From bd5972962d11e523ba726b6e2a206bdd6ce28412 Mon Sep 17 00:00:00 2001
From: sam <sam@samdeMacBook-Air.local>
Date: Fri, 25 Sep 2026 01:27:40 +0800
Subject: [PATCH 1/2] Add allocation-aware Sandbox diagnostics to Dashboard

---
 apps/web/e2e/agents-lifecycle.spec.ts         | 11 ++-
 .../src/features/dashboard/DashboardView.css  | 58 ++++++++++++
 .../features/dashboard/DashboardView.test.tsx |  6 ++
 .../dashboard/RuntimeObservabilityContent.tsx | 32 +++++++
 .../dashboard/RuntimeTrendCharts.test.tsx     | 20 ++---
 .../features/dashboard/RuntimeTrendCharts.tsx | 19 ++--
 .../dashboard/dashboard-model.test.ts         | 51 +++++++++++
 .../src/features/dashboard/dashboard-model.ts | 89 +++++++++++++++++++
 apps/web/src/i18n/locales/en/dashboard.ts     | 10 +++
 apps/web/src/i18n/locales/zh-CN/dashboard.ts  | 10 +++
 contracts/agents-api/runtime-observability.md | 17 +++-
 .../agents-api/system-observability-plan.md   | 79 ++++++++++++++++
 12 files changed, 378 insertions(+), 24 deletions(-)
 create mode 100644 contracts/agents-api/system-observability-plan.md

diff --git a/apps/web/e2e/agents-lifecycle.spec.ts b/apps/web/e2e/agents-lifecycle.spec.ts
index 9de471d89..7b63cb271 100644
--- a/apps/web/e2e/agents-lifecycle.spec.ts
+++ b/apps/web/e2e/agents-lifecycle.spec.ts
@@ -2684,6 +2684,11 @@ test("renders Runtime telemetry as visual snapshot panels with details on demand
   const dashboard = page.locator(".dashboard-page");
   await expect(dashboard.getByRole("heading", { name: "Dashboard", exact: true })).toBeVisible();
   const refresh = dashboard.getByRole("button", { name: "Refresh Dashboard snapshot" });
+  const sandboxDiagnostics = dashboard.getByRole("region", { name: "Sandbox diagnostics" });
+  await expect(sandboxDiagnostics).toBeVisible();
+  await expect(sandboxDiagnostics).toContainText("12 managed targets");
+  await expect(sandboxDiagnostics).toContainText("12 with measured usage and limit");
+  await expect(sandboxDiagnostics).toContainText("Highest memory pressure");
   await expect(dashboard.locator(".dashboard-runtime-sample-count")).toContainText("1 sample ·");
   await expect(dashboard.getByRole("heading", { name: "CPU usage" })).toBeVisible();
   await expect(dashboard.getByRole("heading", { name: "Memory usage" })).toBeVisible();
@@ -2735,7 +2740,7 @@ test("renders Runtime telemetry as visual snapshot panels with details on demand
   await expect(cpuCard.locator(".dashboard-runtime-trend-tooltip")).toBeVisible();
   await cpuChart.click({ position: { x: 260, y: 90 } });
   await expect(cpuCard.locator(".dashboard-runtime-trend-tooltip")).toContainText("Pinned");
-  await expect(cpuCard.locator(".dashboard-runtime-trend-tooltip")).toContainText("10%");
+  await expect(cpuCard.locator(".dashboard-runtime-trend-tooltip")).toContainText("Unavailable");
   await cpuChart.focus();
   await cpuChart.press("ArrowRight");
   await expect(cpuCard.locator(".dashboard-runtime-trend-tooltip")).toContainText("Pinned");
@@ -2930,9 +2935,9 @@ test("restores retained Runtime history after a Dashboard reload", async ({ page
   const durableCpuCard = durableCpuChart.locator("xpath=ancestor::section[contains(@class, 'dashboard-runtime-trend-card')]");
   await durableCpuChart.focus();
   await durableCpuChart.press("ArrowLeft");
-  await expect(durableCpuCard.locator(".dashboard-runtime-trend-tooltip")).toContainText("0%");
+  await expect(durableCpuCard.locator(".dashboard-runtime-trend-tooltip")).toContainText("Unavailable");
   await durableCpuChart.press("ArrowLeft");
-  await expect(durableCpuCard.locator(".dashboard-runtime-trend-tooltip")).toContainText("0%");
+  await expect(durableCpuCard.locator(".dashboard-runtime-trend-tooltip")).toContainText("Unavailable");
   const durableMemoryCard = dashboard.getByRole("region", { name: "Memory usage durable history chart" });
   const durableMemorySpan = await durableMemoryCard.locator("canvas").evaluate((canvas: HTMLCanvasElement) => {
     const context = canvas.getContext("2d");
diff --git a/apps/web/src/features/dashboard/DashboardView.css b/apps/web/src/features/dashboard/DashboardView.css
index 994880b84..5364e670c 100644
--- a/apps/web/src/features/dashboard/DashboardView.css
+++ b/apps/web/src/features/dashboard/DashboardView.css
@@ -633,6 +633,64 @@
   border-bottom: 1px solid var(--line);
 }
 
+.dashboard-sandbox-insights {
+  padding: 16px;
+  border-bottom: 1px solid var(--line);
+}
+
+.dashboard-sandbox-insights > header,
+.dashboard-sandbox-diagnostics {
+  display: flex;
+  justify-content: space-between;
+  gap: 16px;
+}
+
+.dashboard-sandbox-insights h3,
+.dashboard-sandbox-insights h4,
+.dashboard-sandbox-insights p {
+  margin: 0;
+}
+
+.dashboard-sandbox-insights h3 { font-size: 14px; }
+.dashboard-sandbox-insights h4 { font-size: 12px; }
+.dashboard-sandbox-insights p,
+.dashboard-sandbox-insights small { color: var(--fg-muted); font-size: 11px; }
+
+.dashboard-sandbox-insight-grid {
+  display: grid;
+  grid-template-columns: repeat(4, minmax(0, 1fr));
+  gap: 8px;
+  margin-top: 12px;
+}
+
+.dashboard-sandbox-insight-grid > div {
+  display: grid;
+  gap: 3px;
+  padding: 10px 12px;
+  border: 1px solid var(--line);
+  border-radius: 8px;
+  min-width: 0;
+}
+
+.dashboard-sandbox-insight-grid strong {
+  font: 550 20px/1.2 var(--font-mono);
+  font-variant-numeric: tabular-nums;
+}
+
+.dashboard-sandbox-insight-grid span { color: var(--fg-muted); font-size: 10px; }
+.dashboard-sandbox-diagnostics { margin-top: 14px; }
+.dashboard-sandbox-diagnostics > div { flex: 1; min-width: 0; }
+.dashboard-sandbox-diagnostics ul { list-style: none; margin: 7px 0 0; padding: 0; }
+.dashboard-sandbox-diagnostics li { display: flex; justify-content: space-between; gap: 8px; padding: 4px 0; font-size: 11px; }
+.dashboard-sandbox-diagnostics li + li { border-top: 1px solid var(--line); }
+.dashboard-sandbox-diagnostics li button { padding: 0; border: 0; background: none; color: var(--accent); cursor: pointer; text-align: left; }
+.dashboard-sandbox-diagnostics li span { overflow: hidden; text-overflow: ellipsis; white-space: nowrap; }
+
+@media (max-width: 800px) {
+  .dashboard-sandbox-insight-grid { grid-template-columns: repeat(2, minmax(0, 1fr)); }
+  .dashboard-sandbox-diagnostics { flex-direction: column; }
+}
+
 .dashboard-runtime-metric {
   display: grid;
   min-width: 0;
diff --git a/apps/web/src/features/dashboard/DashboardView.test.tsx b/apps/web/src/features/dashboard/DashboardView.test.tsx
index 39f3a23e0..4ad132758 100644
--- a/apps/web/src/features/dashboard/DashboardView.test.tsx
+++ b/apps/web/src/features/dashboard/DashboardView.test.tsx
@@ -247,6 +247,12 @@ describe("Dashboard loaded-result presentation", () => {
     expect(html).toContain("Cumulative CPU / capacity");
     expect(html).toContain("1m 13s / 2 cores");
     expect(html).toContain("512 MiB / 2.00 GiB");
+    expect(html).toContain("Sandbox diagnostics");
+    expect(html).toContain("Memory ≥80%");
+    const withoutMemory = render({
+      runtimeSnapshot: { sessions: [hosted], observations: [{ ...observation, memory: { usage_bytes: null, limit_bytes: null } }], loadedAt: 1_700_000_100_000 },
+    });
+    expect(withoutMemory).toMatch(/Memory ≥80%<\/small><strong>Unavailable<\/strong>/);
     expect(html).toContain('aria-label="Runtime live-window charts"');
     expect(html).toContain("Resource trends");
     expect(html).toContain("Browser-local samples · reset on reload");
diff --git a/apps/web/src/features/dashboard/RuntimeObservabilityContent.tsx b/apps/web/src/features/dashboard/RuntimeObservabilityContent.tsx
index 2c39d9d37..f40f5e62e 100644
--- a/apps/web/src/features/dashboard/RuntimeObservabilityContent.tsx
+++ b/apps/web/src/features/dashboard/RuntimeObservabilityContent.tsx
@@ -26,6 +26,7 @@ import {
 
 import {
   buildRuntimeDashboardModel,
+  buildSandboxInsights,
   formatDashboardBytes,
   formatDashboardDuration,
   formatDashboardTimestamp,
@@ -312,6 +313,9 @@ export function RuntimeObservabilityContent({
     [snapshot, tokenTotals],
   );
   const summary = model.summary;
+  const sandbox = useMemo(() => buildSandboxInsights(model.rows), [model.rows]);
+  const unavailableReasons = Object.entries(sandbox.unavailableReasons)
+    .sort((left, right) => right[1] - left[1]);
   return (
     <>
       <div className="dashboard-runtime-summary" aria-label={t("runtime.resourceSnapshot")}>
@@ -321,6 +325,34 @@ export function RuntimeObservabilityContent({
         <RuntimeMetric icon={<Gauge size={17} />} label={t("runtime.metrics.tokens")} value={summary.totalTokens === null ? t("runtime.filters.unavailable") : formatDashboardTokens(summary.totalTokens, locale)} detail={t("runtime.metrics.tokenDetail", { covered: summary.tokenCoverageCount, total: summary.sessionCount })} />
       </div>
 
+      <section className="dashboard-sandbox-insights" aria-labelledby="dashboard-sandbox-insights-title">
+        <header>
+          <div>
+            <h3 id="dashboard-sandbox-insights-title">{t("sandbox.title")}</h3>
+            <p>{t("sandbox.subtitle")}</p>
+          </div>
+          <small>{t(stale ? "sandbox.retained" : "sandbox.current")}</small>
+        </header>
+        <div className="dashboard-sandbox-insight-grid">
+          <div><small>{t("sandbox.observed")}</small><strong>{sandbox.observed.toLocaleString(locale)}</strong><span>{t("sandbox.observedDetail", { total: summary.managedRuntimeCount })}</span></div>
+          <div><small>{t("sandbox.unavailable")}</small><strong>{sandbox.unavailable.toLocaleString(locale)}</strong><span>{t("sandbox.unavailableDetail")}</span></div>
+          <div><small>{t("sandbox.highMemory")}</small><strong>{sandbox.measuredMemory === 0 ? t("runtime.filters.unavailable") : sandbox.highMemory.toLocaleString(locale)}</strong><span>{t("sandbox.highMemoryDetail", { total: sandbox.measuredMemory })}</span></div>
+          <div><small>{t("sandbox.parked")}</small><strong>{(sandbox.sleeping + sandbox.pending).toLocaleString(locale)}</strong><span>{t("sandbox.parkedDetail", { sleeping: sandbox.sleeping, pending: sandbox.pending })}</span></div>
+        </div>
+        {unavailableReasons.length > 0 || sandbox.highestMemory.length > 0 ? (
+          <div className="dashboard-sandbox-diagnostics">
+            <div>
+              <h4>{t("sandbox.sampleGaps")}</h4>
+              {unavailableReasons.length ? <ul>{unavailableReasons.map(([reason, count]) => <li key={reason}><span>{t(`runtime.status.${reason}` as never)}</span><strong>{count}</strong></li>)}</ul> : <p>{t("sandbox.noGaps")}</p>}
+            </div>
+            <div>
+              <h4>{t("sandbox.memoryLeaders")}</h4>
+              {sandbox.highestMemory.length ? <ul>{sandbox.highestMemory.map((entry) => <li key={entry.allocationId}><button type="button" onClick={() => onOpenSession(entry.sessionId)}>{entry.title}</button><span title={`${formatDashboardBytes(entry.memoryUsageBytes)} / ${formatDashboardBytes(entry.memoryLimitBytes)}`}>{entry.memoryPercent.toLocaleString(locale, { maximumFractionDigits: 1 })}%</span></li>)}</ul> : <p>{t("sandbox.noMemory")}</p>}
+            </div>
+          </div>
+        ) : null}
+      </section>
+
       <RuntimeTrendPanel snapshot={snapshot} stale={stale} loadRuntimeHistory={loadRuntimeHistory} />
 
       <details className="dashboard-runtime-explorer">
diff --git a/apps/web/src/features/dashboard/RuntimeTrendCharts.test.tsx b/apps/web/src/features/dashboard/RuntimeTrendCharts.test.tsx
index eeba94ed4..000f8c416 100644
--- a/apps/web/src/features/dashboard/RuntimeTrendCharts.test.tsx
+++ b/apps/web/src/features/dashboard/RuntimeTrendCharts.test.tsx
@@ -47,7 +47,7 @@ describe("Runtime live-window chart accessibility", () => {
     })).toBe("Memory usage all series hidden; use the legend to show a series");
   });
 
-  it("renders an unavailable current value as zero without retaining a stale value", () => {
+  it("renders an unavailable current value as a gap without retaining a stale value", () => {
     const unavailable = {
       ...sample(120_000, null),
       activeSandboxCount: 0,
@@ -58,13 +58,13 @@ describe("Runtime live-window chart accessibility", () => {
       <RuntimeTrendCharts samples={[sample(60_000, .5), unavailable]} />,
     );
 
-    expect(html).toContain("Runtime worker</th><td>0%</td><td>0</td>");
+    expect(html).toContain("Runtime worker</th><td>Unavailable</td><td>1</td>");
     expect(html).not.toContain("Runtime worker</th><td>50%</td><td>1</td>");
-    expect(html).toContain("used</th><td>0 B</td><td>0</td>");
+    expect(html).toContain("used</th><td>Unavailable</td><td>1</td>");
     expect(html).toContain("active</th><td>0</td><td>0</td>");
   });
 
-  it("renders empty retained buckets as continuous zero-value chart series", () => {
+  it("renders empty retained buckets as missing measurements", () => {
     const empty = (sampledAt: number): RuntimeTrendSample => ({
       ...sample(sampledAt, null),
       targets: [],
@@ -78,13 +78,13 @@ describe("Runtime live-window chart accessibility", () => {
       <RuntimeTrendCharts samples={[empty(60_000), empty(120_000)]} source="durable" />,
     );
 
-    expect(html).toContain("usage</th><td>0%</td><td>0</td>");
-    expect(html).toContain("used</th><td>0 B</td><td>0</td>");
+    expect(html).toContain("usage</th><td>Unavailable</td><td>2</td>");
+    expect(html).toContain("used</th><td>Unavailable</td><td>2</td>");
     expect(html).toContain("active</th><td>0</td><td>0</td>");
-    expect(html).toContain("input</th><td>0/min</td><td>0</td>");
-    expect(html).not.toContain("No retained CPU samples");
-    expect(html).not.toContain("No complete retained memory samples");
-    expect(html).not.toContain("No retained token samples");
+    expect(html).toContain("input</th><td>Unavailable</td><td>2</td>");
+    expect(html).toContain("No retained CPU samples");
+    expect(html).toContain("No retained observed memory samples");
+    expect(html).toContain("No retained token samples");
   });
 
   it("renders uPlot chart mounts and reports trends only after two real samples", () => {
diff --git a/apps/web/src/features/dashboard/RuntimeTrendCharts.tsx b/apps/web/src/features/dashboard/RuntimeTrendCharts.tsx
index c85f06857..d21df864e 100644
--- a/apps/web/src/features/dashboard/RuntimeTrendCharts.tsx
+++ b/apps/web/src/features/dashboard/RuntimeTrendCharts.tsx
@@ -537,13 +537,16 @@ export function RuntimeTrendCharts({
       id,
       label: targetLabel(samples, id),
       tone: tones[index] ?? "blue",
-      points: samples.map((sample) => ({ sampledAt: sample.sampledAt, value: (sample.targets.find((target) => target.seriesId === id)?.cpuRatio ?? 0) * 100 })),
+      points: samples.map((sample) => {
+        const ratio = sample.targets.find((target) => target.seriesId === id)?.cpuRatio;
+        return { sampledAt: sample.sampledAt, value: ratio == null ? null : ratio * 100 };
+      }),
     }));
     if (cpu.length === 0 && samples.length > 0) {
-      cpu.push({ id: "cpu", label: t("charts.usage"), tone: "orange", points: samples.map((sample) => ({ sampledAt: sample.sampledAt, value: 0 })) });
+      cpu.push({ id: "cpu", label: t("charts.usage"), tone: "orange", points: samples.map((sample) => ({ sampledAt: sample.sampledAt, value: null })) });
     }
-    const memoryUsed = samples.map((sample) => ({ sampledAt: sample.sampledAt, value: sample.memoryUsageBytes ?? 0 }));
-    const memoryLimit = samples.map((sample) => ({ sampledAt: sample.sampledAt, value: sample.memoryLimitBytes ?? 0 }));
+    const memoryUsed = samples.map((sample) => ({ sampledAt: sample.sampledAt, value: sample.memoryUsageBytes }));
+    const memoryLimit = samples.map((sample) => ({ sampledAt: sample.sampledAt, value: sample.memoryLimitBytes }));
     const active = [{
       id: "active",
       label: t(activeDisplay === "binary" ? "charts.runtime" : "charts.active.series"),
@@ -552,8 +555,8 @@ export function RuntimeTrendCharts({
       points: samples.map((sample) => ({
         sampledAt: sample.sampledAt,
         value: activeDisplay === "binary"
-          ? (sample.activeSandboxCount ?? 0) > 0 ? 1 : 0
-          : sample.activeSandboxCount ?? 0,
+          ? sample.activeSandboxCount === null ? null : sample.activeSandboxCount > 0 ? 1 : 0
+          : sample.activeSandboxCount,
       })),
     }] satisfies TrendSeries[];
     const throughput = tokenThroughput(samples);
@@ -565,8 +568,8 @@ export function RuntimeTrendCharts({
       ] satisfies TrendSeries[],
       active,
       tokens: [
-        { id: "input", label: t("charts.input"), tone: "orange", points: throughput.map((sample) => ({ sampledAt: sample.sampledAt, value: sample.inputPerMinute ?? 0 })) },
-        { id: "output", label: t("charts.output"), tone: "green", points: throughput.map((sample) => ({ sampledAt: sample.sampledAt, value: sample.outputPerMinute ?? 0 })) },
+        { id: "input", label: t("charts.input"), tone: "orange", points: throughput.map((sample) => ({ sampledAt: sample.sampledAt, value: sample.inputPerMinute })) },
+        { id: "output", label: t("charts.output"), tone: "green", points: throughput.map((sample) => ({ sampledAt: sample.sampledAt, value: sample.outputPerMinute })) },
       ] satisfies TrendSeries[],
     };
   }, [activeDisplay, samples, t]);
diff --git a/apps/web/src/features/dashboard/dashboard-model.test.ts b/apps/web/src/features/dashboard/dashboard-model.test.ts
index 349e6b3ae..8288f1698 100644
--- a/apps/web/src/features/dashboard/dashboard-model.test.ts
+++ b/apps/web/src/features/dashboard/dashboard-model.test.ts
@@ -5,6 +5,7 @@ import type { AgentSession, RuntimeObservation, SavedAgent, TokenUsage } from "@
 import {
   buildDashboardSnapshot,
   buildRuntimeDashboardModel,
+  buildSandboxInsights,
   dashboardEnvironmentLabel,
   dashboardEnvironmentProfile,
   dashboardStatusLabel,
@@ -355,6 +356,56 @@ describe("Dashboard loaded-snapshot model", () => {
       totalTokens: 26,
       tokenCoverageCount: 2,
     });
+    expect(buildSandboxInsights(model.rows)).toMatchObject({
+      observed: 1,
+      unavailable: 0,
+      highMemory: 0,
+      measuredMemory: 1,
+      highestMemory: [{ sessionId: second.id, memoryPercent: 37.5 }],
+    });
+  });
+
+  it("separates sleeping allocations from unexpected sampling gaps", () => {
+    const base = session("11111111-1111-4111-8111-111111111111");
+    const observation: RuntimeObservation = {
+      id: base.id, object: "agent.runtime_observation", session_id: base.id,
+      environment_id: "33333333-3333-4333-8333-333333333333", mode: "openai_hosted",
+      provider_type: "docker",
+      instance: { kind: "managed_allocation", allocation_id: "allocation-1", device_id: null, connection_generation: null },
+      lifecycle_state: "sleeping", status: "unavailable", reason: "runtime_not_running",
+      allocation_created_at: 100, resolved_at: 220, observed_at: null, started_at: null,
+      cpu: null, memory: null,
+    };
+    const failed = { ...observation, id: "other", session_id: "other", instance: { ...observation.instance, allocation_id: "allocation-2" }, lifecycle_state: "active" as const, reason: "sample_timeout" as const };
+    const rows = buildRuntimeDashboardModel([base, session("other")], [observation, failed]).rows;
+    expect(buildSandboxInsights(rows)).toMatchObject({
+      sleeping: 1, pending: 0, unavailable: 1,
+      unavailableReasons: { sample_timeout: 1 },
+    });
+  });
+
+  it("does not pick a conflicting same-second allocation observation by list order", () => {
+    const first = session("first");
+    const second = session("second");
+    const base: RuntimeObservation = {
+      id: first.id, object: "agent.runtime_observation", session_id: first.id,
+      environment_id: "33333333-3333-4333-8333-333333333333", mode: "openai_hosted",
+      provider_type: "docker", instance: { kind: "managed_allocation", allocation_id: "allocation-1", device_id: null, connection_generation: null },
+      lifecycle_state: "active", status: "observed", reason: null,
+      allocation_created_at: 100, resolved_at: 220, observed_at: 210, started_at: 150,
+      cpu: null, memory: { usage_bytes: 900, limit_bytes: 1000 },
+    };
+    const conflict: RuntimeObservation = {
+      ...base, id: second.id, session_id: second.id, status: "unavailable", reason: "sample_timeout",
+      observed_at: null, started_at: null, cpu: null, memory: null,
+    };
+    const insights = (observations: RuntimeObservation[]) => buildSandboxInsights(buildRuntimeDashboardModel([first, second], observations).rows);
+    for (const order of [[base, conflict], [conflict, base]]) {
+      expect(insights(order)).toMatchObject({
+        observed: 0, unavailable: 1, highMemory: 0, measuredMemory: 0,
+        unavailableReasons: { sample_unavailable: 1 },
+      });
+    }
   });
 
   it("holds each Session's last reported tokens in the summary while public usage is null", () => {
diff --git a/apps/web/src/features/dashboard/dashboard-model.ts b/apps/web/src/features/dashboard/dashboard-model.ts
index 9f506aef7..7fd967555 100644
--- a/apps/web/src/features/dashboard/dashboard-model.ts
+++ b/apps/web/src/features/dashboard/dashboard-model.ts
@@ -81,6 +81,26 @@ export interface RuntimeDashboardModel {
   rows: RuntimeDashboardRow[];
 }
 
+export interface SandboxInsight {
+  allocationId: string;
+  sessionId: string;
+  title: string;
+  memoryPercent: number;
+  memoryUsageBytes: number;
+  memoryLimitBytes: number;
+}
+
+export interface SandboxInsights {
+  observed: number;
+  unavailable: number;
+  pending: number;
+  sleeping: number;
+  highMemory: number;
+  measuredMemory: number;
+  unavailableReasons: Partial<Record<NonNullable<RuntimeObservation["reason"]>, number>>;
+  highestMemory: SandboxInsight[];
+}
+
 const sessionStatuses = new Set<SessionStatus>([
   "idle",
   "in_progress",
@@ -475,6 +495,75 @@ export function buildRuntimeDashboardModel(
   };
 }
 
+/** Current managed allocations only; duplicate Session observations share one allocation. */
+export function buildSandboxInsights(rows: readonly RuntimeDashboardRow[]): SandboxInsights {
+  const byAllocation = new Map<string, RuntimeDashboardRow[]>();
+  let unallocatedPending = 0;
+  for (const row of rows) {
+    const allocationId = row.observation.instance.allocation_id;
+    if (row.observation.mode !== "openai_hosted") continue;
+    if (!allocationId) {
+      if (row.observation.lifecycle_state === "pending") unallocatedPending += 1;
+      continue;
+    }
+    const previous = byAllocation.get(allocationId);
+    if (previous) previous.push(row);
+    else byAllocation.set(allocationId, [row]);
+  }
+  const insights: SandboxInsights = {
+    observed: 0, unavailable: 0, pending: unallocatedPending, sleeping: 0,
+    highMemory: 0, measuredMemory: 0, unavailableReasons: {}, highestMemory: [],
+  };
+  for (const [allocationId, candidates] of byAllocation) {
+    const newestTime = candidates.reduce((latest, candidate) => Math.max(latest, candidate.observation.resolved_at), 0);
+    const newest = candidates.filter((candidate) => candidate.observation.resolved_at === newestTime);
+    const signature = (candidate: RuntimeDashboardRow) => {
+      const value = candidate.observation;
+      return JSON.stringify([value.lifecycle_state, value.status, value.reason, value.observed_at, value.memory]);
+    };
+    // Public timestamps have second resolution. Conflicting observations in one
+    // second have no reliable order, so do not publish a pressure measurement.
+    if (new Set(newest.map(signature)).size > 1) {
+      insights.unavailable += 1;
+      insights.unavailableReasons.sample_unavailable = (insights.unavailableReasons.sample_unavailable ?? 0) + 1;
+      continue;
+    }
+    const row = newest.sort((left, right) => left.observation.session_id.localeCompare(right.observation.session_id))[0]!;
+    const observation = row.observation;
+    if (observation.lifecycle_state === "pending") insights.pending += 1;
+    if (observation.lifecycle_state === "sleeping") insights.sleeping += 1;
+    if (observation.status === "unavailable") {
+      // Parked and stopped allocations are not sampling failures.
+      if (observation.lifecycle_state !== "sleeping" && observation.lifecycle_state !== "pending" && observation.lifecycle_state !== "stopped") {
+        insights.unavailable += 1;
+        if (observation.reason) {
+          insights.unavailableReasons[observation.reason] = (insights.unavailableReasons[observation.reason] ?? 0) + 1;
+        }
+      }
+      continue;
+    }
+    if (observation.status !== "observed") continue;
+    insights.observed += 1;
+    const usage = safeNonNegativeInteger(observation.memory?.usage_bytes);
+    const limit = safeNonNegativeInteger(observation.memory?.limit_bytes);
+    if (usage === null || limit === null || limit === 0) continue;
+    insights.measuredMemory += 1;
+    const memoryPercent = usage / limit * 100;
+    if (memoryPercent >= 80) insights.highMemory += 1;
+    insights.highestMemory.push({
+      allocationId,
+      sessionId: row.observation.session_id,
+      title: row.session.title,
+      memoryPercent,
+      memoryUsageBytes: usage,
+      memoryLimitBytes: limit,
+    });
+  }
+  insights.highestMemory.sort((left, right) => right.memoryPercent - left.memoryPercent || left.allocationId.localeCompare(right.allocationId));
+  insights.highestMemory = insights.highestMemory.slice(0, 5);
+  return insights;
+}
+
 export function formatDashboardBytes(value: number | null): string {
   if (value === null) return "Unavailable";
   const units = ["B", "KiB", "MiB", "GiB", "TiB"];
diff --git a/apps/web/src/i18n/locales/en/dashboard.ts b/apps/web/src/i18n/locales/en/dashboard.ts
index 449e5a735..4415511ff 100644
--- a/apps/web/src/i18n/locales/en/dashboard.ts
+++ b/apps/web/src/i18n/locales/en/dashboard.ts
@@ -1,4 +1,14 @@
 export const dashboard = {
+  sandbox: {
+    title: "Sandbox diagnostics", subtitle: "Current allocation identity, sampling coverage, and memory pressure",
+    current: "Current complete snapshot", retained: "Last complete snapshot · refresh failed",
+    observed: "Observed", observedDetail: "of {{total}} managed targets",
+    unavailable: "Sampling gaps", unavailableDetail: "Excludes sleeping, pending and stopped",
+    highMemory: "Memory ≥80%", highMemoryDetail: "of {{total}} with measured usage and limit",
+    parked: "Parked", parkedDetail: "{{sleeping}} sleeping · {{pending}} pending",
+    sampleGaps: "Why samples are missing", noGaps: "No unexpected sampling gaps",
+    memoryLeaders: "Highest memory pressure", noMemory: "No measured memory and limit",
+  },
   runtime: {
     explorer: "Runtime target explorer", targets: "Runtime targets", resourceSnapshot: "Runtime resource snapshot",
     search: "Search Runtime targets", searchPlaceholder: "Search Session, provider, or identity", visible: "{{value}} visible",
diff --git a/apps/web/src/i18n/locales/zh-CN/dashboard.ts b/apps/web/src/i18n/locales/zh-CN/dashboard.ts
index 995f1720b..7c9ca0cf5 100644
--- a/apps/web/src/i18n/locales/zh-CN/dashboard.ts
+++ b/apps/web/src/i18n/locales/zh-CN/dashboard.ts
@@ -1,4 +1,14 @@
 export const dashboard = {
+  sandbox: {
+    title: "Sandbox 诊断", subtitle: "按分配身份查看采样覆盖、生命周期与内存压力",
+    current: "当前完整快照", retained: "上次完整快照 · 刷新失败",
+    observed: "已观测", observedDetail: "共 {{total}} 个托管目标",
+    unavailable: "采样缺口", unavailableDetail: "不含休眠、待分配和已停止状态",
+    highMemory: "内存 ≥80%", highMemoryDetail: "共 {{total}} 个具备用量和上限采样",
+    parked: "暂停/等待", parkedDetail: "{{sleeping}} 个休眠 · {{pending}} 个等待分配",
+    sampleGaps: "采样缺失原因", noGaps: "没有意外采样缺口",
+    memoryLeaders: "内存压力最高", noMemory: "暂无内存用量及上限采样",
+  },
   runtime: {
     explorer: "Runtime 目标浏览器", targets: "Runtime 目标", resourceSnapshot: "Runtime 资源快照",
     search: "搜索 Runtime 目标", searchPlaceholder: "搜索会话、Provider 或标识", visible: "显示 {{value}} 项",
diff --git a/contracts/agents-api/runtime-observability.md b/contracts/agents-api/runtime-observability.md
index 143c45b3b..1962abb51 100644
--- a/contracts/agents-api/runtime-observability.md
+++ b/contracts/agents-api/runtime-observability.md
@@ -83,9 +83,20 @@ Web projects active Runtime state differently by scope. The Dashboard shows one
 summed series of distinct allocation identities: live snapshots count
 `lifecycle_state: active`, while retained buckets count successfully observed
 allocations because lifecycle state is not retained yet. The single-Session view
-collapses the same value to `1` or `0`. Missing or unavailable retained values are
-currently rendered as zero, so this presentation intentionally does not yet
-distinguish sleeping from collection failure.
+collapses the same value to `1` or `0`. Empty retained buckets have no active
+observation; an observed zero in a current lifecycle snapshot remains zero.
+Retained history does not distinguish sleeping from collection failure because it
+does not retain lifecycle state.
+
+The Dashboard's current Sandbox diagnostics deduplicate managed observations by
+allocation ID, report sampling gaps separately from sleeping, pending and stopped
+allocations, and rank measured memory usage against a known limit. A target with
+missing usage or limit has no pressure percentage. Live and retained chart gaps
+remain null rather than fabricated zero; an observed lifecycle count of zero is
+still zero. Conflicting observations of one allocation with the same public
+second-resolution timestamp count as an ambiguous sample gap. The broader
+request, model and tool collection plan is documented in
+[system observability](system-observability-plan.md).
 
 Future automatic suspension requires a separate durable control model, including
 an activity revision and timestamps such as `idle_since` and
diff --git a/contracts/agents-api/system-observability-plan.md b/contracts/agents-api/system-observability-plan.md
new file mode 100644
index 000000000..1f59827c4
--- /dev/null
+++ b/contracts/agents-api/system-observability-plan.md
@@ -0,0 +1,79 @@
+# System observability collection plan
+
+Status: collection inventory and next-phase design. The current Dashboard release
+uses existing Runtime observations and Session Usage only. Request, model, and tool
+panels must not be shown as measured until the instruments below are implemented.
+
+## Existing evidence
+
+| Signal | Source | Current coverage | Dashboard meaning |
+| --- | --- | --- | --- |
+| Sandbox lifecycle | Core allocation state | Managed Docker and microsandbox | Current control state, not execution success |
+| CPU time and capacity | Provider observation | Current and periodic history when sampling is available | Cumulative time; utilization needs two samples from one compute incarnation |
+| Memory usage and limit | Provider observation | Current and periodic history when reported | Point-in-time guest/container memory, not host memory |
+| Token input/output | Measured Session Usage | Public current Session and periodic Runtime history | Reported canonical usage; absent usage remains unknown |
+| Request count, latency, errors | No common metric | Not available | Do not derive from Session count or HTTP page loads |
+| Model distribution and latency | Saved Agent model is available, terminal call records are not aggregated | Configuration only | Do not label configured models as invoked models |
+| Tool calls and failures | Durable Turn events exist, no bounded aggregate read | Partial event evidence | Do not count tools by Agent declarations |
+| Host, database, queue health | No tenant-safe operator projection | Not available | Keep out of Session-scoped charts |
+
+Current Runtime history has a 30-second default periodic collection interval and
+seven-day retention; the public Web ranges are 1h, 6h, and 24h. Its capability
+route distinguishes periodic collection from on-read collection. The Dashboard
+must show the source, age, coverage and missing values. An empty interval is not
+an observed zero.
+
+## Next collection boundary
+
+Add a separate, authenticated operator metrics service behind Core. Keep the
+pinned Agents API resources unchanged. The service should aggregate sanitized
+events server-side and expose a bounded read-only extension through
+`packages/agents-client`. Browser input may select a fixed range and resolution;
+it must not select tenants, storage labels, arbitrary PromQL, endpoints or keys.
+
+1. **Ingress:** count completed HTTP requests by route family, method and coarse
+   outcome (`success`, `client_error`, `server_error`). Record a latency histogram
+   after the response completes. Exclude health polling or show it separately.
+   This is transport health, never terminal Turn success.
+2. **Execution:** emit one terminal Turn outcome from the durable state transition
+   winner, with queue wait and run duration measured from persisted timestamps.
+   Active Turns are a current gauge from Core ownership, not a counter inferred
+   from sampled CPU. Retries and recovery must not double count terminal outcomes.
+3. **Model:** observe actual native model invocation attempts at the harness
+   boundary, including model identifier, terminal attempt outcome, duration,
+   reported input/output tokens and usage coverage. A saved Agent's configured
+   model is not proof of which model ran. Keep bounded model label cardinality.
+4. **Tools:** observe actual tool attempt start and terminal result at the common
+   Runtime adapter boundary. Use a bounded tool category and outcome label;
+   raw tool names, arguments, output and credentials stay out of metric labels.
+   Distinguish model tool selection from completed tool execution.
+5. **Sandbox:** retain allocation identity and compute generation internally.
+   Extend the normalized sample only after Docker and microsandbox values have
+   matching semantics. Candidates are OOM/exit events, disk usage/limit, network
+   bytes, compute restarts, and allocation/compute startup latency. A missing
+   provider value remains null; lifecycle state remains Core-owned.
+6. **Collector health:** count attempted, successful, timed-out and unavailable
+   samples, queue drops and export failures. Display coverage per interval so
+   a quiet chart cannot hide collector failure.
+
+Use bounded histograms for p50/p95 latency and rates from counters over complete
+time buckets. Keep operational cardinality to route family, outcome, provider
+type and bounded model family. Session, allocation, tenant, tool name and native
+identifiers must not become general metrics labels. Drill-down can use authorized
+Core resource IDs through existing tenant-scoped reads. Retention and query limits
+must be explicit, with history storage/export optional for execution.
+
+## Dashboard layout and acceptance
+
+The overview keeps loaded Agent/Session counts and current attention. The Runtime
+section shows allocation-deduplicated lifecycle, coverage, sample gaps, memory
+pressure and existing Live/History trends. A later System section can add request
+rate/error rate/p95, terminal Turn outcomes and durations, actual model usage and
+tool attempts once their collection is qualified. Every panel needs a source and
+freshness label, a coverage denominator, and an unavailable state. It must never
+convert missing measurements to zero or equate a configured Sandbox with a
+completed execution.
+
+Before shipping those new panels, validate counter deduplication under retries,
+provider restarts, missing usage, sampler outages and multi-tenant authorization;
+verify that instrumentation has no effect on dispatch, lifecycle or settlement.

From c4bea994c52da4a9740a24e1e593efb60bf9289f Mon Sep 17 00:00:00 2001
From: sam <sam@samdeMacBook-Air.local>
Date: Fri, 25 Sep 2026 12:28:39 +0800
Subject: [PATCH 2/2] Build operator observability dashboard and telemetry

---
 CONTRIBUTING.md                               |  10 ++
 apps/web/e2e/agents-lifecycle.spec.ts         |  38 +++-
 apps/web/e2e/sandbox-setup.spec.ts            |   4 +-
 apps/web/src/App.tsx                          |   1 +
 .../src/features/dashboard/DashboardView.css  | 101 +++++++++++
 .../features/dashboard/DashboardView.test.tsx |   8 +-
 .../src/features/dashboard/DashboardView.tsx  |  58 +++++--
 .../dashboard/SystemObservabilityContent.tsx  | 140 +++++++++++++++
 .../features/dashboard/SystemRequestChart.tsx |  53 ++++++
 apps/web/src/i18n/locales/en/dashboard.ts     |  16 ++
 apps/web/src/i18n/locales/en/pages.ts         |   4 +
 apps/web/src/i18n/locales/zh-CN/dashboard.ts  |  16 ++
 apps/web/src/i18n/locales/zh-CN/pages.ts      |   4 +
 apps/web/vite.config.ts                       |   1 +
 .../agents-api/sandbox-manager.openapi.yaml   | 127 ++++++++++++++
 .../agents-api/system-observability-plan.md   |  94 +++++-----
 packages/agents-client/src/index.ts           |   1 +
 .../agents-client/src/observability-client.ts |  42 +++++
 services/agents-api/cmd/server/main.go        |  18 +-
 .../cmd/server/observability_cleanup.go       |  30 ++++
 services/agents-api/internal/api/handler.go   |   7 +
 .../internal/api/observability_read.go        |  88 ++++++++++
 .../internal/api/observability_read_test.go   |  62 +++++++
 .../internal/api/observability_requests.go    |  76 ++++++++
 .../internal/db/queries/runtime_history.sql   |   8 +-
 .../agents-api/internal/db/sqlc/models.go     |  85 +++++++--
 .../internal/db/sqlc/runtime_history.sql.go   |  14 +-
 .../agents-api/internal/execution/delivery.go |   2 +-
 .../internal/execution/dispatcher.go          |   5 +
 .../agents-api/internal/execution/journal.go  |  70 ++++++++
 .../execution/observability_tools_test.go     |  60 +++++++
 .../internal/observability/requests.go        | 134 ++++++++++++++
 .../internal/observability/requests_test.go   |  40 +++++
 .../internal/observability/tools.go           |  90 ++++++++++
 .../internal/store/observability_read.go      | 164 ++++++++++++++++++
 .../internal/store/observability_requests.go  |  63 +++++++
 .../store/observability_requests_test.go      |  37 ++++
 .../internal/store/observability_retention.go |  29 ++++
 .../internal/store/observability_tools.go     |  49 ++++++
 .../store/observability_tools_test.go         |  55 ++++++
 .../store/runtime_history_acceptance_test.go  |  12 +-
 .../migrations/000066_observability.sql       |  78 +++++++++
 .../agents-api/tests/runtime_history_live.py  |   2 +-
 services/core-console/sandbox_admin.go        |   5 +
 services/core-console/sandbox_admin_test.go   |   5 +-
 services/core-console/server.go               |   8 +-
 46 files changed, 1916 insertions(+), 98 deletions(-)
 create mode 100644 apps/web/src/features/dashboard/SystemObservabilityContent.tsx
 create mode 100644 apps/web/src/features/dashboard/SystemRequestChart.tsx
 create mode 100644 packages/agents-client/src/observability-client.ts
 create mode 100644 services/agents-api/cmd/server/observability_cleanup.go
 create mode 100644 services/agents-api/internal/api/observability_read.go
 create mode 100644 services/agents-api/internal/api/observability_read_test.go
 create mode 100644 services/agents-api/internal/api/observability_requests.go
 create mode 100644 services/agents-api/internal/execution/observability_tools_test.go
 create mode 100644 services/agents-api/internal/observability/requests.go
 create mode 100644 services/agents-api/internal/observability/requests_test.go
 create mode 100644 services/agents-api/internal/observability/tools.go
 create mode 100644 services/agents-api/internal/store/observability_read.go
 create mode 100644 services/agents-api/internal/store/observability_requests.go
 create mode 100644 services/agents-api/internal/store/observability_requests_test.go
 create mode 100644 services/agents-api/internal/store/observability_retention.go
 create mode 100644 services/agents-api/internal/store/observability_tools.go
 create mode 100644 services/agents-api/internal/store/observability_tools_test.go
 create mode 100644 services/agents-api/migrations/000066_observability.sql

diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index 12cb118d0..5f671aa6d 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -144,6 +144,16 @@ logs. Update this guide when architecture, ownership or generated contracts chan
 Comments and documentation are English. Reuse existing helpers and error mapping;
 split oversized components before extending them. Use `internal/obs/log` for logs.
 
+Operator observability is a side-channel owned by Core. The authenticated
+`/core/v1/observability/summary` extension reads bounded request, terminal Turn,
+tool attempt, and collector aggregates; the Core Web uses it through
+`packages/agents-client`. Runtime samples retain their separate public history
+contract under the `observability_runtime_samples` table. Native model attempts
+have a reserved table but no qualified producer; configured models, cumulative
+Usage, and Turns must not be counted as model calls. Missing collector coverage
+stays unavailable rather than becoming zero. See
+`contracts/agents-api/system-observability-plan.md` for source and retention details.
+
 ## Required checks
 
 Run `make check` before completion. The standalone gate includes all daemon/shared
diff --git a/apps/web/e2e/agents-lifecycle.spec.ts b/apps/web/e2e/agents-lifecycle.spec.ts
index 7b63cb271..e2d653a05 100644
--- a/apps/web/e2e/agents-lifecycle.spec.ts
+++ b/apps/web/e2e/agents-lifecycle.spec.ts
@@ -210,6 +210,36 @@ test("opens Dashboard as the default landing page", async ({ page, request }) =>
   await expect(page.locator(".dashboard-page")).toBeVisible();
 });
 
+test("shows operator tables and keeps incomplete request coverage unavailable", async ({ page, request }, testInfo) => {
+  await resetFixture(request);
+  await page.route("**/console/config", (route) => route.fulfill({ json: {
+    sandbox_admin: true, node_installer: false, node_installer_sha256: "",
+  } }));
+  const end = new Date(Math.floor(Date.now() / 60_000) * 60_000);
+  const start = new Date(end.getTime() - 60 * 60_000);
+  await page.route("**/core/v1/observability/summary?range=1h", (route) => route.fulfill({ json: {
+    generated_at: end.toISOString(), start: start.toISOString(), end: end.toISOString(), step_seconds: 60,
+    requests: [{ start: start.toISOString(), route_family: "sessions", outcome: "server_error", count: 2,
+      latency_sum_ms: 120, latency_bucket_counts: [0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0] }],
+    collector: [{ start: start.toISOString(), source: "request", attempted_count: 2, observed_count: 2,
+      unavailable_count: 0, timeout_count: 0, dropped_count: 0, export_failed_count: 0 }],
+    turns: [{ start: start.toISOString(), status: "failed", count: 1, queue_p95_ms: 50, execution_p95_ms: 400 }],
+    tools: [{ start: start.toISOString(), category: "command", outcome: "error", count: 1,
+      timed_count: 1, duration_p95_ms: 300 }],
+  } }));
+  await page.goto("/");
+  await page.getByRole("tab", { name: "Observability" }).click();
+  const system = page.locator(".dashboard-system");
+  await expect(system.getByRole("heading", { name: "System observability" })).toBeVisible();
+  await expect(system.getByRole("heading", { name: "Terminal Turns" })).toBeVisible();
+  await expect(system.getByRole("heading", { name: "Actual model invocations" })).toBeVisible();
+  await expect(system.getByRole("heading", { name: "Sandbox nodes" })).toBeVisible();
+  await expect(system.getByText("Unavailable", { exact: true }).first()).toBeVisible();
+  await expect(system.getByRole("rowheader", { name: "sessions" })).toBeVisible();
+  await expect(system.getByRole("rowheader", { name: "command" })).toBeVisible();
+  await attachElementScreenshot(system, testInfo, "system-observability-dashboard");
+});
+
 test("explains a 502 Core backend failure and opens copyable Docker recovery steps", async ({ page }) => {
   await page.route("**/v1/agents**", async (route) => {
     await route.fulfill({
@@ -2388,9 +2418,8 @@ test("presents Dashboard page-chain results and System boundaries without extra
   await expect(dashboard).toContainText("Latest complete paginated reads");
   await expect(dashboard.locator(".dashboard-summary > div").filter({ hasText: "Agents" })).toContainText("3");
   await expect(dashboard.locator(".dashboard-summary > div").filter({ hasText: "Sessions" })).toContainText("1");
-  await expect(dashboard.locator(".dashboard-summary > div").filter({ hasText: "In progress" })).toContainText("0");
-  await expect(dashboard.locator(".dashboard-summary > div").filter({ hasText: "Needs attention" })).toContainText("0");
-  await expect(dashboard).not.toContainText("Reported aggregate tokens");
+  await expect(dashboard.locator(".dashboard-summary > div").filter({ hasText: "Active sandboxes" })).toBeVisible();
+  await expect(dashboard.locator(".dashboard-summary > div").filter({ hasText: "Reported tokens" })).toBeVisible();
   await expect(dashboard).toContainText("No Sessions currently need attention.");
   await expect(dashboard.getByRole("button", { name: /Create agent/ })).toBeVisible();
   await expect(dashboard.getByRole("button", { name: /Start session/ })).toBeVisible();
@@ -2683,6 +2712,7 @@ test("renders Runtime telemetry as visual snapshot panels with details on demand
   await page.goto("/");
   const dashboard = page.locator(".dashboard-page");
   await expect(dashboard.getByRole("heading", { name: "Dashboard", exact: true })).toBeVisible();
+  await dashboard.getByRole("tab", { name: "Observability" }).click();
   const refresh = dashboard.getByRole("button", { name: "Refresh Dashboard snapshot" });
   const sandboxDiagnostics = dashboard.getByRole("region", { name: "Sandbox diagnostics" });
   await expect(sandboxDiagnostics).toBeVisible();
@@ -2918,6 +2948,7 @@ test("restores retained Runtime history after a Dashboard reload", async ({ page
   });
 
   await page.goto("/");
+  await page.getByRole("tab", { name: "Observability" }).click();
   const dashboard = page.locator(".dashboard-runtime-panel");
   await expect(dashboard.getByRole("group", { name: "Runtime trend source" })).toHaveCount(0);
   await expect(dashboard.getByLabel(/Durable · 30s; 1 Runtime targets/)).toBeVisible();
@@ -2978,6 +3009,7 @@ test("restores retained Runtime history after a Dashboard reload", async ({ page
   await expect(refreshedDurableCpuChart).toHaveAttribute("data-view-end", String(durableZoomEnd));
 
   await page.reload();
+  await page.getByRole("tab", { name: "Observability" }).click();
   await expect(dashboard.getByRole("group", { name: "Runtime trend source" })).toHaveCount(0);
   await expect(dashboard.getByLabel(/Durable · 30s; 1 Runtime targets/)).toBeVisible();
   await expect(dashboard.getByText("CPU usage durable trend available")).toBeAttached();
diff --git a/apps/web/e2e/sandbox-setup.spec.ts b/apps/web/e2e/sandbox-setup.spec.ts
index c58cd027d..c5fd35b88 100644
--- a/apps/web/e2e/sandbox-setup.spec.ts
+++ b/apps/web/e2e/sandbox-setup.spec.ts
@@ -189,13 +189,15 @@ for (const operation of ["setup", "enrollment"] as const) {
 }
 
 test("unpaired or unavailable consoles show setup guidance without admin credentials or manager requests", async ({ page, request }) => {
+  await expect.poll(async () => ((await (await request.get(`${fixture}/__fixture/sandbox`)).json()) as { calls: unknown[] }).calls.length).toBeGreaterThan(0);
+  const initialCalls = ((await (await request.get(`${fixture}/__fixture/sandbox`)).json()) as { calls: unknown[] }).calls.length;
   for (const body of [{ sandbox_admin: false, node_installer: true }, null]) {
     await page.route("**/console/config", (route) => route.fulfill(body ? { contentType: "application/json", body: JSON.stringify(body) } : { status: 404, body: "Not found" }));
     await page.reload();
     await page.getByRole("button", { name: "Hosted Sandbox Manager", exact: true }).click();
     await expect(page.getByRole("alert")).toContainText("Sandbox administration is not configured");
     await expect(page.getByLabel("Deployment admin key")).toHaveCount(0);
-    expect((await (await request.get(`${fixture}/__fixture/sandbox`)).json()).calls).toHaveLength(0);
+    expect((await (await request.get(`${fixture}/__fixture/sandbox`)).json()).calls).toHaveLength(initialCalls);
   }
   await page.route("**/console/config", (route) => route.fulfill({ contentType: "application/json", body: JSON.stringify({ sandbox_admin: true, node_installer: true, node_installer_sha256: "a".repeat(64) }) }));
   await page.getByRole("button", { name: "Refresh sandbox state" }).click();
diff --git a/apps/web/src/App.tsx b/apps/web/src/App.tsx
index 7765467be..8464c582d 100644
--- a/apps/web/src/App.tsx
+++ b/apps/web/src/App.tsx
@@ -2367,6 +2367,7 @@ export function App() {
               runtimeCollectionError={runtimeCollectionError}
               runtimeCollectionHasSnapshot={runtimeCollectionHasSnapshot}
               loadRuntimeHistory={loadDashboardRuntimeHistory}
+              coreBaseUrl={connection.baseUrl}
               onRefresh={refreshDashboard}
               onCreateAgent={openAgentSetup}
               onStartSession={() => openSessionSetup()}
diff --git a/apps/web/src/features/dashboard/DashboardView.css b/apps/web/src/features/dashboard/DashboardView.css
index 5364e670c..92ff876f4 100644
--- a/apps/web/src/features/dashboard/DashboardView.css
+++ b/apps/web/src/features/dashboard/DashboardView.css
@@ -5,6 +5,79 @@
   overflow: hidden;
 }
 
+.dashboard-system {
+  display: grid;
+  gap: 16px;
+  min-width: 0;
+  margin-top: 18px;
+}
+
+.dashboard-system-header,
+.dashboard-system-controls,
+.dashboard-system-metrics {
+  display: flex;
+  align-items: center;
+}
+
+.dashboard-system-header {
+  justify-content: space-between;
+  gap: 12px;
+}
+
+.dashboard-system-header h2,
+.dashboard-system-tables h3 { margin: 0; }
+.dashboard-system-header p,
+.dashboard-system-source,
+.dashboard-system-tables p { margin: 4px 0 0; color: var(--fg-muted); font-size: 12px; }
+.dashboard-system-controls { gap: 4px; }
+.dashboard-system-controls button {
+  padding: 5px 9px;
+  border: 1px solid var(--line);
+  border-radius: 6px;
+  background: var(--surface);
+  color: var(--fg-muted);
+  cursor: pointer;
+}
+.dashboard-system-controls button[aria-pressed="true"] { color: var(--fg); border-color: var(--fg-muted); }
+.dashboard-system-metrics { align-items: stretch; flex-wrap: wrap; gap: 8px; }
+.dashboard-system-metrics > div {
+  display: grid;
+  flex: 1 1 150px;
+  gap: 5px;
+  min-width: 0;
+  padding: 13px;
+  border: 1px solid var(--line);
+  border-radius: 8px;
+  background: var(--surface);
+}
+.dashboard-system-metrics small,
+.dashboard-system-metrics span { color: var(--fg-muted); font-size: 11px; }
+.dashboard-system-metrics strong { font-size: 20px; font-variant-numeric: tabular-nums; }
+.dashboard-system-tables { display: grid; grid-template-columns: repeat(2, minmax(0, 1fr)); gap: 12px; }
+.dashboard-system-request-chart {
+  min-width: 0;
+  margin: 0;
+  padding: 14px;
+  border: 1px solid var(--line);
+  border-radius: 8px;
+  background: var(--surface);
+}
+.dashboard-system-request-chart figcaption { display: flex; align-items: baseline; justify-content: space-between; gap: 12px; font-size: 13px; }
+.dashboard-system-request-chart figcaption span { color: var(--fg-muted); font-size: 11px; }
+.dashboard-system-request-chart svg { display: block; width: 100%; height: 150px; margin-top: 9px; color: var(--fg); }
+.dashboard-system-tables > section { min-width: 0; padding: 14px; border: 1px solid var(--line); border-radius: 8px; background: var(--surface); }
+.dashboard-system-nodes { grid-column: 1 / -1; }
+.dashboard-system-table-scroll { overflow-x: auto; margin-top: 10px; }
+.dashboard-system-table-scroll table { width: 100%; border-collapse: collapse; font-size: 12px; }
+.dashboard-system-table-scroll th,
+.dashboard-system-table-scroll td { padding: 8px 7px; border-bottom: 1px solid var(--line); text-align: right; white-space: nowrap; }
+.dashboard-system-table-scroll th:first-child { text-align: left; }
+.dashboard-system-table-scroll thead th { color: var(--fg-muted); font-weight: 500; }
+@media (max-width: 760px) {
+  .dashboard-system-header { align-items: flex-start; flex-direction: column; }
+  .dashboard-system-tables { grid-template-columns: 1fr; }
+}
+
 .dashboard-header > div:first-child {
   min-width: 0;
 }
@@ -27,6 +100,34 @@
   overflow-y: auto;
 }
 
+.dashboard-view-tabs {
+  display: flex;
+  width: fit-content;
+  margin-bottom: 14px;
+  padding: 3px;
+  border: 1px solid var(--line);
+  border-radius: 8px;
+  background: var(--surface);
+  gap: 3px;
+}
+
+.dashboard-view-tabs button {
+  padding: 7px 13px;
+  border: 0;
+  border-radius: 5px;
+  background: transparent;
+  color: var(--fg-muted);
+  cursor: pointer;
+  font-size: 12px;
+  font-weight: 600;
+}
+
+.dashboard-view-tabs button[aria-selected="true"] {
+  background: var(--surface-raised, var(--surface));
+  color: var(--fg);
+  box-shadow: var(--shadow-control);
+}
+
 .dashboard-overview,
 .dashboard-panel {
   min-width: 0;
diff --git a/apps/web/src/features/dashboard/DashboardView.test.tsx b/apps/web/src/features/dashboard/DashboardView.test.tsx
index 4ad132758..7acd92461 100644
--- a/apps/web/src/features/dashboard/DashboardView.test.tsx
+++ b/apps/web/src/features/dashboard/DashboardView.test.tsx
@@ -128,13 +128,14 @@ describe("Dashboard loaded-result presentation", () => {
 
   it("keeps a Runtime-only 503 scoped to the optional observation feature", () => {
     const html = render({
+      initialTab: "observability",
       runtimeSnapshot: null,
       runtimeCollectionState: "failed",
       runtimeCollectionError: "Agent core request failed (503).",
       runtimeCollectionHasSnapshot: false,
     });
 
-    expect(html).toContain("Runtime: Agent core request failed (503).");
+    expect(html).toContain("Agent core request failed (503).");
     expect(html).toContain("Runtime observations unavailable");
     expect(html).not.toContain("Agent Core backend is not ready");
     expect(html).not.toContain("Core backend is offline");
@@ -239,6 +240,7 @@ describe("Dashboard loaded-result presentation", () => {
       memory: { usage_bytes: 536_870_912, limit_bytes: 2_147_483_648 },
     };
     const html = render({
+      initialTab: "observability",
       runtimeSnapshot: { sessions: [hosted], observations: [observation], loadedAt: 1_700_000_100_000 },
     });
 
@@ -250,6 +252,7 @@ describe("Dashboard loaded-result presentation", () => {
     expect(html).toContain("Sandbox diagnostics");
     expect(html).toContain("Memory ≥80%");
     const withoutMemory = render({
+      initialTab: "observability",
       runtimeSnapshot: { sessions: [hosted], observations: [{ ...observation, memory: { usage_bytes: null, limit_bytes: null } }], loadedAt: 1_700_000_100_000 },
     });
     expect(withoutMemory).toMatch(/Memory ≥80%<\/small><strong>Unavailable<\/strong>/);
@@ -361,13 +364,14 @@ describe("Dashboard loaded-result presentation", () => {
       memory: null,
     };
     const html = render({
+      initialTab: "observability",
       runtimeSnapshot: { sessions: [runtimeSession], observations: [observation], loadedAt: 1_700_000_100_000 },
       runtimeCollectionState: "failed",
       runtimeCollectionError: "Runtime refresh failed",
       runtimeCollectionHasSnapshot: true,
     });
 
-    expect(html).toContain("Runtime: Runtime refresh failed");
+    expect(html).toContain("Runtime refresh failed");
     expect(html).toContain("Live · loading history");
     expect(html).not.toContain("Live · 30s");
   });
diff --git a/apps/web/src/features/dashboard/DashboardView.tsx b/apps/web/src/features/dashboard/DashboardView.tsx
index 5ac9a7cd2..5a97c4505 100644
--- a/apps/web/src/features/dashboard/DashboardView.tsx
+++ b/apps/web/src/features/dashboard/DashboardView.tsx
@@ -6,21 +6,26 @@ import {
   RefreshCw,
   Rows3,
 } from "lucide-react";
-import { useMemo, type ReactNode } from "react";
+import { useEffect, useMemo, useState, type ReactNode } from "react";
 import { useTranslation } from "react-i18next";
 
-import type { AgentSession, SavedAgent } from "@agents-core-web/agents-client";
+import { SandboxAdminClient, type AgentSession, type SandboxNode, type SavedAgent } from "@agents-core-web/agents-client";
 
 import { StatusIcon, type StatusKind } from "../../components/StatusIcon";
 import { backendFailureStatus } from "../../lib/core-readiness";
+import { isLocalProxyBaseUrl } from "../../lib/connection";
+import { sandboxConsoleConfig } from "../sandbox/console-config";
 import {
   buildDashboardSnapshot,
   buildRuntimeDashboardModel,
+  formatDashboardBytes,
   formatDashboardTimestamp,
+  formatDashboardTokens,
   type DashboardCollectionState,
   type DashboardSessionRow,
 } from "./dashboard-model";
 import { RuntimeObservabilityContent, type RuntimeHistoryLoader } from "./RuntimeObservabilityContent";
+import { SystemObservabilityContent } from "./SystemObservabilityContent";
 import type { RuntimeDashboardSnapshot } from "./runtime-snapshot";
 import "./DashboardView.css";
 
@@ -38,6 +43,8 @@ export interface DashboardViewProps {
   runtimeCollectionError: string | null;
   runtimeCollectionHasSnapshot: boolean;
   loadRuntimeHistory: RuntimeHistoryLoader;
+  coreBaseUrl?: string;
+  initialTab?: "overview" | "observability";
   onRefresh: () => void;
   onCreateAgent: () => void;
   onStartSession: () => void;
@@ -247,6 +254,8 @@ export function DashboardView({
   runtimeCollectionError,
   runtimeCollectionHasSnapshot,
   loadRuntimeHistory,
+  coreBaseUrl = "/v1",
+  initialTab = "overview",
   onRefresh,
   onCreateAgent,
   onStartSession,
@@ -257,6 +266,25 @@ export function DashboardView({
 }: DashboardViewProps) {
   const { t, i18n } = useTranslation("pages");
   const locale = i18n.resolvedLanguage;
+	const [tab, setTab] = useState<"overview" | "observability">(initialTab);
+	const [nodes, setNodes] = useState<SandboxNode[] | null>(null);
+	const [nodeRefresh, setNodeRefresh] = useState(0);
+	useEffect(() => {
+		const timer = window.setInterval(() => setNodeRefresh((value) => value + 1), 30_000);
+		return () => window.clearInterval(timer);
+	}, []);
+	useEffect(() => {
+		if (!isLocalProxyBaseUrl(coreBaseUrl)) { setNodes(null); return; }
+		const controller = new AbortController();
+		void (async () => {
+			const config = await sandboxConsoleConfig(controller.signal);
+			if (!config?.sandbox_admin) { if (!controller.signal.aborted) setNodes(null); return; }
+			const client = new SandboxAdminClient({ baseUrl: "/core/v1/sandbox" });
+			const result = await client.listNodes({ signal: controller.signal });
+			if (!controller.signal.aborted) setNodes(result.data);
+		})().catch(() => { if (!controller.signal.aborted) setNodes(null); });
+		return () => controller.abort();
+	}, [coreBaseUrl, nodeRefresh]);
   const snapshot = useMemo(() => buildDashboardSnapshot(agents, sessions, 6, 5), [agents, sessions]);
   const runtimeModel = useMemo(() => runtimeSnapshot
     ? buildRuntimeDashboardModel(runtimeSnapshot.sessions, runtimeSnapshot.observations)
@@ -302,7 +330,7 @@ export function DashboardView({
           <button
             className="button outline"
             type="button"
-            onClick={onRefresh}
+            onClick={() => { setNodeRefresh((value) => value + 1); onRefresh(); }}
             disabled={refreshing}
             aria-label={t("dashboard.refreshLabel")}
           >
@@ -313,6 +341,11 @@ export function DashboardView({
       </header>
 
       <div className="dashboard-scroll">
+        <div className="dashboard-view-tabs" role="tablist" aria-label={t("dashboard.viewTabs")}>
+          <button type="button" role="tab" aria-selected={tab === "overview"} onClick={() => setTab("overview")}>{t("dashboard.overviewTab")}</button>
+          <button type="button" role="tab" aria-selected={tab === "observability"} onClick={() => setTab("observability")}>{t("dashboard.observabilityTab")}</button>
+        </div>
+        {tab === "overview" ? <>
         <section className="dashboard-overview" aria-labelledby="dashboard-overview-heading">
           <div className="dashboard-snapshot-bar">
             <div className="dashboard-snapshot-copy">
@@ -371,16 +404,16 @@ export function DashboardView({
           <dl className="dashboard-summary" aria-label={t("dashboard.coreSnapshot")}>
             <Metric label={t("dashboard.agents")} value={agentsAvailable ? snapshot.loadedAgentCount.toLocaleString(locale) : t("dashboard.unavailable")} detail={t("dashboard.savedDefinitions")} />
             <Metric label={t("dashboard.sessions")} value={sessionsAvailable ? snapshot.loadedSessionCount.toLocaleString(locale) : t("dashboard.unavailable")} detail={t("dashboard.inSnapshot")} />
-            <Metric label={t("dashboard.inProgress")} value={sessionsAvailable ? snapshot.statusCounts.in_progress.toLocaleString(locale) : t("dashboard.unavailable")} detail={t("dashboard.reportedStatus")} />
-            <Metric
-              label={t("dashboard.needsAttention")}
-              value={sessionsAvailable ? attentionCount.toLocaleString(locale) : t("dashboard.unavailable")}
-              detail={t("dashboard.attentionDetail")}
-              emphasis={sessionsAvailable && attentionCount > 0}
-            />
+            <Metric label={t("dashboard.activeSandboxes")} value={runtimeModel ? runtimeModel.summary.activeSandboxCount.toLocaleString(locale) : t("dashboard.unavailable")} detail={t("dashboard.currentRuntimeSnapshot")} />
+            <Metric label={t("dashboard.totalTokens")} value={runtimeModel && runtimeModel.summary.totalTokens !== null ? formatDashboardTokens(runtimeModel.summary.totalTokens, locale) : t("dashboard.unavailable")} detail={t("dashboard.reportedUsage")} />
+            <Metric label={t("dashboard.cpuCapacity")} value={runtimeModel && runtimeModel.summary.cpuCapacityCores !== null ? runtimeModel.summary.cpuCapacityCores.toLocaleString(locale) : t("dashboard.unavailable")} detail={t("dashboard.measuredCores")} />
+            <Metric label={t("dashboard.memoryLimit")} value={runtimeModel && runtimeModel.summary.memoryLimitBytes !== null ? formatDashboardBytes(runtimeModel.summary.memoryLimitBytes) : t("dashboard.unavailable")} detail={t("dashboard.measuredLimit")} />
+            <Metric label={t("dashboard.readyNodes")} value={nodes === null ? t("dashboard.unavailable") : `${nodes.filter((node) => node.online && node.provider_ready).length}/${nodes.length}`} detail={t("dashboard.nodeSnapshot")} />
           </dl>
         </section>
+        </> : null}
 
+        {tab === "observability" ? <>
         <section className="dashboard-panel dashboard-runtime-panel" aria-labelledby="dashboard-runtime-heading">
           <header>
             <div>
@@ -393,6 +426,7 @@ export function DashboardView({
               </span>
             ) : null}
           </header>
+          {runtimeCollectionState === "failed" && runtimeCollectionError ? <p className="dashboard-runtime-history-error" role="status">{runtimeCollectionError}</p> : null}
           {!runtimeAvailable || !runtimeSnapshot || !runtimeModel ? (
             <p className="dashboard-empty">
               <AlertTriangle size={14} aria-hidden="true" />
@@ -419,7 +453,10 @@ export function DashboardView({
             </>
           )}
         </section>
+        <SystemObservabilityContent coreBaseUrl={coreBaseUrl} nodes={nodes} />
+        </> : null}
 
+        {tab === "overview" ? <>
         <div className="dashboard-primary-grid">
           <section className="dashboard-panel dashboard-attention-panel" aria-labelledby="dashboard-attention-heading">
             <header>
@@ -476,6 +513,7 @@ export function DashboardView({
             <p className="dashboard-empty"><Rows3 size={14} aria-hidden="true" />{t("dashboard.noRecent")}</p>
           )}
         </section>
+        </> : null}
       </div>
     </section>
   );
diff --git a/apps/web/src/features/dashboard/SystemObservabilityContent.tsx b/apps/web/src/features/dashboard/SystemObservabilityContent.tsx
new file mode 100644
index 000000000..cd5391b88
--- /dev/null
+++ b/apps/web/src/features/dashboard/SystemObservabilityContent.tsx
@@ -0,0 +1,140 @@
+import { useEffect, useMemo, useState } from "react";
+import { useTranslation } from "react-i18next";
+import {
+  ObservabilityAdminClient,
+  type OperatorMetricsRange,
+  type OperatorMetricsSummary,
+  type RequestMetricBucket,
+  type SandboxNode,
+} from "@agents-core-web/agents-client";
+
+import { isLocalProxyBaseUrl } from "../../lib/connection";
+import { sandboxConsoleConfig } from "../sandbox/console-config";
+import { formatDashboardBytes } from "./dashboard-model";
+import { SystemRequestChart } from "./SystemRequestChart";
+
+const latencyBounds = [10, 25, 50, 100, 250, 500, 1000, 2500, 5000, 10000] as const;
+
+function percentile95(rows: readonly RequestMetricBucket[]): number | null {
+  const counts = Array.from({ length: 11 }, (_, index) => rows.reduce((sum, row) => sum + (row.latency_bucket_counts[index] ?? 0), 0));
+  const total = counts.reduce((sum, count) => sum + count, 0);
+  if (total === 0) return null;
+  const target = Math.ceil(total * .95);
+  let current = 0;
+  for (let index = 0; index < counts.length; index += 1) {
+    current += counts[index]!;
+    if (current >= target) return latencyBounds[index] ?? 10000;
+  }
+  return null;
+}
+
+interface RouteSummary {
+  family: string;
+  count: number;
+  errors: number;
+  latency: number | null;
+}
+
+function routeSummaries(rows: readonly RequestMetricBucket[]): RouteSummary[] {
+  const families = new Map<string, RequestMetricBucket[]>();
+  for (const row of rows) {
+    const previous = families.get(row.route_family) ?? [];
+    previous.push(row);
+    families.set(row.route_family, previous);
+  }
+  return [...families].map(([family, group]) => ({
+    family,
+    count: group.reduce((sum, row) => sum + row.count, 0),
+    errors: group.filter((row) => row.outcome === "server_error").reduce((sum, row) => sum + row.count, 0),
+    latency: percentile95(group),
+  })).sort((left, right) => right.count - left.count);
+}
+
+export function SystemObservabilityContent({ coreBaseUrl, nodes }: { coreBaseUrl: string; nodes: SandboxNode[] | null }) {
+  const { t, i18n } = useTranslation("dashboard");
+  const locale = i18n.resolvedLanguage;
+  const [range, setRange] = useState<OperatorMetricsRange>("1h");
+  const [refresh, setRefresh] = useState(0);
+  const [summary, setSummary] = useState<OperatorMetricsSummary | null>(null);
+  const [state, setState] = useState<"loading" | "ready" | "unavailable" | "failed">("loading");
+  useEffect(() => {
+    const timer = window.setInterval(() => setRefresh((value) => value + 1), 30_000);
+    return () => window.clearInterval(timer);
+  }, []);
+  useEffect(() => {
+    const controller = new AbortController();
+    if (!isLocalProxyBaseUrl(coreBaseUrl)) {
+      setSummary(null);
+      setState("unavailable");
+      return () => controller.abort();
+    }
+    setState("loading");
+    setSummary(null);
+    void (async () => {
+      const config = await sandboxConsoleConfig(controller.signal);
+      if (!config?.sandbox_admin) {
+        if (!controller.signal.aborted) setState("unavailable");
+        return;
+      }
+      const metricsClient = new ObservabilityAdminClient({ baseUrl: "/core/v1/observability" });
+      const metrics = await metricsClient.retrieveSummary(range, { signal: controller.signal });
+      if (controller.signal.aborted) return;
+      setSummary(metrics);
+      setState("ready");
+    })().catch(() => { if (!controller.signal.aborted) setState("failed"); });
+    return () => controller.abort();
+  }, [coreBaseUrl, range, refresh]);
+
+  const routes = useMemo(() => routeSummaries(summary?.requests ?? []), [summary]);
+  const total = routes.reduce((sum, row) => sum + row.count, 0);
+  const serverErrors = routes.reduce((sum, row) => sum + row.errors, 0);
+  const p95 = percentile95(summary?.requests ?? []);
+  const durationMinutes = range === "1h" ? 60 : range === "6h" ? 360 : 1440;
+  const expectedBuckets = durationMinutes / ((summary?.step_seconds ?? 60) / 60);
+  const heartbeatBuckets = new Set((summary?.collector ?? []).filter((row) => row.source === "request").map((row) => row.start)).size;
+  const completeRequestCoverage = summary !== null && heartbeatBuckets >= expectedBuckets;
+  const requestRate = completeRequestCoverage ? total / durationMinutes : null;
+  const errorRate = completeRequestCoverage && total > 0 ? serverErrors / total * 100 : null;
+  const collector = summary?.collector ?? [];
+  const dropped = collector.reduce((sum, row) => sum + row.dropped_count + row.export_failed_count, 0);
+  const onlineNodes = nodes?.filter((node) => node.online && node.provider_ready).length;
+  const nodeCapacity = nodes?.reduce((sum, node) => sum + node.max_active, 0);
+  const number = (value: number) => value.toLocaleString(locale);
+  const unavailable = t("system.unavailableValue");
+
+  return <section className="dashboard-system" aria-labelledby="dashboard-system-heading">
+    <header className="dashboard-system-header">
+      <div><h2 id="dashboard-system-heading">{t("system.title")}</h2><p>{t("system.subtitle")}</p></div>
+      <div className="dashboard-system-controls" role="group" aria-label={t("system.range")}>
+        {(["1h", "6h", "24h"] as const).map((option) => <button key={option} type="button" aria-pressed={range === option} onClick={() => setRange(option)}>{option}</button>)}
+      </div>
+    </header>
+    <p className="dashboard-system-source" role="status">{state === "loading" ? t("system.loading") : state === "failed" ? t("system.stale") : state === "unavailable" ? t("system.adminUnavailable") : t("system.source")}</p>
+    <div className="dashboard-system-metrics">
+      <div><small>{t("system.requestRate")}</small><strong>{requestRate === null ? unavailable : `${number(Math.round(requestRate))}/min`}</strong><span>{t("system.requestCount", { count: total })}</span></div>
+      <div><small>{t("system.serverErrorRate")}</small><strong>{errorRate === null ? unavailable : `${errorRate.toFixed(1)}%`}</strong><span>{t("system.serverErrorCount", { count: serverErrors })}</span></div>
+      <div><small>{t("system.latencyP95")}</small><strong>{!completeRequestCoverage || p95 === null ? unavailable : p95 === 10000 ? ">10 s" : `${number(p95)} ms`}</strong><span>{t("system.histogramEstimate")}</span></div>
+      <div><small>{t("system.collectorCoverage")}</small><strong>{summary === null ? unavailable : `${number(heartbeatBuckets)}/${number(expectedBuckets)}`}</strong><span>{t("system.collectorLoss", { count: dropped })}</span></div>
+      <div><small>{t("system.nodes")}</small><strong>{nodes === null ? unavailable : `${number(onlineNodes ?? 0)}/${number(nodes.length)}`}</strong><span>{nodeCapacity === undefined ? unavailable : t("system.nodeCapacity", { count: nodeCapacity })}</span></div>
+    </div>
+    <SystemRequestChart summary={summary} />
+    <div className="dashboard-system-tables">
+      <section><h3>{t("system.routeTable")}</h3><p>{t("system.coveredBuckets", { covered: heartbeatBuckets, total: expectedBuckets })}</p>
+        {routes.length ? <div className="dashboard-system-table-scroll"><table><thead><tr><th>{t("system.route")}</th><th>{t("system.requests")}</th><th>{t("system.errors")}</th><th>{t("system.p95")}</th></tr></thead><tbody>{routes.map((row) => <tr key={row.family}><th>{row.family}</th><td>{number(row.count)}</td><td>{number(row.errors)}</td><td>{row.latency === null ? unavailable : row.latency === 10000 ? ">10 s" : `${number(row.latency)} ms`}</td></tr>)}</tbody></table></div> : <p>{t("system.noRequestSamples")}</p>}
+      </section>
+      <section><h3>{t("system.collectorTable")}</h3>
+        {collector.length ? <div className="dashboard-system-table-scroll"><table><thead><tr><th>{t("system.sourceName")}</th><th>{t("system.attempted")}</th><th>{t("system.observed")}</th><th>{t("system.lost")}</th></tr></thead><tbody>{[...new Set(collector.map((row) => row.source))].map((source) => { const rows = collector.filter((row) => row.source === source); return <tr key={source}><th>{source}</th><td>{number(rows.reduce((sum, row) => sum + row.attempted_count, 0))}</td><td>{number(rows.reduce((sum, row) => sum + row.observed_count, 0))}</td><td>{number(rows.reduce((sum, row) => sum + row.dropped_count + row.export_failed_count, 0))}</td></tr>; })}</tbody></table></div> : <p>{t("system.noCollectorSamples")}</p>}
+      </section>
+      <section className="dashboard-system-nodes"><h3>{t("system.turnTable")}</h3>
+        {summary?.turns.length ? <div className="dashboard-system-table-scroll"><table><thead><tr><th>{t("system.turnStatus")}</th><th>{t("system.turnCount")}</th><th>{t("system.queueP95")}</th><th>{t("system.executionP95")}</th></tr></thead><tbody>{[...new Set(summary.turns.map((row) => row.status))].map((status) => { const rows = summary.turns.filter((row) => row.status === status); const count = rows.reduce((sum, row) => sum + row.count, 0); const queue = rows.reduce<number[]>((values, row) => row.queue_p95_ms === null ? values : [...values, row.queue_p95_ms], []); const execution = rows.reduce<number[]>((values, row) => row.execution_p95_ms === null ? values : [...values, row.execution_p95_ms], []); return <tr key={status}><th>{status}</th><td>{number(count)}</td><td>{queue.length ? `${number(Math.round(Math.max(...queue)))} ms` : unavailable}</td><td>{execution.length ? `${number(Math.round(Math.max(...execution)))} ms` : unavailable}</td></tr>; })}</tbody></table></div> : <p>{t("system.noTurns")}</p>}
+      </section>
+      <section><h3>{t("system.toolTable")}</h3>
+        {summary?.tools.length ? <div className="dashboard-system-table-scroll"><table><thead><tr><th>{t("system.toolCategory")}</th><th>{t("system.toolOutcome")}</th><th>{t("system.toolCount")}</th><th>{t("system.toolTiming")}</th></tr></thead><tbody>{summary.tools.map((row, index) => <tr key={`${row.start}:${row.category}:${row.outcome}:${index}`}><th>{row.category}</th><td>{row.outcome}</td><td>{number(row.count)}</td><td>{row.duration_p95_ms === null ? unavailable : `${number(Math.round(row.duration_p95_ms))} ms · ${number(row.timed_count)}/${number(row.count)}`}</td></tr>)}</tbody></table></div> : <p>{t("system.noToolAttempts")}</p>}
+      </section>
+      <section><h3>{t("system.modelTable")}</h3><p>{t("system.modelUnavailable")}</p></section>
+      <section className="dashboard-system-nodes"><h3>{t("system.nodeTable")}</h3>
+        {nodes?.length ? <div className="dashboard-system-table-scroll"><table><thead><tr><th>{t("system.node")}</th><th>{t("system.status")}</th><th>{t("system.active")}</th><th>{t("system.retained")}</th><th>{t("system.cores")}</th><th>{t("system.availableMemory")}</th></tr></thead><tbody>{nodes.map((node) => <tr key={node.id}><th>{node.name}</th><td>{t(node.online && node.provider_ready ? "system.online" : "system.offline")}</td><td>{number(node.active)} / {number(node.max_active)}</td><td>{number(node.retained)}</td><td>{node.cpu_count === null ? unavailable : number(node.cpu_count)}</td><td>{node.available_memory_bytes === null ? unavailable : formatDashboardBytes(node.available_memory_bytes)}</td></tr>)}</tbody></table></div> : <p>{nodes === null ? t("system.nodeUnavailable") : t("system.noNodes")}</p>}
+      </section>
+    </div>
+  </section>;
+}
diff --git a/apps/web/src/features/dashboard/SystemRequestChart.tsx b/apps/web/src/features/dashboard/SystemRequestChart.tsx
new file mode 100644
index 000000000..1822ff44e
--- /dev/null
+++ b/apps/web/src/features/dashboard/SystemRequestChart.tsx
@@ -0,0 +1,53 @@
+import { useMemo } from "react";
+import { useTranslation } from "react-i18next";
+import type { OperatorMetricsSummary } from "@agents-core-web/agents-client";
+
+export function SystemRequestChart({ summary }: { summary: OperatorMetricsSummary | null }) {
+  const { t, i18n } = useTranslation("dashboard");
+  const points = useMemo(() => {
+    if (!summary) return [];
+    const start = Date.parse(summary.start);
+    const end = Date.parse(summary.end);
+    const stepMS = summary.step_seconds * 1_000;
+    const requests = new Map<number, { total: number; errors: number }>();
+    for (const row of summary.requests) {
+      const at = Date.parse(row.start);
+      const prior = requests.get(at) ?? { total: 0, errors: 0 };
+      prior.total += row.count;
+      if (row.outcome === "server_error") prior.errors += row.count;
+      requests.set(at, prior);
+    }
+    const covered = new Set(summary.collector.filter((row) => row.source === "request").map((row) => Date.parse(row.start)));
+    const result = [];
+    for (let at = start; at < end; at += stepMS) {
+      result.push({ at, value: covered.has(at) ? requests.get(at) ?? { total: 0, errors: 0 } : null });
+    }
+    return result;
+  }, [summary]);
+  const known = points.filter((point) => point.value !== null);
+  const maximum = Math.max(1, ...known.map((point) => point.value?.total ?? 0));
+  const width = 780;
+  const height = 150;
+  const left = 28;
+  const baseline = 124;
+  const plotWidth = width - left - 8;
+  const slot = plotWidth / Math.max(1, points.length);
+  return <figure className="dashboard-system-request-chart">
+    <figcaption><strong>{t("system.requestTrend")}</strong><span>{t("system.requestTrendDetail", { covered: known.length, total: points.length })}</span></figcaption>
+    {known.length ? <svg viewBox={`0 0 ${width} ${height}`} preserveAspectRatio="none" role="img" aria-label={t("system.requestTrendLabel", { covered: known.length, total: points.length })}>
+      <line x1={left} y1={baseline} x2={width - 8} y2={baseline} stroke="currentColor" opacity=".25" />
+      {points.map((point, index) => {
+        if (point.value === null) return null;
+        const x = left + index * slot + slot * .12;
+        const barWidth = Math.max(1, slot * .75);
+        const barHeight = Math.max(1, point.value.total / maximum * 100);
+        const errorHeight = Math.max(0, point.value.errors / maximum * 100);
+        return <g key={point.at}>
+          <title>{`${new Date(point.at).toLocaleString(i18n.resolvedLanguage)}: ${point.value.total} ${t("system.requests")}, ${point.value.errors} ${t("system.errors")}`}</title>
+          <rect x={x} y={baseline - barHeight} width={barWidth} height={barHeight} fill="var(--accent, #377db8)" opacity=".8" />
+          {errorHeight > 0 ? <rect x={x} y={baseline - errorHeight} width={barWidth} height={errorHeight} fill="#d66b62" /> : null}
+        </g>;
+      })}
+    </svg> : <p>{t("system.noRequestSamples")}</p>}
+  </figure>;
+}
diff --git a/apps/web/src/i18n/locales/en/dashboard.ts b/apps/web/src/i18n/locales/en/dashboard.ts
index 4415511ff..dcf3d0e15 100644
--- a/apps/web/src/i18n/locales/en/dashboard.ts
+++ b/apps/web/src/i18n/locales/en/dashboard.ts
@@ -1,4 +1,20 @@
 export const dashboard = {
+  system: {
+    title: "System observability", subtitle: "Completed HTTP requests, collection coverage, and sandbox nodes",
+    range: "System metrics range", loading: "Loading operator metrics…", stale: "Operator metrics unavailable · retrying",
+    adminUnavailable: "Deployment administrator metrics are unavailable on this connection.", source: "Core operator metrics · completed time buckets",
+    unavailableValue: "Unavailable", requestRate: "Request rate", requestCount: "{{count}} observed completed requests",
+    serverErrorRate: "HTTP 5xx rate", serverErrorCount: "{{count}} observed server errors", latencyP95: "Latency p95", histogramEstimate: "Estimated from fixed latency buckets",
+    collectorCoverage: "Request collector buckets", collectorLoss: "{{count}} dropped or failed writes", nodes: "Ready nodes", nodeCapacity: "{{count}} active slots configured",
+    routeTable: "Request routes", coveredBuckets: "Observed buckets: {{covered}}/{{total}}", route: "Route family", requests: "Requests", errors: "5xx", p95: "p95",
+    requestTrend: "Request volume", requestTrendDetail: "{{covered}}/{{total}} buckets covered · gaps are unavailable", requestTrendLabel: "Request volume over time; {{covered}} of {{total}} buckets covered",
+    collectorTable: "Collector health", sourceName: "Source", attempted: "Attempted", observed: "Observed", lost: "Dropped / failed",
+    turnTable: "Terminal Turns", turnStatus: "Outcome", turnCount: "Turns", queueP95: "Worst bucket queue p95", executionP95: "Worst bucket run p95", noTurns: "No terminal Turns in this range.",
+    toolTable: "Actual tool attempts", toolCategory: "Category", toolOutcome: "Outcome", toolCount: "Attempts", toolTiming: "Bucket p95 · timed coverage", noToolAttempts: "No terminal tool attempt observations in this range.",
+    modelTable: "Actual model invocations", modelUnavailable: "Invocation-level collection is not yet available from every harness. Configured models and Turn usage are not counted as calls.",
+    nodeTable: "Sandbox nodes", node: "Node", status: "Status", active: "Active / capacity", retained: "Retained", cores: "CPU cores", availableMemory: "Available memory", online: "Ready", offline: "Unavailable",
+    noRequestSamples: "No request buckets collected in this range.", noCollectorSamples: "No collector coverage recorded in this range.", noNodes: "No sandbox nodes registered.", nodeUnavailable: "Node data unavailable.",
+  },
   sandbox: {
     title: "Sandbox diagnostics", subtitle: "Current allocation identity, sampling coverage, and memory pressure",
     current: "Current complete snapshot", retained: "Last complete snapshot · refresh failed",
diff --git a/apps/web/src/i18n/locales/en/pages.ts b/apps/web/src/i18n/locales/en/pages.ts
index 9c6de3a3d..be3773f72 100644
--- a/apps/web/src/i18n/locales/en/pages.ts
+++ b/apps/web/src/i18n/locales/en/pages.ts
@@ -1,5 +1,9 @@
 export const pages = {
   dashboard: {
+    viewTabs: "Dashboard views", overviewTab: "Overview", observabilityTab: "Observability",
+    activeSandboxes: "Active sandboxes", currentRuntimeSnapshot: "Current Runtime snapshot", totalTokens: "Reported tokens", reportedUsage: "Session usage with coverage",
+    cpuCapacity: "Sandbox CPU capacity", measuredCores: "Cores with a current observation", memoryLimit: "Sandbox memory limit", measuredLimit: "Limits with a current observation",
+    readyNodes: "Ready nodes", nodeSnapshot: "Deployment node heartbeat",
     title: "Dashboard", subtitle: "Agents and Sessions that may need your attention.", refresh: "Refresh", refreshing: "Refreshing…", refreshLabel: "Refresh Dashboard snapshot",
     snapshotIncomplete: "Snapshot incomplete", snapshotStale: "Using the last successful snapshot", snapshotRefreshing: "Refreshing snapshot", snapshotReady: "Snapshot ready", snapshotDetail: "Latest complete paginated reads · not a live Core total or runtime-readiness signal", dataSources: "Dashboard data sources",
     agents: "Agents", sessions: "Sessions", runtime: "Runtime", unavailable: "Unavailable", savedDefinitions: "Saved definitions", inSnapshot: "In this snapshot", inProgress: "In progress", reportedStatus: "Core-reported status", needsAttention: "Needs attention", attentionDetail: "Requires action or failed",
diff --git a/apps/web/src/i18n/locales/zh-CN/dashboard.ts b/apps/web/src/i18n/locales/zh-CN/dashboard.ts
index 7c9ca0cf5..010f4a60d 100644
--- a/apps/web/src/i18n/locales/zh-CN/dashboard.ts
+++ b/apps/web/src/i18n/locales/zh-CN/dashboard.ts
@@ -1,4 +1,20 @@
 export const dashboard = {
+  system: {
+    title: "系统观测", subtitle: "已完成的 HTTP 请求、采集覆盖与 Sandbox 节点",
+    range: "系统指标时间范围", loading: "正在加载管理员指标…", stale: "管理员指标不可用 · 正在重试",
+    adminUnavailable: "当前连接无法读取部署管理员指标。", source: "Core 管理员指标 · 已完成的时间桶",
+    unavailableValue: "不可用", requestRate: "请求速率", requestCount: "已观测 {{count}} 个完成请求",
+    serverErrorRate: "HTTP 5xx 比例", serverErrorCount: "已观测 {{count}} 个服务端错误", latencyP95: "延迟 p95", histogramEstimate: "按固定延迟桶估算",
+    collectorCoverage: "请求采集时间桶", collectorLoss: "{{count}} 次丢弃或写入失败", nodes: "就绪节点", nodeCapacity: "配置 {{count}} 个活动名额",
+    routeTable: "请求分类", coveredBuckets: "已观测时间桶：{{covered}}/{{total}}", route: "路由分类", requests: "请求数", errors: "5xx", p95: "p95",
+    requestTrend: "请求量", requestTrendDetail: "覆盖 {{covered}}/{{total}} 个时间桶 · 缺口表示不可用", requestTrendLabel: "请求量趋势；覆盖 {{covered}}/{{total}} 个时间桶",
+    collectorTable: "采集健康", sourceName: "来源", attempted: "尝试", observed: "成功", lost: "丢弃／失败",
+    turnTable: "终态 Turn", turnStatus: "结果", turnCount: "Turn 数", queueP95: "最差时间桶排队 p95", executionP95: "最差时间桶执行 p95", noTurns: "该时段没有终态 Turn。",
+    toolTable: "实际工具调用", toolCategory: "类别", toolOutcome: "结果", toolCount: "调用数", toolTiming: "时间桶 p95 · 时长覆盖", noToolAttempts: "该时段没有终态工具调用观测。",
+    modelTable: "实际模型调用", modelUnavailable: "各 Harness 尚未提供完整的单次调用采集。不会将配置模型或 Turn 用量当作调用次数。",
+    nodeTable: "Sandbox 节点", node: "节点", status: "状态", active: "活动／容量", retained: "保留", cores: "CPU 核数", availableMemory: "可用内存", online: "就绪", offline: "不可用",
+    noRequestSamples: "该时段尚无请求时间桶。", noCollectorSamples: "该时段尚无采集覆盖记录。", noNodes: "尚未注册 Sandbox 节点。", nodeUnavailable: "节点数据不可用。",
+  },
   sandbox: {
     title: "Sandbox 诊断", subtitle: "按分配身份查看采样覆盖、生命周期与内存压力",
     current: "当前完整快照", retained: "上次完整快照 · 刷新失败",
diff --git a/apps/web/src/i18n/locales/zh-CN/pages.ts b/apps/web/src/i18n/locales/zh-CN/pages.ts
index 0a1bdc8c5..9a703e8eb 100644
--- a/apps/web/src/i18n/locales/zh-CN/pages.ts
+++ b/apps/web/src/i18n/locales/zh-CN/pages.ts
@@ -1,5 +1,9 @@
 export const pages = {
   dashboard: {
+    viewTabs: "大盘视图", overviewTab: "概览", observabilityTab: "观测大盘",
+    activeSandboxes: "活跃 Sandbox", currentRuntimeSnapshot: "当前 Runtime 快照", totalTokens: "已报告 Token", reportedUsage: "有覆盖的会话用量",
+    cpuCapacity: "Sandbox CPU 容量", measuredCores: "有当前观测的核数", memoryLimit: "Sandbox 内存上限", measuredLimit: "有当前观测的上限",
+    readyNodes: "就绪节点", nodeSnapshot: "部署节点心跳",
     title: "概览", subtitle: "查看可能需要关注的智能体和会话。", refresh: "刷新", refreshing: "刷新中…", refreshLabel: "刷新概览快照",
     snapshotIncomplete: "快照不完整", snapshotStale: "正在使用最近一次成功快照", snapshotRefreshing: "正在刷新快照", snapshotReady: "快照已就绪", snapshotDetail: "最近一次完整分页读取 · 不代表 Core 实时总量或运行时就绪", dataSources: "概览数据源",
     agents: "智能体", sessions: "会话", runtime: "运行时", unavailable: "不可用", savedDefinitions: "已保存定义", inSnapshot: "当前快照", inProgress: "进行中", reportedStatus: "Core 报告状态", needsAttention: "需要关注", attentionDetail: "需要操作或已失败",
diff --git a/apps/web/vite.config.ts b/apps/web/vite.config.ts
index 6ec76098e..b13b470d6 100644
--- a/apps/web/vite.config.ts
+++ b/apps/web/vite.config.ts
@@ -46,6 +46,7 @@ export default defineConfig(({ command, mode }) => {
     server: {
       proxy: {
         "/core/v1/sandbox": { target, changeOrigin: true },
+        "/core/v1/observability": { target, changeOrigin: true },
         "/v1": {
           target,
           changeOrigin: true,
diff --git a/contracts/agents-api/sandbox-manager.openapi.yaml b/contracts/agents-api/sandbox-manager.openapi.yaml
index e9d7d2807..d74425b6a 100644
--- a/contracts/agents-api/sandbox-manager.openapi.yaml
+++ b/contracts/agents-api/sandbox-manager.openapi.yaml
@@ -7,6 +7,33 @@ definitions:
       rotate:
         type: boolean
     type: object
+  api.OperatorMetricsResponse:
+    properties:
+      collector:
+        items:
+          $ref: '#/definitions/store.CollectorMetricBucket'
+        type: array
+      end:
+        type: string
+      generated_at:
+        type: string
+      requests:
+        items:
+          $ref: '#/definitions/store.RequestMetricBucket'
+        type: array
+      start:
+        type: string
+      step_seconds:
+        type: integer
+      tools:
+        items:
+          $ref: '#/definitions/store.ToolMetricBucket'
+        type: array
+      turns:
+        items:
+          $ref: '#/definitions/store.TurnMetricBucket'
+        type: array
+    type: object
   api.ProjectAPIKeyList:
     properties:
       data:
@@ -102,6 +129,25 @@ definitions:
       revoked_at:
         type: string
     type: object
+  store.CollectorMetricBucket:
+    properties:
+      attempted_count:
+        type: integer
+      dropped_count:
+        type: integer
+      export_failed_count:
+        type: integer
+      observed_count:
+        type: integer
+      source:
+        type: string
+      start:
+        type: string
+      timeout_count:
+        type: integer
+      unavailable_count:
+        type: integer
+    type: object
   store.IssuedExecutorCredential:
     properties:
       environment_id:
@@ -139,6 +185,23 @@ definitions:
       revoked_at:
         type: string
     type: object
+  store.RequestMetricBucket:
+    properties:
+      count:
+        type: integer
+      latency_bucket_counts:
+        items:
+          type: integer
+        type: array
+      latency_sum_ms:
+        type: integer
+      outcome:
+        type: string
+      route_family:
+        type: string
+      start:
+        type: string
+    type: object
   store.ResourceOwner:
     properties:
       api_key:
@@ -300,6 +363,34 @@ definitions:
       maintenance:
         type: boolean
     type: object
+  store.ToolMetricBucket:
+    properties:
+      category:
+        type: string
+      count:
+        type: integer
+      duration_p95_ms:
+        type: number
+      outcome:
+        type: string
+      start:
+        type: string
+      timed_count:
+        type: integer
+    type: object
+  store.TurnMetricBucket:
+    properties:
+      count:
+        type: integer
+      execution_p95_ms:
+        type: number
+      queue_p95_ms:
+        type: number
+      start:
+        type: string
+      status:
+        type: string
+    type: object
   store.WriteOperation:
     properties:
       action:
@@ -451,6 +542,42 @@ paths:
       summary: Revoke an Environment executor credential
       tags:
         - Environment Executor
+  /core/v1/observability/summary:
+    get:
+      description: Deployment administrator only. Returns fixed-range aggregate requests and collector coverage without paths, payloads or project identifiers.
+      parameters:
+        - description: 1h, 6h, or 24h (default 1h)
+          in: query
+          name: range
+          type: string
+      produces:
+        - application/json
+      responses:
+        "200":
+          description: OK
+          schema:
+            $ref: '#/definitions/api.OperatorMetricsResponse'
+        "400":
+          description: Bad Request
+          schema:
+            $ref: '#/definitions/v1.ErrorResponse'
+        "401":
+          description: Unauthorized
+          schema:
+            $ref: '#/definitions/v1.ErrorResponse'
+        "500":
+          description: Internal Server Error
+          schema:
+            $ref: '#/definitions/v1.ErrorResponse'
+        "503":
+          description: Service Unavailable
+          schema:
+            $ref: '#/definitions/v1.ErrorResponse'
+      security:
+        - DeploymentAdminAuth: []
+      summary: Retrieve bounded operator observability metrics
+      tags:
+        - Observability
   /core/v1/project-api-keys/{binding_digest}:
     get:
       description: Core extension requiring deployment administrator authority. The binding digest selects only a statically configured caller; it does not authenticate. Returns safe metadata, never key secrets or token digests. Derived keys freeze that caller's principal and stop authenticating when the static parent is removed or rebound.
diff --git a/contracts/agents-api/system-observability-plan.md b/contracts/agents-api/system-observability-plan.md
index 1f59827c4..0c52c0b1f 100644
--- a/contracts/agents-api/system-observability-plan.md
+++ b/contracts/agents-api/system-observability-plan.md
@@ -1,8 +1,9 @@
 # System observability collection plan
 
-Status: collection inventory and next-phase design. The current Dashboard release
-uses existing Runtime observations and Session Usage only. Request, model, and tool
-panels must not be shown as measured until the instruments below are implemented.
+Status: implementation inventory for the operator Dashboard. Runtime history,
+request buckets, terminal Turns, tool attempts, and node state have independent
+sources. The model-attempt table is reserved but has no writer; its panel must
+remain unavailable until native per-attempt events are qualified.
 
 ## Existing evidence
 
@@ -12,49 +13,64 @@ panels must not be shown as measured until the instruments below are implemented
 | CPU time and capacity | Provider observation | Current and periodic history when sampling is available | Cumulative time; utilization needs two samples from one compute incarnation |
 | Memory usage and limit | Provider observation | Current and periodic history when reported | Point-in-time guest/container memory, not host memory |
 | Token input/output | Measured Session Usage | Public current Session and periodic Runtime history | Reported canonical usage; absent usage remains unknown |
-| Request count, latency, errors | No common metric | Not available | Do not derive from Session count or HTTP page loads |
-| Model distribution and latency | Saved Agent model is available, terminal call records are not aggregated | Configuration only | Do not label configured models as invoked models |
-| Tool calls and failures | Durable Turn events exist, no bounded aggregate read | Partial event evidence | Do not count tools by Agent declarations |
-| Host, database, queue health | No tenant-safe operator projection | Not available | Keep out of Session-scoped charts |
+| Request count, latency, errors | Completed API responses, stored in minute buckets | Fixed route families, methods, outcomes, and latency bins; collector heartbeat tracks quiet intervals | Transport health, not Turn success |
+| Terminal Turn outcome and duration | Persisted `turns` terminal timestamps | Completed, failed, and cancelled Turns | Execution outcome, not HTTP success |
+| Model distribution and latency | Native per-attempt evidence is not normalized across harnesses | Unavailable | Do not label configured models, cumulative Usage, or Turns as invoked models |
+| Tool calls and failures | Sanitized `tool_call` completion events | Attempts with an after event; missing before leaves duration null unless native duration exists | Does not count Agent tool declarations or model selection |
+| Sandbox nodes | Deployment administrator node list | Current online, provider-ready, active, retained and capacity fields | A node being online does not prove a Turn can execute |
+| Host, database, queue health | No qualified operator projection | Not available | Keep out of Session-scoped charts |
 
 Current Runtime history has a 30-second default periodic collection interval and
-seven-day retention; the public Web ranges are 1h, 6h, and 24h. Its capability
-route distinguishes periodic collection from on-read collection. The Dashboard
-must show the source, age, coverage and missing values. An empty interval is not
-an observed zero.
+seven-day retention; the public Web ranges are 1h, 6h, and 24h. Its table is
+renamed in place from `runtime_history_samples` to
+`observability_runtime_samples`. Its capability route distinguishes periodic
+collection from on-read collection. Missing values and empty intervals are not
+observed zeroes.
 
-## Next collection boundary
+## Collection boundary and remaining work
 
-Add a separate, authenticated operator metrics service behind Core. Keep the
-pinned Agents API resources unchanged. The service should aggregate sanitized
-events server-side and expose a bounded read-only extension through
+The authenticated operator metrics service runs behind Core. It keeps the pinned
+Agents API resources unchanged, aggregates sanitized events server-side, and
+exposes a bounded read-only extension through
 `packages/agents-client`. Browser input may select a fixed range and resolution;
 it must not select tenants, storage labels, arbitrary PromQL, endpoints or keys.
 
-1. **Ingress:** count completed HTTP requests by route family, method and coarse
-   outcome (`success`, `client_error`, `server_error`). Record a latency histogram
-   after the response completes. Exclude health polling or show it separately.
-   This is transport health, never terminal Turn success.
-2. **Execution:** emit one terminal Turn outcome from the durable state transition
-   winner, with queue wait and run duration measured from persisted timestamps.
-   Active Turns are a current gauge from Core ownership, not a counter inferred
-   from sampled CPU. Retries and recovery must not double count terminal outcomes.
-3. **Model:** observe actual native model invocation attempts at the harness
+1. **Ingress (implemented):** count completed HTTP requests by route family,
+   method and coarse outcome (`success`, `client_error`, `server_error`). Record a
+   latency histogram after the response completes. Health polling and the metric
+   read itself are excluded. The bounded asynchronous queue records drops and
+   write failures; failed batches are not replayed onto the request path.
+2. **Execution (implemented):** query terminal Turns from durable state with queue
+   wait and run duration measured from persisted timestamps. The query is bounded
+   to a fixed time window and avoids duplicate terminal events. Active Turns
+   remain a current Core gauge, not a counter inferred from sampled CPU.
+3. **Model (pending):** observe actual native model invocation attempts at the harness
    boundary, including model identifier, terminal attempt outcome, duration,
    reported input/output tokens and usage coverage. A saved Agent's configured
    model is not proof of which model ran. Keep bounded model label cardinality.
-4. **Tools:** observe actual tool attempt start and terminal result at the common
-   Runtime adapter boundary. Use a bounded tool category and outcome label;
+4. **Tools (partial):** observe `tool_call` before/after events at the common
+   execution journal boundary. Use a bounded tool category and outcome label;
    raw tool names, arguments, output and credentials stay out of metric labels.
-   Distinguish model tool selection from completed tool execution.
-5. **Sandbox:** retain allocation identity and compute generation internally.
+   Record a completed attempt only after the corresponding event batch has been
+   persisted. Completion is deduplicated by Turn and call ID. A missing after
+   event is not counted as a completed attempt. Some harnesses may not emit
+   these events.
+5. **Sandbox (partial):** retain allocation identity and compute generation internally.
    Extend the normalized sample only after Docker and microsandbox values have
    matching semantics. Candidates are OOM/exit events, disk usage/limit, network
    bytes, compute restarts, and allocation/compute startup latency. A missing
    provider value remains null; lifecycle state remains Core-owned.
-6. **Collector health:** count attempted, successful, timed-out and unavailable
-   samples, queue drops and export failures. Display coverage per interval so
-   a quiet chart cannot hide collector failure.
+6. **Collector health (partial):** request collection writes a minute heartbeat,
+   including quiet minutes. Request and tool queues report drops and write
+   failures. Runtime sampler and model-attempt collector health are not yet wired
+   into this operator projection; an absent row is unavailable, not zero.
+
+The operator migration adds `observability_request_minute_buckets`,
+`observability_model_attempts`, `observability_tool_attempts`, and
+`observability_collector_minute_buckets`. Request and collector buckets retain
+30 days; model and tool attempts retain seven days. Model attempts remain empty
+until a native per-attempt producer is implemented. The deployment administrator
+summary supports only 1h, 6h, and 24h and returns no tenant or raw identifiers.
 
 Use bounded histograms for p50/p95 latency and rates from counters over complete
 time buckets. Keep operational cardinality to route family, outcome, provider
@@ -65,14 +81,14 @@ must be explicit, with history storage/export optional for execution.
 
 ## Dashboard layout and acceptance
 
-The overview keeps loaded Agent/Session counts and current attention. The Runtime
-section shows allocation-deduplicated lifecycle, coverage, sample gaps, memory
-pressure and existing Live/History trends. A later System section can add request
-rate/error rate/p95, terminal Turn outcomes and durations, actual model usage and
-tool attempts once their collection is qualified. Every panel needs a source and
-freshness label, a coverage denominator, and an unavailable state. It must never
-convert missing measurements to zero or equate a configured Sandbox with a
-completed execution.
+The overview keeps a small current snapshot: Agents, Sessions, active Sandboxes,
+reported tokens, measured CPU/memory capacity, and ready nodes. The Observability
+tab leads with the four existing Runtime trends, followed by request, Turn, tool,
+collector, and node panels. The model panel states that collection is unavailable. The
+request rate, error rate, and range latency require complete collector heartbeat
+coverage; otherwise they are unavailable. Request route tables show counts only
+for covered buckets. Every panel must preserve missing measurements and must not equate
+a configured Sandbox with a completed execution.
 
 Before shipping those new panels, validate counter deduplication under retries,
 provider restarts, missing usage, sampler outages and multi-tenant authorization;
diff --git a/packages/agents-client/src/index.ts b/packages/agents-client/src/index.ts
index 20dd3eaf1..f19125310 100644
--- a/packages/agents-client/src/index.ts
+++ b/packages/agents-client/src/index.ts
@@ -4,3 +4,4 @@ export { createSSEDecoder } from "./sse";
 export type { SSEDecoder, SSEMessage } from "./sse";
 export type * from "./types";
 export * from "./sandbox-client";
+export * from "./observability-client";
diff --git a/packages/agents-client/src/observability-client.ts b/packages/agents-client/src/observability-client.ts
new file mode 100644
index 000000000..1eb4791b4
--- /dev/null
+++ b/packages/agents-client/src/observability-client.ts
@@ -0,0 +1,42 @@
+import { OpenAIAgentsClient } from "./client";
+import type { ReadOptions } from "./types";
+
+export type OperatorMetricsRange = "1h" | "6h" | "24h";
+
+export interface RequestMetricBucket {
+  start: string;
+  route_family: string;
+  outcome: "success" | "client_error" | "server_error";
+  count: number;
+  latency_sum_ms: number;
+  latency_bucket_counts: number[];
+}
+
+export interface CollectorMetricBucket {
+  start: string;
+  source: string;
+  attempted_count: number;
+  observed_count: number;
+  unavailable_count: number;
+  timeout_count: number;
+  dropped_count: number;
+  export_failed_count: number;
+}
+
+export interface OperatorMetricsSummary {
+  generated_at: string;
+  start: string;
+  end: string;
+  step_seconds: number;
+  requests: RequestMetricBucket[];
+  collector: CollectorMetricBucket[];
+  turns: Array<{ start: string; status: "completed" | "failed" | "cancelled"; count: number; queue_p95_ms: number | null; execution_p95_ms: number | null }>;
+  tools: Array<{ start: string; category: string; outcome: "success" | "error" | "cancelled" | "unknown"; count: number; timed_count: number; duration_p95_ms: number | null }>;
+}
+
+/** Deployment administrator extension; never uses a project credential. */
+export class ObservabilityAdminClient extends OpenAIAgentsClient {
+  retrieveSummary(range: OperatorMetricsRange, options?: ReadOptions): Promise<OperatorMetricsSummary> {
+    return this.request(`/summary?range=${range}`, { signal: options?.signal }, undefined, false);
+  }
+}
diff --git a/services/agents-api/cmd/server/main.go b/services/agents-api/cmd/server/main.go
index 985126151..6a75ed261 100644
--- a/services/agents-api/cmd/server/main.go
+++ b/services/agents-api/cmd/server/main.go
@@ -35,6 +35,7 @@ import (
 	"github.com/MiniMax-AI-Dev/parsar/internal/obs/log"
 	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/api"
 	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/execution"
+	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/observability"
 	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/runtime"
 	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/runtimeenrollment"
 	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/runtimehistory"
@@ -103,6 +104,9 @@ func run() error {
 		return err
 	}
 	executionStore := store.NewWithCredentialCipherAndOAuthRefresh(pool, credentialKey, oauthClient)
+	toolMetricsCtx, cancelToolMetrics := context.WithCancel(ctx)
+	toolMetrics := observability.NewToolRecorder(toolMetricsCtx, executionStore)
+	defer func() { cancelToolMetrics(); <-toolMetrics.Done() }()
 	if err := executionStore.EnsureProjectScopes(ready, auth.ProjectScopes()); err != nil {
 		return err
 	}
@@ -187,6 +191,7 @@ func run() error {
 		}
 	}
 	options = append(options, api.WithProjectAPIKeys(executionStore, keyAdmin), api.WithWriteAudit(executionStore, keyAdmin))
+	options = append(options, api.WithOperatorMetrics(executionStore, keyAdmin))
 	if history.Reader != nil {
 		historyResolver, resolverErr := historystoreresolver.NewResolver(executionStore)
 		if resolverErr != nil {
@@ -210,7 +215,7 @@ func run() error {
 	}
 	if registry != nil {
 		dispatcher := &execution.Dispatcher{Store: executionStore, Registry: registry,
-			ManagedRuntimes: managed, Options: transientOptions}
+			ManagedRuntimes: managed, Options: transientOptions, ToolRecorder: toolMetrics}
 
 		worker, err = execution.StartWorker(ctx, dispatcher)
 		if err != nil {
@@ -264,6 +269,17 @@ func run() error {
 		startupManaged = managedNodes.setup.selected.Load()
 	}
 	options = append(options, api.WithStartupConfiguration(coreStartupConfiguration(engine, kinds, registry != nil, modelProviderEndpoints, managedRuntimeProviderKind(startupManaged), startupManaged)))
+	requestMetricsCtx, cancelRequestMetrics := context.WithCancel(ctx)
+	requestMetrics := observability.NewRequestRecorder(requestMetricsCtx, executionStore)
+	defer func() { cancelRequestMetrics(); <-requestMetrics.Done() }()
+	options = append(options, api.WithRequestMetrics(requestMetrics))
+	operatorCleanupCtx, cancelOperatorCleanup := context.WithCancel(ctx)
+	operatorCleanupDone := make(chan struct{})
+	go func() {
+		defer close(operatorCleanupDone)
+		runOperatorMetricsCleanup(operatorCleanupCtx, executionStore)
+	}()
+	defer func() { cancelOperatorCleanup(); <-operatorCleanupDone }()
 	handler, err := api.NewHandler(executionStore, auth, engine, options...)
 	if err != nil {
 		return err
diff --git a/services/agents-api/cmd/server/observability_cleanup.go b/services/agents-api/cmd/server/observability_cleanup.go
new file mode 100644
index 000000000..ed835f2fc
--- /dev/null
+++ b/services/agents-api/cmd/server/observability_cleanup.go
@@ -0,0 +1,30 @@
+package main
+
+import (
+	"context"
+	"time"
+
+	"github.com/MiniMax-AI-Dev/parsar/internal/obs/log"
+)
+
+type operatorMetricsPruner interface {
+	PruneOperatorMetrics(context.Context, time.Time) error
+}
+
+func runOperatorMetricsCleanup(ctx context.Context, pruner operatorMetricsPruner) {
+	ticker := time.NewTicker(time.Minute)
+	defer ticker.Stop()
+	for {
+		pruneCtx, cancel := context.WithTimeout(ctx, 3*time.Second)
+		err := pruner.PruneOperatorMetrics(pruneCtx, time.Now().UTC())
+		cancel()
+		if err != nil && ctx.Err() == nil {
+			log.Ctx(ctx).Warn("Operator metrics retention cleanup failed")
+		}
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+		}
+	}
+}
diff --git a/services/agents-api/internal/api/handler.go b/services/agents-api/internal/api/handler.go
index 7cb048e89..4db24b692 100644
--- a/services/agents-api/internal/api/handler.go
+++ b/services/agents-api/internal/api/handler.go
@@ -13,6 +13,7 @@ import (
 	"github.com/MiniMax-AI-Dev/parsar/internal/obs/log"
 	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/execution"
 	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/identity"
+	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/observability"
 	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/store"
 	"github.com/go-chi/chi/v5"
 	"github.com/go-chi/chi/v5/middleware"
@@ -61,6 +62,8 @@ type Handler struct {
 	subagents             SubagentStore
 	runtimeObservations   RuntimeObservationService
 	runtimeHistory        RuntimeHistoryService
+	requestMetrics        *observability.RequestRecorder
+	operatorMetrics       OperatorMetricsStore
 	startup               *v1.CoreStartupConfiguration
 }
 
@@ -92,6 +95,9 @@ func NewHandler(s ResourceStore, auth *Authenticator, engine string, options ...
 func (h *Handler) routes() *chi.Mux {
 	router := chi.NewRouter()
 	router.Use(agentsResponseHeaders, log.HTTPMiddleware, middleware.GetHead)
+	if h.requestMetrics != nil {
+		router.Use(h.recordRequestMetrics)
+	}
 	router.MethodNotAllowed(methodNotAllowed)
 	router.Get("/healthz", func(w http.ResponseWriter, _ *http.Request) {
 		writeJSON(w, http.StatusOK, map[string]string{"status": "ok"})
@@ -109,6 +115,7 @@ func (h *Handler) routes() *chi.Mux {
 	h.registerSandboxManagerRoutes(router)
 	h.registerProjectAPIKeyRoutes(router)
 	h.registerWriteAuditRoutes(router)
+	h.registerOperatorMetricsRoutes(router)
 	h.registerEnvironmentExecutorRoutes(router)
 	router.Route("/v1", func(r chi.Router) {
 		r.Use(h.authenticate)
diff --git a/services/agents-api/internal/api/observability_read.go b/services/agents-api/internal/api/observability_read.go
new file mode 100644
index 000000000..fd79ec262
--- /dev/null
+++ b/services/agents-api/internal/api/observability_read.go
@@ -0,0 +1,88 @@
+package api
+
+import (
+	"context"
+	"net/http"
+	"time"
+
+	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/store"
+	"github.com/go-chi/chi/v5"
+)
+
+type OperatorMetricsStore interface {
+	ReadOperatorMetrics(context.Context, time.Time, time.Time, time.Duration) (store.OperatorMetrics, error)
+}
+
+type OperatorMetricsResponse struct {
+	GeneratedAt time.Time                     `json:"generated_at"`
+	Start       time.Time                     `json:"start"`
+	End         time.Time                     `json:"end"`
+	StepSeconds int64                         `json:"step_seconds"`
+	Requests    []store.RequestMetricBucket   `json:"requests"`
+	Collector   []store.CollectorMetricBucket `json:"collector"`
+	Turns       []store.TurnMetricBucket      `json:"turns"`
+	Tools       []store.ToolMetricBucket      `json:"tools"`
+}
+
+func WithOperatorMetrics(s OperatorMetricsStore, auth *DeploymentAuthenticator) Option {
+	return func(h *Handler) {
+		h.operatorMetrics = s
+		if auth != nil {
+			h.deploymentAuth = auth
+		}
+	}
+}
+
+func (h *Handler) registerOperatorMetricsRoutes(r chi.Router) {
+	if h.operatorMetrics == nil || h.deploymentAuth == nil {
+		return
+	}
+	r.Group(func(r chi.Router) {
+		r.Use(h.deploymentAuth.authenticate)
+		r.Get("/core/v1/observability/summary", h.getOperatorMetrics)
+		r.Head("/core/v1/observability/summary", methodNotAllowed)
+	})
+}
+
+// @Summary Retrieve bounded operator observability metrics
+// @Description Deployment administrator only. Returns fixed-range aggregate requests and collector coverage without paths, payloads or project identifiers.
+// @Tags Observability
+// @Produce json
+// @Security DeploymentAdminAuth
+// @Param range query string false "1h, 6h, or 24h (default 1h)"
+// @Success 200 {object} api.OperatorMetricsResponse
+// @Failure 400,401,500,503 {object} v1.ErrorResponse
+// @Router /core/v1/observability/summary [get]
+func (h *Handler) getOperatorMetrics(w http.ResponseWriter, r *http.Request) {
+	values := r.URL.Query()
+	for key, entries := range values {
+		if key != "range" || len(entries) != 1 {
+			writeStoreError(w, r, store.ErrInvalidInput)
+			return
+		}
+	}
+	span, step := time.Hour, time.Minute
+	switch values.Get("range") {
+	case "", "1h":
+	case "6h":
+		span, step = 6*time.Hour, 5*time.Minute
+	case "24h":
+		span, step = 24*time.Hour, 15*time.Minute
+	default:
+		writeStoreError(w, r, store.ErrInvalidInput)
+		return
+	}
+	end := time.Now().UTC().Truncate(step)
+	start := end.Add(-span)
+	ctx, cancel := context.WithTimeout(r.Context(), 5*time.Second)
+	defer cancel()
+	metrics, err := h.operatorMetrics.ReadOperatorMetrics(ctx, start, end, step)
+	if err != nil {
+		writeStoreError(w, r, err)
+		return
+	}
+	writeJSON(w, http.StatusOK, OperatorMetricsResponse{
+		GeneratedAt: time.Now().UTC(), Start: start, End: end, StepSeconds: int64(step / time.Second),
+		Requests: metrics.Requests, Collector: metrics.Collector, Turns: metrics.Turns, Tools: metrics.Tools,
+	})
+}
diff --git a/services/agents-api/internal/api/observability_read_test.go b/services/agents-api/internal/api/observability_read_test.go
new file mode 100644
index 000000000..5f9e5ee47
--- /dev/null
+++ b/services/agents-api/internal/api/observability_read_test.go
@@ -0,0 +1,62 @@
+package api
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/MiniMax-AI-Dev/parsar/internal/agentdaemon/device"
+	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/store"
+)
+
+type operatorMetricsStub struct{ calls int }
+
+func (s *operatorMetricsStub) ReadOperatorMetrics(_ context.Context, start, end time.Time, step time.Duration) (store.OperatorMetrics, error) {
+	s.calls++
+	if end.Sub(start) != time.Hour || step != time.Minute {
+		return store.OperatorMetrics{}, store.ErrInvalidInput
+	}
+	return store.OperatorMetrics{Requests: []store.RequestMetricBucket{}, Collector: []store.CollectorMetricBucket{}, Turns: []store.TurnMetricBucket{}, Tools: []store.ToolMetricBucket{}}, nil
+}
+
+func TestOperatorMetricsRequireAdministratorAndBoundedRange(t *testing.T) {
+	project, err := NewAuthenticator([]APIKey{callerBinding()})
+	if err != nil {
+		t.Fatal(err)
+	}
+	admin, err := NewDeploymentAuthenticator([]string{device.HashCredential("administrator")})
+	if err != nil {
+		t.Fatal(err)
+	}
+	metrics := &operatorMetricsStub{}
+	h, err := NewHandler(&recordingStore{}, project, "codex", WithOperatorMetrics(metrics, admin))
+	if err != nil {
+		t.Fatal(err)
+	}
+	request := func(path, key string) *httptest.ResponseRecorder {
+		req := httptest.NewRequest(http.MethodGet, path, nil)
+		req.Header.Set("Authorization", "Bearer "+key)
+		response := httptest.NewRecorder()
+		h.ServeHTTP(response, req)
+		return response
+	}
+	for _, key := range []string{"caller", ""} {
+		response := request("/core/v1/observability/summary?range=1h", key)
+		if response.Code != http.StatusUnauthorized || metrics.calls != 0 {
+			t.Fatal("project or missing key reached operator metrics", response.Code, metrics.calls)
+		}
+	}
+	for _, path := range []string{"/core/v1/observability/summary?range=30d", "/core/v1/observability/summary?range=1h&range=24h", "/core/v1/observability/summary?tenant_id=x"} {
+		response := request(path, "administrator")
+		if response.Code != http.StatusBadRequest || metrics.calls != 0 {
+			t.Fatal("unbounded metrics query admitted", response.Code, metrics.calls)
+		}
+	}
+	response := request("/core/v1/observability/summary?range=1h", "administrator")
+	if response.Code != http.StatusOK || metrics.calls != 1 || !strings.Contains(response.Body.String(), `"requests":[]`) {
+		t.Fatal("administrator metrics read failed", response.Code, metrics.calls, response.Body.String())
+	}
+}
diff --git a/services/agents-api/internal/api/observability_requests.go b/services/agents-api/internal/api/observability_requests.go
new file mode 100644
index 000000000..7c420df60
--- /dev/null
+++ b/services/agents-api/internal/api/observability_requests.go
@@ -0,0 +1,76 @@
+package api
+
+import (
+	"net/http"
+	"strings"
+	"time"
+
+	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/observability"
+	"github.com/go-chi/chi/v5"
+	"github.com/go-chi/chi/v5/middleware"
+)
+
+func WithRequestMetrics(recorder *observability.RequestRecorder) Option {
+	return func(h *Handler) { h.requestMetrics = recorder }
+}
+
+func (h *Handler) recordRequestMetrics(next http.Handler) http.Handler {
+	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path == "/healthz" || strings.HasPrefix(r.URL.Path, "/core/v1/observability/") {
+			next.ServeHTTP(w, r)
+			return
+		}
+		start := time.Now()
+		wrapped := middleware.NewWrapResponseWriter(w, r.ProtoMajor)
+		next.ServeHTTP(wrapped, r)
+		status := wrapped.Status()
+		if status == 0 {
+			status = http.StatusOK
+		}
+		outcome := "success"
+		if status >= 500 {
+			outcome = "server_error"
+		} else if status >= 400 {
+			outcome = "client_error"
+		}
+		method := r.Method
+		switch method {
+		case http.MethodGet, http.MethodPost, http.MethodPut, http.MethodPatch, http.MethodDelete, http.MethodHead:
+		default:
+			method = "OTHER"
+		}
+		pattern := ""
+		if route := chi.RouteContext(r.Context()); route != nil {
+			pattern = route.RoutePattern()
+		}
+		h.requestMetrics.Record(observability.Request{
+			CompletedAt: time.Now().UTC(), RouteFamily: requestRouteFamily(pattern),
+			Method: method, Outcome: outcome, Latency: time.Since(start),
+		})
+	})
+}
+
+func requestRouteFamily(pattern string) string {
+	switch {
+	case strings.HasPrefix(pattern, "/core/v1/sandbox"):
+		return "sandbox_admin"
+	case strings.HasPrefix(pattern, "/core/v1"):
+		return "core_admin"
+	case strings.HasPrefix(pattern, "/v1/agents/sessions"):
+		return "sessions"
+	case strings.HasPrefix(pattern, "/v1/agents/runtime"):
+		return "runtime"
+	case strings.HasPrefix(pattern, "/v1/agents/environments"):
+		return "environments"
+	case strings.HasPrefix(pattern, "/v1/agents"):
+		return "agents"
+	case strings.HasPrefix(pattern, "/v1/files"):
+		return "files"
+	case strings.HasPrefix(pattern, "/v1/vaults"):
+		return "vaults"
+	case strings.HasPrefix(pattern, "/v1/skills"):
+		return "skills"
+	default:
+		return "other"
+	}
+}
diff --git a/services/agents-api/internal/db/queries/runtime_history.sql b/services/agents-api/internal/db/queries/runtime_history.sql
index d0c019024..6a20f0a00 100644
--- a/services/agents-api/internal/db/queries/runtime_history.sql
+++ b/services/agents-api/internal/db/queries/runtime_history.sql
@@ -1,5 +1,5 @@
 -- name: InsertRuntimeHistorySample :exec
-INSERT INTO runtime_history_samples (
+INSERT INTO observability_runtime_samples (
     tenant_id, session_id, environment_id, resolved_at_ns, allocation_id, provider_type,
     status, observed_at_ns, started_at_ns, cpu_usage_seconds, cpu_capacity_cores,
     memory_usage_bytes, memory_limit_bytes, input_tokens, output_tokens
@@ -12,7 +12,7 @@ WHERE s.tenant_id = sqlc.arg(tenant_id) AND s.id = sqlc.arg(session_id) AND e.id
 ON CONFLICT (tenant_id, session_id, environment_id, resolved_at_ns) DO NOTHING;
 
 -- name: ListRuntimeHistorySamples :many
-SELECT * FROM runtime_history_samples
+SELECT * FROM observability_runtime_samples
 WHERE tenant_id = sqlc.arg(tenant_id) AND session_id = sqlc.arg(session_id) AND environment_id = sqlc.arg(environment_id)
   AND resolved_at_ns >= sqlc.arg(start_ns) AND resolved_at_ns < sqlc.arg(end_ns)
 ORDER BY resolved_at_ns
@@ -20,10 +20,10 @@ LIMIT sqlc.arg(row_limit);
 
 -- name: PruneRuntimeHistorySamples :execrows
 WITH expired AS (
-    SELECT p.tenant_id, p.session_id, p.environment_id, p.resolved_at_ns FROM runtime_history_samples p
+    SELECT p.tenant_id, p.session_id, p.environment_id, p.resolved_at_ns FROM observability_runtime_samples p
     WHERE p.resolved_at_ns < sqlc.arg(before_ns)
     ORDER BY p.resolved_at_ns LIMIT 256 FOR UPDATE SKIP LOCKED
 )
-DELETE FROM runtime_history_samples h USING expired e
+DELETE FROM observability_runtime_samples h USING expired e
 WHERE h.tenant_id = e.tenant_id AND h.session_id = e.session_id
   AND h.environment_id = e.environment_id AND h.resolved_at_ns = e.resolved_at_ns;
diff --git a/services/agents-api/internal/db/sqlc/models.go b/services/agents-api/internal/db/sqlc/models.go
index 0c64280b3..3cf0776df 100644
--- a/services/agents-api/internal/db/sqlc/models.go
+++ b/services/agents-api/internal/db/sqlc/models.go
@@ -136,6 +136,73 @@ type InitialEnvironmentFile struct {
 	Contents  []byte      `json:"contents"`
 }
 
+type ObservabilityCollectorMinuteBucket struct {
+	BucketStart       pgtype.Timestamptz `json:"bucket_start"`
+	Source            string             `json:"source"`
+	AttemptedCount    int64              `json:"attempted_count"`
+	ObservedCount     int64              `json:"observed_count"`
+	UnavailableCount  int64              `json:"unavailable_count"`
+	TimeoutCount      int64              `json:"timeout_count"`
+	DroppedCount      int64              `json:"dropped_count"`
+	ExportFailedCount int64              `json:"export_failed_count"`
+}
+
+type ObservabilityModelAttempt struct {
+	ID           pgtype.UUID        `json:"id"`
+	TenantID     pgtype.UUID        `json:"tenant_id"`
+	SessionID    pgtype.UUID        `json:"session_id"`
+	TurnID       pgtype.UUID        `json:"turn_id"`
+	StartedAt    pgtype.Timestamptz `json:"started_at"`
+	FinishedAt   pgtype.Timestamptz `json:"finished_at"`
+	ModelFamily  string             `json:"model_family"`
+	ProviderType string             `json:"provider_type"`
+	Outcome      string             `json:"outcome"`
+	DurationMs   int64              `json:"duration_ms"`
+	InputTokens  pgtype.Int8        `json:"input_tokens"`
+	OutputTokens pgtype.Int8        `json:"output_tokens"`
+	UsageStatus  string             `json:"usage_status"`
+}
+
+type ObservabilityRequestMinuteBucket struct {
+	BucketStart         pgtype.Timestamptz `json:"bucket_start"`
+	RouteFamily         string             `json:"route_family"`
+	Method              string             `json:"method"`
+	Outcome             string             `json:"outcome"`
+	RequestCount        int64              `json:"request_count"`
+	LatencySumMs        int64              `json:"latency_sum_ms"`
+	LatencyBucketCounts []int64            `json:"latency_bucket_counts"`
+}
+
+type ObservabilityRuntimeSample struct {
+	TenantID         pgtype.UUID   `json:"tenant_id"`
+	SessionID        pgtype.UUID   `json:"session_id"`
+	EnvironmentID    pgtype.UUID   `json:"environment_id"`
+	ResolvedAtNs     int64         `json:"resolved_at_ns"`
+	AllocationID     pgtype.UUID   `json:"allocation_id"`
+	ProviderType     string        `json:"provider_type"`
+	Status           string        `json:"status"`
+	ObservedAtNs     pgtype.Int8   `json:"observed_at_ns"`
+	StartedAtNs      pgtype.Int8   `json:"started_at_ns"`
+	CpuUsageSeconds  pgtype.Float8 `json:"cpu_usage_seconds"`
+	CpuCapacityCores pgtype.Float8 `json:"cpu_capacity_cores"`
+	MemoryUsageBytes pgtype.Int8   `json:"memory_usage_bytes"`
+	MemoryLimitBytes pgtype.Int8   `json:"memory_limit_bytes"`
+	InputTokens      pgtype.Int8   `json:"input_tokens"`
+	OutputTokens     pgtype.Int8   `json:"output_tokens"`
+}
+
+type ObservabilityToolAttempt struct {
+	ID           pgtype.UUID        `json:"id"`
+	TenantID     pgtype.UUID        `json:"tenant_id"`
+	SessionID    pgtype.UUID        `json:"session_id"`
+	TurnID       pgtype.UUID        `json:"turn_id"`
+	StartedAt    pgtype.Timestamptz `json:"started_at"`
+	FinishedAt   pgtype.Timestamptz `json:"finished_at"`
+	ToolCategory string             `json:"tool_category"`
+	Outcome      string             `json:"outcome"`
+	DurationMs   pgtype.Int8        `json:"duration_ms"`
+}
+
 type ProjectApiKey struct {
 	ID             pgtype.UUID        `json:"id"`
 	Name           string             `json:"name"`
@@ -213,24 +280,6 @@ type RuntimeDeviceAuthority struct {
 	CredentialHash string      `json:"credential_hash"`
 }
 
-type RuntimeHistorySample struct {
-	TenantID         pgtype.UUID   `json:"tenant_id"`
-	SessionID        pgtype.UUID   `json:"session_id"`
-	EnvironmentID    pgtype.UUID   `json:"environment_id"`
-	ResolvedAtNs     int64         `json:"resolved_at_ns"`
-	AllocationID     pgtype.UUID   `json:"allocation_id"`
-	ProviderType     string        `json:"provider_type"`
-	Status           string        `json:"status"`
-	ObservedAtNs     pgtype.Int8   `json:"observed_at_ns"`
-	StartedAtNs      pgtype.Int8   `json:"started_at_ns"`
-	CpuUsageSeconds  pgtype.Float8 `json:"cpu_usage_seconds"`
-	CpuCapacityCores pgtype.Float8 `json:"cpu_capacity_cores"`
-	MemoryUsageBytes pgtype.Int8   `json:"memory_usage_bytes"`
-	MemoryLimitBytes pgtype.Int8   `json:"memory_limit_bytes"`
-	InputTokens      pgtype.Int8   `json:"input_tokens"`
-	OutputTokens     pgtype.Int8   `json:"output_tokens"`
-}
-
 type RuntimeNode struct {
 	ID                 pgtype.UUID        `json:"id"`
 	InstallationID     pgtype.UUID        `json:"installation_id"`
diff --git a/services/agents-api/internal/db/sqlc/runtime_history.sql.go b/services/agents-api/internal/db/sqlc/runtime_history.sql.go
index 1c6a84eac..24567a81c 100644
--- a/services/agents-api/internal/db/sqlc/runtime_history.sql.go
+++ b/services/agents-api/internal/db/sqlc/runtime_history.sql.go
@@ -12,7 +12,7 @@ import (
 )
 
 const insertRuntimeHistorySample = `-- name: InsertRuntimeHistorySample :exec
-INSERT INTO runtime_history_samples (
+INSERT INTO observability_runtime_samples (
     tenant_id, session_id, environment_id, resolved_at_ns, allocation_id, provider_type,
     status, observed_at_ns, started_at_ns, cpu_usage_seconds, cpu_capacity_cores,
     memory_usage_bytes, memory_limit_bytes, input_tokens, output_tokens
@@ -65,7 +65,7 @@ func (q *Queries) InsertRuntimeHistorySample(ctx context.Context, arg InsertRunt
 }
 
 const listRuntimeHistorySamples = `-- name: ListRuntimeHistorySamples :many
-SELECT tenant_id, session_id, environment_id, resolved_at_ns, allocation_id, provider_type, status, observed_at_ns, started_at_ns, cpu_usage_seconds, cpu_capacity_cores, memory_usage_bytes, memory_limit_bytes, input_tokens, output_tokens FROM runtime_history_samples
+SELECT tenant_id, session_id, environment_id, resolved_at_ns, allocation_id, provider_type, status, observed_at_ns, started_at_ns, cpu_usage_seconds, cpu_capacity_cores, memory_usage_bytes, memory_limit_bytes, input_tokens, output_tokens FROM observability_runtime_samples
 WHERE tenant_id = $1 AND session_id = $2 AND environment_id = $3
   AND resolved_at_ns >= $4 AND resolved_at_ns < $5
 ORDER BY resolved_at_ns
@@ -81,7 +81,7 @@ type ListRuntimeHistorySamplesParams struct {
 	RowLimit      int32       `json:"row_limit"`
 }
 
-func (q *Queries) ListRuntimeHistorySamples(ctx context.Context, arg ListRuntimeHistorySamplesParams) ([]RuntimeHistorySample, error) {
+func (q *Queries) ListRuntimeHistorySamples(ctx context.Context, arg ListRuntimeHistorySamplesParams) ([]ObservabilityRuntimeSample, error) {
 	rows, err := q.db.Query(ctx, listRuntimeHistorySamples,
 		arg.TenantID,
 		arg.SessionID,
@@ -94,9 +94,9 @@ func (q *Queries) ListRuntimeHistorySamples(ctx context.Context, arg ListRuntime
 		return nil, err
 	}
 	defer rows.Close()
-	items := []RuntimeHistorySample{}
+	items := []ObservabilityRuntimeSample{}
 	for rows.Next() {
-		var i RuntimeHistorySample
+		var i ObservabilityRuntimeSample
 		if err := rows.Scan(
 			&i.TenantID,
 			&i.SessionID,
@@ -126,11 +126,11 @@ func (q *Queries) ListRuntimeHistorySamples(ctx context.Context, arg ListRuntime
 
 const pruneRuntimeHistorySamples = `-- name: PruneRuntimeHistorySamples :execrows
 WITH expired AS (
-    SELECT p.tenant_id, p.session_id, p.environment_id, p.resolved_at_ns FROM runtime_history_samples p
+    SELECT p.tenant_id, p.session_id, p.environment_id, p.resolved_at_ns FROM observability_runtime_samples p
     WHERE p.resolved_at_ns < $1
     ORDER BY p.resolved_at_ns LIMIT 256 FOR UPDATE SKIP LOCKED
 )
-DELETE FROM runtime_history_samples h USING expired e
+DELETE FROM observability_runtime_samples h USING expired e
 WHERE h.tenant_id = e.tenant_id AND h.session_id = e.session_id
   AND h.environment_id = e.environment_id AND h.resolved_at_ns = e.resolved_at_ns
 `
diff --git a/services/agents-api/internal/execution/delivery.go b/services/agents-api/internal/execution/delivery.go
index 52b6d59dd..9e97f14c3 100644
--- a/services/agents-api/internal/execution/delivery.go
+++ b/services/agents-api/internal/execution/delivery.go
@@ -65,7 +65,7 @@ func (d *Dispatcher) deliver(ctx context.Context, tenantID, sessionID string, pe
 		}
 	}()
 	journal := &journal{store: d.Store, tenant: tenantID, session: sessionID, turn: request.RunID, next: 1,
-		observeSubagents: request.ObserveSubagentIdentities}
+		observeSubagents: request.ObserveSubagentIdentities, toolRecorder: d.ToolRecorder}
 	defer func() {
 		finishCtx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
 		defer cancel()
diff --git a/services/agents-api/internal/execution/dispatcher.go b/services/agents-api/internal/execution/dispatcher.go
index 3d9c98e03..4370da884 100644
--- a/services/agents-api/internal/execution/dispatcher.go
+++ b/services/agents-api/internal/execution/dispatcher.go
@@ -11,6 +11,7 @@ import (
 	v1 "github.com/MiniMax-AI-Dev/parsar/contracts/agents-api/v1"
 	"github.com/MiniMax-AI-Dev/parsar/internal/agentdaemon/gateway"
 	"github.com/MiniMax-AI-Dev/parsar/internal/agentdaemon/proto"
+	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/observability"
 	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/store"
 )
 
@@ -36,6 +37,10 @@ type Dispatcher struct {
 	Options func(context.Context, store.Session) (map[string]any, error)
 	// ManagedRuntimes is optional internal provisioning; it does not admit hosted API requests.
 	ManagedRuntimes *RuntimeProvider
+	// ToolRecorder receives sanitized terminal observations without affecting execution.
+	ToolRecorder interface {
+		Record(observability.ToolAttempt)
+	}
 }
 
 type Result struct {
diff --git a/services/agents-api/internal/execution/journal.go b/services/agents-api/internal/execution/journal.go
index ae22830de..0835a42c2 100644
--- a/services/agents-api/internal/execution/journal.go
+++ b/services/agents-api/internal/execution/journal.go
@@ -6,7 +6,9 @@ import (
 	"time"
 
 	"github.com/MiniMax-AI-Dev/parsar/internal/agentdaemon/proto"
+	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/observability"
 	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/store"
+	"github.com/google/uuid"
 )
 
 type journal struct {
@@ -14,9 +16,14 @@ type journal struct {
 	tenant, session, turn string
 	next                  int32
 	batch                 []store.ExecutionEvent
+	batchTimes            []time.Time
 	bytes                 int
 	pendingCount          int
 	observeSubagents      bool
+	toolRecorder          interface {
+		Record(observability.ToolAttempt)
+	}
+	toolStarts map[string]time.Time
 }
 
 type eventWriter interface {
@@ -67,10 +74,66 @@ func (j *journal) enqueue(env proto.Envelope) error {
 		return store.ErrEventLimit
 	}
 	j.batch = append(j.batch, store.ExecutionEvent{Kind: env.Type, Payload: env.Payload})
+	j.batchTimes = append(j.batchTimes, time.Now().UTC())
 	j.bytes += len(env.Payload)
 	return nil
 }
 
+func (j *journal) observeToolAttempt(raw json.RawMessage, observedAt time.Time) {
+	if j.toolRecorder == nil {
+		return
+	}
+	var call proto.ToolCallPayload
+	if json.Unmarshal(raw, &call) != nil || call.ID == "" {
+		return
+	}
+	if call.Stage == "before" {
+		if j.toolStarts == nil {
+			j.toolStarts = make(map[string]time.Time)
+		}
+		if _, exists := j.toolStarts[call.ID]; !exists {
+			j.toolStarts[call.ID] = observedAt
+		}
+		return
+	}
+	if call.Stage != "after" {
+		return
+	}
+	finished := observedAt
+	var started *time.Time
+	if value, exists := j.toolStarts[call.ID]; exists {
+		started = &value
+		delete(j.toolStarts, call.ID)
+	}
+	category, outcome := "other", "unknown"
+	var duration *int64
+	if call.Observation != nil {
+		switch call.Observation.Kind {
+		case "command", "mcp", "function", "web_search", "file":
+			category = call.Observation.Kind
+		}
+		switch call.Observation.Status {
+		case "completed":
+			outcome = "success"
+		case "failed":
+			outcome = "error"
+		}
+		if call.Observation.DurationMS != nil && *call.Observation.DurationMS >= 0 {
+			value := *call.Observation.DurationMS
+			duration = &value
+		}
+	}
+	if duration == nil && started != nil {
+		value := finished.Sub(*started).Milliseconds()
+		duration = &value
+	}
+	j.toolRecorder.Record(observability.ToolAttempt{
+		ID:       uuid.NewSHA1(uuid.NameSpaceOID, []byte(j.turn+":"+call.ID)).String(),
+		TenantID: j.tenant, SessionID: j.session, TurnID: j.turn,
+		StartedAt: started, FinishedAt: finished, Category: category, Outcome: outcome, DurationMS: duration,
+	})
+}
+
 func (j *journal) flush(ctx context.Context) error {
 	if len(j.batch) == 0 {
 		return nil
@@ -96,12 +159,19 @@ func (j *journal) flush(ctx context.Context) error {
 		if err := j.store.AppendTurnEvents(ctx, j.tenant, j.session, j.turn, j.next, j.batch[:count]); err != nil {
 			return err
 		}
+		for index, event := range j.batch[:count] {
+			if event.Kind == proto.TypeToolCall {
+				j.observeToolAttempt(event.Payload, j.batchTimes[index])
+			}
+		}
 		j.pendingCount = 0
 		j.next += int32(count)
 		j.bytes -= size
 		j.batch = j.batch[count:]
+		j.batchTimes = j.batchTimes[count:]
 	}
 	j.batch = nil
+	j.batchTimes = nil
 	return nil
 }
 
diff --git a/services/agents-api/internal/execution/observability_tools_test.go b/services/agents-api/internal/execution/observability_tools_test.go
new file mode 100644
index 000000000..2a4cabf28
--- /dev/null
+++ b/services/agents-api/internal/execution/observability_tools_test.go
@@ -0,0 +1,60 @@
+package execution
+
+import (
+	"context"
+	"errors"
+	"testing"
+
+	"github.com/MiniMax-AI-Dev/parsar/internal/agentdaemon/proto"
+	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/observability"
+	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/store"
+)
+
+type toolCapture struct{ attempts []observability.ToolAttempt }
+
+func (c *toolCapture) Record(value observability.ToolAttempt) { c.attempts = append(c.attempts, value) }
+
+type toolEventWriter struct{ err error }
+
+func (w *toolEventWriter) AppendTurnEvents(_ context.Context, _, _, _ string, _ int32, _ []store.ExecutionEvent) error {
+	return w.err
+}
+
+func TestToolAttemptProjectsOnlyTerminalSanitizedEvidence(t *testing.T) {
+	capture := &toolCapture{}
+	writer := &toolEventWriter{}
+	j := &journal{store: writer, tenant: "tenant", session: "session", turn: "turn", toolRecorder: capture}
+	before, err := proto.NewEnvelope(proto.TypeToolCall, "turn", proto.ToolCallPayload{
+		ID: "native-call", Stage: "before", Name: "secret-tool-name",
+		Observation: &proto.ToolObservation{Kind: "command", Status: "in_progress", Command: "secret command"},
+	})
+	if err != nil || j.enqueue(before) != nil || len(capture.attempts) != 0 {
+		t.Fatalf("tool start should not count as terminal: %v", err)
+	}
+	after, err := proto.NewEnvelope(proto.TypeToolCall, "turn", proto.ToolCallPayload{
+		ID: "native-call", Stage: "after", Name: "secret-tool-name",
+		Observation: &proto.ToolObservation{Kind: "command", Status: "failed", Command: "secret command"},
+	})
+	if err != nil || j.enqueue(after) != nil || len(capture.attempts) != 0 {
+		t.Fatalf("tool attempt must wait for durable event append: %v", err)
+	}
+	writer.err = errors.New("event persistence failed")
+	if j.flush(context.Background()) == nil || len(capture.attempts) != 0 {
+		t.Fatal("failed event append produced a tool metric")
+	}
+	writer.err = nil
+	if err := j.flush(context.Background()); err != nil || len(capture.attempts) != 1 {
+		t.Fatalf("durable terminal tool observation missing: %v", err)
+	}
+	got := capture.attempts[0]
+	if got.Category != "command" || got.Outcome != "error" || got.StartedAt == nil || got.DurationMS == nil ||
+		got.ID == "native-call" || got.ID == "secret-tool-name" || got.TurnID != "turn" {
+		t.Fatalf("tool attempt lost its bounded projection: %+v", got)
+	}
+	if err := j.enqueue(after); err != nil {
+		t.Fatal(err)
+	}
+	if err := j.flush(context.Background()); err != nil || capture.attempts[1].ID != got.ID {
+		t.Fatalf("retry did not retain deterministic attempt identity: %v", err)
+	}
+}
diff --git a/services/agents-api/internal/observability/requests.go b/services/agents-api/internal/observability/requests.go
new file mode 100644
index 000000000..d871906de
--- /dev/null
+++ b/services/agents-api/internal/observability/requests.go
@@ -0,0 +1,134 @@
+// Package observability collects bounded, side-channel operator metrics.
+package observability
+
+import (
+	"context"
+	"sync/atomic"
+	"time"
+)
+
+// Request contains only bounded labels; it never carries a URL, principal, or payload.
+type Request struct {
+	CompletedAt time.Time
+	RouteFamily string
+	Method      string
+	Outcome     string
+	Latency     time.Duration
+}
+
+type RequestBucket struct {
+	Start         time.Time
+	RouteFamily   string
+	Method        string
+	Outcome       string
+	Count         int64
+	LatencySumMS  int64
+	LatencyCounts [11]int64
+}
+
+type RequestSink interface {
+	WriteRequestBuckets(context.Context, []RequestBucket, int64, int64) error
+}
+
+// RequestRecorder batches metrics away from the HTTP response path.
+type RequestRecorder struct {
+	queue   chan Request
+	sink    RequestSink
+	dropped atomic.Int64
+	failed  atomic.Int64
+	done    chan struct{}
+}
+
+func NewRequestRecorder(ctx context.Context, sink RequestSink) *RequestRecorder {
+	r := &RequestRecorder{queue: make(chan Request, 2048), sink: sink, done: make(chan struct{})}
+	go r.run(ctx)
+	return r
+}
+
+func (r *RequestRecorder) Record(value Request) {
+	select {
+	case r.queue <- value:
+	default:
+		r.dropped.Add(1)
+	}
+}
+
+func (r *RequestRecorder) Done() <-chan struct{} { return r.done }
+
+func (r *RequestRecorder) run(ctx context.Context) {
+	defer close(r.done)
+	ticker := time.NewTicker(time.Second)
+	defer ticker.Stop()
+	pending := make(map[RequestBucket]*RequestBucket)
+	var lastHeartbeat time.Time
+	flush := func() {
+		buckets := make([]RequestBucket, 0, len(pending))
+		var pendingCount int64
+		for _, value := range pending {
+			buckets = append(buckets, *value)
+			pendingCount += value.Count
+		}
+		dropped := r.dropped.Swap(0)
+		failed := r.failed.Swap(0)
+		minute := time.Now().UTC().Truncate(time.Minute)
+		if len(buckets) == 0 && dropped == 0 && failed == 0 && minute.Equal(lastHeartbeat) {
+			return
+		}
+		writeCtx, cancel := context.WithTimeout(context.Background(), 2*time.Second)
+		err := r.sink.WriteRequestBuckets(writeCtx, buckets, dropped, failed)
+		cancel()
+		if err != nil {
+			r.failed.Add(1)
+			r.dropped.Add(dropped + pendingCount)
+		} else {
+			lastHeartbeat = minute
+		}
+		clear(pending)
+	}
+	add := func(value Request) {
+		if value.CompletedAt.IsZero() {
+			return
+		}
+		start := value.CompletedAt.UTC().Truncate(time.Minute)
+		key := RequestBucket{Start: start, RouteFamily: value.RouteFamily, Method: value.Method, Outcome: value.Outcome}
+		bucket := pending[key]
+		if bucket == nil {
+			bucket = &key
+			pending[key] = bucket
+		}
+		bucket.Count++
+		ms := value.Latency.Milliseconds()
+		if ms < 0 {
+			ms = 0
+		}
+		bucket.LatencySumMS += ms
+		bucket.LatencyCounts[latencyIndex(ms)]++
+	}
+	for {
+		select {
+		case value := <-r.queue:
+			add(value)
+		case <-ticker.C:
+			flush()
+		case <-ctx.Done():
+			for {
+				select {
+				case value := <-r.queue:
+					add(value)
+				default:
+					flush()
+					return
+				}
+			}
+		}
+	}
+}
+
+func latencyIndex(ms int64) int {
+	for i, bound := range [...]int64{10, 25, 50, 100, 250, 500, 1000, 2500, 5000, 10000} {
+		if ms <= bound {
+			return i
+		}
+	}
+	return 10
+}
diff --git a/services/agents-api/internal/observability/requests_test.go b/services/agents-api/internal/observability/requests_test.go
new file mode 100644
index 000000000..c6af0950f
--- /dev/null
+++ b/services/agents-api/internal/observability/requests_test.go
@@ -0,0 +1,40 @@
+package observability
+
+import (
+	"context"
+	"testing"
+	"time"
+)
+
+type capturedRequests struct {
+	buckets []RequestBucket
+	dropped int64
+}
+
+func (s *capturedRequests) WriteRequestBuckets(_ context.Context, buckets []RequestBucket, dropped, _ int64) error {
+	s.buckets = append(s.buckets, buckets...)
+	s.dropped += dropped
+	return nil
+}
+
+func TestRequestRecorderAggregatesBoundedLatencyBuckets(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	sink := &capturedRequests{}
+	recorder := NewRequestRecorder(ctx, sink)
+	at := time.Date(2026, 9, 25, 10, 20, 30, 0, time.UTC)
+	recorder.Record(Request{CompletedAt: at, RouteFamily: "sessions", Method: "GET", Outcome: "success", Latency: 50 * time.Millisecond})
+	recorder.Record(Request{CompletedAt: at.Add(3 * time.Second), RouteFamily: "sessions", Method: "GET", Outcome: "success", Latency: 11 * time.Second})
+	cancel()
+	select {
+	case <-recorder.Done():
+	case <-time.After(3 * time.Second):
+		t.Fatal("request recorder did not drain on shutdown")
+	}
+	if len(sink.buckets) != 1 {
+		t.Fatalf("expected one aggregate bucket, got %d", len(sink.buckets))
+	}
+	bucket := sink.buckets[0]
+	if bucket.Count != 2 || bucket.Start != at.Truncate(time.Minute) || bucket.LatencySumMS != 11050 || bucket.LatencyCounts[2] != 1 || bucket.LatencyCounts[10] != 1 || sink.dropped != 0 {
+		t.Fatalf("unexpected request aggregate: %+v, dropped=%d", bucket, sink.dropped)
+	}
+}
diff --git a/services/agents-api/internal/observability/tools.go b/services/agents-api/internal/observability/tools.go
new file mode 100644
index 000000000..ab4d7a34f
--- /dev/null
+++ b/services/agents-api/internal/observability/tools.go
@@ -0,0 +1,90 @@
+package observability
+
+import (
+	"context"
+	"sync/atomic"
+	"time"
+)
+
+// ToolAttempt is a sanitized terminal execution observation.
+type ToolAttempt struct {
+	ID         string
+	TenantID   string
+	SessionID  string
+	TurnID     string
+	StartedAt  *time.Time
+	FinishedAt time.Time
+	Category   string
+	Outcome    string
+	DurationMS *int64
+}
+
+type ToolSink interface {
+	WriteToolAttempts(context.Context, []ToolAttempt, int64, int64) error
+}
+
+type ToolRecorder struct {
+	queue   chan ToolAttempt
+	sink    ToolSink
+	dropped atomic.Int64
+	failed  atomic.Int64
+	done    chan struct{}
+}
+
+func NewToolRecorder(ctx context.Context, sink ToolSink) *ToolRecorder {
+	r := &ToolRecorder{queue: make(chan ToolAttempt, 1024), sink: sink, done: make(chan struct{})}
+	go r.run(ctx)
+	return r
+}
+
+func (r *ToolRecorder) Record(value ToolAttempt) {
+	select {
+	case r.queue <- value:
+	default:
+		r.dropped.Add(1)
+	}
+}
+
+func (r *ToolRecorder) Done() <-chan struct{} { return r.done }
+
+func (r *ToolRecorder) run(ctx context.Context) {
+	defer close(r.done)
+	ticker := time.NewTicker(time.Second)
+	defer ticker.Stop()
+	pending := make([]ToolAttempt, 0, 128)
+	flush := func() {
+		dropped, failed := r.dropped.Swap(0), r.failed.Swap(0)
+		if len(pending) == 0 && dropped == 0 && failed == 0 {
+			return
+		}
+		writeCtx, cancel := context.WithTimeout(context.Background(), 2*time.Second)
+		err := r.sink.WriteToolAttempts(writeCtx, pending, dropped, failed)
+		cancel()
+		if err != nil {
+			r.failed.Add(1)
+			r.dropped.Add(dropped + int64(len(pending)))
+		}
+		pending = pending[:0]
+	}
+	for {
+		select {
+		case value := <-r.queue:
+			pending = append(pending, value)
+			if len(pending) >= 128 {
+				flush()
+			}
+		case <-ticker.C:
+			flush()
+		case <-ctx.Done():
+			for {
+				select {
+				case value := <-r.queue:
+					pending = append(pending, value)
+				default:
+					flush()
+					return
+				}
+			}
+		}
+	}
+}
diff --git a/services/agents-api/internal/store/observability_read.go b/services/agents-api/internal/store/observability_read.go
new file mode 100644
index 000000000..211a0835d
--- /dev/null
+++ b/services/agents-api/internal/store/observability_read.go
@@ -0,0 +1,164 @@
+package store
+
+import (
+	"context"
+	"time"
+)
+
+type RequestMetricBucket struct {
+	Start         time.Time `json:"start"`
+	RouteFamily   string    `json:"route_family"`
+	Outcome       string    `json:"outcome"`
+	Count         int64     `json:"count"`
+	LatencySumMS  int64     `json:"latency_sum_ms"`
+	LatencyCounts []int64   `json:"latency_bucket_counts"`
+}
+
+type CollectorMetricBucket struct {
+	Start             time.Time `json:"start"`
+	Source            string    `json:"source"`
+	AttemptedCount    int64     `json:"attempted_count"`
+	ObservedCount     int64     `json:"observed_count"`
+	UnavailableCount  int64     `json:"unavailable_count"`
+	TimeoutCount      int64     `json:"timeout_count"`
+	DroppedCount      int64     `json:"dropped_count"`
+	ExportFailedCount int64     `json:"export_failed_count"`
+}
+
+type OperatorMetrics struct {
+	Requests  []RequestMetricBucket   `json:"requests"`
+	Collector []CollectorMetricBucket `json:"collector"`
+	Turns     []TurnMetricBucket      `json:"turns"`
+	Tools     []ToolMetricBucket      `json:"tools"`
+}
+
+type ToolMetricBucket struct {
+	Start         time.Time `json:"start"`
+	Category      string    `json:"category"`
+	Outcome       string    `json:"outcome"`
+	Count         int64     `json:"count"`
+	TimedCount    int64     `json:"timed_count"`
+	DurationP95MS *float64  `json:"duration_p95_ms"`
+}
+
+type TurnMetricBucket struct {
+	Start          time.Time `json:"start"`
+	Status         string    `json:"status"`
+	Count          int64     `json:"count"`
+	QueueP95MS     *float64  `json:"queue_p95_ms"`
+	ExecutionP95MS *float64  `json:"execution_p95_ms"`
+}
+
+const requestMetricSelect = `SELECT
+  to_timestamp(floor(extract(epoch FROM bucket_start) / $3) * $3) AS bucket,
+  route_family, outcome, sum(request_count)::bigint, sum(latency_sum_ms)::bigint,
+  ARRAY[
+    sum(latency_bucket_counts[1])::bigint, sum(latency_bucket_counts[2])::bigint,
+    sum(latency_bucket_counts[3])::bigint, sum(latency_bucket_counts[4])::bigint,
+    sum(latency_bucket_counts[5])::bigint, sum(latency_bucket_counts[6])::bigint,
+    sum(latency_bucket_counts[7])::bigint, sum(latency_bucket_counts[8])::bigint,
+    sum(latency_bucket_counts[9])::bigint, sum(latency_bucket_counts[10])::bigint,
+    sum(latency_bucket_counts[11])::bigint
+  ] AS latency_bucket_counts
+FROM observability_request_minute_buckets
+WHERE bucket_start >= $1 AND bucket_start < $2
+GROUP BY bucket, route_family, outcome ORDER BY bucket, route_family, outcome`
+
+const collectorMetricSelect = `SELECT
+  to_timestamp(floor(extract(epoch FROM bucket_start) / $3) * $3) AS bucket,
+  source, sum(attempted_count)::bigint, sum(observed_count)::bigint,
+  sum(unavailable_count)::bigint, sum(timeout_count)::bigint,
+  sum(dropped_count)::bigint, sum(export_failed_count)::bigint
+FROM observability_collector_minute_buckets
+WHERE bucket_start >= $1 AND bucket_start < $2
+GROUP BY bucket, source ORDER BY bucket, source`
+
+const turnMetricSelect = `SELECT
+  to_timestamp(floor(extract(epoch FROM completed_at) / $3) * $3) AS bucket,
+  status, count(*)::bigint,
+  (percentile_disc(0.95) WITHIN GROUP (ORDER BY greatest(extract(epoch FROM started_at - created_at) * 1000, 0))
+    FILTER (WHERE started_at IS NOT NULL))::double precision,
+  (percentile_disc(0.95) WITHIN GROUP (ORDER BY greatest(extract(epoch FROM completed_at - started_at) * 1000, 0))
+    FILTER (WHERE started_at IS NOT NULL))::double precision
+FROM turns
+WHERE completed_at >= $1 AND completed_at < $2
+  AND status IN ('completed', 'failed', 'cancelled')
+GROUP BY bucket, status ORDER BY bucket, status`
+
+const toolMetricSelect = `SELECT
+  to_timestamp(floor(extract(epoch FROM finished_at) / $3) * $3) AS bucket,
+  tool_category, outcome, count(*)::bigint, count(duration_ms)::bigint,
+  (percentile_disc(0.95) WITHIN GROUP (ORDER BY duration_ms)
+    FILTER (WHERE duration_ms IS NOT NULL))::double precision
+FROM observability_tool_attempts
+WHERE finished_at >= $1 AND finished_at < $2
+GROUP BY bucket, tool_category, outcome ORDER BY bucket, tool_category, outcome`
+
+// ReadOperatorMetrics returns only bounded labels and aggregate measurements.
+func (s *Store) ReadOperatorMetrics(ctx context.Context, start, end time.Time, step time.Duration) (OperatorMetrics, error) {
+	result := OperatorMetrics{Requests: []RequestMetricBucket{}, Collector: []CollectorMetricBucket{}, Turns: []TurnMetricBucket{}, Tools: []ToolMetricBucket{}}
+	seconds := int64(step / time.Second)
+	rows, err := s.pool.Query(ctx, requestMetricSelect, start, end, seconds)
+	if err != nil {
+		return result, err
+	}
+	for rows.Next() {
+		var row RequestMetricBucket
+		if err := rows.Scan(&row.Start, &row.RouteFamily, &row.Outcome, &row.Count, &row.LatencySumMS, &row.LatencyCounts); err != nil {
+			rows.Close()
+			return result, err
+		}
+		result.Requests = append(result.Requests, row)
+	}
+	err = rows.Err()
+	rows.Close()
+	if err != nil {
+		return result, err
+	}
+	rows, err = s.pool.Query(ctx, collectorMetricSelect, start, end, seconds)
+	if err != nil {
+		return result, err
+	}
+	defer rows.Close()
+	for rows.Next() {
+		var row CollectorMetricBucket
+		if err := rows.Scan(&row.Start, &row.Source, &row.AttemptedCount, &row.ObservedCount, &row.UnavailableCount,
+			&row.TimeoutCount, &row.DroppedCount, &row.ExportFailedCount); err != nil {
+			return result, err
+		}
+		result.Collector = append(result.Collector, row)
+	}
+	if err := rows.Err(); err != nil {
+		return result, err
+	}
+	rows.Close()
+	rows, err = s.pool.Query(ctx, turnMetricSelect, start, end, seconds)
+	if err != nil {
+		return result, err
+	}
+	defer rows.Close()
+	for rows.Next() {
+		var row TurnMetricBucket
+		if err := rows.Scan(&row.Start, &row.Status, &row.Count, &row.QueueP95MS, &row.ExecutionP95MS); err != nil {
+			return result, err
+		}
+		result.Turns = append(result.Turns, row)
+	}
+	if err := rows.Err(); err != nil {
+		return result, err
+	}
+	rows.Close()
+	rows, err = s.pool.Query(ctx, toolMetricSelect, start, end, seconds)
+	if err != nil {
+		return result, err
+	}
+	defer rows.Close()
+	for rows.Next() {
+		var row ToolMetricBucket
+		if err := rows.Scan(&row.Start, &row.Category, &row.Outcome, &row.Count, &row.TimedCount, &row.DurationP95MS); err != nil {
+			return result, err
+		}
+		result.Tools = append(result.Tools, row)
+	}
+	return result, rows.Err()
+}
diff --git a/services/agents-api/internal/store/observability_requests.go b/services/agents-api/internal/store/observability_requests.go
new file mode 100644
index 000000000..eebff6722
--- /dev/null
+++ b/services/agents-api/internal/store/observability_requests.go
@@ -0,0 +1,63 @@
+package store
+
+import (
+	"context"
+	"time"
+
+	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/observability"
+)
+
+const upsertRequestBucket = `INSERT INTO observability_request_minute_buckets AS current
+    (bucket_start, route_family, method, outcome, request_count, latency_sum_ms, latency_bucket_counts)
+VALUES ($1, $2, $3, $4, $5, $6, $7)
+ON CONFLICT (bucket_start, route_family, method, outcome) DO UPDATE SET
+    request_count = current.request_count + EXCLUDED.request_count,
+    latency_sum_ms = current.latency_sum_ms + EXCLUDED.latency_sum_ms,
+    latency_bucket_counts = ARRAY[
+        current.latency_bucket_counts[1] + EXCLUDED.latency_bucket_counts[1],
+        current.latency_bucket_counts[2] + EXCLUDED.latency_bucket_counts[2],
+        current.latency_bucket_counts[3] + EXCLUDED.latency_bucket_counts[3],
+        current.latency_bucket_counts[4] + EXCLUDED.latency_bucket_counts[4],
+        current.latency_bucket_counts[5] + EXCLUDED.latency_bucket_counts[5],
+        current.latency_bucket_counts[6] + EXCLUDED.latency_bucket_counts[6],
+        current.latency_bucket_counts[7] + EXCLUDED.latency_bucket_counts[7],
+        current.latency_bucket_counts[8] + EXCLUDED.latency_bucket_counts[8],
+        current.latency_bucket_counts[9] + EXCLUDED.latency_bucket_counts[9],
+        current.latency_bucket_counts[10] + EXCLUDED.latency_bucket_counts[10],
+        current.latency_bucket_counts[11] + EXCLUDED.latency_bucket_counts[11]
+    ]`
+
+// WriteRequestBuckets persists asynchronous HTTP aggregates and collector coverage
+// together. It is never called by the request handler's response path.
+func (s *Store) WriteRequestBuckets(ctx context.Context, buckets []observability.RequestBucket, dropped, failed int64) error {
+	tx, err := s.pool.Begin(ctx)
+	if err != nil {
+		return err
+	}
+	defer tx.Rollback(ctx)
+	var written int64
+	for _, bucket := range buckets {
+		if bucket.Count <= 0 {
+			continue
+		}
+		_, err = tx.Exec(ctx, upsertRequestBucket, bucket.Start, bucket.RouteFamily, bucket.Method, bucket.Outcome,
+			bucket.Count, bucket.LatencySumMS, bucket.LatencyCounts[:])
+		if err != nil {
+			return err
+		}
+		written += bucket.Count
+	}
+	_, err = tx.Exec(ctx, `INSERT INTO observability_collector_minute_buckets AS current
+			(bucket_start, source, attempted_count, observed_count, dropped_count, export_failed_count)
+			VALUES ($1, 'request', $2, $3, $4, $5)
+			ON CONFLICT (bucket_start, source) DO UPDATE SET
+			attempted_count = current.attempted_count + EXCLUDED.attempted_count,
+			observed_count = current.observed_count + EXCLUDED.observed_count,
+			dropped_count = current.dropped_count + EXCLUDED.dropped_count,
+			export_failed_count = current.export_failed_count + EXCLUDED.export_failed_count`,
+		time.Now().UTC().Truncate(time.Minute), written+dropped, written, dropped, failed)
+	if err != nil {
+		return err
+	}
+	return tx.Commit(ctx)
+}
diff --git a/services/agents-api/internal/store/observability_requests_test.go b/services/agents-api/internal/store/observability_requests_test.go
new file mode 100644
index 000000000..a886a65ba
--- /dev/null
+++ b/services/agents-api/internal/store/observability_requests_test.go
@@ -0,0 +1,37 @@
+package store
+
+import (
+	"testing"
+	"time"
+
+	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/observability"
+)
+
+func TestPostgresOperatorRequestBuckets(t *testing.T) {
+	s, _ := testStore(t)
+	start := time.Now().UTC().Truncate(time.Minute).Add(-time.Minute)
+	first := observability.RequestBucket{Start: start, RouteFamily: "sessions", Method: "GET", Outcome: "success", Count: 2, LatencySumMS: 70}
+	first.LatencyCounts[1], first.LatencyCounts[2] = 1, 1
+	second := observability.RequestBucket{Start: start, RouteFamily: "sessions", Method: "GET", Outcome: "success", Count: 1, LatencySumMS: 300}
+	second.LatencyCounts[5] = 1
+	if err := s.WriteRequestBuckets(t.Context(), []observability.RequestBucket{first}, 0, 0); err != nil {
+		t.Fatal(err)
+	}
+	if err := s.WriteRequestBuckets(t.Context(), []observability.RequestBucket{second}, 1, 0); err != nil {
+		t.Fatal(err)
+	}
+	got, err := s.ReadOperatorMetrics(t.Context(), start, start.Add(time.Minute), time.Minute)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(got.Requests) != 1 || got.Requests[0].Count != 3 || got.Requests[0].LatencySumMS != 370 ||
+		got.Requests[0].LatencyCounts[1] != 1 || got.Requests[0].LatencyCounts[2] != 1 || got.Requests[0].LatencyCounts[5] != 1 {
+		t.Fatalf("request bucket lost concurrent-shape additions: %+v", got.Requests)
+	}
+	// Collector health is written in the current minute, separate from the
+	// completed request's event-time bucket.
+	health, err := s.ReadOperatorMetrics(t.Context(), start, start.Add(2*time.Minute), time.Minute)
+	if err != nil || len(health.Collector) == 0 {
+		t.Fatalf("collector health unavailable: %+v, %v", health.Collector, err)
+	}
+}
diff --git a/services/agents-api/internal/store/observability_retention.go b/services/agents-api/internal/store/observability_retention.go
new file mode 100644
index 000000000..ff65fa4ea
--- /dev/null
+++ b/services/agents-api/internal/store/observability_retention.go
@@ -0,0 +1,29 @@
+package store
+
+import (
+	"context"
+	"time"
+)
+
+// PruneOperatorMetrics deletes small batches so retention never holds a long lock.
+func (s *Store) PruneOperatorMetrics(ctx context.Context, now time.Time) error {
+	cutoffs := []struct {
+		query  string
+		before time.Time
+	}{
+		{`DELETE FROM observability_request_minute_buckets WHERE ctid IN
+			(SELECT ctid FROM observability_request_minute_buckets WHERE bucket_start < $1 LIMIT 256 FOR UPDATE SKIP LOCKED)`, now.Add(-30 * 24 * time.Hour)},
+		{`DELETE FROM observability_collector_minute_buckets WHERE ctid IN
+			(SELECT ctid FROM observability_collector_minute_buckets WHERE bucket_start < $1 LIMIT 256 FOR UPDATE SKIP LOCKED)`, now.Add(-30 * 24 * time.Hour)},
+		{`DELETE FROM observability_model_attempts WHERE id IN
+			(SELECT id FROM observability_model_attempts WHERE started_at < $1 LIMIT 256 FOR UPDATE SKIP LOCKED)`, now.Add(-7 * 24 * time.Hour)},
+		{`DELETE FROM observability_tool_attempts WHERE id IN
+			(SELECT id FROM observability_tool_attempts WHERE finished_at < $1 LIMIT 256 FOR UPDATE SKIP LOCKED)`, now.Add(-7 * 24 * time.Hour)},
+	}
+	for _, cutoff := range cutoffs {
+		if _, err := s.pool.Exec(ctx, cutoff.query, cutoff.before); err != nil {
+			return err
+		}
+	}
+	return nil
+}
diff --git a/services/agents-api/internal/store/observability_tools.go b/services/agents-api/internal/store/observability_tools.go
new file mode 100644
index 000000000..f6ef4f728
--- /dev/null
+++ b/services/agents-api/internal/store/observability_tools.go
@@ -0,0 +1,49 @@
+package store
+
+import (
+	"context"
+	"time"
+
+	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/observability"
+)
+
+const insertToolAttempt = `INSERT INTO observability_tool_attempts
+    (id, tenant_id, session_id, turn_id, started_at, finished_at, tool_category, outcome, duration_ms)
+SELECT $1::uuid, s.tenant_id, s.id, t.id, $5::timestamptz, $6, $7, $8, $9::bigint
+FROM sessions s JOIN turns t ON t.session_id = s.id
+WHERE s.tenant_id = $2::uuid AND s.id = $3::uuid AND t.id = $4::uuid
+ON CONFLICT (id) DO NOTHING`
+
+// WriteToolAttempts stores only sanitized terminal tool evidence. Duplicate
+// native completion events have one deterministic attempt ID.
+func (s *Store) WriteToolAttempts(ctx context.Context, attempts []observability.ToolAttempt, dropped, failed int64) error {
+	tx, err := s.pool.Begin(ctx)
+	if err != nil {
+		return err
+	}
+	defer tx.Rollback(ctx)
+	var inserted int64
+	for _, attempt := range attempts {
+		result, err := tx.Exec(ctx, insertToolAttempt, attempt.ID, attempt.TenantID, attempt.SessionID, attempt.TurnID,
+			attempt.StartedAt, attempt.FinishedAt, attempt.Category, attempt.Outcome, attempt.DurationMS)
+		if err != nil {
+			return err
+		}
+		inserted += result.RowsAffected()
+	}
+	if len(attempts) > 0 || dropped > 0 || failed > 0 {
+		_, err = tx.Exec(ctx, `INSERT INTO observability_collector_minute_buckets AS current
+            (bucket_start, source, attempted_count, observed_count, dropped_count, export_failed_count)
+            VALUES ($1, 'tool', $2, $3, $4, $5)
+            ON CONFLICT (bucket_start, source) DO UPDATE SET
+              attempted_count = current.attempted_count + EXCLUDED.attempted_count,
+              observed_count = current.observed_count + EXCLUDED.observed_count,
+              dropped_count = current.dropped_count + EXCLUDED.dropped_count,
+              export_failed_count = current.export_failed_count + EXCLUDED.export_failed_count`,
+			time.Now().UTC().Truncate(time.Minute), int64(len(attempts))+dropped, inserted, dropped, failed)
+		if err != nil {
+			return err
+		}
+	}
+	return tx.Commit(ctx)
+}
diff --git a/services/agents-api/internal/store/observability_tools_test.go b/services/agents-api/internal/store/observability_tools_test.go
new file mode 100644
index 000000000..a98c61a70
--- /dev/null
+++ b/services/agents-api/internal/store/observability_tools_test.go
@@ -0,0 +1,55 @@
+package store
+
+import (
+	"testing"
+	"time"
+
+	"github.com/MiniMax-AI-Dev/parsar/services/agents-api/internal/observability"
+	"github.com/google/uuid"
+)
+
+func TestPostgresToolAttemptIsIdempotentAndTenantScoped(t *testing.T) {
+	s, pool := testStore(t)
+	tenant, session, turn, attemptID := uuid.NewString(), uuid.NewString(), uuid.NewString(), uuid.NewString()
+	_, err := pool.Exec(t.Context(), `INSERT INTO sessions (id, tenant_id, engine, idempotency_key, request_hash)
+		VALUES ($1, $2, 'codex', $3, 'observability-test')`, session, tenant, uuid.NewString())
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() {
+		_, _ = pool.Exec(t.Context(), `DELETE FROM turns WHERE id=$1`, turn)
+		_, _ = pool.Exec(t.Context(), `DELETE FROM sessions WHERE id=$1`, session)
+	})
+	_, err = pool.Exec(t.Context(), `INSERT INTO turns (id, session_id, status, completed_at)
+		VALUES ($1, $2, 'completed', clock_timestamp())`, turn, session)
+	if err != nil {
+		t.Fatal(err)
+	}
+	finished := time.Now().UTC()
+	duration := int64(120)
+	value := observability.ToolAttempt{ID: attemptID, TenantID: tenant, SessionID: session, TurnID: turn,
+		FinishedAt: finished, Category: "command", Outcome: "success", DurationMS: &duration}
+	if err := s.WriteToolAttempts(t.Context(), []observability.ToolAttempt{value}, 0, 0); err != nil {
+		t.Fatal(err)
+	}
+	if err := s.WriteToolAttempts(t.Context(), []observability.ToolAttempt{value}, 0, 0); err != nil {
+		t.Fatal(err)
+	}
+	var count int
+	var startedAt *time.Time
+	if err := pool.QueryRow(t.Context(), `SELECT count(*), max(started_at) FROM observability_tool_attempts WHERE id=$1`, attemptID).Scan(&count, &startedAt); err != nil || count != 1 || startedAt != nil {
+		t.Fatalf("tool attempt deduplication or unknown start failed: count=%d start=%v err=%v", count, startedAt, err)
+	}
+	value.ID = uuid.NewString()
+	value.TenantID = uuid.NewString()
+	if err := s.WriteToolAttempts(t.Context(), []observability.ToolAttempt{value}, 0, 0); err != nil {
+		t.Fatal(err)
+	}
+	if err := pool.QueryRow(t.Context(), `SELECT count(*) FROM observability_tool_attempts WHERE id=$1`, value.ID).Scan(&count); err != nil || count != 0 {
+		t.Fatalf("foreign tenant wrote an attempt: count=%d err=%v", count, err)
+	}
+	metrics, err := s.ReadOperatorMetrics(t.Context(), finished.Add(-time.Minute), finished.Add(time.Minute), time.Minute)
+	if err != nil || len(metrics.Tools) == 0 || len(metrics.Turns) == 0 {
+		t.Fatalf("terminal operator metrics unavailable: tools=%+v turns=%+v err=%v", metrics.Tools, metrics.Turns, err)
+	}
+}
diff --git a/services/agents-api/internal/store/runtime_history_acceptance_test.go b/services/agents-api/internal/store/runtime_history_acceptance_test.go
index 23cc3aa7f..2c45688cb 100644
--- a/services/agents-api/internal/store/runtime_history_acceptance_test.go
+++ b/services/agents-api/internal/store/runtime_history_acceptance_test.go
@@ -148,10 +148,10 @@ func TestPostgresRuntimeHistoryDenseReadAndBoundedRetention(t *testing.T) {
 	}
 	// Seed the equivalent of a full day of five-second periodic samples in one
 	// fixture statement; production inserts still go through the exporter.
-	_, err := pool.Exec(t.Context(), `INSERT INTO runtime_history_samples
+	_, err := pool.Exec(t.Context(), `INSERT INTO observability_runtime_samples
  (tenant_id,session_id,environment_id,resolved_at_ns,allocation_id,provider_type,status,observed_at_ns,started_at_ns,cpu_usage_seconds,cpu_capacity_cores,memory_usage_bytes,input_tokens,output_tokens)
  SELECT tenant_id,session_id,environment_id,resolved_at_ns+n*5000000000,allocation_id,provider_type,status,observed_at_ns+n*5000000000,started_at_ns,n::double precision,cpu_capacity_cores,memory_usage_bytes,input_tokens+n,output_tokens+n
- FROM runtime_history_samples CROSS JOIN generate_series(1,17279) n WHERE tenant_id=$1 AND session_id=$2`, scope.TenantID, scope.SessionID)
+ FROM observability_runtime_samples CROSS JOIN generate_series(1,17279) n WHERE tenant_id=$1 AND session_id=$2`, scope.TenantID, scope.SessionID)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -173,10 +173,10 @@ func TestPostgresRuntimeHistoryDenseReadAndBoundedRetention(t *testing.T) {
 	if err := s.InsertRuntimeHistorySample(t.Context(), old); err != nil {
 		t.Fatal(err)
 	}
-	_, err = pool.Exec(t.Context(), `INSERT INTO runtime_history_samples
+	_, err = pool.Exec(t.Context(), `INSERT INTO observability_runtime_samples
  (tenant_id,session_id,environment_id,resolved_at_ns,allocation_id,provider_type,status,observed_at_ns,started_at_ns)
  SELECT tenant_id,session_id,environment_id,resolved_at_ns-n,allocation_id,provider_type,'unavailable',NULL,NULL
- FROM runtime_history_samples CROSS JOIN generate_series(1,4100) n WHERE tenant_id=$1 AND session_id=$2 AND resolved_at_ns=$3`, scope.TenantID, scope.SessionID, old.ResolvedAt.UnixNano())
+ FROM observability_runtime_samples CROSS JOIN generate_series(1,4100) n WHERE tenant_id=$1 AND session_id=$2 AND resolved_at_ns=$3`, scope.TenantID, scope.SessionID, old.ResolvedAt.UnixNano())
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -188,13 +188,13 @@ func TestPostgresRuntimeHistoryDenseReadAndBoundedRetention(t *testing.T) {
 		t.Fatal(err)
 	}
 	var remaining int
-	if err := pool.QueryRow(t.Context(), `SELECT count(*) FROM runtime_history_samples WHERE tenant_id=$1 AND resolved_at_ns<$2`, scope.TenantID, end.Add(-7*24*time.Hour).UnixNano()).Scan(&remaining); err != nil || remaining != 5 {
+	if err := pool.QueryRow(t.Context(), `SELECT count(*) FROM observability_runtime_samples WHERE tenant_id=$1 AND resolved_at_ns<$2`, scope.TenantID, end.Add(-7*24*time.Hour).UnixNano()).Scan(&remaining); err != nil || remaining != 5 {
 		t.Fatal("cleanup exceeded bounded batch", remaining, err)
 	}
 	if err := reader.Prune(t.Context()); err != nil {
 		t.Fatal(err)
 	}
-	if err := pool.QueryRow(t.Context(), `SELECT count(*) FROM runtime_history_samples WHERE tenant_id=$1 AND resolved_at_ns<$2`, scope.TenantID, end.Add(-7*24*time.Hour).UnixNano()).Scan(&remaining); err != nil || remaining != 0 {
+	if err := pool.QueryRow(t.Context(), `SELECT count(*) FROM observability_runtime_samples WHERE tenant_id=$1 AND resolved_at_ns<$2`, scope.TenantID, end.Add(-7*24*time.Hour).UnixNano()).Scan(&remaining); err != nil || remaining != 0 {
 		t.Fatal("retention cleanup incomplete", remaining, err)
 	}
 	cancelled, cancel := context.WithCancel(t.Context())
diff --git a/services/agents-api/migrations/000066_observability.sql b/services/agents-api/migrations/000066_observability.sql
new file mode 100644
index 000000000..263db79d3
--- /dev/null
+++ b/services/agents-api/migrations/000066_observability.sql
@@ -0,0 +1,78 @@
+-- +goose Up
+-- Runtime samples remain a bounded telemetry projection, not execution or Usage authority.
+ALTER TABLE runtime_history_samples RENAME TO observability_runtime_samples;
+ALTER INDEX runtime_history_samples_pkey RENAME TO observability_runtime_samples_pkey;
+ALTER INDEX runtime_history_samples_retention_idx RENAME TO observability_runtime_samples_retention_idx;
+CREATE INDEX turns_completed_time_idx ON turns (completed_at) WHERE completed_at IS NOT NULL;
+
+-- Histogram bins: <=10, 25, 50, 100, 250, 500, 1000, 2500, 5000,
+-- 10000 milliseconds, and >10000 milliseconds. Counts are disjoint.
+CREATE TABLE observability_request_minute_buckets (
+    bucket_start timestamptz NOT NULL,
+    route_family text NOT NULL CHECK (length(route_family) BETWEEN 1 AND 64),
+    method text NOT NULL CHECK (method IN ('GET', 'POST', 'PUT', 'PATCH', 'DELETE', 'HEAD', 'OTHER')),
+    outcome text NOT NULL CHECK (outcome IN ('success', 'client_error', 'server_error')),
+    request_count bigint NOT NULL DEFAULT 0 CHECK (request_count >= 0),
+    latency_sum_ms bigint NOT NULL DEFAULT 0 CHECK (latency_sum_ms >= 0),
+    latency_bucket_counts bigint[] NOT NULL DEFAULT array_fill(0::bigint, ARRAY[11])
+        CHECK (array_length(latency_bucket_counts, 1) = 11),
+    PRIMARY KEY (bucket_start, route_family, method, outcome)
+);
+
+CREATE TABLE observability_model_attempts (
+    id uuid PRIMARY KEY,
+    tenant_id uuid NOT NULL,
+    session_id uuid NOT NULL REFERENCES sessions(id) ON DELETE CASCADE,
+    turn_id uuid NOT NULL REFERENCES turns(id) ON DELETE CASCADE,
+    started_at timestamptz NOT NULL,
+    finished_at timestamptz NOT NULL CHECK (finished_at >= started_at),
+    model_family text NOT NULL CHECK (length(model_family) BETWEEN 1 AND 64),
+    provider_type text NOT NULL CHECK (length(provider_type) BETWEEN 1 AND 64),
+    outcome text NOT NULL CHECK (outcome IN ('success', 'error', 'cancelled', 'unknown')),
+    duration_ms bigint NOT NULL CHECK (duration_ms >= 0),
+    input_tokens bigint CHECK (input_tokens >= 0),
+    output_tokens bigint CHECK (output_tokens >= 0),
+    usage_status text NOT NULL CHECK (usage_status IN ('observed', 'unavailable', 'unsupported')),
+    CHECK ((input_tokens IS NULL AND output_tokens IS NULL) OR usage_status = 'observed')
+);
+CREATE INDEX observability_model_attempts_tenant_time_idx ON observability_model_attempts (tenant_id, started_at DESC);
+CREATE INDEX observability_model_attempts_session_time_idx ON observability_model_attempts (session_id, started_at DESC);
+CREATE INDEX observability_model_attempts_retention_idx ON observability_model_attempts (started_at);
+
+CREATE TABLE observability_tool_attempts (
+    id uuid PRIMARY KEY,
+    tenant_id uuid NOT NULL,
+    session_id uuid NOT NULL REFERENCES sessions(id) ON DELETE CASCADE,
+    turn_id uuid NOT NULL REFERENCES turns(id) ON DELETE CASCADE,
+    started_at timestamptz,
+    finished_at timestamptz NOT NULL,
+    tool_category text NOT NULL CHECK (length(tool_category) BETWEEN 1 AND 64),
+    outcome text NOT NULL CHECK (outcome IN ('success', 'error', 'cancelled', 'unknown')),
+    duration_ms bigint CHECK (duration_ms >= 0),
+    CHECK (started_at IS NULL OR finished_at >= started_at)
+);
+CREATE INDEX observability_tool_attempts_tenant_time_idx ON observability_tool_attempts (tenant_id, finished_at DESC);
+CREATE INDEX observability_tool_attempts_session_time_idx ON observability_tool_attempts (session_id, finished_at DESC);
+CREATE INDEX observability_tool_attempts_retention_idx ON observability_tool_attempts (finished_at);
+
+CREATE TABLE observability_collector_minute_buckets (
+    bucket_start timestamptz NOT NULL,
+    source text NOT NULL CHECK (source IN ('runtime', 'request', 'model', 'tool', 'export')),
+    attempted_count bigint NOT NULL DEFAULT 0 CHECK (attempted_count >= 0),
+    observed_count bigint NOT NULL DEFAULT 0 CHECK (observed_count >= 0),
+    unavailable_count bigint NOT NULL DEFAULT 0 CHECK (unavailable_count >= 0),
+    timeout_count bigint NOT NULL DEFAULT 0 CHECK (timeout_count >= 0),
+    dropped_count bigint NOT NULL DEFAULT 0 CHECK (dropped_count >= 0),
+    export_failed_count bigint NOT NULL DEFAULT 0 CHECK (export_failed_count >= 0),
+    PRIMARY KEY (bucket_start, source)
+);
+
+-- +goose Down
+DROP TABLE observability_collector_minute_buckets;
+DROP TABLE observability_tool_attempts;
+DROP TABLE observability_model_attempts;
+DROP TABLE observability_request_minute_buckets;
+DROP INDEX turns_completed_time_idx;
+ALTER INDEX observability_runtime_samples_retention_idx RENAME TO runtime_history_samples_retention_idx;
+ALTER INDEX observability_runtime_samples_pkey RENAME TO runtime_history_samples_pkey;
+ALTER TABLE observability_runtime_samples RENAME TO runtime_history_samples;
diff --git a/services/agents-api/tests/runtime_history_live.py b/services/agents-api/tests/runtime_history_live.py
index ba06f92fc..3e0c44017 100644
--- a/services/agents-api/tests/runtime_history_live.py
+++ b/services/agents-api/tests/runtime_history_live.py
@@ -33,7 +33,7 @@ def request(self, *args, **kwargs):
         encoded = json.dumps(value)
         base.require(not any(secret in encoded for secret in self.secrets), "public_response_exposed_secret")
         base.require(not any(marker in encoded for marker in (
-            "postgres://", "postgresql://", "runtime_history_samples", "token_sha256",
+            "postgres://", "postgresql://", "observability_runtime_samples", "token_sha256",
             "unix:///", "host.microsandbox.internal", "/home/parsar-acceptance/")),
             "public_response_exposed_backend_configuration")
         return value
diff --git a/services/core-console/sandbox_admin.go b/services/core-console/sandbox_admin.go
index cfcdd48fc..6daa3e05a 100644
--- a/services/core-console/sandbox_admin.go
+++ b/services/core-console/sandbox_admin.go
@@ -39,6 +39,11 @@ func sandboxAdminRequest(r *http.Request) bool {
 	return false
 }
 
+// Operator metrics expose only the fixed, read-only deployment summary.
+func observabilityAdminRequest(r *http.Request) bool {
+	return r.URL.Path == "/core/v1/observability/summary" && r.Method == http.MethodGet
+}
+
 func explicitBearer(r *http.Request) bool {
 	parts := strings.Fields(r.Header.Get("Authorization"))
 	return len(r.Header.Values("Authorization")) == 1 && len(parts) == 2 && strings.EqualFold(parts[0], "Bearer")
diff --git a/services/core-console/sandbox_admin_test.go b/services/core-console/sandbox_admin_test.go
index 886427fd2..14f6c5e1c 100644
--- a/services/core-console/sandbox_admin_test.go
+++ b/services/core-console/sandbox_admin_test.go
@@ -44,6 +44,7 @@ func TestSandboxAdminProxyPreservesIndependentCredential(t *testing.T) {
 		{"DELETE", "/core/v1/sandbox/nodes/node-id"},
 		{"GET", "/core/v1/sandbox/nodes/node-id/allocations"},
 		{"POST", "/core/v1/sandbox/enrollment-tokens"},
+		{"GET", "/core/v1/observability/summary?range=1h"},
 	} {
 		t.Run(tc.method+tc.path, func(t *testing.T) {
 			r := adminRequest(t, server.URL, tc.method, tc.path, "separate-admin-key")
@@ -56,7 +57,7 @@ func TestSandboxAdminProxyPreservesIndependentCredential(t *testing.T) {
 			}
 		})
 	}
-	if calls.Load() != 6 {
+	if calls.Load() != 7 {
 		t.Fatalf("proxied %d operations", calls.Load())
 	}
 }
@@ -129,6 +130,8 @@ func TestSandboxNodeControlAndUnknownCoreRoutesAreNotProxied(t *testing.T) {
 		{"POST", "/core/v1/sandbox/nodes"},
 		{"GET", "/core/v1/sandbox/enrollment-tokens"},
 		{"GET", "/core/v1/sandbox/nodes/id/allocations/extra"},
+		{"POST", "/core/v1/observability/summary"},
+		{"GET", "/core/v1/observability/summary/extra"},
 	} {
 		t.Run(tc.method+tc.path, func(t *testing.T) {
 			r := adminRequest(t, server.URL, tc.method, tc.path, "separate-admin-key")
diff --git a/services/core-console/server.go b/services/core-console/server.go
index 293c1193a..9b0018d34 100644
--- a/services/core-console/server.go
+++ b/services/core-console/server.go
@@ -72,9 +72,9 @@ func newConsole(c config) (*console, error) {
 			r.Out.Header.Del("Cookie")
 			r.Out.Header.Del("Origin")
 			r.Out.Header.Del("Referer")
-			if ((publicAPIRequest(r.In) || projectExtensionRequest(r.In)) && explicitBearer(r.In)) || nodeTransportRequest(r.In) || (sandboxAdminRequest(r.In) && c.adminToken == "") {
+			if ((publicAPIRequest(r.In) || projectExtensionRequest(r.In)) && explicitBearer(r.In)) || nodeTransportRequest(r.In) || ((sandboxAdminRequest(r.In) || observabilityAdminRequest(r.In)) && c.adminToken == "") {
 				r.Out.Header.Set("Authorization", r.In.Header.Get("Authorization"))
-			} else if sandboxAdminRequest(r.In) || projectKeyAdminRequest(r.In) {
+			} else if sandboxAdminRequest(r.In) || observabilityAdminRequest(r.In) || projectKeyAdminRequest(r.In) {
 				r.Out.Header.Set("Authorization", "Bearer "+c.adminToken)
 			} else {
 				r.Out.Header.Set("Authorization", "Bearer "+c.token)
@@ -170,7 +170,7 @@ func (h *console) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 	if (r.URL.Path == "/core" || strings.HasPrefix(r.URL.Path, "/core/")) && h.adminToken == "" && !projectExtensionRequest(r) {
-		if !sandboxAdminRequest(r) {
+		if !sandboxAdminRequest(r) && !observabilityAdminRequest(r) {
 			http.NotFound(w, r)
 			return
 		}
@@ -206,7 +206,7 @@ func (h *console) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 	if r.URL.Path == "/core" || strings.HasPrefix(r.URL.Path, "/core/") {
-		if !sandboxAdminRequest(r) && !projectExtensionRequest(r) {
+		if !sandboxAdminRequest(r) && !observabilityAdminRequest(r) && !projectExtensionRequest(r) {
 			http.NotFound(w, r)
 			return
 		}