diff --git a/__tests__/unit/health.test.js b/__tests__/unit/health.test.js
new file mode 100644
index 0000000..6419b41
--- /dev/null
+++ b/__tests__/unit/health.test.js
@@ -0,0 +1,115 @@
+import { evaluateHealth } from '../../src/health.js';
+
+describe('evaluateHealth', () => {
+  it('reports healthy when no probes are provided', () => {
+    const result = evaluateHealth({});
+    expect(result).toEqual({ ok: true, status: 'healthy', checks: {} });
+  });
+
+  describe('scheduler probe', () => {
+    function fakeScheduler({ running, lastTickAt, intervalMs = 60_000 } = {}) {
+      return {
+        isRunning: () => running,
+        getLastTickAt: () => lastTickAt,
+        getIntervalMs: () => intervalMs
+      };
+    }
+
+    it('passes when not running (treated as not-yet-started)', () => {
+      const result = evaluateHealth({ scheduler: fakeScheduler({ running: false }) });
+      expect(result.status).toBe('healthy');
+      expect(result.checks.scheduler.ok).toBe(true);
+    });
+
+    it('reports degraded when running but no tick has completed yet', () => {
+      const result = evaluateHealth({
+        scheduler: fakeScheduler({ running: true, lastTickAt: null })
+      });
+      expect(result.status).toBe('degraded');
+      expect(result.checks.scheduler.status).toBe('warn');
+      expect(result.ok).toBe(true);
+    });
+
+    it('passes when last tick is within 2× the configured interval', () => {
+      const ageMs = 30_000; // less than 2 * 60_000
+      const result = evaluateHealth({
+        scheduler: fakeScheduler({ running: true, lastTickAt: Date.now() - ageMs })
+      });
+      expect(result.status).toBe('healthy');
+      expect(result.checks.scheduler.ok).toBe(true);
+      expect(result.checks.scheduler.detail.ageMs).toBeGreaterThanOrEqual(ageMs);
+    });
+
+    it('flags unhealthy when last tick is older than 2× the interval', () => {
+      const result = evaluateHealth({
+        scheduler: fakeScheduler({
+          running: true,
+          lastTickAt: Date.now() - 5 * 60_000, // way past 2× the 60s interval
+          intervalMs: 60_000
+        })
+      });
+      expect(result.status).toBe('unhealthy');
+      expect(result.ok).toBe(false);
+      expect(result.checks.scheduler.status).toBe('fail');
+    });
+
+    it('reports degraded when scheduler probe itself throws', () => {
+      const result = evaluateHealth({
+        scheduler: {
+          isRunning: () => { throw new Error('scheduler exploded'); },
+          getLastTickAt: () => Date.now()
+        }
+      });
+      expect(result.status).toBe('degraded');
+      expect(result.checks.scheduler.status).toBe('error');
+      expect(result.checks.scheduler.detail).toBe('scheduler exploded');
+    });
+  });
+
+  describe('kv probe', () => {
+    it('passes when ping() succeeds', () => {
+      const result = evaluateHealth({ kv: { ping: () => {} } });
+      expect(result.status).toBe('healthy');
+      expect(result.checks.kv.ok).toBe(true);
+    });
+
+    it('flags unhealthy when ping() throws', () => {
+      const result = evaluateHealth({
+        kv: { ping: () => { throw new Error('SQLITE_BUSY'); } }
+      });
+      expect(result.status).toBe('unhealthy');
+      expect(result.ok).toBe(false);
+      expect(result.checks.kv.status).toBe('fail');
+      expect(result.checks.kv.detail).toBe('SQLITE_BUSY');
+    });
+
+    it('skips the check when ping is not callable', () => {
+      const result = evaluateHealth({ kv: {} });
+      expect(result.checks.kv).toBeUndefined();
+    });
+  });
+
+  describe('disk probe', () => {
+    it('passes for a real directory with plenty of free space', () => {
+      // /tmp is virtually always > 100 MB free on dev/CI hosts.
+      const result = evaluateHealth({ dataDir: '/tmp' });
+      expect(result.checks.disk.ok).toBe(true);
+      expect(result.checks.disk.detail.freeBytes).toBeGreaterThan(0);
+    });
+
+    it('reports degraded when statfs throws (e.g. nonexistent path)', () => {
+      const result = evaluateHealth({ dataDir: '/this/path/does/not/exist/temper' });
+      expect(result.checks.disk.status).toBe('error');
+      expect(result.status).toBe('degraded');
+    });
+  });
+
+  it('combines unhealthy + degraded → unhealthy (most-severe wins)', () => {
+    const result = evaluateHealth({
+      kv: { ping: () => { throw new Error('down'); } },     // unhealthy
+      dataDir: '/this/does/not/exist'                          // degraded
+    });
+    expect(result.status).toBe('unhealthy');
+    expect(result.ok).toBe(false);
+  });
+});
diff --git a/src/app.js b/src/app.js
index db1446d..62dd8d4 100644
--- a/src/app.js
+++ b/src/app.js
@@ -30,12 +30,18 @@ import { initTaskStore } from './task-store.js';
 import { createScheduler } from './scheduler.js';
 import { initKVStore } from './persistent-kv.js';
 import { provisionRepo, parseIssueFormBody, validateProvisionRequest } from './provisioning.js';
+import { evaluateHealth } from './health.js';
 
 const __dirname = path.dirname(fileURLToPath(import.meta.url));
 
 // Module-level state for the task store and scheduler
 let _taskStore = null;
 let _scheduler = null;
+// Bug #8: /health uses _dedupStore.ping() as a SQLite liveness probe.
+// Same instance that backs setIdempotencyStore — exposed here so the
+// /health handlers can access it without reaching into idempotency.js.
+let _dedupStore = null;
+let _dataDir = null;
 
 function getTaskStore() { return _taskStore; }
 function getScheduler() { return _scheduler; }
@@ -194,11 +200,17 @@ function createCustomRoutesHandler() {
 
     if (req.method === 'GET' && pathname === '/health') {
       applySecurityHeaders(res);
+      const probe = evaluateHealth({
+        scheduler: _scheduler,
+        kv: _dedupStore,
+        dataDir: _dataDir
+      });
       const healthData = {
-        status: 'healthy',
+        status: probe.status,
         timestamp: new Date().toISOString(),
         version: DEPLOY_SHA || '1.0.0',
-        queue: defaultQueue.stats()
+        queue: defaultQueue.stats(),
+        checks: probe.checks
       };
       if (_taskStore) {
         healthData.tasks = _taskStore.getStats();
@@ -207,7 +219,8 @@ function createCustomRoutesHandler() {
         healthData.scheduler = { running: _scheduler.isRunning() };
       }
       const body = JSON.stringify(healthData);
-      res.writeHead(200, { 'Content-Type': 'application/json' });
+      // Bug #8: 503 lets PM2 / external uptime monitors restart on hang.
+      res.writeHead(probe.ok ? 200 : 503, { 'Content-Type': 'application/json' });
       res.end(body);
       return true;
     }
@@ -982,11 +995,17 @@ function registerApp(app, options = {}) {
     applySecurityMiddleware(router);
 
     router.get('/health', (req, res) => {
+      const probe = evaluateHealth({
+        scheduler: _scheduler,
+        kv: _dedupStore,
+        dataDir: _dataDir
+      });
       const healthData = {
-        status: 'healthy',
+        status: probe.status,
         timestamp: new Date().toISOString(),
         version: DEPLOY_SHA || '1.0.0',
-        queue: defaultQueue.stats()
+        queue: defaultQueue.stats(),
+        checks: probe.checks
       };
       if (_taskStore) {
         healthData.tasks = _taskStore.getStats();
@@ -994,7 +1013,7 @@ function registerApp(app, options = {}) {
       if (_scheduler) {
         healthData.scheduler = { running: _scheduler.isRunning() };
       }
-      res.status(200).json(healthData);
+      res.status(probe.ok ? 200 : 503).json(healthData);
     });
 
     router.get('/webhook', (req, res) => {
@@ -1035,11 +1054,13 @@ function initPersistence() {
   try {
     const dataDir = getDataDir();
     if (!fs.existsSync(dataDir)) fs.mkdirSync(dataDir, { recursive: true });
+    _dataDir = dataDir;
 
     const dedupStore = initKVStore(path.join(dataDir, 'dedup.db'), 'webhook_dedup', {
       ttlMs: 60 * 60 * 1000
     });
     setIdempotencyStore(dedupStore);
+    _dedupStore = dedupStore;
 
     const rateStore = initKVStore(path.join(dataDir, 'dedup.db'), 'ai_rate_limits', {
       ttlMs: 5 * 60 * 1000
diff --git a/src/health.js b/src/health.js
new file mode 100644
index 0000000..d416681
--- /dev/null
+++ b/src/health.js
@@ -0,0 +1,120 @@
+/**
+ * Liveness probe for /health (Bug #8).
+ *
+ * The previous /health endpoint returned 200 unconditionally — which means
+ * PM2's restart-on-failed-healthcheck and any external uptime monitor were
+ * both blind to a hung scheduler, a locked SQLite DB, or a full data disk.
+ *
+ * The checks here are *liveness* signals (am I able to do my job RIGHT NOW),
+ * not readiness (have I finished initialising). All checks are best-effort:
+ * if a check itself throws unexpectedly, that check is reported as
+ * `'error'` and the overall verdict is `degraded` — we'd rather report
+ * something slightly wrong than mask a real outage.
+ *
+ *   ok:        every check passed
+ *   degraded:  a check is missing or threw; not actionable but worth
+ *              flagging in dashboards. /health returns 200.
+ *   unhealthy: a check has FAILED in a way that means the bot can't do
+ *              its job (DB ping throws, scheduler hung, disk full).
+ *              /health returns 503 so PM2 / uptime monitors restart us.
+ */
+
+import fs from 'node:fs';
+
+const MIN_FREE_DISK_BYTES = 100 * 1024 * 1024; // 100 MB — below this we
+// risk losing the next sweep / commit. Threshold low because Netcup VM
+// only has ~30 GB total.
+
+const SCHEDULER_STALE_FACTOR = 2; // a tick should land within 2× the
+// configured interval — beyond that, the scheduler is hung.
+
+/**
+ * Synchronously evaluate liveness. Cheap (millisecond-range) by design —
+ * /health is hit by PM2 every few seconds.
+ *
+ * @param {object} probes
+ * @param {object} [probes.scheduler] - { isRunning(), getLastTickAt(), getIntervalMs() }
+ * @param {object} [probes.kv] - { ping() } — typically the dedup KV store
+ * @param {string} [probes.dataDir] - filesystem path checked for free space
+ * @returns {{
+ *   ok: boolean,
+ *   status: 'healthy'|'degraded'|'unhealthy',
+ *   checks: Record<string, { ok: boolean, status: 'ok'|'warn'|'fail'|'error', detail?: any }>
+ * }}
+ */
+export function evaluateHealth(probes = {}) {
+  const checks = {};
+  let unhealthy = false;
+  let degraded = false;
+
+  // 1. Scheduler tick freshness ---------------------------------------------
+  if (probes.scheduler) {
+    try {
+      const isRunning = probes.scheduler.isRunning();
+      const lastTickAt = probes.scheduler.getLastTickAt?.() ?? null;
+      const intervalMs = probes.scheduler.getIntervalMs?.() ?? 5 * 60 * 1000;
+      const staleAfter = intervalMs * SCHEDULER_STALE_FACTOR;
+
+      if (!isRunning) {
+        // Scheduler was never started (or has been stopped). Not necessarily
+        // an outage — `start()` may not have been called yet (early in
+        // probot bootstrap). Don't flunk /health for it.
+        checks.scheduler = { ok: true, status: 'ok', detail: { running: false } };
+      } else if (lastTickAt === null) {
+        // Started but no tick has completed yet — first tick fires
+        // immediately on start(), so we expect a value within milliseconds.
+        // Don't fail here either; report as warn so dashboards see it.
+        checks.scheduler = { ok: true, status: 'warn', detail: { running: true, lastTickAt: null } };
+        degraded = true;
+      } else {
+        const ageMs = Date.now() - lastTickAt;
+        const stale = ageMs > staleAfter;
+        checks.scheduler = {
+          ok: !stale,
+          status: stale ? 'fail' : 'ok',
+          detail: { running: true, ageMs, staleAfterMs: staleAfter }
+        };
+        if (stale) unhealthy = true;
+      }
+    } catch (err) {
+      checks.scheduler = { ok: false, status: 'error', detail: err.message };
+      degraded = true;
+    }
+  }
+
+  // 2. SQLite ping ---------------------------------------------------------
+  if (probes.kv && typeof probes.kv.ping === 'function') {
+    try {
+      probes.kv.ping();
+      checks.kv = { ok: true, status: 'ok' };
+    } catch (err) {
+      checks.kv = { ok: false, status: 'fail', detail: err.message };
+      unhealthy = true;
+    }
+  }
+
+  // 3. Free disk on the data directory -------------------------------------
+  if (probes.dataDir && typeof fs.statfsSync === 'function') {
+    try {
+      const stat = fs.statfsSync(probes.dataDir);
+      const freeBytes = Number(stat.bavail) * Number(stat.bsize);
+      const ok = freeBytes >= MIN_FREE_DISK_BYTES;
+      checks.disk = {
+        ok,
+        status: ok ? 'ok' : 'fail',
+        detail: { freeBytes, thresholdBytes: MIN_FREE_DISK_BYTES }
+      };
+      if (!ok) unhealthy = true;
+    } catch (err) {
+      checks.disk = { ok: false, status: 'error', detail: err.message };
+      degraded = true;
+    }
+  }
+
+  let status;
+  if (unhealthy) status = 'unhealthy';
+  else if (degraded) status = 'degraded';
+  else status = 'healthy';
+
+  return { ok: !unhealthy, status, checks };
+}
diff --git a/src/persistent-kv.js b/src/persistent-kv.js
index 86b505d..c38591c 100644
--- a/src/persistent-kv.js
+++ b/src/persistent-kv.js
@@ -77,5 +77,12 @@ export function initKVStore(dbPath, name, opts = {}) {
     db.close();
   }
 
-  return { has, get, set, sweep, close, _db: db };
+  // Bug #8: /health uses ping() as a SQLite liveness probe. Throws if
+  // the connection is dead, the file is locked beyond timeout, or disk
+  // I/O is jammed. SELECT 1 doesn't touch any table — minimum latency.
+  function ping() {
+    db.prepare('SELECT 1 AS ok').get();
+  }
+
+  return { has, get, set, sweep, close, ping, _db: db };
 }
diff --git a/src/scheduler.js b/src/scheduler.js
index 7977bcb..827d6fc 100644
--- a/src/scheduler.js
+++ b/src/scheduler.js
@@ -23,6 +23,9 @@ function createScheduler(store, octokitOrFactory, options = {}) {
   const handlers = new Map();
   let timer = null;
   let running = false;
+  // Bug #8: /health uses this to detect a hung scheduler. Updated at the
+  // *end* of every tick (success, no-op, or error caught inside tick).
+  let lastTickAt = null;
 
   async function getOctokit() {
     if (typeof octokitOrFactory === 'function') {
@@ -98,6 +101,7 @@ function createScheduler(store, octokitOrFactory, options = {}) {
       }
     } finally {
       running = false;
+      lastTickAt = Date.now();
     }
   }
 
@@ -128,6 +132,17 @@ function createScheduler(store, octokitOrFactory, options = {}) {
     /** Check if the scheduler is running. */
     isRunning() {
       return timer !== null;
+    },
+
+    /** Timestamp (ms since epoch) of the last completed tick, or null
+     *  before the first tick finishes. Used by /health for liveness. */
+    getLastTickAt() {
+      return lastTickAt;
+    },
+
+    /** Configured interval; /health uses 2× this as the staleness threshold. */
+    getIntervalMs() {
+      return intervalMs;
     }
   };
 }