and structured-data types. averageAgentSemanticSummaryScore tracks whether agent.semanticSummary and top-level agent.semantic* shortcuts preserve semantic tree counts, role-group counts, heading/landmark outline flow with parent context, keyboard shortcut hints, in-page links, top role/heading/landmark/named role fields with direct paths, interactive and focusable state, link URL, button description, image, table/list structure, form-field input hints, description, value, resolved relation targets, choice, state, and unavailable-subtree shortcuts for quick page-shape routing. State scoring includes parsed top-state fields so agents do not need to parse ARIA state strings. averageAgentBarrierShortcutScore tracks whether top-level agent.topBarrier* shortcuts mirror the highest-priority page barrier. averageAgentStructuredShortcutScore tracks whether top-level structured content counts and top* shortcuts mirror the first table, FAQ, code block, resource, media item, section paths/selectors, navigation/media structure, provenance, offer, dataset, identity, timeline, contact point, and best structured read-target shortcut. averageAgentReadTargetScore tracks whether agent.readTargets points to payload fields that actually exist and are worth reading, and whether read-current actions mark the matching target as primary. averageAgentTopReadTargetShortcutScore tracks whether agent.topReadTarget* mirrors the first read-target entry for fast routing without scanning agent.readTargets. averageAgentAlternativeActionShortcutScore tracks whether top-level agent.alternativeAction* shortcuts mirror the first non-primary action. averageAgentHandoffScore and averageAgentBriefExecutorStepScore also cover handoff detail preservation. Search handoffs must keep executable result/source choices with snippets and command args; answer handoffs must keep selected evidence text/reasons; read handoffs for forms and action targets must keep URL templates, fields, selectors, methods, and encoding; diagnostic handoffs must keep selected signals and quality gates. This prevents the compact handoff from turning into an opaque "retry/open this" instruction. averageAgentResultChoiceScore, averageAgentSourceChoiceScore, and averageAgentActionListScore cover the same problem outside the handoff. Search choices keep snippets, freshness dates, and sitelinks; source choices keep source-link text, snippets, selectors, and executable commands; source-link actions keep sourceLinkRef so agents can jump back to the exact pageCheck.sourceLinks[n] item. averageAgentTopActionShortcutScore tracks whether agent.topAction* mirrors the first action candidate, including execution, priority, command/read target, URL, and source-link reference. averageAgentResultCountScore tracks whether agent.resultCount is zero for non-search pages and at least the compact result count for search pages. averageAgentChoiceCountScore tracks whether executable choice-count shortcuts match their result, form, action-target, and source-link source counts. averageAgentTopChoiceShortcutScore tracks whether agent.topChoiceKind, path, label, URL, and command arguments mirror the first executable result, source, form, or action-target choice for fast subagent routing. averageAgentTopResultChoiceShortcutScore tracks whether agent.topResultChoice* mirrors the first search result choice, including URL, rank, open-result value, command arguments, source quality, freshness/relevance, matches, sitelinks, and selection reason. averageAgentTopSourceChoiceShortcutScore tracks whether source-link specific top-level shortcuts mirror the first executable source choice, including source type, hints, score, and selection reason. averageAgentEvidenceCountShortcutScore tracks citation, answer-evidence, read-target, and action count shortcuts against their agent arrays. averageAgentTopCitationShortcutScore tracks whether agent.topCitation* mirrors the first citation item, including path, kind, confidence, reason, URL, and score. averageAgentSignalCountShortcutScore tracks signal severity and failing quality-gate count shortcuts against agent.signals and agent.qualityGates. averageAgentTopQualityShortcutScore tracks whether agent.topSignal* and agent.topQualityGate* mirror the first signal and quality gate for fast accept/block routing without scanning diagnostic arrays. averageAgentProblemShortcutScore tracks whether agent.problemSignalKind, severity, message, and agent.failingQualityGate* mirror the first warning/error signal and first failing quality gate, including gate severity and score, so agents can explain blocked pages without scanning diagnostic arrays. averageAgentSourceLinkCountScore tracks whether agent.sourceLinkCount is zero for search pages and matches compact pageCheck.sourceLinks for ordinary content pages. averageAgentFormActionCountScore tracks whether top-level agent.formCount and agent.actionTargetCount match compact pageCheck.forms and pageCheck.actionTargets, so agents can detect hidden forms and JSON-LD/OpenSearch actions before scanning nested page-check arrays. averageAgentFormActionChoiceScore tracks whether agent.formChoices and agent.actionTargetChoices preserve the compact form/action target IDs, paths, selectors, URL templates, query inputs, submit text, first-field hints, and methods needed for subagent form execution and selection loops. averageAgentTopFormActionChoiceShortcutScore tracks whether top-level form/action-target shortcuts mirror the first executable form and action target. averagePageLinkCommandScore tracks whether compact pageCheck.primaryLinks and pageCheck.sourceLinks include direct command and commandArgs, so agents can open page links without reconstructing fetch flags. averageAgentBrowserNeedScore tracks whether agent.needsBrowserHtml agrees with the primary action: browser HTML retry actions should require browser HTML, while URL search recovery, alternate-result recovery, read-current, and retry-later actions should not. averageAgentBrowserHtmlScore tracks whether browser HTML fallback payloads preserve capture file/script data plus nested reason, command, and command-args fields across agent.next, execution plans, executor steps, and handoff steps. averageAgentPageKindScore tracks whether agent.pageKind mirrors the root payload kind, so agents can route from the top-level agent object without re-reading analysis.kind or the envelope root. averageAgentAlternativeActionCountScore tracks whether agent.alternativeActionCount matches the deduplicated compact follow-up actions left outside agent.primaryAction, so agents can know whether a page has useful alternatives before scanning nested action arrays. averageAgentUsabilityScoreConsistency tracks whether agent.usabilityScore matches the documented compact quality heuristic derived from status, readability, confidence, evidence, search results, source links, and verification status. averageAgentEvidenceQualityScoreConsistency and averageAgentSourceQualityScoreConsistency track whether top-level evidence and source quality scores match the compact evidence/source arrays, so agents can compare payload quality before reading every candidate item. averageAgentBestReadTargetScore tracks whether agent.bestReadTarget and its count, score, primary flag, and reason match the primary or highest-scored agent.readTargets entry, so agents can start reading the best compact field without sorting candidates. averageAgentDiagnosticCountScore tracks whether top-level diagnostic severity counts and agent.topDiagnostic* match the compact diagnostics array, so agents can distinguish warnings from hard errors before drilling into diagnostic messages. averageAgentVerificationCountScore tracks whether top-level verification requested/found/missing counts match the compact verification object, so agents can decide whether requested evidence is complete before reading details. averageAgentVerificationQueryScore tracks whether agent.verificationFoundQueries and agent.verificationMissingQueries preserve the exact matched and missing --find query lists and whether the top matched or missing query shortcuts mirror the first items; agent.handoff and agent.executor carry the same lists for brief subagent loops. averageAgentResponseMetadataScore tracks whether agent.responseStatus, agent.responseOk, agent.responseContentType, and agent.finalUrlChanged mirror the compact envelope response fields, so agents can judge fetch health from the top-level agent object. averageAgentHiddenSignalScore tracks whether hidden pageCheck groups such as hydration, API endpoints, app config, app/mobile hints, provenance, policies, JSON-LD facts, and resource metadata are present at valid payload paths and discoverable through at least one read target when available. This is the executor-focused counterweight to raw accessibility-tree overlap: these signals are often useful to subagents but absent from browser accessibility snapshots. averageAgentHiddenSignalCountScore tracks whether top-level agent.hiddenSignalCount, agent.hiddenReadTargetCount, and agent.bestHiddenReadTarget* match those hidden groups and read-target shortcuts. averageAgentTopHiddenSignalShortcutScore tracks whether agent.topHiddenSignal* mirrors the first hidden metadata, API, config, or provenance signal. averageAgentHiddenCommandShortcutScore tracks whether hidden hydration, safe GET-like API, and app-hint URLs keep executable top-level follow-up commands. averageAgentBrowserAdvantageScore tracks whether those hidden pageCheck signals create a concrete agent-browser advantage when they exist, rather than only matching visible accessibility-tree roles. The higher-level CLI agent score also credits hidden pageCheck signal groups and recoverable browser-HTML retry actions. A page with little visible text can still be useful to a subagent when it exposes metadata read targets or a runnable browser-capture handoff. When ax-grep produces a ready, high-scoring agent payload with content evidence, the static comparison treats the page as usable before applying raw tree-size failure classes such as thin-reference challenge or over-collection. This keeps agent-browser advantage cases visible in the gate even when the raw static tree is larger than the browser snapshot. averageAgentCanContinueScore tracks whether agent.canContinue agrees with the primary action execution class, so recoverable errors with runnable actions do not look terminal and usage/input errors without actions do not look actionable. averageAgentPrimaryExecutionScore tracks whether agent.primaryExecution matches agent.primaryAction.execution, so agents can route from the shortcut field without rereading the full action object. averageAgentPrimaryShortcutScore tracks whether agent.primaryActionName, reason, priority, command, URL, rank, read-from, source-link reference, and browser shortcuts mirror agent.primaryAction, so agents can continue from top-level routing fields. averageAgentExecutorShortcutScore tracks whether agent.executorActionName, decision, mode, operation, confidence, terminal/continue flags, command arguments, read-from, URL, target, and expected-outcome shortcuts mirror agent.executor, so subagents can route the next step without parsing the full executor object. averageAgentHandoffShortcutScore tracks whether agent.handoffActionName, decision, mode, operation, answer status, confidence, terminal/continue flags, priority, command arguments, read-from, URL, target, and expected-outcome shortcuts mirror agent.handoff, so brief loops can run from top-level fields when they do not need the full handoff object. averageAgentAnswerShortcutScore tracks whether agent.answerPlanStatus, confidence, reason, next action, gap count, citation IDs, first answer-evidence metadata, command arguments, after-interaction command, read-from, and URL shortcuts mirror agent.answerPlan and agent.answerEvidence, so agents can decide whether to answer or continue without parsing the full plan object. averageAgentSourceSearchProvenanceScore tracks whether opened-result payloads with sourceSearch.selectedResult or sourceSearch.alternateResults expose matching agent.readTargets, so agents can inspect original SERP provenance before trusting or recovering from an opened page. averageAgentSourceSearchShortcutScore tracks whether top-level agent.sourceSearchQuery, locale, verification-query count/top query, engine/search URL, selected rank/title/URL, selected command, and first alternate command mirror the source-search payload for quick SERP recovery decisions. The command shortcuts are exposed as sourceSearchSelectedCommandArgs and sourceSearchAlternateCommandArgs. averageAgentRecommendedMetadataScore tracks whether search pages with a recommendedResult repeat its URL, title, rank, source, relevance, official-source hint, selection reason, and command args on the top-level agent object for quick routing. averageAgentSearchDecisionScore and averageAgentPageDecisionScore also check top-level agent.searchDecision* and agent.pageDecision* shortcuts, so agents can route without reopening the nested decision objects. Terminal actions such as read-content and use-evidence are treated as usable without executable commands when execution is read-current and a readFrom pointer is present, because the compact payload already contains the evidence an agent should read. Browser-interaction actions are also valid without commands when execution is interact-browser; in those cases another static fetch would not advance the page state.

The --agent payload intentionally removes repeated routing data: top-level diagnostics are represented as agent.diagnosticCodes, repeated primary actions are omitted from suggestedActions and verification/page-check action slots, page-level alternatives are suppressed when verification has selected use-evidence, search-page link/action follow-ups are omitted when searchResults already carry the decision surface, and search output is capped to the first five results plus any out-of-window recommended result. Opened-result payloads omit engine attempts once sourceSearch records the selected result, but keep compact selected/alternate SERP candidates with executable open commands for failure recovery, preserving custom fetch options such as --timeout and --user-agent. Page checks also skip common global-navigation headings, links, and buttons, and omit extra external primary links when source links are already present, so repository and documentation pages route agents toward page content instead of site chrome. Fetch failures that still have a target URL emit an executable browser-HTML retry command, preserving --find checks for the next run. When browser HTML is already supplied through --html-file or --stdin, the compact agent payload does not ask for another browser retry. Parsed search result pages keep agent.canUseFetchedHtml true even when their page-readability score is low, because the result cards remain usable for open/refine routing. Captured blocker pages still keep challenge/login/paywall diagnostics; the follow-up action changes to browser-state inspection instead of another capture loop. HTTP error actions are status-aware, so missing URLs and transient server errors no longer all collapse into a browser-HTML retry; missing opened search results and opened-result verification failures can route directly to an alternate original SERP candidate when one matches the missing --find text.

The default baseline does not force a viewport. A shared viewport can be tested with AX_LITE_COMPARE_VIEWPORT=WIDTHxHEIGHT, but the default run is kept stable to avoid changing the benchmark shape unexpectedly.

For state-sensitive pages, AX_LITE_COMPARE_SETUP=path/to/setup.js evaluates a setup script in both Puppeteer and agent-browser before extraction. This keeps exact-match scoring intact while making page state explicit.

Run with pnpm compare:korea.

Run with pnpm compare:static URL.... Run with pnpm compare:static:agent for the smaller executor-focused regression set that exercises readable pages, listings, forum-style links, and a search diagnostic while tracking averageAgentExecutorScore. In the current run, the gate summary includes 8 targets and excludes 4 diagnostics; averageAgentExecutorScore is 1.00 and averageAgentHiddenSignalScore is 1.00 for the included executor targets. averageAgentBrowserAdvantageScore is also tracked so the hidden-metadata fixture proves more than raw accessibility-tree overlap. The set includes a synthetic hidden-metadata gate whose browser snapshot only contains a visible heading while pageCheck exposes 13 hidden head, script, policy, app-link, provenance, and JSON-LD signals.

This path fetches HTML and runs extract(html) from the static entry without Chrome, jsdom, WebView, layout, or script execution. agent-browser is used only as the reference snapshot for comparison.

Run with pnpm compare:static:diverse.

Run with pnpm compare:tokens URL....

This serializes both browser-injected and static SSR extraction into compact agent prompt text and estimates token cost with cl100k_base. It also measures the recommended --agent compact JSON payload, so the benchmark can compare raw tree prompts with the actual CLI payload agents should use. The prompt text includes role, name, state/value, and selectors for interactive nodes. Token gate averages skip browser references that are only a tiny shell while static or agent output contains substantially more inspectable payload. Those thin browser snapshots are counted separately as excludedThinBrowserReference instead of distorting static/browser and agent/browser ratios.

Run with pnpm compare:tokens:diverse.

Run with pnpm compare:static:korea-social and pnpm compare:tokens:korea-social.

This target set covers Clien, Ruliweb, DCInside, Google/Bing/Startpage Search, X/Twitter, and Instagram. The static comparison benchmark first tries plain HTML fetch. If the response looks like a bot challenge, login shell, or empty client-rendered shell, it falls back to agent-browser rendered HTML through document.documentElement.outerHTML before running the static extractor.

Search and social targets stay in the benchmark as diagnostics, but are not included in the gate summary because their logged-out public views are anti-bot, hydration, and personalization sensitive. In this run, the gate summary includes 4 targets and excludes 5 diagnostics; the average gate agent score is 0.698 and the average static/browser token ratio is 0.395.

Notes:

• Ruliweb can require rendered HTML fallback in some runs, but this run fetched useful static HTML directly.
• Clien matching improved after benchmark normalization started stripping icon font private-use glyphs and leading menu bullets from comparable names.
• DCInside preserved action/navigation signals, moved from over-collected to usable, and lowered static/browser token ratio below 0.50 after compact static extraction started pruning unnamed leaf wrappers.
• Google Search returned a bot/interstitial shell in the browser reference path.
• Bing Search is volatile in this environment: fetch or rendered HTML can expose useful search UI or unrelated image-search affordances, but the exact reference comparison is not stable enough for a gate yet. Search diagnostics can be classified as volatile instead of usable.
• Startpage can return useful fetch HTML, but this run hit a suspended-connection captcha page in the browser-derived reference path. Embedded CSS-in-JS text is now excluded from static names, but the target remains a reference-challenge fixture.
• Instagram can alternate between login-only and fuller logged-out shells in this environment; keep it diagnostic even when a run scores as usable.

Run with pnpm compare:static:china-japan and pnpm compare:tokens:china-japan.

This target set covers Chinese and Japanese encyclopedia, news, portal, forum, developer, search, and video/social pages. Search, video/social, and pages whose reference navigation fails in this environment stay in diagnostics. In this run, the gate summary includes 7 targets and excludes 6 diagnostics; the average gate agent score is 0.654 and the average static/browser token ratio is 0.544.

Notes:

• China Wikipedia became usable after the benchmark stopped treating Wikipedia table-of-contents section numbers as part of comparable link names and static extraction started auto-detecting wiki-like HTML to preserve more article links by default.
• Xinhua and Douban are the strongest Chinese gate targets in this run.
• People China fetches usable HTML, but agent-browser navigation is blocked in this environment, so the target is diagnostic until a stable reference path is available.
• Baidu search is unstable across runs. It can collapse to a tiny feedback shell or expose a larger fetched search page; keep it diagnostic.
• Japan Wikipedia is usable but still has low exact content recall on the large article body.
• NHK fetches static HTML, but Puppeteer and agent-browser both hit HTTP/2 navigation failures in this environment. Token ratio is reported as n/a when the browser reference is unavailable.
• Qiita and Hatena are useful Japanese gate targets; Hatena remains a token-cost stress case.
• 5ch became usable after reference comparison hardening, forum thread metadata normalization, auto-detected forum link-farm limits, and pruning redundant listitem wrappers around links/buttons. It remains a token-cost stress case at roughly parity with browser injection.

• Simple static pages line up well. example.com matched the important named roles exactly.
• Wikipedia exposes a large language <select>. ax-grep can still unroll options for agent operation, but the comparison harness now disables option unrolling to match agent-browser snapshot more closely.
• Wikipedia language links use both visible article-count text and descriptive title attributes. ax-grep now follows accessible-name priority more closely by using link contents before title fallback.
• MDN uses many custom elements. ax-grep now prunes simple custom-element wrappers, but host elements that expose state, ids, or shadow content still need deeper handling.
• MDN ad-like placements can be excluded in comparison mode with excludeLikelyAds. The general extractor keeps this off by default so callers do not silently lose content.
• A shared comparison viewport is available through AX_LITE_COMPARE_VIEWPORT=WIDTHxHEIGHT, but it is opt-in because responsive pages can change the benchmark shape significantly.
• Hacker News relies on layout tables. The comparison harness normalizes Chrome's LayoutTableCell role to cell and removes punctuation-adjacent whitespace, improving overlap from 0.64 to 0.75.
• The comparison harness normalizes common role vocabulary differences such as image vs img, paragraph vs p, and StaticText vs text.
• libraries.io/npm/typescript is the stable package-registry-like sample.
• The new agent-facing metrics show a different picture than raw overlap on Wikipedia and Libraries.io: static-text recall is low, but actionable and navigation targets are mostly preserved. That distinction better matches the goal of making pages tractable for agents.
• Korean samples cover a large encyclopedia article, two news-like pages, and a public portal. The Korean Wikipedia page is intentionally heavy and is kept in compare:korea rather than the default sample script.
• hani.co.kr timed out waiting for Puppeteer network idle during the baseline run and used the DOMContentLoaded state. Keep it as a news-site stress case, but do not treat it as a tightly stable target yet.
• Korean live pages can shift by a few nodes or snapshot lines between runs as headlines, ads, and embedded widgets update.
• yonhapnewstv.co.kr currently lines up best among the Korean samples across exact overlap, content recall, and agent score.
• Static SSR extraction is viable for simple and server-rendered pages. It works especially well on Hacker News and reasonably on Yonhap News TV without any browser runtime.
• Static SSR extraction can prune some non-exposed menu content from HTML alone. The most important signal so far is a collapsed control with aria-expanded="false" and aria-controls; pruning the controlled subtree reduced Wikipedia static tokens from 11,183 to 1,292 and improved exact overlap from 0.05 to 0.57.
• Static SSR extraction now skips non-semantic payload tags, summarizes large child lists, and collapses repeated template-like subtrees. This keeps raw SSR payloads from turning into unbounded prompt input, while preserving an explicit note that nodes were omitted.
• Compact static extraction prunes unnamed leaf wrappers such as decorative spans, emphasis tags, empty inputs, and line breaks. Ancestor accessible names are computed before pruning, so useful link/button names are preserved while prompt-only wrapper noise is removed.
• Static SSR extraction cannot account for computed CSS, responsive layout, client-only rendering, open shadow roots, iframe documents, or post-load DOM mutation. Treat it as a lightweight agent parsing fallback, not an AXTree replacement.
• Static SSR extraction is not automatically cheaper in prompt tokens, but it can be competitive when collapsed controlled regions are pruned. It is now slightly cheaper than browser injection on Wikipedia and still cheaper on Hacker News and Yonhap News TV.
• Token cost needs its own benchmark gate. Agent-readiness can be acceptable while prompt cost is unacceptable, especially on SSR pages with large hidden menus, language selectors, or template payloads.
• Diverse targets show why benchmark categories matter. Government, ecommerce, and article pages preserve useful action/navigation signals; large news portals are good stress tests; Reddit/X/Instagram are better treated as social/challenge fixtures because public logged-out views often collapse to shell, login, or bot-protection states.
• Diverse token results show static extraction is often cheaper on server rendered news, government, and ecommerce pages. Social sites are inconsistent: X's fetched shell is tiny compared with the browser view, old Reddit is the opposite in this environment, and Instagram exposes enough SSR payload to make static more expensive than the rendered shell.
• Shell/challenge classification is required because exact overlap and agent score can look deceptively good when both static and reference snapshots are nearly empty.
• AP News and Ars Technica were tested as additional candidates but omitted from compare:static:diverse because the reference snapshot timed out in this environment. Reuters returned HTTP 401 from plain fetch and is also omitted from the automated diverse set.
• npmjs.com currently serves a Cloudflare challenge in the sample environment. The baseline is useful as a challenge-page fixture, not as a package-page content fixture.

• Improve custom-element/shadow-host pruning without losing useful selector targets.
• Add explicit benchmark gates for actionable and navigation recall once a stable target set is chosen.
• Compare browser and static extraction side-by-side on the same target set to decide when the Worker-compatible path is good enough.
• Tune static pruning controls for hidden menus, select/options, and repeated template regions against the diverse benchmark set.
• Support authenticated/cached sessions for npmjs.com if the real npm package page remains useful as a target.
• Add more real WebView smoke tests once Android/iOS host projects exist.