diff --git a/plugins/dashboard/server.ts b/plugins/dashboard/server.ts
index 5b29384..c0bde9e 100644
--- a/plugins/dashboard/server.ts
+++ b/plugins/dashboard/server.ts
@@ -27,7 +27,7 @@ import {
   ListToolsRequestSchema,
   CallToolRequestSchema,
 } from '@modelcontextprotocol/sdk/types.js'
-import { readFileSync, mkdirSync, readdirSync, unlinkSync, watch, chmodSync, copyFileSync } from 'fs'
+import { readFileSync, mkdirSync, readdirSync, unlinkSync, watch, chmodSync, copyFileSync, openSync, closeSync, writeSync, statSync } from 'fs'
 import { homedir } from 'os'
 import { join } from 'path'
 import { installLifecycle, recordLifecycle } from './lifecycle.ts'
@@ -47,6 +47,16 @@ const AGENT_INBOX_DIR = join(STATE_DIR, 'agent-inbox')
 // purpose: a nudge carries no payload and must never be mistaken for a message
 // drop-file, so the two ingesters can never read each other's files.
 const COLLECT_NOW_DIR = join(STATE_DIR, 'collect-now')
+// DIVE-3809: the drain's `draining`/`rerun` guard below is PER-PROCESS. Two
+// plugin processes for the same agent (an overlapping restart, a stray
+// supervisor respawn) each fetch the SAME pending rows and push every message
+// into the session twice, because the ack lands only after the notifications
+// are sent. This file is the cross-process arm of that guard: O_EXCL create
+// wins the drain, everyone else skips this pass and picks it up on the next
+// sweep. Refuted as the CAUSE of the loss DIVE-3806 observed (lifecycle.log
+// showed exactly one live process across that window) — it is still a real
+// race, and it is scope 3 of this row.
+const DRAIN_LOCK = join(STATE_DIR, 'pending-drain.lock')
 mkdirSync(AGENT_INBOX_DIR, { recursive: true, mode: 0o700 })
 mkdirSync(COLLECT_NOW_DIR, { recursive: true, mode: 0o700 })
 
@@ -302,10 +312,17 @@ function startCollectNowWatch(): void {
 // DIVE-848 offline heal: a message sent while this box was unreachable never
 // produced a drop file — it sits in the control plane with delivered_at NULL.
 // Pull those on boot (and on a slow sweep), push them into the session, then
-// ack so they stamp delivered. Ack only AFTER the notifications are sent; a
+// ack so they stamp COLLECTED. Ack only AFTER the notifications are sent; a
 // crash in between redelivers rather than losing the message. A row whose
-// drop landed but whose delivered-stamp write failed may arrive twice — rare
+// drop landed but whose collected-stamp write failed may arrive twice — rare
 // and preferable to silence.
+// DIVE-3809: the ack no longer stamps delivered_at. It could never attest
+// delivery, and stamping the column `/pending` reads meant one wrong ack
+// deleted the only copy. A collected row is now merely hidden for a TTL and
+// comes back, bounded by an attempt count. Note the consequence for the
+// empty-text branch below: it acks a row it never pushed, so such a row is
+// re-offered until the attempt bound retires it — bounded and visible, where
+// before it was silently destroyed.
 // DIVE-3574: drainPending is now reachable from three places (boot, the 5-min
 // timer, and a collect nudge that can fire several times a second while someone
 // types in the dashboard) where it used to be reachable from two that could
@@ -316,11 +333,63 @@ function startCollectNowWatch(): void {
 // landed after the fetch from waiting out the full timer.
 let draining = false
 let rerun = false
+
+// A drain that dies mid-flight (SIGKILL, box reboot) leaves the lock file
+// behind, and a stale lock that nothing clears would wedge the collect path
+// permanently — the exact failure shape this row exists to remove. So the lock
+// is TIME-BOUNDED: older than this and it is treated as abandoned and broken.
+// One drain is a fetch + N notifications + an ack, all with short timeouts;
+// two minutes is far past any healthy pass.
+const DRAIN_LOCK_STALE_MS = 2 * 60_000
+
+// Returns true if THIS process now holds the lock. Never throws: a filesystem
+// that cannot support the lock must degrade to today's per-process-only
+// behaviour, not stop the customer's message from being collected.
+function acquireDrainLock(): boolean {
+  for (let attempt = 0; attempt < 2; attempt++) {
+    try {
+      const fd = openSync(DRAIN_LOCK, 'wx')
+      try { writeSync(fd, `${process.pid} ${new Date().toISOString()}\n`) } catch {}
+      closeSync(fd)
+      return true
+    } catch (err) {
+      if ((err as NodeJS.ErrnoException).code !== 'EEXIST') {
+        process.stderr.write(`dashboard channel: drain lock unavailable (${err}); per-process guard only\n`)
+        return true
+      }
+      // Held. Break it only if it is provably stale, then retry the create
+      // once — if another process wins that race, we simply skip this pass.
+      try {
+        const age = Date.now() - statSync(DRAIN_LOCK).mtimeMs
+        if (age <= DRAIN_LOCK_STALE_MS) return false
+        process.stderr.write(`dashboard channel: breaking stale drain lock (age ${Math.round(age / 1000)}s)\n`)
+        unlinkSync(DRAIN_LOCK)
+      } catch { return false }
+    }
+  }
+  return false
+}
+
+function releaseDrainLock(): void {
+  try { unlinkSync(DRAIN_LOCK) } catch {}
+}
+
 async function drainPending(): Promise<void> {
   if (draining) { rerun = true; return }
   draining = true
   try {
-    await drainPendingOnce()
+    // Cross-process (DIVE-3809). Skipping is safe and NOT a lost message: the
+    // holder is draining the same rows right now, and anything it misses is
+    // re-offered by the control plane once the collect TTL expires.
+    if (!acquireDrainLock()) {
+      process.stderr.write('dashboard channel: another process holds the drain lock; skipping this pass\n')
+      return
+    }
+    try {
+      await drainPendingOnce()
+    } finally {
+      releaseDrainLock()
+    }
   } finally {
     draining = false
   }
@@ -362,7 +431,7 @@ async function drainPendingOnce(): Promise<void> {
       })
       acked.push(m.id)
     } catch (err) {
-      process.stderr.write(`dashboard channel: pending deliver failed for ${m.id}: ${err}\n`)
+      process.stderr.write(`dashboard channel: pending push failed for ${m.id}: ${err}\n`)
     }
   }
   if (acked.length === 0) return
@@ -372,10 +441,18 @@ async function drainPendingOnce(): Promise<void> {
       headers: { 'content-type': 'application/json' },
       body: JSON.stringify({ agent: AGENT, ids: acked }),
     })
+    // DIVE-3810: a non-2xx ack must reach the catch below. Without this the
+    // rows are logged as collected while the control plane still holds them
+    // uncollected — the exact split this row exists to close.
     if (!ack.ok) throw new Error(`${ack.status}`)
-    process.stderr.write(`dashboard channel: healed ${acked.length} undelivered message(s)\n`)
+    // DIVE-3809: "collected", not "delivered" or "healed". This ack attests
+    // that the notification's bytes entered the stdout pipe — the SDK's send()
+    // has no reject path, and a client with nothing subscribed drops the
+    // notification silently — so it can never say the session displayed it.
+    // The control plane now re-offers a collected row whose TTL expires.
+    process.stderr.write(`dashboard channel: collected ${acked.length} pending message(s) (collection is not display)\n`)
   } catch (err) {
-    process.stderr.write(`dashboard channel: pending ack failed (will redeliver next boot): ${err}\n`)
+    process.stderr.write(`dashboard channel: pending ack failed (row stays uncollected; re-offered next sweep): ${err}\n`)
   }
 }
 
diff --git a/test/dive3809-drain-lock.test.ts b/test/dive3809-drain-lock.test.ts
new file mode 100644
index 0000000..a53b07f
--- /dev/null
+++ b/test/dive3809-drain-lock.test.ts
@@ -0,0 +1,168 @@
+// DIVE-3809 scope 3: the drain guard must serialise ACROSS PROCESSES.
+//
+// server.ts's `draining`/`rerun` pair is module state — it serialises the three
+// callers inside ONE process and is blind to a second one. Two plugin processes
+// for the same agent (an overlapping restart, a stray respawn) each fetch the
+// SAME pending rows and push every message into the session TWICE, because the
+// ack only lands after the notifications are sent.
+//
+// DIVE-3806 REFUTED this as the cause of the loss it observed — lifecycle.log
+// showed exactly one live process across that window — so this is not a fix for
+// that. It is a real race, it is scope 3, and dashboard-collect-now.test.ts
+// cannot see it: that harness runs a single process, so it grades the
+// per-process half and would stay green with the lock deleted.
+//
+// The arm therefore runs TWO real server.ts processes against ONE stub control
+// plane and ONE shared state dir, and counts pushes across BOTH.
+import { describe, test, expect } from 'bun:test'
+import { mkdtempSync, rmSync, mkdirSync, writeFileSync, existsSync, utimesSync } from 'node:fs'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+
+const SERVER = join(import.meta.dir, '..', 'plugins', 'dashboard', 'server.ts')
+const BOOT_MS = 5_000
+
+function spawnPlugin(dir: string, port: number, delivered: string[]) {
+  const proc = Bun.spawn(['bun', SERVER], {
+    env: {
+      ...process.env,
+      DASHBOARD_STATE_DIR: dir,
+      DASHBOARD_API_BASE: `http://127.0.0.1:${port}`,
+      CONNECTORD_TOKEN: 'test-token-abcdefghijkl',
+      USER: 'agent-dev',
+    },
+    stdin: 'pipe',
+    stdout: 'pipe',
+    stderr: 'pipe',
+  })
+  proc.stdin.write(
+    JSON.stringify({
+      jsonrpc: '2.0', id: 1, method: 'initialize',
+      params: { protocolVersion: '2024-11-05', capabilities: {}, clientInfo: { name: 'lock-test', version: '0' } },
+    }) + '\n'
+  )
+  proc.stdin.flush()
+  void (async () => {
+    const reader = proc.stdout.getReader()
+    const dec = new TextDecoder()
+    let buf = ''
+    for (;;) {
+      const { done, value } = await reader.read()
+      if (done) return
+      buf += dec.decode(value, { stream: true })
+      const lines = buf.split('\n')
+      buf = lines.pop() ?? ''
+      for (const line of lines) {
+        if (!line.trim()) continue
+        try {
+          const msg = JSON.parse(line)
+          if (msg.method === 'notifications/claude/channel') delivered.push(String(msg.params?.content ?? ''))
+          if (msg.id === 1) {
+            proc.stdin.write(JSON.stringify({ jsonrpc: '2.0', method: 'notifications/initialized' }) + '\n')
+            proc.stdin.flush()
+          }
+        } catch {}
+      }
+    }
+  })()
+  return proc
+}
+
+async function waitFor(pred: () => boolean, ms: number) {
+  const deadline = Date.now() + ms
+  while (Date.now() < deadline) {
+    if (pred()) return true
+    await Bun.sleep(25)
+  }
+  return pred()
+}
+
+describe('cross-process drain lock (DIVE-3809)', () => {
+  test('two plugin processes on one state dir push each message EXACTLY once', async () => {
+    const dir = mkdtempSync(join(tmpdir(), 'drain-lock-'))
+    const delivered: string[] = []
+    let queue = [
+      { id: 1, text: 'first' },
+      { id: 2, text: 'second' },
+    ]
+    let pendingHits = 0
+    // Snapshot at REQUEST time then delay, exactly as dashboard-collect-now
+    // does: a stub that re-reads after the delay hands the second drain the
+    // state AFTER the first one's ack and silently removes the race.
+    const api = Bun.serve({
+      port: 0,
+      async fetch(req) {
+        const url = new URL(req.url)
+        if (url.pathname === '/server/messages/pending') {
+          pendingHits++
+          const snapshot = queue
+          await Bun.sleep(900)
+          return Response.json({ pending: snapshot })
+        }
+        if (url.pathname === '/server/messages/pending/ack') {
+          const body = (await req.json()) as { ids: number[] }
+          queue = queue.filter(m => !body.ids.includes(m.id))
+          return Response.json({ ok: true })
+        }
+        return new Response('not found', { status: 404 })
+      },
+    })
+    const a = spawnPlugin(dir, api.port, delivered)
+    const b = spawnPlugin(dir, api.port, delivered)
+    try {
+      // Both boot drains fire ~5s after start, i.e. genuinely overlapping.
+      await waitFor(() => delivered.length >= 2, BOOT_MS + 10_000)
+      // Settle: give a second (unserialised) drain every chance to double-push.
+      await Bun.sleep(3_000)
+      const counts = new Map<string, number>()
+      for (const d of delivered) counts.set(d, (counts.get(d) ?? 0) + 1)
+      expect(delivered.length).toBeGreaterThanOrEqual(2)
+      expect([...counts.entries()].filter(([, n]) => n > 1)).toEqual([])
+      expect(pendingHits).toBeGreaterThan(0)
+    } finally {
+      a.kill(); b.kill(); api.stop(true)
+      rmSync(dir, { recursive: true, force: true })
+    }
+  }, 40_000)
+
+  test('a stale lock left by a killed drain is broken, not honoured forever', async () => {
+    // The failure this forecloses: a lock file is the classic way to turn an
+    // intermittent loss into a permanent one. A drain killed mid-flight leaves
+    // the file behind, and if nothing breaks it the collect path is wedged for
+    // the life of the box — strictly worse than the bug being fixed.
+    const dir = mkdtempSync(join(tmpdir(), 'stale-lock-'))
+    mkdirSync(dir, { recursive: true })
+    const lock = join(dir, 'pending-drain.lock')
+    writeFileSync(lock, '999999 stale\n')
+    // Age it well past the 2-minute staleness bound. utimesSync, not `touch
+    // -d` — bun's built-in shell does not implement that flag.
+    const old = new Date(Date.now() - 10 * 60_000)
+    utimesSync(lock, old, old)
+    const delivered: string[] = []
+    let queue = [{ id: 7, text: 'after a crash' }]
+    const api = Bun.serve({
+      port: 0,
+      async fetch(req) {
+        const url = new URL(req.url)
+        if (url.pathname === '/server/messages/pending') return Response.json({ pending: queue })
+        if (url.pathname === '/server/messages/pending/ack') {
+          const body = (await req.json()) as { ids: number[] }
+          queue = queue.filter(m => !body.ids.includes(m.id))
+          return Response.json({ ok: true })
+        }
+        return new Response('not found', { status: 404 })
+      },
+    })
+    const p = spawnPlugin(dir, api.port, delivered)
+    try {
+      const got = await waitFor(() => delivered.length >= 1, BOOT_MS + 10_000)
+      expect(got).toBe(true)
+      expect(delivered[0]).toBe('after a crash')
+      // And the drain cleaned up after itself, so the next pass is not blocked.
+      expect(existsSync(lock)).toBe(false)
+    } finally {
+      p.kill(); api.stop(true)
+      rmSync(dir, { recursive: true, force: true })
+    }
+  }, 40_000)
+})