Add background claim investigation actions

Author: devjoe

docs/plans/general-page-reader.md

@@ -8,7 +8,7 @@ session-only multi-page switching are implemented on this branch. The
 live-DOM review has been summarized in
 `general-page-reader-quality-findings-2026-07-03-live-dom.md`. Merge-readiness evidence is indexed in
 `general-page-reader-merge-readiness.md`.
-Last updated: 2026-07-14
+Last updated: 2026-07-16
 
 ## Decision
 
@@ -714,21 +714,37 @@ next candidate gate closed; no fresh holdout should be created yet.
 
 ### Phase 4: Session-only Claim Investigation
 
-- Status: implementation and fail-closed contract completed on the feature
-  branch, but candidates v1 and v2 did not clear their private holdout gates and
-  the v3 development probe did not clear the coverage/runtime-stability boundary.
+- Status: the session-only prepared-action vertical slice and its fail-closed
+  contract are implemented on the feature branch. Candidates v1 and v2 did not
+  clear their private holdout gates and the v3 development probe did not clear
+  the coverage/runtime-stability boundary, so this remains unreleased.
 - A grounded `claims.q` is preferred; a bounded natural-question fallback from
   `claim.c + claim.need` is used only when the model question is missing or
   locally rejected. URLs, domains, search-engine instructions, vague references,
   and likely compound claims fail closed instead of bypassing the guard.
-- `查核選項` only expands a bounded, session-only intent in the current Page or
-  Focus scope. It does not start the future Truly Agent, open a tab, send
-  another model request, persist history, or assign a verdict.
-- The intent is compiled into two distinct external payloads: concise claim and
-  source keywords for standard Google Search, and a natural-language evidence
-  request for Google AI Mode. Copy and original-source actions remain explicit.
-  Page navigation, reread, a new analysis key, and a new Focus target clear
-  stale task state.
+- When a completed Page or Focus reading contains a candidate claim, the
+  service worker schedules one lower-priority `derived` adapter request. The
+  claim stays visible with a compact preparing status, then changes directly to
+  the prepared question and actions. Abstention, malformed output, a stale
+  analysis key, or an unavailable side panel quietly falls back to the original
+  claim. This preparation is ephemeral and never creates durable history.
+- Model work shares one resource-aware scheduler: explicit user work is
+  `user_blocking`, current reading is `foreground`, prepared actions are
+  `derived`, and speculative work is `prefetch`. Each model resource executes
+  one request at a time; deduplication, supersession, and a bounded foreground
+  burst keep Page preparation from starving Feed work without increasing the
+  number of model calls.
+- A prepared intent exposes three explicit actions: standard Google Search,
+  Google AI Mode, and copy. Standard Search receives concise claim/source
+  keywords only. AI Mode receives a natural-language evidence request and may
+  receive the current HTTP(S) URL as metadata; the URL is never treated as
+  evidence or copied into model output. The former original-source action is
+  intentionally absent because it duplicated the page the user is already on.
+- The adapter requires an exact `sourceQuote` grounding span, preserves source
+  language for the atomic claim/question, tolerates harmless schema-version and
+  optional-attribution drift, and applies the existing deterministic eligibility
+  guard after model output. Page navigation, reread, a new analysis key, and a
+  new Focus target clear stale task state; Page and Focus keep separate slots.
 - The future Truly Agent uses a separate non-runtime semantic Case draft. The
   model selects document families, source roles, authority hints, and numbered
   question coverage; local code owns IDs, question linkage, verification
@@ -738,14 +754,23 @@ next candidate gate closed; no fresh holdout should be created yet.
 
 ### Phase 5: Runtime and UX Gate
 
-- Status: implementation verification completed; product-quality holdout gates
-  failed for candidates v1 and v2, while v3 remains a development-only probe,
-  so the investigation action remains unreleased.
-- Focused unit coverage validates query sanitization, deterministic fallback,
-  fail-closed eligibility, Page/Focus state isolation, and the two-step UI.
-- The CDP UI audit validates that preparing a task opens no browser target and
-  captures the expanded card at 430px alongside Page, Focus, screenshot
-  recovery, loading, and ready states.
+- Status: implementation and live UX verification completed; product-quality
+  holdout gates failed for candidates v1 and v2, while v3 remains a
+  development-only probe, so the investigation action remains unreleased.
+- Focused unit coverage validates scheduler priority/fairness, adapter parsing
+  and grounding, query sanitization, deterministic fallback, fail-closed
+  eligibility, Page/Focus race isolation, and the automatic
+  preparing-to-ready/fallback transitions. The 180 ms height/fade transition is
+  skipped for reduced motion and never delays the underlying session update.
+- A 2026-07-16 no-focus CDP check used dev build
+  `1784200373031-0ae1017-dirty` on a real Financial Times page. It observed an
+  automatic `preparing -> ready` transition with no manual click, no redundant
+  label, and Google Search / Gemini / Copy actions at 430 px. A separate live
+  run exercised `preparing -> fallback`, confirming that unavailable model
+  output clears the loading state and restores the original claim.
+- Review screenshots remain local-only:
+  `/private/tmp/truly-auto-investigation-preparing-430-2026-07-16.png` and
+  `/private/tmp/truly-auto-investigation-ready-430-2026-07-16.png`.
 - Real-content paired audit artifacts remain private under `tmp/`; only
   anonymized aggregate findings may be copied into tracked documentation.
 

package.json

@@ -95,7 +95,7 @@
     "audit:general-page-model-integration": "vitest run tests/audit/general-page-model-integration-audit.test.ts",
     "audit:release-bundle": "node scripts/audit-release-bundle.mjs --dist dist",
     "test:contract:public": "vitest run tests/contract/current-region-targeting-contract.test.ts tests/contract/general-page-analysis-contract.test.ts tests/contract/general-page-extraction-contract.test.ts tests/contract/general-page-model-context-contract.test.ts tests/contract/general-page-parser-advisor-contract.test.ts tests/contract/model-response-contract.test.ts tests/contract/post-context-boundary-contract.test.ts tests/contract/reading-action-contract.test.ts tests/contract/zhtw-segmentation-contract.test.ts",
-    "test:gpr": "vitest run tests/audit/general-page-model-integration-audit.test.ts tests/contract/current-region-targeting-contract.test.ts tests/contract/general-page-analysis-contract.test.ts tests/contract/general-page-extraction-contract.test.ts tests/contract/general-page-model-context-contract.test.ts tests/contract/general-page-parser-advisor-contract.test.ts tests/contract/model-response-contract.test.ts tests/contract/reading-action-contract.test.ts tests/unit/cdp-client.test.mjs tests/unit/web-focus-continuity-scenario.test.mjs tests/unit/meaningful-navigation-scenario.test.mjs tests/unit/dev-build-freshness.test.mjs tests/unit/general-page-audit-runtime-reload.test.mjs tests/unit/general-page-host-permission.test.ts tests/unit/general-page-real-world-sanitizer.test.mjs tests/unit/private-general-page-eval.test.mjs tests/unit/i18n.test.ts tests/unit/i18n-hardcoded-strings.test.ts tests/unit/page-reader-content-script.test.ts tests/unit/page-reader-tab-transport.test.ts tests/unit/page-reading-analysis-coordinator.test.ts tests/unit/page-reading-export.test.ts tests/unit/page-reading-presentation.test.ts tests/unit/page-reading-runtime.test.ts tests/unit/page-reading-session.test.ts tests/unit/page-readability.test.ts tests/unit/page-url-identity.test.ts tests/unit/reading-brief-question-list.test.ts tests/unit/reading-command-envelope.test.ts tests/unit/runtime-message.test.ts tests/unit/screenshot-data-url.test.ts",
+    "test:gpr": "vitest run tests/audit/general-page-model-integration-audit.test.ts tests/contract/current-region-targeting-contract.test.ts tests/contract/general-page-analysis-contract.test.ts tests/contract/general-page-extraction-contract.test.ts tests/contract/general-page-model-context-contract.test.ts tests/contract/general-page-parser-advisor-contract.test.ts tests/contract/model-response-contract.test.ts tests/contract/reading-action-contract.test.ts tests/unit/cdp-client.test.mjs tests/unit/web-focus-continuity-scenario.test.mjs tests/unit/meaningful-navigation-scenario.test.mjs tests/unit/dev-build-freshness.test.mjs tests/unit/general-page-audit-runtime-reload.test.mjs tests/unit/general-page-host-permission.test.ts tests/unit/general-page-investigation-adapter.test.ts tests/unit/general-page-investigation-background.test.ts tests/unit/general-page-real-world-sanitizer.test.mjs tests/unit/model-work-scheduler.test.ts tests/unit/private-general-page-eval.test.mjs tests/unit/i18n.test.ts tests/unit/i18n-hardcoded-strings.test.ts tests/unit/page-claim-investigation.test.ts tests/unit/page-reader-content-script.test.ts tests/unit/page-reader-tab-transport.test.ts tests/unit/page-reading-analysis-coordinator.test.ts tests/unit/page-reading-export.test.ts tests/unit/page-reading-presentation.test.ts tests/unit/page-reading-runtime.test.ts tests/unit/page-reading-session.test.ts tests/unit/page-readability.test.ts tests/unit/page-url-identity.test.ts tests/unit/reading-brief-question-list.test.ts tests/unit/reading-command-envelope.test.ts tests/unit/runtime-message-router.test.ts tests/unit/runtime-message.test.ts tests/unit/screenshot-data-url.test.ts",
     "test:unit:public": "vitest run tests/unit/cdp-client.test.mjs tests/unit/web-focus-continuity-scenario.test.mjs tests/unit/meaningful-navigation-scenario.test.mjs tests/unit/cdp-page-source.test.mjs tests/unit/dev-build-freshness.test.mjs tests/unit/feed-boundary.test.ts tests/unit/general-page-audit-runtime-reload.test.mjs tests/unit/general-page-host-permission.test.ts tests/unit/general-page-real-world-sanitizer.test.mjs tests/unit/private-general-page-eval.test.mjs tests/unit/heads-up-chip-policy.test.ts tests/unit/heads-up-panel-status.test.ts tests/unit/i18n.test.ts tests/unit/i18n-hardcoded-strings.test.ts tests/unit/investigation-actions-renderer.test.ts tests/unit/logging-policy.test.ts tests/unit/model-display.test.ts tests/unit/model-source-config.test.ts tests/unit/ollama-cloud-capabilities.test.ts tests/unit/openai-api-key-mock.test.ts tests/unit/page-reader-content-script.test.ts tests/unit/page-reader-tab-transport.test.ts tests/unit/page-reading-analysis-coordinator.test.ts tests/unit/page-reading-export.test.ts tests/unit/page-reading-presentation.test.ts tests/unit/page-reading-runtime.test.ts tests/unit/page-reading-session.test.ts tests/unit/page-readability.test.ts tests/unit/page-url-identity.test.ts tests/unit/provider-capabilities.test.ts tests/unit/reading-brief-question-list.test.ts tests/unit/reading-command-envelope.test.ts tests/unit/reading-question-policy.test.ts tests/unit/request-auth.test.ts tests/unit/runtime-message.test.ts tests/unit/screenshot-data-url.test.ts tests/unit/snapshot-redaction.test.ts tests/unit/sponsored-memory.test.ts tests/unit/ssr-cache-bridge.test.ts tests/unit/tabs.test.ts tests/unit/theme-mode.test.ts tests/unit/tier-b-vision-probe.test.ts tests/unit/trusted-model-runtime.test.ts",
     "test:unit:watch": "vitest",
     "check:public": "npm run check:public-boundary && npm run check:release-metadata && npm run check:general-page && npm run check:type && npm run test:contract:public && npm run test:unit:public && npm run test:gpr:investigation && npm run build && npm run audit:release-bundle",

scripts/audit-general-page-reader.mjs

@@ -328,6 +328,8 @@ async function startMockOpenAiEndpoint() {
     const hasImageUrl = JSON.stringify(userContent).includes('"image_url"');
     const kind = /parser recovery classifier/i.test(systemText)
       ? "parser-advisor"
+      : /prepare one candidate fact-check action/i.test(systemText)
+      ? "investigation-adapter"
       : hasImageUrl
       ? "screenshot-brief"
       : /dominant color/i.test(systemText)
@@ -355,6 +357,20 @@ async function startMockOpenAiEndpoint() {
         riskTags: ["needs_visual_grounding"],
         rationale: "The synthetic fixture needs visible screenshot grounding.",
       });
+    } else if (kind === "investigation-adapter") {
+      content = JSON.stringify({
+        schemaVersion: 1,
+        decision: "prepared",
+        reason: "actionable",
+        claim: {
+          c: "The analyzed content is synthetic.",
+          why: "The UI check must not depend on live page content.",
+          need: "Confirm the expected scope.",
+          q: "Is the analyzed content synthetic?",
+          atom: { s: "The analyzed content", p: "is", o: "synthetic" },
+          policy: { claimKind: "fact", consequence: "public_interest" },
+        },
+      });
     } else {
       const targetKind = /targetKind:\s*selection/i.test(userText)
         ? "selection"
@@ -1458,6 +1474,12 @@ async function auditSuccessfulRead(extensionId, allowedBase) {
     })()`);
 
     const pageBrief = await observePageBrief(side, "page-analysis-ready.png");
+    await waitFor(
+      side,
+      `Boolean(document.querySelector('#page-pane .page-claim-start'))`,
+      5000,
+      "background claim investigation preparation",
+    );
     const claimInvestigation = await observeClaimInvestigation(side);
     const initialLoadTimeline = await side.evaluateJson(`(() => {
       const timeline = globalThis.__trulyPagePaneTimeline;
@@ -3068,7 +3090,10 @@ function claimActionPayloadContract(result) {
         aiModeUrl.searchParams.get("udm") === "50" &&
         standardQuery && aiModePrompt && standardQuery !== aiModePrompt &&
         !/Please verify this claim|請協助查核以下說法/u.test(standardQuery) &&
-        /Evidence needed|需要的證據/u.test(aiModePrompt)
+        !/https?:\/\//u.test(standardQuery) &&
+        /Evidence needed|需要的證據/u.test(aiModePrompt) &&
+        /Source URL \(metadata\)|來源網址（metadata）/u.test(aiModePrompt) &&
+        /127\.0\.0\.1/u.test(aiModePrompt)
       ),
       standardQueryLength: standardQuery.length,
       aiModePromptLength: aiModePrompt.length,

scripts/lib/private-general-page-eval.mjs

@@ -2,6 +2,12 @@ import fs from "node:fs";
 import path from "node:path";
 
 export const PRIVATE_EVAL_SURFACES = ["facebook", "news"];
+const SOURCE_CONTEXT_LIMITS = {
+  title: 100,
+  sourceName: 60,
+  publishedAt: 32,
+};
+const SOURCE_CONTEXT_ARTIFACT_RE = /https?:\/\/|\[[^\]]+\]\([^\)]+\)|(?:^|\s)(?:curl|wget|npm|pnpm|brew|git)\s/i;
 
 export function parsePrivateEvalJsonl(text) {
   return String(text)
@@ -45,6 +51,31 @@ export function privateEvalInputErrors(rows, expectedCount, declaredCategories)
     if (!['zh-TW', 'en'].includes(row.language)) errors.push(`${label}: invalid language`);
     if (typeof row.text !== "string" || row.text.trim().length < 80 || row.text.length > 12000) errors.push(`${label}: text must be 80-12000 characters`);
     if (typeof row.sourceSha256 !== "string" || !/^[a-f0-9]{64}$/.test(row.sourceSha256)) errors.push(`${label}: invalid sourceSha256`);
+    if (row.sourceContext !== undefined) {
+      if (!row.sourceContext || typeof row.sourceContext !== "object" || Array.isArray(row.sourceContext)) {
+        errors.push(`${label}: sourceContext must be an object`);
+      } else {
+        for (const [key, limit] of Object.entries(SOURCE_CONTEXT_LIMITS)) {
+          const value = row.sourceContext[key];
+          if (value === undefined) continue;
+          if (typeof value !== "string" || !value.trim() || value.length > limit || SOURCE_CONTEXT_ARTIFACT_RE.test(value)) {
+            errors.push(`${label}: invalid sourceContext.${key}`);
+          }
+        }
+        if (row.sourceContext.url !== undefined) {
+          try {
+            const url = new URL(row.sourceContext.url);
+            if (!/^https?:$/.test(url.protocol) || url.username || url.password || row.sourceContext.url.length > 320) {
+              errors.push(`${label}: invalid sourceContext.url`);
+            }
+          } catch {
+            errors.push(`${label}: invalid sourceContext.url`);
+          }
+        }
+        const unknownKeys = Object.keys(row.sourceContext).filter((key) => !(key in SOURCE_CONTEXT_LIMITS) && key !== "url");
+        if (unknownKeys.length > 0) errors.push(`${label}: unsupported sourceContext fields`);
+      }
+    }
   }
   if ([...actual].some((category) => !declared.has(category)) || [...declared].some((category) => !actual.has(category))) {
     errors.push(`data categories mismatch: declared ${[...declared].sort().join(",")}; actual ${[...actual].sort().join(",")}`);

scripts/private-general-page-eval-entry.ts

@@ -27,6 +27,12 @@ interface InputRow {
   language: "zh-TW" | "en";
   sourceSha256: string;
   text: string;
+  sourceContext?: {
+    title?: string;
+    sourceName?: string;
+    publishedAt?: string;
+    url?: string;
+  };
 }
 
 function option(name: string, fallback?: string): string | undefined {
@@ -133,6 +139,7 @@ async function evaluateRow(row: InputRow) {
       claimIndex: 0,
       claim,
       groundingText: row.text,
+      source: row.sourceContext,
     }) : undefined;
     return {
       schemaVersion: 1,
@@ -149,6 +156,8 @@ async function evaluateRow(row: InputRow) {
         eligibilityReason: eligibility && !eligibility.ok ? eligibility.reason : undefined,
         questionSource: task ? (modelQuestion ? "model" : "deterministic_fallback") : "none",
         question: task?.intent.question,
+        googleKeywords: task?.googleKeywords,
+        aiModePrompt: task?.aiModePrompt,
       },
       raw: response.raw,
     };