Stop reader mode passing a site's cookie notice off as the article

Reported live: reader mode on a New Republic feature showed nothing but
the cookie disclaimer, and had cached it.

Reproduced outside the app by running ReaderExtractor's own injected
script (Readability.js and all) against the real page. With the article
body present, extraction is perfect. With the body absent from the DOM at
the moment the script runs, Readability's first pass -- which discards
"unlikely candidates", and the Didomi banner's class is
didomi-notice-banner -- comes back under its own 500-character threshold,
so it retries with that stripping switched off. The banner's single
paragraph is 510 characters, i.e. just over the threshold, so it comes
back as the article with the right title attached.

Nothing on the Kotlin side looked at that. Any non-blank content counted
as ExtractionResult.Success, was written to articles.extractedContent,
and was never reconsidered -- so one bad load became a permanent cookie
disclaimer, which is the "extractedContent cached: true" in the report.

ConsentNotice recognises such a result: a hard plain-text length cap
first (nothing long is ever rejected, however much consent boilerplate it
quotes), then either a known CMP container id -- Readability strips
classes but keeps ids -- or two or more distinct consent phrases.

The accept/retry/give-up decision moves out of the WebView callback into
decideEvalOutcome so it can be tested without a device. A consent notice
is now "not ready yet", which routes it into the retry the extractor
already does for an empty result; if it survives all six attempts that is
what the page serves, and it fails as ExtractionFailure.ConsentInterstitial
rather than being accepted.

A fixed extractor cannot repair what is already on disk, so
loadCachedExtraction drops a cached notice on read and re-extracts, in
both ReaderScreen and LinkReaderScreen. No schema change: the repair
happens when the article is next opened. Reader mode is still never
turned on heuristically -- replacing a discarded notice does not flip it.

The bug report now also carries the last extraction error, without which
a report about bad reader content cannot say whether extraction failed.

Tests: 59 -> 78.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018AwnALZL6bD4VBtATDVv3B

7 files changed, 284 insertions, 25 deletions

diff --git a/app/src/main/java/org/terst/feedreader/data/CachedExtraction.kt b/app/src/main/java/org/terst/feedreader/data/CachedExtraction.kt
new file mode 100644
index 0000000..313d467
--- /dev/null
+++ b/app/src/main/java/org/terst/feedreader/data/CachedExtraction.kt
@@ -0,0 +1,41 @@
+package org.terst.feedreader.data
+
+import org.terst.feedreader.readability.ConsentNotice
+
+/** What a cached `articles.extractedContent` value turned out to be worth. */
+data class CachedExtraction(
+    /** The content to show, or null if there is nothing usable. */
+    val content: String?,
+    /**
+     * True when a cached consent notice was thrown away. The caller has nothing to
+     * show and should extract again -- the cache is the reason the article looked
+     * broken, so dropping it silently would just leave a blank reader.
+     */
+    val discardedConsentNotice: Boolean
+)
+
+/**
+ * Reads a cached extraction, dropping it if it is the site's cookie/consent notice
+ * rather than the article.
+ *
+ * Needed because the extractor's own guard can only keep a bad extraction from
+ * being written in future. The reported bug ("reader mode just extracts cookie
+ * disclaimer", with `extractedContent cached: true`) was already on disk: every
+ * subsequent open read it straight back out, so no amount of fixing the extractor
+ * would have changed what the user saw. A poisoned row is cleared here rather than
+ * in a schema migration, so it is repaired whenever the article is next opened --
+ * including for anything extracted before this check existed.
+ *
+ * [clearCache] is passed in rather than taking the DAO so this stays a plain
+ * decision that can be tested without a database.
+ */
+suspend fun loadCachedExtraction(
+    cached: String?,
+    clearCache: suspend () -> Unit
+): CachedExtraction {
+    if (!ConsentNotice.looksLikeConsentNotice(cached)) {
+        return CachedExtraction(content = cached, discardedConsentNotice = false)
+    }
+    clearCache()
+    return CachedExtraction(content = null, discardedConsentNotice = true)
+}
diff --git a/app/src/main/java/org/terst/feedreader/readability/ConsentNotice.kt b/app/src/main/java/org/terst/feedreader/readability/ConsentNotice.kt
new file mode 100644
index 0000000..5699595
--- /dev/null
+++ b/app/src/main/java/org/terst/feedreader/readability/ConsentNotice.kt
@@ -0,0 +1,132 @@
+package org.terst.feedreader.readability
+
+/**
+ * Recognises an extraction result that is the page's cookie/consent notice rather
+ * than its article.
+ *
+ * Why this is needed at all: Readability's first pass discards "unlikely
+ * candidates" by class/id, which a consent banner normally is. But when that pass
+ * comes back under Readability's own 500-character threshold -- which is what
+ * happens whenever the real article body is not in the DOM at the moment the
+ * script runs -- it retries with that stripping switched OFF. A consent banner is
+ * then a perfectly ordinary candidate, and the ones sites actually ship are long
+ * enough to clear the threshold (the reported New Republic/Didomi notice is 510
+ * characters). Readability hands it back as the article, with the right title
+ * attached, and nothing downstream could tell the difference.
+ *
+ * The bar for rejecting something has to be high: throwing away a real article is
+ * worse than showing a cookie notice. So two independent, deliberately narrow
+ * signals, and a hard length cap on both -- a real article body is far longer than
+ * any banner, so nothing long is ever rejected no matter what it says.
+ */
+object ConsentNotice {
+
+    /**
+     * Plain-text length past which a result is never treated as a banner, however
+     * much consent boilerplate it contains. Readability's own `charThreshold` is
+     * 500; this leaves generous room above it for the longest real banners
+     * (multi-paragraph "we and our 847 partners" notices) while staying far below
+     * any actual article.
+     */
+    private const val MAX_NOTICE_TEXT_LENGTH = 2_000
+
+    /** How many distinct consent phrases a result needs before wording alone convicts it. */
+    private const val MIN_PHRASE_MATCHES = 2
+
+    /**
+     * Container ids belonging to consent-management platforms. Readability strips
+     * class attributes but preserves ids, so the CMP's own container id survives
+     * into the extracted HTML -- the most precise signal available, since no real
+     * article body is ever inside one of these.
+     */
+    private val CMP_CONTAINER_IDS = listOf(
+        "didomi-host",
+        "didomi-notice",
+        "didomi-popup",
+        "onetrust-banner-sdk",
+        "onetrust-consent-sdk",
+        "onetrust-policy-text",
+        "optanon-alert-box-wrapper",
+        "cybotcookiebotdialog",
+        "qc-cmp2-container",
+        "sp_message_container",
+        "truste-consent-track",
+        "usercentrics-root",
+        "usercentrics-cmp-ui",
+        "osano-cm-dialog",
+        "iubenda-cs-banner",
+        "cookie-law-info-bar",
+        "cookie-notice",
+        "cookie-banner",
+        "cookieconsent",
+        "cmpbox",
+    )
+
+    /**
+     * Phrases that only really occur in consent boilerplate. Individually weak --
+     * an article can mention any one of them -- so [MIN_PHRASE_MATCHES] are
+     * required, and only within the length cap.
+     */
+    private val CONSENT_PHRASES = listOf(
+        "uses cookies",
+        "use cookies",
+        "using cookies",
+        "tracking technologies",
+        "by continuing to use this",
+        "by continuing to browse",
+        "accept all cookies",
+        "reject all",
+        "manage your preferences",
+        "manage preferences",
+        "cookie preferences",
+        "cookie policy",
+        "cookie settings",
+        "we and our partners",
+        "our advertising and analytics",
+        "your privacy choices",
+        "consent to our use of cookies",
+        "personalise content and ads",
+        "personalize content and ads",
+    )
+
+    /**
+     * True when [html] is a consent/cookie notice rather than article content.
+     *
+     * Takes the stored HTML rather than Readability's `textContent` so the same
+     * check works on a value already cached in `articles.extractedContent`, where
+     * the text form was never kept.
+     */
+    fun looksLikeConsentNotice(html: String?): Boolean {
+        if (html.isNullOrBlank()) return false
+
+        // Length is judged on the text, not the markup: a banner wrapped in enough
+        // nested divs and inline styles clears 2KB of characters while still being
+        // three sentences of text. Checked first, so nothing long is ever rejected
+        // however much consent boilerplate it happens to quote.
+        val text = plainText(html)
+        val withinNoticeLength = text.isNotBlank() && text.length <= MAX_NOTICE_TEXT_LENGTH
+
+        return withinNoticeLength && (hasCmpContainer(html) || readsAsConsentBoilerplate(text))
+    }
+
+    private fun hasCmpContainer(html: String): Boolean {
+        val markup = html.lowercase()
+        return CMP_CONTAINER_IDS.any { markup.contains("id=\"$it") || markup.contains("id='$it") }
+    }
+
+    private fun readsAsConsentBoilerplate(text: String): Boolean {
+        val lowerText = text.lowercase()
+        return CONSENT_PHRASES.count { lowerText.contains(it) } >= MIN_PHRASE_MATCHES
+    }
+
+    /**
+     * Crude tag strip -- enough to measure length and match phrases. Deliberately
+     * not a parser: this runs on every article open, and getting the character
+     * count roughly right is all the decision needs.
+     */
+    private fun plainText(html: String): String =
+        html.replace(TAG_REGEX, " ").replace(WHITESPACE_REGEX, " ").trim()
+
+    private val TAG_REGEX = Regex("<[^>]*>")
+    private val WHITESPACE_REGEX = Regex("\\s+")
+}
diff --git a/app/src/main/java/org/terst/feedreader/readability/EvalDecision.kt b/app/src/main/java/org/terst/feedreader/readability/EvalDecision.kt
new file mode 100644
index 0000000..ed0033d
--- /dev/null
+++ b/app/src/main/java/org/terst/feedreader/readability/EvalDecision.kt
@@ -0,0 +1,61 @@
+package org.terst.feedreader.readability
+
+/** What to do with one `evaluateJavascript` result. */
+sealed interface EvalDecision {
+    /** Nothing more to try: this is the extraction's outcome. */
+    data class Done(val result: ExtractionResult) : EvalDecision
+
+    /** The page probably was not ready; inject the script again. */
+    data object Retry : EvalDecision
+}
+
+/**
+ * Decides, for one script result, whether extraction succeeded, should be
+ * retried, or has failed and why.
+ *
+ * Lifted out of the WebView callback it used to live inside so it can be
+ * exercised without a device. The bug that prompted that: a result consisting
+ * only of the page's cookie-consent banner was non-blank and error-free, so it
+ * counted as a success, was cached to `articles.extractedContent`, and was never
+ * reconsidered. Reported live as "reader mode just extracts cookie disclaimer".
+ */
+internal fun decideEvalOutcome(
+    article: ExtractedArticle?,
+    rawResult: String,
+    attempt: Int,
+    maxAttempts: Int
+): EvalDecision {
+    val readabilityError = article?.error?.takeIf { it.isNotBlank() }
+    // A banner is exactly what a half-loaded page produces -- the consent script
+    // runs early, the article body may not be in the DOM yet -- so this is not
+    // usable, which routes it into the same retry the extractor already does for
+    // an empty result rather than being accepted.
+    val isConsentNotice = article != null &&
+        readabilityError == null &&
+        ConsentNotice.looksLikeConsentNotice(article.content)
+    val usable = article != null &&
+        readabilityError == null &&
+        !article.content.isNullOrBlank() &&
+        !isConsentNotice
+
+    return when {
+        usable -> EvalDecision.Done(ExtractionResult.Success(article!!))
+
+        // Readability ran and objected: that verdict will not change.
+        readabilityError != null ->
+            EvalDecision.Done(ExtractionResult.Failure(ExtractionFailure.ReadabilityError(readabilityError)))
+
+        attempt < maxAttempts -> EvalDecision.Retry
+
+        // Every attempt came back with nothing but the consent notice, so this is
+        // what the page actually serves -- report it as its own failure rather
+        // than passing a cookie disclaimer off as the article.
+        isConsentNotice ->
+            EvalDecision.Done(ExtractionResult.Failure(ExtractionFailure.ConsentInterstitial))
+
+        article == null ->
+            EvalDecision.Done(ExtractionResult.Failure(ExtractionFailure.BadResult(rawResult)))
+
+        else -> EvalDecision.Done(ExtractionResult.Failure(ExtractionFailure.NoContent))
+    }
+}
diff --git a/app/src/main/java/org/terst/feedreader/readability/ExtractionFailure.kt b/app/src/main/java/org/terst/feedreader/readability/ExtractionFailure.kt
index 532b109..63fbdb2 100644
--- a/app/src/main/java/org/terst/feedreader/readability/ExtractionFailure.kt
+++ b/app/src/main/java/org/terst/feedreader/readability/ExtractionFailure.kt
@@ -28,6 +28,12 @@ sealed interface ExtractionFailure {
 
     /** The JS result could not be decoded. */
     data class BadResult(val raw: String) : ExtractionFailure
+
+    /**
+     * Readability returned content, but it was the page's cookie/consent notice,
+     * not the article -- see [ConsentNotice].
+     */
+    data object ConsentInterstitial : ExtractionFailure
 }
 
 private const val MILLIS_PER_SECOND = 1_000
@@ -47,6 +53,7 @@ fun ExtractionFailure.userMessage(): String = when (this) {
     is ExtractionFailure.ReadabilityError -> "Couldn't parse the article ($message)"
     ExtractionFailure.NoContent -> "No article text found on the page"
     is ExtractionFailure.BadResult -> "Unexpected result from the page"
+    ExtractionFailure.ConsentInterstitial -> "Only the site's cookie notice came through, not the article"
 }
 
 /**
@@ -61,4 +68,6 @@ fun ExtractionFailure.diagnostic(): String = when (this) {
     is ExtractionFailure.ReadabilityError -> "ReadabilityError $message"
     ExtractionFailure.NoContent -> "NoContent (Readability returned no usable body)"
     is ExtractionFailure.BadResult -> "BadResult raw=${raw.take(RAW_DIAGNOSTIC_CHARS)}"
+    ExtractionFailure.ConsentInterstitial ->
+        "ConsentInterstitial (extracted content was the page's cookie/consent notice)"
 }
diff --git a/app/src/main/java/org/terst/feedreader/readability/ReaderExtractor.kt b/app/src/main/java/org/terst/feedreader/readability/ReaderExtractor.kt
index 5061444..21633b4 100644
--- a/app/src/main/java/org/terst/feedreader/readability/ReaderExtractor.kt
+++ b/app/src/main/java/org/terst/feedreader/readability/ReaderExtractor.kt
@@ -436,10 +436,9 @@ object ReaderExtractor {
                  * Runs the extraction script, retrying while the result looks like
                  * the page simply was not ready yet.
                  *
-                 * A Readability error is final -- retrying will not change what the
-                 * parser thinks of the markup. A null result or empty content is
-                 * not: both are what an injection against a half-loaded document
-                 * produces, and the next attempt usually succeeds.
+                 * The accept/retry/give-up decision itself lives in
+                 * [decideEvalOutcome], which is testable without a WebView; this
+                 * only decodes the payload and carries the decision out.
                  */
                 private fun evaluate(view: WebView, script: String, attempt: Int) {
                     if (finished) return
@@ -451,24 +450,18 @@ object ReaderExtractor {
                         val article = unwrapped?.let {
                             runCatching { json.decodeFromString<ExtractedArticle>(it) }.getOrNull()
                         }
-                        val usable = article != null &&
-                            article.error.isNullOrBlank() &&
-                            !article.content.isNullOrBlank()
 
-                        when {
-                            usable -> finish(ExtractionResult.Success(article!!))
-
-                            // Readability ran and objected: that verdict will not change.
-                            article != null && !article.error.isNullOrBlank() ->
-                                fail(ExtractionFailure.ReadabilityError(article.error))
-
-                            attempt < MAX_EVAL_ATTEMPTS ->
+                        when (
+                            val decision = decideEvalOutcome(
+                                article = article,
+                                rawResult = rawResult ?: "null",
+                                attempt = attempt,
+                                maxAttempts = MAX_EVAL_ATTEMPTS
+                            )
+                        ) {
+                            is EvalDecision.Done -> finish(decision.result)
+                            EvalDecision.Retry ->
                                 view.postDelayed({ evaluate(view, script, attempt + 1) }, EVAL_RETRY_DELAY_MS)
-
-                            article == null ->
-                                fail(ExtractionFailure.BadResult(rawResult ?: "null"))
-
-                            else -> fail(ExtractionFailure.NoContent)
                         }
                     }
                 }
diff --git a/app/src/main/java/org/terst/feedreader/ui/LinkReaderScreen.kt b/app/src/main/java/org/terst/feedreader/ui/LinkReaderScreen.kt
index 4ed0722..72f6c1f 100644
--- a/app/src/main/java/org/terst/feedreader/ui/LinkReaderScreen.kt
+++ b/app/src/main/java/org/terst/feedreader/ui/LinkReaderScreen.kt
@@ -51,6 +51,7 @@ import org.terst.feedreader.data.DEFAULT_READER_FONT_SIZE_PX
 import org.terst.feedreader.data.Settings
 import org.terst.feedreader.data.ensureStandaloneArticle
 import org.terst.feedreader.data.saveArticleToStash
+import org.terst.feedreader.data.loadCachedExtraction
 import org.terst.feedreader.data.urlToArticleId
 import org.terst.feedreader.readability.ReaderExtractor
 import java.net.URLEncoder
@@ -95,9 +96,15 @@ fun LinkReaderScreen(url: String, onBack: () -> Unit, onOpenLink: (String) -> Un
     LaunchedEffect(url) {
         val existing = db.articleDao().getById(articleId)
         article = existing
-        if (existing?.extractedContent != null) {
-            extractedTitle = existing.title
-            extractedContent = existing.extractedContent
+        // A cached extraction can be the site's cookie-consent notice rather than the
+        // article (see CachedExtraction) -- drop it and extract again instead of showing
+        // a cookie disclaimer as the page.
+        val cached = loadCachedExtraction(existing?.extractedContent) {
+            db.articleDao().clearExtractedContent(articleId)
+        }
+        if (cached.content != null) {
+            extractedTitle = existing?.title
+            extractedContent = cached.content
         } else {
             val extracted = ReaderExtractor.extract(context, url)
             val content = extracted?.content
diff --git a/app/src/main/java/org/terst/feedreader/ui/ReaderScreen.kt b/app/src/main/java/org/terst/feedreader/ui/ReaderScreen.kt
index 6240b77..f08088b 100644
--- a/app/src/main/java/org/terst/feedreader/ui/ReaderScreen.kt
+++ b/app/src/main/java/org/terst/feedreader/ui/ReaderScreen.kt
@@ -68,6 +68,7 @@ import org.terst.feedreader.data.Settings
 import org.terst.feedreader.data.ensureStandaloneArticle
 import org.terst.feedreader.data.saveArticleToStash
 import org.terst.feedreader.data.setReaderModeEnabled
+import org.terst.feedreader.data.loadCachedExtraction
 import org.terst.feedreader.readability.ReaderExtractor
 import java.net.URLEncoder
 
@@ -219,15 +220,26 @@ private fun ReaderContent(
         feed = loadedFeed
 
         feedContent = loaded.feedContentHtml
-        readerContent = loaded.extractedContent
+        // A previously-cached extraction can be the site's cookie-consent notice rather
+        // than the article (reported live on a New Republic feature). That is dropped
+        // from the cache on read, which leaves nothing to show -- so re-extract, or the
+        // repair would just turn a wrong article into a blank one.
+        val cachedExtraction = loadCachedExtraction(loaded.extractedContent) {
+            db.articleDao().clearExtractedContent(articleId)
+        }
+        readerContent = cachedExtraction.content
 
         // Reader mode is NEVER heuristic/automatic (no more "content looks thin, switch
         // for you") -- it strictly follows the feed's persisted, explicitly-set opt-in.
         // Once a feed has it on, articles from it default to the extracted view and
         // auto-extract on open (that's the point of persisting it per-feed: no re-tapping
-        // per article), but nothing here ever turns it on by itself.
+        // per article), but nothing here ever turns it on by itself. Replacing a discarded
+        // consent notice does not turn it on either -- for a feed-less saved article there
+        // is no feed to opt in, and the reader falls back to the extracted content anyway.
         if (loadedFeed?.readerModeEnabled == true) {
             isReaderMode = true
+        }
+        if (loadedFeed?.readerModeEnabled == true || cachedExtraction.discardedConsentNotice) {
             if (readerContent == null) {
                 isExtracting = true
                 val result = ReaderExtractor.extractResult(context, loaded.link)
@@ -509,6 +521,10 @@ private fun ReaderContent(
                             appendLine("Session isReaderMode: $isReaderMode")
                             appendLine("isSaved: ${a?.isSaved}, stashArtifactId: ${a?.stashArtifactId}")
                             appendLine("extractedContent cached: ${readerContent != null}")
+                            // Without this, a report about bad/absent reader content could
+                            // not say whether extraction had failed, let alone how -- which
+                            // is exactly what the cookie-disclaimer report ran into.
+                            appendLine("Last extraction error: ${extractionError ?: "none"}")
                         }
                         val clipboard = context.getSystemService(Context.CLIPBOARD_SERVICE) as ClipboardManager
                         clipboard.setPrimaryClip(ClipData.newPlainText("Bug report", report))