pause, resume and restart a background job without asking the agent

Three buttons in the expanded task panel, beside Stop. Not on the chip: the panel
is where you decide what to do with a job, and a row of unlabelled glyphs on a pill
is where you misclick.

── Pause / Resume ──

SIGSTOP and SIGCONT — the only pair that means "stop doing things and keep
everything". The process keeps its memory, its open files and its bound port, and
picks up mid-instruction. Signals go to every pid holding the job's output file,
because for a shell job that is the shell AND what it spawned, and freezing one of
them would be worse than freezing neither.

Verified by watching a job that writes a line a second: 3 lines, frozen at 3 across
three seconds of pause, 6 after resume.

Worth knowing before pressing it: a paused server does not REFUSE connections, it
stops answering them, so a browser pointed at it hangs rather than erroring. That is
what pausing is, not a shortcoming here.

── Restart, and the cost it accepts ──

The command is recovered by joining the transcript to itself: the harness records
the `tool_use` carrying `command`, and separately its result — the text
`Command running in background with ID: <id>`. Neither alone is enough; together
they are a join on `tool_use_id`, and that is the only place the command survives.
The transcript is found FROM the output path rather than searched for, since a task
file at `<tmp>/claude-<uid>/<slug>/<session>/tasks/<id>.output` names both.

The respawned process appends to that same output file. That is what keeps
everything else working untouched — the log view reads that file, and Stop finds the
process by asking which pids hold it open.

The real cost, chosen deliberately: the platform becomes a SECOND owner of a task.
The harness stops watching when its own child dies, so the platform supervises what
it spawned and emits the same `task:started` / `task:notification` pair everything
downstream is built on. `emitToSession` is the seam, and it is the only one — nothing
else learns that a task can have two kinds of parent.

A restart re-announces an id already in the transcript, so `task:started` now RESETS
an existing task row rather than appending a second one. Without that the tray showed
the job twice, once finished and once running.

Restart is offered for a finished job too, which is the case that matters most: a dev
server that fell over is exactly what you want back without typing a sentence at an
agent that may be mid-turn.

Verified end to end against a planted task: command recovered, old process killed,
new one spawned, output APPENDED (`first run output` then `RESTARTED-OK`), exit
reported through the callback, supervision registry released afterwards.

`[open]` A restarted job does not survive an `officer` restart — the platform is its
parent now. The harness's own jobs never did either.

tsgo clean, frontend builds, 890 pass / 7 fail unchanged.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

src/servers/api/chat/chat.ts

@@ -1,4 +1,6 @@
 import type { Context } from 'hono';
+import { emitToSession } from './websocket';
+import { pauseBackgroundTask, restartBackgroundTask } from './task-restart';
 import { createRouter } from '../../create-router';
 import * as errors from '@@/custom-errors';
 import * as sidecar from '@@/sidecar-registry';
@@ -281,6 +283,49 @@ chatRouter.delete('/tasks/:id', async (ctx) => {
   return ctx.json(result, result.stopped ? 200 : 400);
 });
 
+// POST /chat/tasks/:id/restart — run this job's command again, without involving the agent.
+//
+// The platform supervises what it respawns, because the harness stopped watching when its own child died.
+// It emits the same two events a harness-owned task emits, so the tray, the log view and reconnects need to
+// know nothing about the difference. `sessionId` comes from the client because that is who knows which
+// conversation the chip is being looked at in.
+chatRouter.post('/tasks/:id/restart', async (ctx) => {
+  const who = await chatIdentity(ctx.get('user'));
+  const taskId = ctx.req.param('id');
+  const body = await ctx.req.json<{ sessionId?: string }>().catch(() => ({}) as { sessionId?: string });
+
+  const result = await restartBackgroundTask(who, taskId, (code: number) => {
+    if (!body.sessionId) return;
+    emitToSession(body.sessionId, {
+      type: 'task:notification',
+      taskId,
+      status: code === 0 ? 'completed' : 'failed',
+      summary: code === 0 ? 'Restarted job finished' : `Restarted job exited with code ${code}`,
+    });
+  });
+
+  // Announced only on success, and only after the process exists — a `task:started` for something that
+  // failed to spawn would leave a chip running forever with nothing behind it.
+  if (result.ok && body.sessionId) {
+    emitToSession(body.sessionId, {
+      type: 'task:started',
+      taskId,
+      description: result.command?.split('\n')[0]?.slice(0, 80) ?? 'restarted job',
+      taskType: 'local_bash',
+    });
+  }
+
+  return ctx.json(result, result.ok ? 200 : 400);
+});
+
+// POST /chat/tasks/:id/pause — freeze or thaw a running job, keeping everything it holds.
+chatRouter.post('/tasks/:id/pause', async (ctx) => {
+  const who = await chatIdentity(ctx.get('user'));
+  const body = await ctx.req.json<{ paused?: boolean }>().catch(() => ({}) as { paused?: boolean });
+  const result = await pauseBackgroundTask(who, ctx.req.param('id'), body.paused !== false);
+  return ctx.json(result, result.ok ? 200 : 400);
+});
+
 // GET /chat/models — Claude tiers only (the runner is the `claude` CLI).
 chatRouter.get('/models', async (ctx: Context) => {
   const who = await chatIdentity(ctx.get('user'));

src/servers/api/chat/claude-sessions.ts

@@ -711,7 +711,7 @@ export type BackgroundTaskDetail =
  * `<projects>/<slug>/<session>/subagents/agent-<id>.jsonl` (4). `taskId` is `TASK_ID_RE`-checked above, so
  * it carries no glob or shell metacharacter into either.
  */
-function findTaskOutput(who: ChatIdentity, taskId: string): string | null {
+export function findTaskOutput(who: ChatIdentity, taskId: string): string | null {
   // The member's `claude` runs under `--reset-env`, so its TMPDIR is the passwd default, not ours. Both
   // resolve to /tmp on Linux; a server started with a TMPDIR of its own would look in the wrong place.
   const tmpRoot = process.env.TMPDIR ?? '/tmp';
@@ -733,6 +733,23 @@ function tailFile(osUser: AsUser, filePath: string, bytes: number): { text: stri
   return { text: truncated ? text.slice(text.indexOf('\n') + 1) : text, truncated };
 }
 
+/**
+ * The pids with this file open — which is how a job with no recorded pid is found.
+ *
+ * `fuser` reads `/proc/<pid>/fd`, which only the process's own user and root may open, so it runs AS the
+ * job's owner. Run as the platform it would find nothing for a member's job and report it already stopped.
+ */
+export async function pidsHoldingFile(who: ChatIdentity, path: string): Promise<number[]> {
+  const argv = who.osUser ? runAsArgv(who.osUser, ['fuser', path]) : ['fuser', path];
+  const proc = Bun.spawn(argv, { stdout: 'pipe', stderr: 'ignore' });
+  const out = await new Response(proc.stdout).text();
+  await proc.exited;
+  return out
+    .split(/\s+/)
+    .map((token) => Number(token.replace(/\D/g, '')))
+    .filter((n) => Number.isInteger(n) && n > 1);
+}
+
 /**
  * Stop a running background job.
  *
@@ -762,15 +779,7 @@ export async function stopBackgroundTask(
   // stopped. And the kill would then be cross-user: EPERM, silently swallowed by the catch below.
   const asUser = (command: string[]) => (who.osUser ? runAsArgv(who.osUser, command) : command);
 
-  const proc = Bun.spawn(asUser(['fuser', found]), { stdout: 'pipe', stderr: 'ignore' });
+  const pids = await pidsHoldingFile(who, found);
-  const out = await new Response(proc.stdout).text();
-  await proc.exited;
-
-  const pids = out
-    .split(/\s+/)
-    .map((t) => Number(t.replace(/\D/g, '')))
-    .filter((n) => Number.isInteger(n) && n > 1);
-
   if (pids.length === 0) return { stopped: false, reason: 'already stopped' };
 
   if (who.osUser) {

src/servers/api/chat/task-restart.ts

@@ -0,0 +1,202 @@
+import { openSync } from 'node:fs';
+import { basename, dirname, join } from 'node:path';
+import { runAsArgv } from '../../os-user';
+import { readTextAs } from '../../read-as-user';
+import { findTaskOutput, pidsHoldingFile, type ChatIdentity } from './claude-sessions';
+
+/**
+ * Run a background job's command again, without asking the agent.
+ *
+ * ── What this owns, and what it does not ──
+ *
+ * A background task normally belongs entirely to the harness: it spawns the process, watches it, and emits
+ * `task:started` and `task:notification`. Everything downstream — the tray, the log view, reconnects — is
+ * built on exactly those two events.
+ *
+ * A restart cannot be that. The agent is not involved, so nothing would emit either event, and the harness
+ * stops watching the moment its own child dies. So the platform takes over supervision of the process it
+ * spawned: it emits the same two events for it, and watches for the exit itself.
+ *
+ * That is a second owner for the same kind of thing, and it is a real cost — accepted deliberately, so
+ * that restarting a dev server is a button rather than a sentence typed at an agent that may be mid-turn.
+ * The seam is contained here: nothing else learns that a task can have two kinds of parent.
+ *
+ * ── The output file is the anchor ──
+ *
+ * The respawned process appends to the SAME `<task-id>.output` the harness created. That is what keeps
+ * everything else working untouched: the log view reads that file, and Stop finds the process by asking
+ * which pids hold it open. A new file would have meant teaching both about a second location.
+ */
+
+/** In-flight restarts, so a second click stops the first rather than racing it. */
+const restarted = new Map<string, { proc: Bun.Subprocess }>();
+
+export type RestartResult = { ok: boolean; command?: string; reason?: string };
+
+/**
+ * The command that started a task, recovered by joining the transcript back to itself.
+ *
+ * The harness records the tool call (`tool_use`, carrying `command`) and, separately, its result — which
+ * is the text `Command running in background with ID: <task id>`. Neither alone is enough; together they
+ * are a join on `tool_use_id`, and that is the only place the command survives.
+ *
+ * The transcript is found from the OUTPUT PATH rather than by searching: a task file lives at
+ * `<tmp>/claude-<uid>/<project-slug>/<session-uuid>/tasks/<id>.output`, so the slug and session name the
+ * transcript exactly. Guessing among a member's transcripts would be both slower and occasionally wrong.
+ */
+export async function commandForTask(who: ChatIdentity, taskId: string): Promise<string | null> {
+  const output = findTaskOutput(who, taskId);
+  if (!output || !output.endsWith('.output')) return null;
+
+  const sessionDir = dirname(dirname(output));
+  const transcript = join(
+    who.home,
+    '.claude',
+    'projects',
+    basename(dirname(sessionDir)),
+    `${basename(sessionDir)}.jsonl`,
+  );
+
+  let text: string;
+  try {
+    text = readTextAs(who.osUser, transcript);
+  } catch {
+    return null;
+  }
+
+  const commands = new Map<string, string>();
+  let toolUseId: string | null = null;
+
+  for (const line of text.split('\n')) {
+    if (!line) continue;
+    let entry: unknown;
+    try {
+      entry = JSON.parse(line);
+    } catch {
+      continue;
+    }
+    const content = (entry as { message?: { content?: unknown } })?.message?.content;
+    if (!Array.isArray(content)) continue;
+
+    for (const block of content as Record<string, unknown>[]) {
+      if (block.type === 'tool_use' && block.name === 'Bash') {
+        const command = (block.input as { command?: unknown } | undefined)?.command;
+        if (typeof block.id === 'string' && typeof command === 'string') commands.set(block.id, command);
+      }
+      if (block.type === 'tool_result') {
+        const body = typeof block.content === 'string' ? block.content : JSON.stringify(block.content ?? '');
+        // Both phrasings the harness produces carry the id; this is the one that names a shell job.
+        if (body.includes(`background with ID: ${taskId}`) && typeof block.tool_use_id === 'string') {
+          toolUseId = block.tool_use_id;
+        }
+      }
+    }
+  }
+
+  return toolUseId ? (commands.get(toolUseId) ?? null) : null;
+}
+
+/**
+ * Stop whatever is running for this task and start its command again.
+ *
+ * `onExit` is how the caller emits `task:notification` — this module deliberately knows nothing about
+ * websockets or sessions.
+ */
+export async function restartBackgroundTask(
+  who: ChatIdentity,
+  taskId: string,
+  onExit: (code: number) => void,
+): Promise<RestartResult> {
+  const output = findTaskOutput(who, taskId);
+  if (!output) return { ok: false, reason: 'no such task' };
+
+  const command = await commandForTask(who, taskId);
+  if (!command) {
+    // Honest about which half failed: the file is there, the command is not recoverable. That happens when
+    // the transcript has been pruned, and no amount of retrying will change it.
+    return { ok: false, reason: 'cannot tell what command started this job — its transcript is gone' };
+  }
+
+  // Ours first, then anything else holding the file. A second click must replace the first restart, not
+  // leave two dev servers fighting over a port.
+  const previous = restarted.get(taskId);
+  if (previous) {
+    try {
+      previous.proc.kill();
+    } catch {
+      /* already dead */
+    }
+    restarted.delete(taskId);
+  }
+
+  for (const pid of await pidsHoldingFile(who, output)) {
+    const argv = who.osUser ? runAsArgv(who.osUser, ['kill', '-TERM', String(pid)]) : null;
+    try {
+      if (argv) Bun.spawn(argv, { stdout: 'ignore', stderr: 'ignore' });
+      else process.kill(pid, 'SIGTERM');
+    } catch {
+      /* gone between listing and killing */
+    }
+  }
+  // A dev server holding a port needs a moment to release it, and the whole point of a restart is that the
+  // new one binds. Not a fix for a process that ignores SIGTERM — that is what the second click is for.
+  await Bun.sleep(300);
+
+  // Appended, never truncated: the log view is reading this file, and a restart is a new chapter in a job's
+  // life rather than a new job.
+  const fd = openSync(output, 'a');
+  const argv = who.osUser ? runAsArgv(who.osUser, ['sh', '-c', command]) : ['sh', '-c', command];
+  const proc = Bun.spawn(argv, { stdout: fd, stderr: fd, stdin: 'ignore' });
+  restarted.set(taskId, { proc });
+
+  void proc.exited.then((code) => {
+    // Only report the exit of the process we are still tracking — a later restart owns the task by then.
+    if (restarted.get(taskId)?.proc === proc) {
+      restarted.delete(taskId);
+      onExit(code ?? 0);
+    }
+  });
+
+  return { ok: true, command };
+}
+
+/** Whether the platform (rather than the harness) is currently supervising this task. */
+export const isRestarted = (taskId: string): boolean => restarted.has(taskId);
+
+/**
+ * Freeze or thaw a job, without ending it.
+ *
+ * `SIGSTOP` and `SIGCONT`, which is the only pair that means "stop doing things and keep everything" — the
+ * process keeps its memory, its open files and its bound ports, and picks up mid-instruction.
+ *
+ * Worth knowing before pressing it: a paused server does not REFUSE connections, it stops answering them.
+ * A browser pointed at it hangs rather than erroring, and an HMR client may give up and reconnect on
+ * resume. That is inherent to what pausing is, not a shortcoming of this implementation.
+ *
+ * Signals go to every pid holding the output file — for a shell job that is the shell and whatever it
+ * spawned, and freezing only one of them would be worse than not freezing at all.
+ */
+export async function pauseBackgroundTask(
+  who: ChatIdentity,
+  taskId: string,
+  paused: boolean,
+): Promise<{ ok: boolean; reason?: string }> {
+  const output = findTaskOutput(who, taskId);
+  if (!output) return { ok: false, reason: 'no such task' };
+
+  const pids = await pidsHoldingFile(who, output);
+  if (pids.length === 0) return { ok: false, reason: 'nothing is running for this job' };
+
+  const signal = paused ? '-STOP' : '-CONT';
+  for (const pid of pids) {
+    const argv = who.osUser ? runAsArgv(who.osUser, ['kill', signal, String(pid)]) : null;
+    try {
+      if (argv) Bun.spawn(argv, { stdout: 'ignore', stderr: 'ignore' });
+      else process.kill(pid, paused ? 'SIGSTOP' : 'SIGCONT');
+    } catch {
+      // Gone between listing and signalling. Nothing to do, and nothing to report: the caller asked for a
+      // state, and "not running" satisfies "paused" closely enough to not be worth an error.
+    }
+  }
+  return { ok: true };
+}

src/servers/api/chat/websocket.ts

@@ -318,6 +318,25 @@ function createMessageHandler(sessionId: string, model: string) {
   };
 }
 
+/**
+ * Push an event into a live session from OUTSIDE a turn.
+ *
+ * The restart button is the only caller: the platform respawns a process the harness is no longer watching,
+ * so nothing else would emit the `task:started` / `task:notification` pair that the tray, the log view and
+ * every reconnect are built on. Folded into the session first and then sent, in that order, so a browser
+ * that reconnects a second later sees the same thing as one that was already listening.
+ *
+ * Returns false when the session is not live — a restart is still legitimate then, it simply has nobody to
+ * tell.
+ */
+export function emitToSession(sessionId: string, msg: TurnMessage): boolean {
+  const session = sessionManager.getSession(sessionId);
+  if (!session) return false;
+  foldIntoSession(session, msg, session.meta.model ?? '');
+  for (const socket of session.sockets) sendToClient(socket as ServerWebSocket<WSData>, msg);
+  return true;
+}
+
 async function handleChat(
   ws: ServerWebSocket<WSData>,
   msg: {

src/workspaces/officerdev/src/apps/Chat/components/BackgroundTaskTray.tsx

@@ -2,7 +2,7 @@ import type { ReactNode } from 'react';
 import { useEffect, useRef, useState } from 'react';
 import { toast } from 'sonner';
 import { useClient } from 'hooks/useClient';
-import { Check, ChevronDown, CircleSlash, Loader2, Pin, Square, X } from 'lucide-react';
+import { Check, ChevronDown, CircleSlash, Loader2, Pause, Pin, Play, RotateCw, Square, X } from 'lucide-react';
 import type { Tone } from '@/components/Data';
 import { toneText } from '@/components/Data';
 import type { ChatMessage } from '../types';
@@ -13,6 +13,8 @@ import { markUserBackgroundJob } from '../../../background-job-origin';
 
 type BackgroundTaskTrayProps = {
   messages: ChatMessage[];
+  /** Which conversation the chips belong to. A restart emits its events back into this session. */
+  sessionId: string | null;
 };
 
 /** Same vocabulary as the transcript's own task rows — a completed task must not be green here and grey there. */
@@ -31,7 +33,7 @@ function statusTone(status: BackgroundTask['status']): Tone {
  * Clicking a chip opens what it is actually doing right now: a subagent's own tool-by-tool trace, or the
  * tail of a backgrounded shell's log, both read from the file Claude Code streams the task into.
  */
-export const BackgroundTaskTray = ({ messages }: BackgroundTaskTrayProps) => {
+export const BackgroundTaskTray = ({ messages, sessionId }: BackgroundTaskTrayProps) => {
   const { tasks, pinned, unpinned, running, clearable, togglePin, dismiss, dismissFinished } =
     useBackgroundTasks(messages);
   const [openId, setOpenId] = useState<string | null>(null);
@@ -56,6 +58,46 @@ export const BackgroundTaskTray = ({ messages }: BackgroundTaskTrayProps) => {
     }
   };
 
+  /**
+   * Run this job's command again, without asking the agent.
+   *
+   * The platform recovers the command from the transcript, stops whatever is running, and respawns it
+   * appending to the same output file — so the log keeps growing rather than starting a new one, and Stop
+   * keeps working because it finds the process by who holds that file open.
+   *
+   * Offered for a FINISHED job too, which is the case that matters most: a dev server that fell over is
+   * exactly the thing you want back without typing a sentence at an agent that may be mid-turn.
+   */
+  const restartTask = async (taskId: string) => {
+    try {
+      const result = await client.post<{ ok: boolean; reason?: string }>(`/chat/tasks/${taskId}/restart`, {
+        sessionId,
+      });
+      if (!result.ok) toast.error(result.reason ?? 'Could not restart that job');
+      else markUserBackgroundJob(taskId);
+    } catch (err) {
+      toast.error(err instanceof Error ? err.message : 'Could not restart that job');
+    }
+  };
+
+  /**
+   * Freeze or thaw the process, without ending it.
+   *
+   * SIGSTOP/SIGCONT, so a dev server keeps its port, its memory and its place — a paused one does not
+   * refuse connections, it stops answering them, which is the honest thing for the button to do and worth
+   * knowing before you wonder why the page hangs.
+   */
+  const pauseTask = async (taskId: string, paused: boolean): Promise<boolean> => {
+    try {
+      const result = await client.post<{ ok: boolean; reason?: string }>(`/chat/tasks/${taskId}/pause`, { paused });
+      if (!result.ok) toast.error(result.reason ?? 'Could not pause that job');
+      return result.ok;
+    } catch (err) {
+      toast.error(err instanceof Error ? err.message : 'Could not pause that job');
+      return false;
+    }
+  };
+
   // Derived, not stored: dismissing the chip that is open closes the panel without a second piece of state
   // having to be kept in step with the first.
   const open = tasks.find((t) => t.taskId === openId) ?? null;
@@ -80,6 +122,8 @@ export const BackgroundTaskTray = ({ messages }: BackgroundTaskTrayProps) => {
           task={open}
           onClose={() => setOpenId(null)}
           onStop={open.status ? undefined : () => void stopTask(open.taskId)}
+          onPause={(next) => pauseTask(open.taskId, next)}
+          onRestart={() => void restartTask(open.taskId)}
         />
       )}
 
@@ -234,12 +278,20 @@ const TaskChip = ({ task, active, pinned, onClick, onTogglePin, onDismiss, onSto
 type TaskPanelProps = {
   /** Present only while the task is running — the panel shows a real button rather than a bare glyph. */
   onStop?: () => void;
+  /** Freeze and thaw the process. Present while it is running; the label follows the state. */
+  onPause?: (paused: boolean) => Promise<boolean>;
+  /** Run the command again, whether it is running or long finished. */
+  onRestart?: () => void;
   task: BackgroundTask;
   onClose: () => void;
 };
 
-const TaskPanel = ({ task, onClose, onStop }: TaskPanelProps) => {
+const TaskPanel = ({ task, onClose, onStop, onPause, onRestart }: TaskPanelProps) => {
   const running = !task.status;
+  // Local, because a paused process is a short-lived state somebody is watching. It resets on reload,
+  // which is honest enough: SIGCONT is idempotent, and the button says what it will do rather than what
+  // it believes has happened.
+  const [paused, setPaused] = useState(false);
   const { data, isLoading } = useTaskDetail(task.taskId, running);
   const bodyRef = useRef<HTMLDivElement | null>(null);
   const pinnedRef = useRef(true);
@@ -264,6 +316,33 @@ const TaskPanel = ({ task, onClose, onStop }: TaskPanelProps) => {
         {/* Labelled here, unlike the chip's icon-only version. The panel is where you come to decide
             whether to keep something running, so the control that ends it should not be a glyph you
             have to hover to identify. */}
+        {/* Pause and Restart sit beside Stop rather than on the chip: the panel is where you decide what
+            to do with a job, and a row of unlabelled glyphs on a pill is where you misclick. */}
+        {running && onPause && (
+          <button
+            type="button"
+            onClick={async () => {
+              const next = !paused;
+              if (await onPause(next)) setPaused(next);
+            }}
+            title={paused ? 'Let it continue where it left off' : 'Freeze the process, keeping its state'}
+            className="flex shrink-0 cursor-pointer items-center gap-1 rounded border border-border px-2 py-1 text-xs text-muted-foreground transition-colors hover:border-warning/40 hover:text-warning"
+          >
+            {paused ? <Play className="h-3 w-3" /> : <Pause className="h-3 w-3" />}
+            {paused ? 'Resume' : 'Pause'}
+          </button>
+        )}
+        {onRestart && (
+          <button
+            type="button"
+            onClick={onRestart}
+            title="Stop it and run the same command again"
+            className="flex shrink-0 cursor-pointer items-center gap-1 rounded border border-border px-2 py-1 text-xs text-muted-foreground transition-colors hover:border-duck-teal/40 hover:text-duck-teal"
+          >
+            <RotateCw className="h-3 w-3" />
+            Restart
+          </button>
+        )}
         {running && onStop && (
           <button
             type="button"

src/workspaces/officerdev/src/apps/Chat/components/InputArea.tsx

@@ -45,6 +45,7 @@ export const InputArea = ({ manager }: InputAreaProps) => {
     sendQueuedNow,
     unqueue,
     sendWithHiddenPrefix,
+    sessionId,
   } = manager;
 
   const { recording, transcribing, toggleRecording } = useAudioRecording(appendToInput);
@@ -85,7 +86,7 @@ export const InputArea = ({ manager }: InputAreaProps) => {
         <div className="mb-2 px-3 py-1.5 text-xs text-duck-teal bg-duck-teal/10 rounded-md">{commandFeedback}</div>
       )}
 
-      <BackgroundTaskTray messages={messages} />
+      <BackgroundTaskTray messages={messages} sessionId={sessionId} />
 
       <QueuedList queued={queued} onRemove={unqueue} onSendNow={sendQueuedNow} />
 

src/workspaces/officerdev/src/hooks/useChat.ts

@@ -436,10 +436,23 @@ export function useChat(initialSessionId?: string, initialModel?: string | null,
         break;
 
       case 'task:started':
-        setMessages((prev) => [
+        setMessages((prev) => {
-          ...prev,
+          // A RESTART re-announces an id that is already in the transcript — the platform respawned the
-          { role: 'task', taskId: msg.taskId, description: msg.description, taskType: msg.taskType },
+          // job and emits the same two events the harness would. Resetting that row rather than appending
-        ]);
+          // a second one is what makes the chip go back to running instead of the tray showing the job
+          // twice, once finished and once not.
+          const at = prev.findIndex((m) => m.role === 'task' && m.taskId === msg.taskId);
+          const row: ChatMessage = {
+            role: 'task',
+            taskId: msg.taskId,
+            description: msg.description,
+            taskType: msg.taskType,
+          };
+          if (at < 0) return [...prev, row];
+          const next = [...prev];
+          next[at] = row;
+          return next;
+        });
         // Only jobs the user started from the dialog. An agent starts background jobs constantly while
         // working — tests, builds, greps — and reloading a preview on those is the same churn this was
         // built to avoid, only with a different trigger. See `background-job-origin`.