akm-cli 0.9.0-beta.5 → 0.9.0-beta.51

package/dist/commands/health/html-report.js

@@ -6,7 +6,7 @@
  * (#582). Ports the external akm-health-report skill's collect.py + render.py
  * to TypeScript so the report is generated in-process (no python, no
  * shell-out). The template (`src/assets/templates/html/health.html`) is a
- * verbatim copy of the skill's report.html; this module computes the 17
+ * strict superset of the skill's report.html; this module computes the 17
  * `%%TOKEN%%` replacements it consumes.
  *
  * Determinism: nothing here depends on Date.now()/Math.random(). Runs are
@@ -17,13 +17,44 @@ import fs from "node:fs";
 import path from "node:path";
 import { escapeHtml } from "../../output/html-render.js";
 import { getDirname } from "../../runtime.js";
+import { pkgVersion } from "../../version.js";
+/**
+ * Distill skip-reasons hidden from the breakdown chart. `no new signal since
+ * last proposal` is the steady-state "nothing changed, nothing to do" outcome —
+ * it dominates the histogram and drowns out the actionable reasons, so it is
+ * intentionally excluded from the chart (the count still lives in the data).
+ */
+const DISTILL_REASONS_HIDDEN = new Set(["no new signal since last proposal"]);
 const ECHARTS_CDN = "https://cdn.jsdelivr.net/npm/echarts@5/dist/echarts.min.js";
 const ECHARTS_VENDOR_PATH = path.join(getDirname(import.meta.url), "../../assets/templates/html/vendor/echarts.min.js");
 // ── Small formatters (ports of render.py helpers) ───────────────────────────
 const esc = escapeHtml;
+/** Emit a <time> element that the browser's JS will reformat to the viewer's local timezone. */
+function isoTimeTag(iso) {
+    const fallback = iso.slice(0, 16).replace("T", " ");
+    return `<time data-iso="${esc(iso)}">${esc(fallback)}</time>`;
+}
 function num(value) {
     return Math.round(value).toLocaleString("en-US");
 }
+/** Compact token/count formatter, e.g. 127345 → "127K", 1_500_000 → "1.5M". */
+function compact(value) {
+    const v = Math.round(value);
+    if (Math.abs(v) >= 1_000_000)
+        return `${(v / 1_000_000).toFixed(1).replace(/\.0$/, "")}M`;
+    if (Math.abs(v) >= 1_000)
+        return `${(v / 1_000).toFixed(1).replace(/\.0$/, "")}K`;
+    return String(v);
+}
+/** Humanize a camelCase / kebab / snake enum into reader-facing text. */
+function humanize(raw) {
+    return raw
+        .replace(/[-_]/g, " ")
+        .replace(/([a-z])([A-Z])/g, "$1 $2")
+        .replace(/\s+/g, " ")
+        .trim()
+        .replace(/^./, (c) => c.toUpperCase());
+}
 function fmtMs(ms) {
     return ms ? `${(ms / 60000).toFixed(1)}m` : "—";
 }
@@ -60,6 +91,7 @@ function reshapeRun(r) {
     const geMs = ge.durationMs || 0;
     return {
         id: r.id,
+        taskId: r.taskId ?? "manual",
         startedAt: r.startedAt,
         completedAt: r.completedAt,
         wallTimeMs: wall,
@@ -144,11 +176,14 @@ function deltaPill(deltas, key, lowerIsBetter = false) {
     const signed = `${raw > 0 ? "+" : ""}${Math.round(raw)}%`;
     return `<span class="trend-pill ${cls}">${arrow} ${signed}</span>`;
 }
-// ── Advisory / watch-item cards ──────────────────────────────────────────────
-function advisoryCard(cls, icon, title, descHtml) {
-    return (`<div class="advisory ${cls}"><div class="advisory-icon">${icon}</div>` +
-        `<div class="advisory-body"><div class="title">${title}</div>` +
-        `<div class="desc">${descHtml}</div></div></div>`);
+const PRIO_RANK = { P1: 0, P2: 1, P3: 2 };
+function actionItemCard(item) {
+    const icon = item.cls === "fail" ? "🔴" : item.prio === "P3" ? "🟡" : "⚠️";
+    const remedy = item.remedy ? `<div class="remedy">Fix: <code>${esc(item.remedy)}</code></div>` : "";
+    return (`<div class="advisory ${item.cls}"><div class="advisory-icon">${icon}</div>` +
+        `<div class="advisory-body">` +
+        `<div class="title"><span class="prio ${item.prio.toLowerCase()}">${item.prio}</span>${item.title}</div>` +
+        `<div class="desc">${item.descHtml}</div>${remedy}</div></div>`);
 }
 function passCard(title, desc) {
     return ('<div class="advisory" style="border-left:3px solid var(--green);">' +
@@ -168,6 +203,43 @@ function readSemSearch(advisories) {
     return { blocked, detail: `${num(entries)} entries, ${num(embeddings)} embeddings` };
 }
 // ── ECharts delivery ─────────────────────────────────────────────────────────
+/** Parse an akm window string (`24h`, `7d`, `30m`, `2w`) to milliseconds; 0 if unparseable. */
+function windowToMs(window) {
+    const m = /^(\d+)\s*([mhdw])$/i.exec(window.trim());
+    if (!m)
+        return 0;
+    const n = Number(m[1]);
+    const mult = { m: 60_000, h: 3_600_000, d: 86_400_000, w: 604_800_000 }[m[2].toLowerCase()] ?? 0;
+    return n * mult;
+}
+/**
+ * Build the time-slice `<option>`s for the report's filter bar, DERIVED from the
+ * actual report window so the choices always make sense (the old hard-coded
+ * 1d–21d list was useless on a 24h or 7d report). "All" is the default; the
+ * sub-window options carry their cutoff in milliseconds (consumed by
+ * filteredRuns), largest first, only those strictly shorter than the window.
+ */
+function buildSliceOptions(window) {
+    const windowMs = windowToMs(window);
+    const HOUR = 3_600_000;
+    const DAY = 86_400_000;
+    const candidates = [
+        [6 * HOUR, "6h"],
+        [12 * HOUR, "12h"],
+        [DAY, "1d"],
+        [3 * DAY, "3d"],
+        [7 * DAY, "7d"],
+        [14 * DAY, "14d"],
+    ];
+    const subs = candidates
+        .filter(([ms]) => windowMs > 0 && ms < windowMs)
+        .sort((a, b) => b[0] - a[0])
+        .slice(0, 4);
+    const opts = [`<option value="all" selected>All (${esc(window)})</option>`];
+    for (const [ms, label] of subs)
+        opts.push(`<option value="${ms}">Last ${label}</option>`);
+    return opts.join("\n      ");
+}
 function buildEchartsTag(opts) {
     const mode = opts.echarts ?? (process.env.AKM_ECHARTS === "cdn" ? "cdn" : "inline");
     if (mode === "cdn")
@@ -182,7 +254,8 @@ function buildEchartsTag(opts) {
  * Compute all 17 `%%TOKEN%%` replacements for the health HTML template.
  * There is deliberately NO standalone `%%OVERALL_STATUS%%` token — the
  * overall status is embedded in the pre-rendered badge / exec-summary
- * fragments, matching the skill template.
+ * fragments, matching the skill template. Advisories and "what to watch" are
+ * merged + de-duplicated into a single prioritized `%%ACTION_ITEMS_HTML%%`.
  */
 export function buildHealthHtmlReplacements(result, opts) {
     const deltas = opts.deltas ?? {};
@@ -196,6 +269,31 @@ export function buildHealthHtmlReplacements(result, opts) {
     const mi = improve.memoryInference;
     const ge = improve.graphExtraction;
     const wallTime = improve.wallTime;
+    // WS-5: perf telemetry, coverage, and degradation metrics.
+    const perf = improve.perfTelemetry ?? {
+        dedupPoolSize: 0,
+        llmPoolSize: 0,
+        judgedCacheSkipped: 0,
+        embedMs: 0,
+        embedCacheHits: 0,
+        embedCacheMisses: 0,
+        overBudgetRuns: 0,
+        runsWithTelemetry: 0,
+    };
+    const coverage = improve.coverage;
+    const degradation = improve.degradation;
+    // #576: real per-stage LLM token/time accounting (replaces the GPU-time
+    // proxy). Optional-guarded so reports built from older health JSON without
+    // the aggregate still render.
+    const llm = result.metrics.llmUsage ?? {
+        calls: 0,
+        totalDurationMs: 0,
+        promptTokens: 0,
+        completionTokens: 0,
+        totalTokens: 0,
+        reasoningTokens: 0,
+        byStage: {},
+    };
     const totalRuns = runs.length;
     const failedRuns = runs.filter((r) => !r.ok).length;
     const invoked = improve.invoked || totalRuns;
@@ -212,11 +310,23 @@ export function buildHealthHtmlReplacements(result, opts) {
     // ── Meta (collect.py steps 10-11) ──────────────────────────────────────────
     const sinceIso = result.since;
     const reportDate = sinceIso.slice(0, 10);
-    const sinceHuman = sinceIso ? `${sinceIso.slice(0, 16).replace("T", " ")} UTC → now` : `last ${opts.window}`;
+    const sinceHuman = sinceIso ? `${isoTimeTag(sinceIso)} → now` : `last ${esc(opts.window)}`;
     const reportTitle = reportDate ? `AKM Health Report — ${reportDate}` : "AKM Health Report";
     const lastRun = runs[runs.length - 1];
     const generatedAt = lastRun ? lastRun.completedAt || lastRun.startedAt || sinceIso : sinceIso;
     const latest = [...runs].reverse().find((r) => r.ok) ?? lastRun;
+    // Freshness: surface the newest run's timestamp + the generated-at anchor in
+    // the exec card. Staleness is computed deterministically (no Date.now()) from
+    // the gap between the window start (`since`) and the newest run we have: if no
+    // run landed in the window, or the newest run sits in the first 25% of a
+    // window wider than the 6h threshold (i.e. a long idle tail), we flag stale.
+    const STALE_MS = 6 * 60 * 60 * 1000;
+    const latestRunMs = lastRun ? Date.parse(lastRun.completedAt || lastRun.startedAt) : NaN;
+    const sinceMs = Date.parse(sinceIso);
+    const generatedMs = Date.parse(generatedAt);
+    const idleTailMs = Number.isFinite(latestRunMs) && Number.isFinite(generatedMs) ? Math.max(0, generatedMs - latestRunMs) : 0;
+    const isStale = totalRuns === 0 || (Number.isFinite(latestRunMs) && Number.isFinite(sinceMs) && idleTailMs > STALE_MS);
+    const latestRunHuman = lastRun ? isoTimeTag(lastRun.completedAt || lastRun.startedAt) : "—";
     // ── Status badges ──────────────────────────────────────────────────────────
     const badgeByStatus = {
         pass: { badge: "badge-pass", dot: "dot-pass", label: "PASS" },
@@ -238,7 +348,7 @@ export function buildHealthHtmlReplacements(result, opts) {
         li("MI yield rate", miYieldRate),
         li("MI written", num(miWritten)),
         li("Consolidation promoted", num(cons.promoted)),
-        li("Consolidation judgedNoAction", num(cons.judgedNoAction)),
+        li("Consolidation judged: no action", `<abbr title="Candidates the consolidator reviewed but intentionally left unchanged (the 'judgedNoAction' field).">${num(cons.judgedNoAction)}</abbr>`),
         li("Chunk failure", chunkFail),
         li("Median wall time", fmtMs(wallTime.medianMs)),
         li("P95 wall time", fmtMs(wallTime.p95Ms)),
@@ -259,12 +369,12 @@ export function buildHealthHtmlReplacements(result, opts) {
     const snapRows = latest
         ? [
             li("Run id", `<code>${esc(latest.id.slice(0, 28))}</code>`),
-            li("Completed", `${esc((latest.completedAt || latest.startedAt).slice(0, 16).replace("T", " "))} UTC`),
+            li("Completed", isoTimeTag(latest.completedAt || latest.startedAt)),
             li("Status", latest.ok ? "✅ ok" : "❌ failed"),
             li("Wall time", fmtMs(latest.wallTimeMs)),
             li("Reflect ok/fail", `${latest.reflectOk} / ${latest.reflectFailed}`),
             li("Promoted", String(latest.promoted)),
-            li("judgedNoAction", String(latest.judgedNoAction)),
+            li("Judged: no action", `<abbr title="Candidates reviewed but intentionally left unchanged on this run.">${latest.judgedNoAction}</abbr>`),
             li("MI written", String(latest.miWritten)),
             li("Graph entities/relations", `${latest.geEntities} / ${latest.geRelations}`),
         ].join("")
@@ -273,16 +383,50 @@ export function buildHealthHtmlReplacements(result, opts) {
         li("Report window", esc(opts.window)),
         li("Compare window", esc(opts.compare)),
         li("Runs", `${num(totalRuns)} (${failedRuns} failed)`),
-        li("Stash derived", num(improve.memorySummary.derived)),
-        li("Stash eligible", num(improve.memorySummary.eligible)),
+        li("Stash derived", `<abbr title="Whole-stash recount of derived assets at report time — not a per-run sum.">${num(improve.memorySummary.derived)}</abbr>`),
+        li("Stash eligible", `<abbr title="Whole-stash recount of eligible assets at report time — not a per-run sum.">${num(improve.memorySummary.eligible)}</abbr>`),
         li("Pending proposals", String(proposals.length)),
         li("Semantic search", sem.blocked ? "BLOCKED" : "OK"),
     ].join("");
     const overallEmoji = trend.overall === "improving" ? "📈" : trend.overall === "degrading" ? "📉" : "↔️";
+    // ── Synthesized verdict (one sentence + 2-3 drivers) ───────────────────────
+    // The verdict WORD is the authoritative health status (hard checks). Concerns
+    // (failed improve runs, blocked search, pending proposals) are advisory and do
+    // NOT gate that word, so we must not pair a green PASS with red "drivers" as
+    // if they were failures. When the status is PASS we frame concerns as "watch:"
+    // alongside the trend; only a degraded status (WARN/FAIL) leads with them.
+    const concerns = [];
+    if (failedRuns > 0)
+        concerns.push(`${failedRuns} failed run${failedRuns === 1 ? "" : "s"}`);
+    if (sem.blocked)
+        concerns.push("semantic search blocked");
+    if (proposals.length > 0)
+        concerns.push(`${proposals.length} pending proposal${proposals.length === 1 ? "" : "s"}`);
+    const trendClause = trend.overall === "improving"
+        ? "throughput and latency improving"
+        : trend.overall === "degrading"
+            ? "throughput or latency degrading"
+            : "throughput and latency steady";
+    const verdictWord = badge.label;
+    let verdictRest;
+    if (result.status === "pass") {
+        const watch = concerns.length > 0 ? `; watch: ${concerns.slice(0, 3).join(", ")}` : "";
+        verdictRest = `healthy, ${trendClause}${watch}`;
+    }
+    else {
+        const lead = concerns.length > 0 ? concerns.slice(0, 3).join("; ") : trendClause;
+        verdictRest = trend.overall === "degrading" && concerns.length > 0 ? `${lead}; ${trendClause}` : lead;
+    }
+    const verdictSentence = `${verdictWord} — ${esc(verdictRest)}.`;
+    const verdictHtml = `<div class="verdict ${result.status}"><b>Verdict:</b> ${verdictSentence}</div>`;
+    // ── Freshness line ─────────────────────────────────────────────────────────
+    const freshnessHtml = `<div class="freshness${isStale ? " stale" : ""}">${isStale ? "⚠️ Stale: " : ""}Latest run ${latestRunHuman} &nbsp;·&nbsp; generated ${isoTimeTag(generatedAt)}${isStale ? " — no recent activity (newest run older than the 6h freshness threshold)." : "."}</div>`;
     const execSummary = `
     <h2>${overallEmoji} Executive Summary
         <span class="badge-pill ${badge.badge}" style="font-size:11px;">
         <span class="dot ${badge.dot}"></span>${badge.label}</span></h2>
+    ${verdictHtml}
+    ${freshnessHtml}
     <div class="exec-grid">
       <div>
         <h4>Quick Numbers</h4>
@@ -306,143 +450,341 @@ export function buildHealthHtmlReplacements(result, opts) {
     <div class="overall">Overall trend: <b>${esc(trend.overall)}</b> ${overallEmoji}
       &nbsp;·&nbsp; based on decision quality, output volume, failures, and latency vs the prior window.</div>`.trim();
     // ── KPI cards ──────────────────────────────────────────────────────────────
+    // Color is a health SIGNAL, not decoration: green/yellow/red where a card has
+    // a meaningful threshold; "neutral" for purely-informational counts. Cards are
+    // ordered by operator priority (failures first, informational counts last).
     const semValue = sem.blocked ? "BLOCKED" : "OK";
-    const semColor = sem.blocked ? "yellow" : "green";
+    const semColor = sem.blocked ? "red" : "green";
     const semStyle = sem.blocked ? "font-size:18px;" : "";
+    const completionPct = invoked ? (100 * completed) / invoked : 100;
+    const completionColor = completionPct >= 99 ? "green" : completionPct >= 90 ? "yellow" : "red";
+    const llmTokensCompact = compact(llm.totalTokens);
     const kpiCard = (color, label, value, sub, valueStyle = "") => `<div class="kpi-card ${color}">
       <div class="label">${label}</div>
       <div class="value"${valueStyle ? ` style="${valueStyle}"` : ""}>${value}</div>
       <div class="sub">${sub}</div>
     </div>`;
     const kpiCards = [
-        kpiCard(failedRuns === 0 ? "green" : "yellow", "Completion Rate", completionRate, `${num(completed)} / ${num(invoked)} invoked`),
         kpiCard(failedRuns === 0 ? "green" : "red", "Failed Runs", String(failedRuns), `of ${num(totalRuns)} runs · ${taskFailRate} task fail`),
+        kpiCard(completionColor, "Completion Rate", completionRate, `${num(completed)} / ${num(invoked)} invoked`),
+        kpiCard("neutral", "Median Duration", `${medianDurMin}m`, `p95 = ${p95DurMin}m`),
         kpiCard("blue", "Total Promoted", num(cons.promoted), `avg ${avgPromoted} / run`),
         kpiCard("blue", "MI Written", num(miWritten), `${miYieldRate} yield rate`),
         kpiCard("purple", "Graph Entities", num(ge.entities), `+${num(ge.relations)} relations`),
-        kpiCard("green", "Stash Derived", num(improve.memorySummary.derived), `of ${num(improve.memorySummary.eligible)} eligible`),
-        kpiCard("yellow", "Median Duration", `${medianDurMin}m`, `p95 = ${p95DurMin}m`),
+        kpiCard("neutral", "Stash Derived", num(improve.memorySummary.derived), `of ${num(improve.memorySummary.eligible)} eligible (whole-stash)`),
+        // #576: real LLM work — duration leads, tokens compact, not a GPU proxy.
+        kpiCard(llm.calls > 0 ? "purple" : "neutral", "🧠 LLM Work", `${llmTokensCompact} tok`, `${fmtMs(llm.totalDurationMs)} · ${num(llm.calls)} calls · ${compact(llm.reasoningTokens)} reasoning`),
         kpiCard(semColor, "Semantic Search", semValue, esc(sem.detail), semStyle),
-        kpiCard("yellow", "Pending Proposals", String(proposals.length), `from ${esc(opts.window)} batch`),
+        kpiCard(proposals.length > 0 ? "yellow" : "neutral", "Pending Proposals", String(proposals.length), `from ${esc(opts.window)} batch`),
     ].join("\n");
     // ── Chart payload ──────────────────────────────────────────────────────────
-    const distillReasons = [...new Set(runs.flatMap((r) => Object.keys(r.distillByReason)))].sort();
+    const distillReasons = [...new Set(runs.flatMap((r) => Object.keys(r.distillByReason)))]
+        .filter((reason) => !DISTILL_REASONS_HIDDEN.has(reason))
+        .sort();
     const runsJsConst = `const RUNS = ${JSON.stringify(runs)};`;
     // ── Summary table rows ─────────────────────────────────────────────────────
+    // Optional 4th element = a glossary tooltip rendered as <abbr title>.
     const summaryRows = [
         ["Task fail rate", taskFailRate, "flat"],
         ["Agent fail rate", agentFailRate, "flat"],
         ["Improve completion", `${num(completed)} / ${num(invoked)}`, "flat"],
-        ["MI yield rate", miYieldRate, trend.decisionQuality],
-        ["MI written", num(miWritten), trend.outputVolume],
+        [
+            "MI yield rate",
+            miYieldRate,
+            trend.decisionQuality,
+            "Memory-inference yield: share of considered candidates that produced a written fact.",
+        ],
+        ["MI written", num(miWritten), trend.outputVolume, "Memory-inference: facts written this window."],
         ["Consolidation promoted", num(cons.promoted), trend.outputVolume],
         ["Consolidation merged", num(cons.merged), "flat"],
         ["Consolidation deleted", num(cons.deleted), "flat"],
         ["Consolidation contradicted", num(cons.contradicted), "flat"],
-        ["Consolidation judgedNoAction", num(cons.judgedNoAction), "flat"],
+        [
+            "Consolidation judged: no action",
+            num(cons.judgedNoAction),
+            "flat",
+            "Candidates reviewed but intentionally left unchanged (the 'judgedNoAction' field).",
+        ],
         ["Chunk failure", chunkFail, "flat"],
         ["Graph entities", num(ge.entities), "up"],
         ["Graph relations", num(ge.relations), "up"],
-        ["Stash derived", num(improve.memorySummary.derived), "up"],
+        [
+            "Stash derived",
+            num(improve.memorySummary.derived),
+            "up",
+            "Whole-stash recount of derived assets at report time — not a per-run sum.",
+        ],
         ["Median wall time", fmtMs(wallTime.medianMs), trend.latency],
         ["P95 wall time", fmtMs(wallTime.p95Ms), trend.latency],
+        // #576: real LLM accounting (replaces the GPU-time proxy).
+        ["LLM calls", num(llm.calls), "flat"],
+        ["LLM total tokens", num(llm.totalTokens), "flat"],
+        ["LLM prompt tokens", num(llm.promptTokens), "flat"],
+        ["LLM completion tokens", num(llm.completionTokens), "flat"],
+        [
+            "LLM reasoning tokens",
+            num(llm.reasoningTokens),
+            "flat",
+            "Tokens spent on model reasoning/thinking, billed separately from prompt and completion.",
+        ],
+        ["LLM wall time", fmtMs(llm.totalDurationMs), trend.latency],
     ];
+    // #612 — auto-accept gate calibration. Only surface when the gate actually
+    // acted on proposals in the window (samples > 0); a default ungated install
+    // reports an empty summary and we omit the rows to keep the table parity-clean.
+    const calibration = improve.calibration;
+    if (calibration && calibration.samples > 0) {
+        summaryRows.push([
+            "Calibration samples",
+            num(calibration.samples),
+            "flat",
+            "Auto-accept gate decisions (auto-accepted + auto-rejected) the calibration join measured this window.",
+        ], [
+            "Calibration accept rate",
+            String(calibration.overallAcceptRate),
+            "flat",
+            "Realized accept rate of acted-on gate decisions (auto-accepted / total acted-on).",
+        ], [
+            "Calibration gap",
+            String(calibration.calibrationGap),
+            "flat",
+            "Mean predicted confidence minus realized accept rate. Positive = the gate is over-confident.",
+        ]);
+    }
+    // WS-5: denominator-fixed coverage rows (only when we have real data).
+    if (coverage && !Number.isNaN(coverage.rate)) {
+        summaryRows.push([
+            "Coverage rate",
+            pct(coverage.rate, 1),
+            "flat",
+            "Accepted proposals / total stash assets (denominator-fixed). Shows what fraction of the corpus has been touched.",
+        ], [
+            "Eligible fraction",
+            pct(coverage.eligibleFraction, 1),
+            "flat",
+            "Eligible assets / total stash assets. Fraction the improve pipeline actively considers.",
+        ], [
+            "Coverage accepted",
+            num(coverage.acceptedProposals),
+            "up",
+            "Total accepted proposals used for the denominator-fixed coverage rate.",
+        ]);
+    }
+    // WS-5: perf telemetry rows (only when at least one run reported telemetry).
+    if (perf.runsWithTelemetry > 0) {
+        const embedCacheTotal = perf.embedCacheHits + perf.embedCacheMisses;
+        const embedCacheHitRate = embedCacheTotal > 0 ? pct(perf.embedCacheHits / embedCacheTotal, 1) : "—";
+        summaryRows.push([
+            "Embed cache hit rate",
+            embedCacheHitRate,
+            "flat",
+            "Fraction of embedding lookups served from cache (>95% is healthy). Aggregated across WS-5 runs.",
+        ], [
+            "Embed wall time",
+            fmtMs(perf.embedMs),
+            "flat",
+            "Cumulative embedding wall-clock time across consolidation runs in the window.",
+        ], [
+            "Judged-cache skipped",
+            num(perf.judgedCacheSkipped),
+            "flat",
+            "Candidates skipped by the judged-cache (not sent to LLM). Higher = more efficient reuse of prior judgments.",
+        ], [
+            "Dedup pool size",
+            num(perf.dedupPoolSize),
+            "flat",
+            "Average memory pool size after deduplication (before judged-cache narrowing). WS-5 perf telemetry.",
+        ], [
+            "Over-budget consolidation runs",
+            String(perf.overBudgetRuns),
+            perf.overBudgetRuns > 0 ? "down" : "flat",
+            "Runs where consolidation alone exceeded the total run budget (estimatedBudgetFractionUsed > 1.0).",
+        ]);
+    }
+    // WS-5: degradation metrics rows.
+    if (degradation) {
+        summaryRows.push([
+            "Corpus diversity (Gini)",
+            num(degradation.corpusCentroidDistance),
+            degradation.entrenchmentFlagged ? "down" : "flat",
+            "Gini coefficient of retrieval_salience for top-100 ranked assets. High Gini (>0.35) = entrenchment risk.",
+        ], [
+            "Merge fidelity contradiction rate",
+            pct(degradation.mergeFidelityContradictionRate, 1),
+            "flat",
+            "Fraction of consolidated proposals that involved a contradiction, from consolidation result envelopes.",
+        ], [
+            "High-generation fraction",
+            pct(degradation.highGenerationFraction, 1),
+            "flat",
+            "Fraction of assets with consecutive_no_ops >= 2 (proxy for high-generation assets in the salience table).",
+        ]);
+    }
     const summaryRowsHtml = summaryRows
-        .map(([label, value, t]) => `            <tr><td>${esc(label)}</td><td>${esc(value)}</td>` +
-        `<td class="trend ${trendClass(t)}">${trendLabel(t)}</td></tr>`)
+        .map(([label, value, t, tip]) => {
+        const labelHtml = tip ? `<abbr title="${esc(tip)}">${esc(label)}</abbr>` : esc(label);
+        return (`            <tr><td>${labelHtml}</td><td>${esc(value)}</td>` +
+            `<td class="trend ${trendClass(t)}">${trendLabel(t)}</td></tr>`);
+    })
         .join("\n");
-    // ── Advisory cards ─────────────────────────────────────────────────────────
-    const advisoryParts = [];
+    // ── Action Items (merged + de-duplicated advisories + what-to-watch) ────────
+    // Advisories and the old "what to watch" cards were built from the same data
+    // (result.advisories, sem-blocked, proposals, tail-latency, failed-runs). We
+    // collapse them into ONE prioritized, de-duplicated list (P1/P2/P3 + a
+    // remediation command per item), keyed so each concern appears exactly once.
+    const items = [];
+    const seen = new Set();
+    const pushItem = (item) => {
+        if (seen.has(item.key))
+            return;
+        seen.add(item.key);
+        items.push(item);
+    };
+    // Hard advisories from the health check engine (own remediation in message).
     for (const a of result.advisories) {
         if (a.status !== "warn" && a.status !== "fail")
             continue;
-        advisoryParts.push(advisoryCard(a.status === "fail" ? "fail" : "warn", a.status === "fail" ? "🔴" : "⚠️", esc(a.name), esc(a.message)));
+        pushItem({
+            key: `advisory:${a.name}`,
+            prio: a.status === "fail" ? "P1" : "P2",
+            cls: a.status === "fail" ? "fail" : "warn",
+            title: esc(humanize(a.name)),
+            descHtml: esc(a.message),
+        });
+    }
+    // Failed runs in window.
+    if (failedRuns > 0) {
+        pushItem({
+            key: "failed-runs",
+            prio: "P1",
+            cls: "fail",
+            title: `${failedRuns} failed run${failedRuns === 1 ? "" : "s"} in window`,
+            descHtml: `Task fail rate ${esc(taskFailRate)}. Inspect failed runs (ok=false) for early-exit or harness errors.`,
+            remedy: `akm health --since=${opts.window} --group-by run`,
+        });
     }
+    // Semantic search blocked.
     if (sem.blocked) {
-        advisoryParts.push(advisoryCard("warn", "⚠️", "Semantic search blocked", `Embedding provider unreachable. ${esc(sem.detail)}. Curate falls back to keyword search — relevance scoring degraded.`));
+        pushItem({
+            key: "semantic-search-blocked",
+            prio: "P2",
+            cls: "warn",
+            title: "Semantic search blocked",
+            descHtml: `Embedding provider unreachable. ${esc(sem.detail)}. Curate falls back to keyword search — relevance scoring degraded.`,
+            remedy: "akm config show",
+        });
     }
+    // Pending proposals to drain.
     if (proposals.length > 0) {
-        advisoryParts.push(advisoryCard("warn", "⚠️", `${proposals.length} proposals pending (drain needed)`, "Run <code>akm proposal list</code> to review and drain."));
+        const bySource = new Map();
+        for (const p of proposals)
+            bySource.set(p.source, (bySource.get(p.source) ?? 0) + 1);
+        const srcSummary = [...bySource.entries()]
+            .sort(([a], [b]) => a.localeCompare(b))
+            .map(([source, count]) => `${count} via ${esc(source)}`)
+            .join(", ");
+        pushItem({
+            key: "drain-proposals",
+            prio: "P2",
+            cls: "warn",
+            title: `Drain ${proposals.length} pending proposal${proposals.length === 1 ? "" : "s"}`,
+            descHtml: `Proposals generated this batch (${srcSummary}). Review before the queue grows further.`,
+            remedy: "akm proposal list",
+        });
+    }
+    // High tail latency.
+    if (wallTime.p95Ms && wallTime.medianMs && wallTime.p95Ms / wallTime.medianMs > 2.5) {
+        pushItem({
+            key: "tail-latency",
+            prio: "P3",
+            cls: "warn",
+            title: `High tail latency: p95=${fmtMs(wallTime.p95Ms)}, median=${fmtMs(wallTime.medianMs)}`,
+            descHtml: "P95 is well above median. Consolidation/LLM phase dominates wall time on slow runs. " +
+                "Check for slow chunks or LLM rate limiting.",
+        });
+    }
+    // Stale freshness.
+    if (isStale && totalRuns > 0) {
+        pushItem({
+            key: "stale",
+            prio: "P3",
+            cls: "warn",
+            title: "No recent improve runs",
+            descHtml: `Newest run is ${esc(latestRunHuman)} — older than the 6h freshness threshold. Check the improve scheduler/cron.`,
+        });
     }
+    // WS-5: corpus entrenchment flag.
+    if (degradation?.entrenchmentFlagged) {
+        pushItem({
+            key: "corpus-entrenchment",
+            prio: "P2",
+            cls: "warn",
+            title: "Corpus entrenchment risk: retrieval_salience Gini > 0.35",
+            descHtml: "A small set of assets dominates retrieval — retrieval diversity is low. " +
+                "Review top-ranked assets for stale or over-represented content. " +
+                `Corpus diversity proxy: ${esc(String(degradation.corpusCentroidDistance))}.`,
+            remedy: "akm health --format json | jq '.improve.degradation'",
+        });
+    }
+    // WS-5: over-budget consolidation advisory.
+    if (perf.overBudgetRuns > 0) {
+        pushItem({
+            key: "over-budget-consolidation",
+            prio: "P2",
+            cls: "warn",
+            title: `${perf.overBudgetRuns} consolidation run${perf.overBudgetRuns === 1 ? "" : "s"} exceeded budget`,
+            descHtml: "Consolidation phase wall time exceeded the total run budget on these runs. " +
+                "Consider increasing the timeout or reducing the consolidation pool via profile config.",
+            remedy: "akm config show",
+        });
+    }
+    // Session-log notes (informational, lowest priority).
     if (result.sessionLogAdvisories.length > 0) {
         const patterns = result.sessionLogAdvisories
             .slice(0, 6)
             .map((p) => `<li>${esc(p.topic)}</li>`)
             .join("");
-        advisoryParts.push('<div class="advisory" style="border-left:3px solid var(--accent);">' +
-            '<div class="advisory-icon">ℹ️</div><div class="advisory-body">' +
-            `<div class="title">${result.sessionLogAdvisories.length} session-log note(s) (informational)</div>` +
-            `<div class="desc"><ul style="margin:4px 0 0 16px;padding:0;">${patterns}</ul></div></div></div>`);
+        pushItem({
+            key: "session-log-notes",
+            prio: "P3",
+            cls: "warn",
+            title: `${result.sessionLogAdvisories.length} session-log note(s) (informational)`,
+            descHtml: `<ul style="margin:4px 0 0 16px;padding:0;">${patterns}</ul>`,
+        });
     }
-    const advisoryCardsHtml = advisoryParts.length > 0
-        ? advisoryParts.join("\n")
-        : passCard("No active advisories", "All checks passed for this window.");
+    items.sort((a, b) => PRIO_RANK[a.prio] - PRIO_RANK[b.prio]);
+    const actionItemsHtml = items.length > 0
+        ? items.map(actionItemCard).join("\n")
+        : passCard("No action items", "All checks passed and nothing needs attention for this window.");
     // ── Proposal rows ──────────────────────────────────────────────────────────
     const proposalRowsHtml = proposals.length > 0
         ? proposals
             .map((p, i) => {
             const tagCls = p.source === "extract" ? "tag-extract" : "tag-consolidate";
-            const ts = p.createdAt.slice(0, 16).replace("T", " ");
             return (`<tr><td>${i + 1}</td><td><code>${esc(p.ref)}</code></td>` +
                 `<td><span class="tag ${tagCls}">${esc(p.source)}</span></td>` +
-                `<td>${esc(ts)}</td></tr>`);
+                `<td>${isoTimeTag(p.createdAt)}</td></tr>`);
         })
             .join("\n")
         : '<tr><td colspan="4" style="text-align:center;color:var(--muted);">No pending proposals</td></tr>';
-    // ── What to watch ──────────────────────────────────────────────────────────
-    const watchParts = [];
-    for (const a of result.advisories) {
-        if (a.status !== "warn" && a.status !== "fail")
-            continue;
-        const prio = a.status === "fail" ? "P1" : "P2";
-        watchParts.push(advisoryCard(a.status === "fail" ? "fail" : "warn", a.status === "fail" ? "🔴" : "🟡", `${esc(a.name)} (${prio})`, esc(a.message)));
-    }
-    if (sem.blocked) {
-        watchParts.push(advisoryCard("warn", "🟡", "Embedding server unreachable (P2)", "Curate quality and semantic ranking are degraded. Check the embedding endpoint configured in config.json."));
-    }
-    if (proposals.length > 0) {
-        const bySource = new Map();
-        for (const p of proposals)
-            bySource.set(p.source, (bySource.get(p.source) ?? 0) + 1);
-        const srcSummary = [...bySource.entries()]
-            .sort(([a], [b]) => a.localeCompare(b))
-            .map(([source, count]) => `${count} via ${esc(source)}`)
-            .join(", ");
-        watchParts.push(advisoryCard("warn", "🟡", `Drain ${proposals.length} pending proposals (P2)`, `Proposals generated this batch (${srcSummary}). Run <code>akm proposal list</code> before the queue grows further.`));
-    }
-    if (wallTime.p95Ms && wallTime.medianMs && wallTime.p95Ms / wallTime.medianMs > 2.5) {
-        watchParts.push(advisoryCard("warn", "🟡", `High tail latency (P3): p95=${fmtMs(wallTime.p95Ms)}, median=${fmtMs(wallTime.medianMs)}`, "P95 is well above median. Consolidation/LLM phase dominates wall time on slow runs. " +
-            "Check for slow chunks or LLM rate limiting."));
-    }
-    if (failedRuns > 0) {
-        watchParts.push(advisoryCard("warn", "🟡", `${failedRuns} failed run(s) in window (P2)`, `Task fail rate ${taskFailRate}. Inspect failed runs (ok=false) for early-exit or harness errors.`));
-    }
-    const watchItemsHtml = watchParts.length > 0
-        ? watchParts.join("\n")
-        : passCard("Nothing critical to watch", "All indicators are within normal range.");
-    // ── Commands used ──────────────────────────────────────────────────────────
-    const commandsHtml = [
-        `      <div><span>akm health --since=${esc(opts.window)} --group-by run --format json</span></div>`,
-        `      <div><span>akm health --since=${esc(opts.window)} --window-compare=${esc(opts.compare)} --format json</span></div>`,
-        "      <div><span>akm proposal list</span></div>",
-    ].join("\n");
     return {
         "%%ECHARTS_TAG%%": buildEchartsTag(opts),
         "%%REPORT_TITLE%%": esc(reportTitle),
         "%%WINDOW%%": esc(opts.window),
-        "%%SINCE_HUMAN%%": esc(sinceHuman),
+        "%%SINCE_HUMAN%%": sinceHuman,
         "%%RUN_COUNT%%": num(totalRuns),
         "%%STATUS_BADGE_HTML%%": `${statusBadge}\n    ${failBadge}`,
         "%%EXEC_SUMMARY_HTML%%": execSummary,
         "%%KPI_CARDS_HTML%%": kpiCards,
         "%%RUNS_JS_CONST%%": runsJsConst,
         "%%DISTILL_REASONS_JSON%%": JSON.stringify(distillReasons),
+        "%%SLICE_OPTIONS_HTML%%": buildSliceOptions(opts.window),
+        "%%LLM_BY_STAGE_JSON%%": JSON.stringify(llm.byStage ?? {}),
         "%%SUMMARY_ROWS_HTML%%": summaryRowsHtml,
-        "%%ADVISORY_CARDS_HTML%%": advisoryCardsHtml,
+        "%%ACTION_ITEMS_HTML%%": actionItemsHtml,
         "%%PROPOSAL_ROWS_HTML%%": proposalRowsHtml,
         "%%PROPOSAL_COUNT%%": String(proposals.length),
-        "%%WATCH_ITEMS_HTML%%": watchItemsHtml,
-        "%%COMMANDS_HTML%%": commandsHtml,
         "%%GENERATED_AT%%": esc(generatedAt),
+        "%%AKM_VERSION%%": esc(pkgVersion),
     };
 }