@gobing-ai/knowledge-kit 0.0.29 → 0.0.30

package/plugins/kk/plugin.json

@@ -1,5 +1,5 @@
 {
     "name": "kk",
-    "version": "0.0.29",
+    "version": "0.0.30",
     "description": "knowledge-kit Claude Code plugin: skills, commands, subagents, hooks, and rules"
 }

package/plugins/kk/scripts/daily-voice-style.zh.md

@@ -25,10 +25,10 @@
 
 ## 段落角色
 
-- `intro`：开场导语，串起今天几个主题，接在固定问候之后，不超过 120 字。
-- `lead`：一句话口语桥段，引出下一个主题的【标题】；只做铺垫，不得复述标题本身，不超过 60 字。
-- `insight`：只基于这一条自身可核事实的 1–2 句点评，说明它为什么重要或怎么用得上；不得挪用其他条目的话题，不超过 160 字。
-- `outro`：简短收束语，接在固定致谢之前，不超过 120 字。
+- `intro`：开场导语，串起今天几个主题，接在固定问候之后，不超过 150 字。
+- `lead`：一句话口语桥段，引出下一个主题的【标题】；只做铺垫，不得复述标题本身，不超过 80 字。
+- `insight`：只基于这一条自身可核事实的点评（3–6 句），展开这一条自身的机制→意义→实用角度；不得挪用其他条目的话题，不超过 380 字。时长契约：全片按 15–20 分钟口播设计，insight 应写到接近上限——宁可深挖这一条的事实，也不要浅尝辄止。字数按配音驱动折算：gemini ≈3.3 字/秒（正文约 4300–5800 字）；Voicebox robin-news ≈6.3 字/秒（正文约 5700–7600 字，或增加条目数）。
+- `outro`：简短收束语，接在固定致谢之前，不超过 150 字。
 
 ## 机器可读契约
 
@@ -39,7 +39,7 @@
 
 ```json
 {
-  "caps": { "intro": 120, "outro": 120, "lead": 60, "insight": 160 },
+  "caps": { "intro": 150, "outro": 150, "lead": 80, "insight": 380 },
   "banned": [
     "在各种技术光环背后，正是这样专注而不知疲倦的底层攻坚者，为今天整个大模型生态奠定了最坚实的技术基石。",
     "这其实释放了一个非常明确的信号：AI竞争的下半场，已经不再只是单纯拼模型参数，更是电力、能源和全栈工业基础设施的硬核较量。",

package/plugins/kk/scripts/itc-stages.ts

@@ -131,7 +131,18 @@ export function itcReport(args: string[]): void {
     const [workDir = '', maxRevisions = ''] = args;
     if (!workDir) fail('usage: itc-report <work_dir> [max_revisions]');
     const verdictPath = join(workDir, '.itc-verdict.json');
-    const verdict = existsSync(verdictPath) ? ((readJson(verdictPath) as { verdict?: unknown }).verdict ?? '') : '';
+    const verdictJson = existsSync(verdictPath)
+        ? (readJson(verdictPath) as { verdict?: unknown; confidence?: unknown })
+        : {};
+    const verdict = verdictJson.verdict ?? '';
+    // Same enum as the workflow guards and `daily-safety-gate`: a PASS without a HIGH/MEDIUM/LOW
+    // confidence never cleared the judge gate, so the report must not claim a publish-ready draft.
+    // An absent verdict file stays the dry-run PASS convention (verdict '').
+    const unusablePass =
+        verdict === 'PASS' &&
+        verdictJson.confidence !== 'HIGH' &&
+        verdictJson.confidence !== 'MEDIUM' &&
+        verdictJson.confidence !== 'LOW';
     // Auto runs count on .auto-revise-count; interactive runs keep the existing .revise-count.
     const revisions = readCount(join(workDir, '.auto-revise-count')) ?? readCount(join(workDir, '.revise-count')) ?? 0;
     // Exhaustion only fails the run when the last verdict still asks for another revision —
@@ -186,7 +197,8 @@ export function itcReport(args: string[]): void {
     }
 
     const publishFailed = targets.some((t) => !t.ok);
-    const outcome = verdict === 'FAIL' || exhausted || publishFailed ? 'failed' : declined ? 'declined' : 'done';
+    const outcome =
+        verdict === 'FAIL' || unusablePass || exhausted || publishFailed ? 'failed' : declined ? 'declined' : 'done';
     const reportDir = join(workDir, '5-report');
     const jsonPath = join(reportDir, 'report.json');
     const publish = hasPlan ? { targets, retryCommands } : undefined;
@@ -198,7 +210,13 @@ export function itcReport(args: string[]): void {
         );
         writeFileSync(
             join(reportDir, 'report.md'),
-            `# kk-itc run report\n\n- Outcome: ${outcome}\n- Verdict: ${verdict === '' ? 'PASS (empty)' : verdict}\n- Revisions: ${revisions}\n` +
+            `# kk-itc run report\n\n- Outcome: ${outcome}\n- Verdict: ${
+                verdict === ''
+                    ? 'PASS (empty)'
+                    : unusablePass
+                      ? 'PASS (no HIGH/MEDIUM/LOW confidence — the judge gate never cleared)'
+                      : verdict
+            }\n- Revisions: ${revisions}\n` +
                 (declined
                     ? '\n## Publish\n\n- Declined by the operator — nothing was sent.\n'
                     : targets.length > 0

package/plugins/kk/scripts/kk-workflow-stages.ts

@@ -840,6 +840,7 @@ function dailyQcContent(args: string[]): void {
         minImportance = '',
         minUrgency = '',
         minImpact = '',
+        minEpisodeItems = '',
     ] = args;
     const candidates = join(workDir, '21-plan', `${runDate}_04_quality-control-content_candidates.json`);
     const annotated = join(workDir, '21-plan', `${runDate}_04_quality-control-content_annotated.json`);
@@ -870,6 +871,7 @@ function dailyQcContent(args: string[]): void {
                     QC_MIN_IMPORTANCE: minImportance,
                     QC_MIN_URGENCY: minUrgency,
                     QC_MIN_IMPACT: minImpact,
+                    QC_MIN_EPISODE_ITEMS: minEpisodeItems,
                 },
             },
         );
@@ -1524,16 +1526,21 @@ function safetyPath(workDir: string, runDate: string, name: string): string {
  */
 function safetyDigest(workDir: string, runDate: string): string {
     const hash = createHash('sha256');
-    for (const path of [
-        join(workDir, '22-article', `${runDate}_06_article_content.json`),
-        join(workDir, '24-script', `${runDate}_08_script_voicescript.yaml`),
-    ]) {
+    for (const path of safetyDraftPaths(workDir, runDate)) {
         if (!isFile(path)) failWith(1, `safety-review: required draft not found: ${path}`);
         hash.update(readFileSync(path)).update('\0');
     }
     return hash.digest('hex');
 }
 
+/** The drafts under review — the digest's bytes, and the quarantine set on a quality block. */
+function safetyDraftPaths(workDir: string, runDate: string): [string, string] {
+    return [
+        join(workDir, '22-article', `${runDate}_06_article_content.json`),
+        join(workDir, '24-script', `${runDate}_08_script_voicescript.yaml`),
+    ];
+}
+
 /**
  * Open a content-safety review of the current drafts: drop every artifact of an earlier review
  * (a retry must be re-reviewed on its final version) and write the request the two agent.run
@@ -1581,17 +1588,39 @@ function dailySafetyGate(args: string[]): void {
         const digest = safetyDigest(workDir, runDate);
         let status = 'block';
         let reason = 'verdict PASS on unchanged drafts, no BLOCK/RISK';
-        if (request?.digest !== digest) reason = 'drafts changed after the review was requested';
-        else if (decisions.length === 0 || decisions.some((d) => d !== 'OK' && d !== 'RISK' && d !== 'BLOCK'))
+        // A digest-mismatch block means the drafts were deliberately edited (HITL): keep them so
+        // a re-run of safety-prep re-reviews the edited version. Any other block means the
+        // content itself is rejected — quarantine the drafts below.
+        let edited = false;
+        if (request?.digest !== digest) {
+            reason = 'drafts changed after the review was requested';
+            edited = true;
+        } else if (decisions.length === 0 || decisions.some((d) => d !== 'OK' && d !== 'RISK' && d !== 'BLOCK'))
             reason = 'content-safety report missing or malformed';
         else if (decisions.includes('BLOCK')) reason = 'content-safety report carries a BLOCK decision';
         else if (verdict?.verdict !== 'PASS')
             reason = `content-judge verdict is ${typeof verdict?.verdict === 'string' ? verdict.verdict : 'missing or unreadable'}`;
+        else if (verdict?.confidence !== 'HIGH' && verdict?.confidence !== 'MEDIUM' && verdict?.confidence !== 'LOW')
+            reason = 'content-judge verdict is missing a HIGH/MEDIUM/LOW confidence level';
         else if (decisions.includes('RISK') || report?.escalate === true) {
             status = 'escalate';
             reason = 'RISK decision or escalation — human review required';
         } else status = 'pass';
         if (status !== 'block') sted('write', stampFile, () => writeJsonText(stampFile, { digest, status }));
+        if (status === 'block' && !edited) {
+            // A quality block poisons the isFile() draft cache: every same-day retry would
+            // re-fail on unchanged bytes (20261002 dogfood — two judge-fail runs replayed until
+            // the drafts were moved by hand). Quarantine (mv, never rm) so the next run
+            // regenerates from scratch; the 09_* verdict artifacts stay in 24-script for audit.
+            const q = join(workDir, '24-script', `.quarantine-${runDate}-judge-block`);
+            mkdirSync(q, { recursive: true });
+            for (const path of safetyDraftPaths(workDir, runDate)) {
+                if (!isFile(path)) continue;
+                let dest = join(q, basename(path));
+                for (let n = 2; isFile(dest); n += 1) dest = join(q, `${n}-${basename(path)}`);
+                renameSync(path, dest);
+            }
+        }
         sted('write', statusFile, () => writeRaw(statusFile, status));
         console.log(`safety: status=${status} (${reason}); routing is transition-owned`);
     });
@@ -1657,6 +1686,54 @@ function dailyGenerate(args: string[]): void {
  * `intro_music` empty -> skip. Writes `<stem>.intro.wav` (original render kept), points
  * `audioPath` at the merged file, extends `duration`, marks `introMerged` (idempotent).
  */
+/**
+ * Episode duration gate (standing requirement, 20261002): a daily episode must run 15–20 minutes
+ * (duration is authoritative; the 8–10-item figure is only an estimate). Reads the final audio
+ * content's `metadata.duration` — after the intro sting merged, this is the true episode length —
+ * and fails loud outside `[min, max]` seconds-worth of minutes. A missing duration is a failure,
+ * not a pass: the gate must never silently let an under-length episode through again.
+ */
+function dailyEpisodeDuration(args: string[]): void {
+    const [workDir = '', runDate = '', minArg = '', maxArg = ''] = args;
+    if (!workDir) fail('missing work_dir');
+    if (!runDate) fail('missing run_date');
+    const minMinutes = Number(minArg);
+    const maxMinutes = Number(maxArg);
+    if (!Number.isFinite(minMinutes) || minMinutes <= 0) fail(`invalid min_minutes: ${minArg || '(empty)'}`);
+    if (!Number.isFinite(maxMinutes) || maxMinutes < minMinutes) fail(`invalid max_minutes: ${maxArg || '(empty)'}`);
+    const contentPath = join(workDir, '30-audio', `${runDate}_11_generate_content.json`);
+    if (!isFile(contentPath)) failWith(1, `episode-duration: audio content not found: ${contentPath}`);
+    let raw: unknown;
+    try {
+        const parsed = JSON.parse(readFileSync(contentPath, 'utf-8')) as { metadata?: { duration?: unknown } };
+        raw = parsed.metadata?.duration;
+    } catch (error) {
+        if ((error as { code?: string }).code !== undefined) {
+            sted('read', contentPath, () => readFileSync(contentPath, 'utf-8'));
+        }
+        raw = undefined;
+    }
+    const seconds = typeof raw === 'number' ? raw : typeof raw === 'string' && raw.trim() !== '' ? Number(raw) : NaN;
+    if (!Number.isFinite(seconds) || seconds <= 0) {
+        failWith(1, `episode-duration: audio metadata.duration missing or not a number in ${contentPath}`);
+    }
+    const minutes = seconds / 60;
+    const shown = minutes.toFixed(1);
+    if (minutes < minMinutes) {
+        failWith(
+            1,
+            `episode-duration: episode runs ${shown} min, below the ${minMinutes}–${maxMinutes} min requirement (duration is authoritative). The candidate pool is too thin — deepen per-item treatment or add items; do not ship a short episode silently.`,
+        );
+    }
+    if (minutes > maxMinutes) {
+        failWith(
+            1,
+            `episode-duration: episode runs ${shown} min, above the ${minMinutes}–${maxMinutes} min requirement. Trim items or per-item treatment.`,
+        );
+    }
+    console.log(`episode-duration ok: ${shown} min within ${minMinutes}–${maxMinutes} min`);
+}
+
 function dailyIntro(args: string[]): void {
     const [workDir = '', runDate = '', introSrc = ''] = args;
     if (!introSrc || introSrc.trim() === '') {
@@ -1928,8 +2005,47 @@ function dailyShowNotes(args: string[]): void {
     timed(workDir, 'show-notes', () => {
         const effective = readJson(article)?.title;
         const titleArg = title !== '' ? title : typeof effective === 'string' ? effective : '';
+        // Boundary-excluded docs must not render into show notes (judge finding 2026-10-02): the
+        // article body's 未写入 note names the dropped ids; filter the plan copy podcast-pub sees.
+        // Inlined parse — this file must not import @gobing-ai/* (header contract); the same note
+        // grammar lives in daily-article-gen's excludedIdsFromNote.
         const postPath = firstPostPath(readJson(result), 'surfdash-pub');
-        kk(['executor', 'run', 'podcast-pub', '--in', plan, '--out', out, ...pluginsArgs(pluginsPath)], {
+        const articleBody = readJson(article)?.body;
+        const excluded = new Set<string>();
+        if (typeof articleBody === 'string') {
+            for (const note of articleBody.matchAll(/未写入[：:](.*)/g)) {
+                for (const id of note[1].matchAll(/(?:[a-zA-Z0-9_-]+:)?[0-9a-f]{16}/g)) {
+                    excluded.add(id[0]);
+                    const colon = id[0].indexOf(':');
+                    if (colon !== -1) excluded.add(id[0].slice(colon + 1));
+                }
+            }
+        }
+        let planIn = plan;
+        if (excluded.size > 0) {
+            const parsed: unknown = readJson(plan);
+            if (Array.isArray(parsed)) {
+                const kept = parsed.filter(
+                    (d) =>
+                        !(
+                            d !== null &&
+                            typeof d === 'object' &&
+                            (excluded.has(String((d as Record<string, unknown>).id)) ||
+                                (String((d as Record<string, unknown>).id).includes(':') &&
+                                    excluded.has(
+                                        String((d as Record<string, unknown>).id).slice(
+                                            String((d as Record<string, unknown>).id).indexOf(':') + 1,
+                                        ),
+                                    )))
+                        ),
+                );
+                if (kept.length !== parsed.length) {
+                    planIn = plan.replace(/\.json$/, '-shownotes.json');
+                    writeJsonText(planIn, kept);
+                }
+            }
+        }
+        kk(['executor', 'run', 'podcast-pub', '--in', planIn, '--out', out, ...pluginsArgs(pluginsPath)], {
             env: {
                 PODCAST_PUB_TITLE: titleArg,
                 PODCAST_PUB_RUN_DATE: runDate,
@@ -2018,7 +2134,22 @@ function dailyPublishClassify(args: string[]): void {
         const status = okCount > 0 && okCount === total ? 'full' : okCount > 0 ? 'partial' : 'none';
         sted('write', result, () => writeJqText(result, { ok: okCount === total && total > 0, targets }));
         sted('write', statusFile, () => writeRaw(statusFile, status));
+        const completeness = publishCompleteness(workDir, runDate);
+        const completenessFile = join(workDir, '31-publish', `${runDate}_17_publish_completeness.json`);
+        sted('write', completenessFile, () => writeJqText(completenessFile, { runDate, ...completeness }));
         console.log(`publish: status=${status} (${okCount}/${total} targets published); routing is transition-owned`);
+        console.log(`publish-completeness: ${completeness.rows.map((row) => `${row.channel}=${row.state}`).join(' ')}`);
+        if (completeness.missing.length > 0) {
+            const details = completeness.rows
+                .filter((row) => row.state !== 'done')
+                .map((row) => `${row.channel}=${row.state}${row.detail === '' ? '' : ` (${row.detail})`}`)
+                .join('; ');
+            console.error(
+                `WARNING: publish completeness — ${completeness.missing.join(', ')} did not complete: ${details}. ` +
+                    'Independent channels never enter publish_status; re-run their stages for this work dir (see ' +
+                    `${completenessFile})`,
+            );
+        }
     });
 }
 
@@ -2029,6 +2160,59 @@ function targetsOf(path: string): Record<string, unknown>[] {
         : [];
 }
 
+/**
+ * The publish tail's independent channels (locales, XHS, WeChat) and the surfdash CMS re-sync never
+ * enter `publish_status` — they are non-blocking by design (ADR-022/025/026). That is also why a
+ * tail driven by hand can skip every one of them and still leave a `full` status file behind:
+ * 20261002 shipped zh-only surfdash with no XHS and no WeChat draft, and nothing in the run's own
+ * artifacts said so. Record each channel's state beside the status file so the gap is visible.
+ */
+function publishCompleteness(
+    workDir: string,
+    runDate: string,
+): { rows: { channel: string; state: string; detail: string }[]; missing: string[] } {
+    const at = (name: string): string => join(workDir, '31-publish', `${runDate}_${name}`);
+    const markerFilled = (path: string): boolean => isFile(path) && readFileSync(path, 'utf-8').trim() !== '';
+    const rows: { channel: string; state: string; detail: string }[] = [];
+    const localesDone = ['en', 'ja'].every((locale) =>
+        markerFilled(at(`20_translate-localize_${locale}-post-path.txt`)),
+    );
+    rows.push({
+        channel: 'locales',
+        state: localesDone ? 'done' : 'not-run',
+        detail: localesDone ? 'en/ja drafts scaffolded' : 'translate markers absent (en/ja never scaffolded)',
+    });
+    const resultChannels: [string, string, string][] = [
+        ['xhs', '19_publish-xhs_result.json', ''],
+        ['wechat', '20_publish-wechat_result.json', ''],
+        ['cms-resync', '23_patch-surfdash_sync-result.json', '23_patch-surfdash_emdash-resync.txt'],
+    ];
+    for (const [channel, resultName, skipMarkerName] of resultChannels) {
+        const result = isFile(at(resultName)) ? readJson(at(resultName)) : null;
+        if (result !== null) {
+            const metadata = (result.metadata ?? {}) as Record<string, unknown>;
+            rows.push({
+                channel,
+                state: result.ok === true ? 'done' : 'failed',
+                detail:
+                    result.ok === true
+                        ? typeof metadata.publishStatus === 'string'
+                            ? metadata.publishStatus
+                            : 'ok'
+                        : String(result.error ?? 'ok=false'),
+            });
+            continue;
+        }
+        const skipMarker = skipMarkerName === '' ? '' : at(skipMarkerName);
+        rows.push(
+            skipMarker !== '' && isFile(skipMarker)
+                ? { channel, state: 'skipped', detail: readFileSync(skipMarker, 'utf-8').trim() }
+                : { channel, state: 'not-run', detail: 'no result artifact' },
+        );
+    }
+    return { rows, missing: rows.filter((row) => row.state !== 'done').map((row) => row.channel) };
+}
+
 function dailyPublishPartial(args: string[]): void {
     const [workDir = '', runDate = ''] = args;
     const result = join(workDir, '31-publish', `${runDate}_17_publish_result.json`);
@@ -2168,6 +2352,11 @@ function dailyPatchSurfdash(args: string[]): void {
         // changed===0 early-return so re-runs heal rows an earlier commit left stale
         // (20260925/20260929). Fixture tests pass no plugins-path and must not shell out.
         if (pluginsPath === '') {
+            // A missing plugins-path silently drops the CMS re-sync, and the CMS cards then keep the
+            // pre-patch `./assets/cover.png` the deployed worker cannot serve (20261002 lost its
+            // cover this way). Leave a marker for the classify step's completeness report.
+            const skipMarker = join(workDir, '31-publish', `${runDate}_23_patch-surfdash_emdash-resync.txt`);
+            sted('write', skipMarker, () => writeRaw(skipMarker, 'skipped: no plugins-path\n'));
             console.log('patch-surfdash: no plugins-path — skip emdash re-sync');
         } else {
             const syncIn = join(workDir, '31-publish', `${runDate}_23_patch-surfdash_sync.json`);
@@ -2460,6 +2649,9 @@ if (import.meta.main)
         case 'daily-intro':
             dailyIntro(rest);
             break;
+        case 'daily-episode-duration':
+            dailyEpisodeDuration(rest);
+            break;
         case 'daily-quality-report':
             dailyQualityReport(rest);
             break;

package/plugins/kk/scripts/topic-boundary.json

@@ -1,11 +1,11 @@
 {
     "version": 1,
-    "note": "Task 0189 R2 topic admission for the AI 乱炖 column. Excludes items whose mainline is political/diplomatic-trade, military, public-policy dispute, or a social public event; records every exclusion with id, rule, reason and the matched keyword. Sources are never deleted or modified. Scan covers the title plus the first scanChars chars of the body. Precision first: these keywords are a floor, not a sieve — a rare false negative is caught by the script-write/article-write topic-boundary prompts and the content-safety skill review, and a suspected false positive is restored by hand with a recorded reason (manual review), never silently.",
+    "note": "Task 0189 R2 topic admission for the AI 乱炖 column, with the 2026-10-02 operator scope correction: the political-diplomatic rule is China-only (中国政治/外交主线). Foreign public-affairs items are NOT floor-excluded — they flow to the plan and are handled by the article-write/script-write/cover 选题边界 prompts (skip + note / plan `excluded`) and the content-safety review. Excluded items are recorded with id, rule, reason and the matched keyword. Sources are never deleted or modified. Scan covers the title plus the first scanChars chars of the body. Precision first: these keywords are a floor, not a sieve — a rare false negative is caught by the prompt-layer boundaries and the content-safety skill review, and a suspected false positive is restored by hand with a recorded reason (manual review), never silently.",
     "scanChars": 200,
     "rules": [
         {
             "id": "political-diplomatic",
-            "reason": "政治/外交主线：涉及政治人物、政府机构、选举、外交或双边关系的条目，不属于 AI 乱炖的技术分享范围（0189 R2）。",
+            "reason": "中国政治/外交主线：涉及中国党政机构、领导人、外交部门或台海/中美关系的条目，不属于 AI 乱炖的技术分享范围（0189 R2；2026-10-02 操作者裁定本规则仅限中国，外国政治条目由选题边界提示与内容安全评审处理）。",
             "keywords": [
                 "领导人",
                 "主席访",
@@ -16,18 +16,9 @@
                 "国事访问",
                 "政治局",
                 "中南海",
-                "总统",
-                "白宫",
-                "国会",
-                "议会",
-                "选举",
                 "台海",
                 "两岸关系",
-                "中美关系",
-                "president",
-                "parliament",
-                "election",
-                "diplomat"
+                "中美关系"
             ]
         },
         {

package/plugins/kk/skills/content-judge/SKILL.md

@@ -30,6 +30,7 @@ Every evaluation writes a JSON verdict file. The default path is
 ```json
 {
   "verdict": "PASS | FAIL | NEEDS_REVISION",
+  "confidence": "HIGH | MEDIUM | LOW",
   "rubric": "<rubric-id>",
   "score": 0.0,
   "summary": "<one-paragraph reasoning summary>",
@@ -57,6 +58,35 @@ Every evaluation writes a JSON verdict file. The default path is
 `PASS` is the only verdict that unblocks publish. The shell guard treats `FAIL` and
 `NEEDS_REVISION` identically (exit 1); the distinction is for the feedback loop, not the gate.
 
+### Confidence level
+
+Every verdict carries a `confidence` level — `HIGH`, `MEDIUM`, or `LOW` — self-assessed by the
+judge from evidence coverage:
+
+- `HIGH` — every criterion grounded in directly verified evidence (sources opened, claims
+  cross-checked); no unverified rule text or assumption entered the reasoning.
+- `MEDIUM` — verdict grounded but with known evidence gaps (e.g. unverified platform rule text,
+  single-source claims), none strong enough to flip the verdict.
+- `LOW` — verdict rests on assumptions or unverified evidence that could plausibly flip it.
+
+Every judge-verdict **consumer** enforces the enum, so a `PASS` whose `confidence` is missing or
+outside `HIGH|MEDIUM|LOW` is not a usable verdict anywhere:
+
+- `kk stage daily-safety-gate` blocks the run (`missing a HIGH/MEDIUM/LOW confidence level`).
+- The spur workflow gates (`kk-itc`, `judge-gated-publish-example`) read the verdict through a
+  confidence-aware `jq` filter that turns such a `PASS` into an out-of-contract value (`PASS!`),
+  which no accept guard matches and the state machine's catch-all routes to `failed` — the import
+  neither publishes nor deadlocks.
+- `kk stage itc-report` reports `outcome: failed` (and says why in `report.md`) for a `PASS` without
+  a valid confidence; an absent verdict file keeps the dry-run convention.
+
+Add that filter to any new judge gate; `references/workflow-integration.md` carries the pattern and
+`plugins/kk/tests/verdict-confidence.test.ts` fails when a workflow accepts a judge `PASS` without
+checking the enum.
+
+The level itself does not change routing — escalation stays owned by the review's `RISK` decisions
+and `escalate` flag; treat `LOW` as a signal to demand human review before publishing.
+
 ## Evaluation procedure
 
 1. **Read the Content** from the path given in the evaluation request (the `content.json` written

package/plugins/kk/skills/content-judge/references/workflow-integration.md

@@ -33,7 +33,10 @@ if [ ! -f "$verdict_path" ]; then
   echo "judge: verdict file not found at $verdict_path"
   exit 1
 fi
-verdict=$(jq -r '.verdict' "$verdict_path")
+# Confidence-aware read: a PASS without a HIGH|MEDIUM|LOW confidence becomes `PASS!`, which no
+# accept guard matches and the catch-all below rejects. Keep this filter byte-identical wherever a
+# guard reads a judge verdict (`plugins/kk/tests/verdict-confidence.test.ts` enforces that).
+verdict=$(jq -r 'if (.verdict // "") == "PASS" then (.verdict + (if ((.confidence // "") | test("^(HIGH|MEDIUM|LOW)$")) then "" else "!" end)) else (.verdict // "") end' "$verdict_path")
 case "$verdict" in
   PASS)
     echo "judge: PASS"
@@ -45,7 +48,7 @@ case "$verdict" in
     exit 1
     ;;
   *)
-    echo "judge: unrecognized verdict '$verdict' in $verdict_path"
+    echo "judge: unrecognized verdict or a PASS without a valid confidence ('$verdict') in $verdict_path"
     exit 1
     ;;
 esac
@@ -57,6 +60,10 @@ esac
 2. **Unrecognized value → block.** Only `PASS` unblocks; anything else exits 1.
 3. **`FAIL` and `NEEDS_REVISION` both block** — the distinction is for the feedback loop, not the
    gate. The workflow topology decides whether to retry (NEEDS_REVISION) or halt (FAIL).
+4. **A `PASS` without a `HIGH | MEDIUM | LOW` confidence is not a verdict.** The confidence-aware
+   read above makes it arrive as `PASS!`, so it fails rule 2 like any other out-of-contract value.
+5. **Leave a catch-all.** A guard set with no matching transition deadlocks the machine; every
+   judge state needs a terminal `*`/catch-all edge for verdicts outside the contract.
 
 ## Retry topology (NEEDS_REVISION)
 

package/plugins/kk/skills/publish/SKILL.md

@@ -50,7 +50,7 @@ Overrides: `--profile <dir>`, `--timeout <ms>`; env fallback `XHS_PROFILE_DIR` /
 **ADR-025 (2026-09-28, after an XHS warning about automated activity).** Publish now runs this same
 check as a **preflight** before opening anything headed — so a missing or expired session is a fast
 refusal (`ok: false`) rather than a headed window sitting on the login gate for 120 s — and both
-console visits are capped at **one headed publish cycle + one session check per 24 h**
+console visits are capped at **one headed publish cycle + one session check per 2 h**
 (`~/.config/kk/xhs-automation.json`, override `KK_XHS_AUTOMATION_STATE`; a manual `check` draws on
 the same `check` budget). A capped or sessionless publish fails with `ok: false` and a named reason.
 `login` is never automated and never capped. Do not schedule XHS publishing: publish deliberately,

package/plugins/kk/workflows/judge-gated-publish-example.yaml

@@ -25,7 +25,7 @@ states:
           # Seed the verdict so the judge gate guards can be evaluated in --dry-run
           # (dry-run skips actions but still evaluates shell guards). A real run's
           # judge step overwrites this file before the gate is read.
-          command: 'echo ''{"verdict":"PASS","score":0.0}'' > ${vars.verdictFile} && echo ''generating content...'''
+          command: 'echo ''{"verdict":"PASS","confidence":"HIGH","score":0.0}'' > ${vars.verdictFile} && echo ''generating content...'''
   - id: judge
     description: "Evaluate generated content via the judge skill (agent.run)"
     onEnter:
@@ -51,7 +51,7 @@ states:
   - id: done
     description: "Publish completed successfully"
   - id: failed
-    description: "Judge rejected content (FAIL) or retries exhausted (NEEDS_REVISION)"
+    description: "Judge rejected content (FAIL or an out-of-contract verdict) or retries exhausted (NEEDS_REVISION)"
 
 transitions:
   - from: generate
@@ -63,14 +63,18 @@ transitions:
   # PASS -> publish.
   - from: judge
     to: publish
-    description: "Judge verdict is PASS (or file absent under --dry-run) -> proceed to publish"
+    description: "Judge verdict is PASS with a HIGH/MEDIUM/LOW confidence (or file absent under --dry-run) -> proceed to publish"
     guard:
       kind: shell
       options:
-        # Fail-closed on an explicit non-PASS verdict; fail-open only when the file is
-        # absent, which happens under --dry-run (actions are skipped, so the judge
-        # never writes its verdict). A real run always has the file.
-        command: 'v="$(jq -r .verdict ${vars.verdictFile} 2>/dev/null)"; test -z "$v" -o "$v" = PASS'
+        # Fail-closed on an explicit non-PASS verdict or a PASS without a valid confidence level;
+        # fail-open only when the file is absent, which happens under --dry-run (actions are skipped,
+        # so the judge never writes its verdict). A real run always has the file.
+        # The confidence-aware read appends `!` to an out-of-contract PASS, so the guards below treat
+        # it as a rejection instead of letting it through.
+        command: >-
+          v="$(jq -r 'if (.verdict // "") == "PASS" then (.verdict + (if ((.confidence // "") | test("^(HIGH|MEDIUM|LOW)$")) then "" else "!" end)) else (.verdict // "") end' ${vars.verdictFile} 2>/dev/null)";
+          test -z "$v" -o "$v" = PASS
 
   # NEEDS_REVISION -> back to generate with a bounded retry budget.
   - from: judge
@@ -84,14 +88,18 @@ transitions:
         # responsible for incrementing attempt on each loop in a real run.
         command: 'v="$(jq -r .verdict ${vars.verdictFile} 2>/dev/null)"; test "$v" = NEEDS_REVISION -a "$attempt" -lt "$maxRetries"'
 
-  # FAIL, or NEEDS_REVISION with no retries left -> failed.
+  # FAIL, out-of-contract verdict, or NEEDS_REVISION with no retries left -> failed.
   - from: judge
     to: failed
-    description: "Judge verdict is FAIL, or NEEDS_REVISION with retries exhausted -> block publish"
+    description: "Judge verdict is FAIL, out-of-contract, or NEEDS_REVISION with retries exhausted -> block publish"
     guard:
       kind: shell
       options:
-        command: 'v="$(jq -r .verdict ${vars.verdictFile} 2>/dev/null)"; test -n "$v" && { test "$v" = FAIL || test "$v" = NEEDS_REVISION -a "$attempt" -ge "$maxRetries"; }'
+        # Anything that is neither a usable PASS nor a retryable NEEDS_REVISION fails closed.
+        command: >-
+          v="$(jq -r 'if (.verdict // "") == "PASS" then (.verdict + (if ((.confidence // "") | test("^(HIGH|MEDIUM|LOW)$")) then "" else "!" end)) else (.verdict // "") end' ${vars.verdictFile} 2>/dev/null)";
+          { test -n "$v" -a "$v" != PASS -a "$v" != NEEDS_REVISION; } ||
+          { test "$v" = NEEDS_REVISION -a "$attempt" -ge "$maxRetries"; }
 
   - from: publish
     to: done