diff --git a/cli/src/slack-active-pill.test.ts b/cli/src/slack-active-pill.test.ts
new file mode 100644
index 0000000..c02ddc7
--- /dev/null
+++ b/cli/src/slack-active-pill.test.ts
@@ -0,0 +1,83 @@
+import { test, describe, before, after } from "node:test";
+import { strict as assert } from "node:assert";
+import { mkdtempSync, rmSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+
+// Point kanbanHome() at a tmp dir BEFORE importing the module under test so the
+// persistence files land somewhere disposable per run.
+let tmpHome: string;
+
+before(() => {
+  tmpHome = mkdtempSync(join(tmpdir(), "kanban-active-pill-"));
+  process.env.KANBAN_CODE_HOME = tmpHome;
+});
+after(() => {
+  rmSync(tmpHome, { recursive: true, force: true });
+  delete process.env.KANBAN_CODE_HOME;
+});
+
+describe("active-pill persistence", () => {
+  test("write -> read round-trips the pill state", async () => {
+    const { writeActivePill, readActivePill } = await import("./slack/active-pill.js");
+    writeActivePill("agent-a", {
+      channelId: "C123",
+      threadTs: "1730000000.000100",
+      label: "is working…",
+      lastSetMs: 1730000000000,
+    });
+    const got = readActivePill("agent-a");
+    assert.deepEqual(got, {
+      channelId: "C123",
+      threadTs: "1730000000.000100",
+      label: "is working…",
+      lastSetMs: 1730000000000,
+    });
+  });
+
+  test("read returns undefined for an agent that has never had a pill", async () => {
+    const { readActivePill } = await import("./slack/active-pill.js");
+    assert.equal(readActivePill("never-existed"), undefined);
+  });
+
+  test("clear removes the file so a later read sees nothing", async () => {
+    const { writeActivePill, readActivePill, clearActivePill } = await import("./slack/active-pill.js");
+    writeActivePill("agent-b", {
+      channelId: "C2",
+      threadTs: "1.2",
+      label: "is working…",
+      lastSetMs: 1,
+    });
+    assert.notEqual(readActivePill("agent-b"), undefined);
+    clearActivePill("agent-b");
+    assert.equal(readActivePill("agent-b"), undefined);
+  });
+
+  test("clear on a missing file is a no-op (does not throw)", async () => {
+    const { clearActivePill } = await import("./slack/active-pill.js");
+    assert.doesNotThrow(() => clearActivePill("never-existed-also"));
+  });
+
+  test("read tolerates a corrupted file (returns undefined)", async () => {
+    const { writeFileSync, mkdirSync } = await import("node:fs");
+    const { join } = await import("node:path");
+    const dir = join(tmpHome, "active-pills");
+    mkdirSync(dir, { recursive: true });
+    writeFileSync(join(dir, "agent-corrupt"), "not json");
+    const { readActivePill } = await import("./slack/active-pill.js");
+    assert.equal(readActivePill("agent-corrupt"), undefined);
+  });
+
+  test("read rejects partial records (missing required fields)", async () => {
+    const { writeFileSync, mkdirSync } = await import("node:fs");
+    const { join } = await import("node:path");
+    const dir = join(tmpHome, "active-pills");
+    mkdirSync(dir, { recursive: true });
+    writeFileSync(
+      join(dir, "agent-partial"),
+      JSON.stringify({ channelId: "C", threadTs: "1.0" }), // missing label + lastSetMs
+    );
+    const { readActivePill } = await import("./slack/active-pill.js");
+    assert.equal(readActivePill("agent-partial"), undefined);
+  });
+});
diff --git a/cli/src/slack/active-pill.ts b/cli/src/slack/active-pill.ts
new file mode 100644
index 0000000..e71312f
--- /dev/null
+++ b/cli/src/slack/active-pill.ts
@@ -0,0 +1,66 @@
+import { mkdirSync, writeFileSync, readFileSync, renameSync, unlinkSync } from "node:fs";
+import { dirname, join } from "node:path";
+import { kanbanHome } from "../paths.js";
+
+/// On-disk record of the "is working…" pill the bridge has set on an
+/// agent's current channel-root anchor. The bridge keeps an in-memory
+/// `active` map already; this file mirrors that map so a bridge restart
+/// (config-sync triggers one on every applied bundle) doesn't drop the
+/// pill until the agent's NEXT text post relights it. An agent that
+/// goes long on tools with no intermediate text — common when it's
+/// debugging or planning — would otherwise have its channel look dead
+/// for the rest of the turn after a restart.
+///
+/// Same pattern as thread-root: atomic write (tmp + rename) per slug,
+/// stored under ~/.kanban-code/active-pills/<slug>. We keep the record
+/// even after a pill is cleared on a non-terminal turn-end so the
+/// restart path always has SOMETHING to re-light to.
+
+export interface PersistedPill {
+  channelId: string;
+  threadTs: string;
+  label: string;
+  /// Wall-clock when the pill was last set/refreshed. Used by the
+  /// restore path: if a pill is much older than Slack's own idle
+  /// behaviour suggests an agent is still working (say, 10 minutes
+  /// without any refresh), we don't blindly re-light it on restart —
+  /// the agent probably finished and Slack has already cleared it.
+  lastSetMs: number;
+}
+
+function pillPath(slug: string): string {
+  return join(kanbanHome(), "active-pills", slug);
+}
+
+export function writeActivePill(slug: string, pill: PersistedPill): void {
+  if (!slug) return;
+  const path = pillPath(slug);
+  mkdirSync(dirname(path), { recursive: true });
+  const tmp = `${path}.tmp`;
+  writeFileSync(tmp, JSON.stringify(pill));
+  renameSync(tmp, path);
+}
+
+export function readActivePill(slug: string): PersistedPill | undefined {
+  try {
+    const raw = readFileSync(pillPath(slug), "utf-8");
+    const parsed = JSON.parse(raw) as Partial<PersistedPill>;
+    if (!parsed.channelId || !parsed.threadTs || !parsed.label) return undefined;
+    if (typeof parsed.lastSetMs !== "number") return undefined;
+    return parsed as PersistedPill;
+  } catch {
+    return undefined;
+  }
+}
+
+/// Best-effort delete. Used when the bridge explicitly clears a pill
+/// (e.g. a terminal text post) so the next restart doesn't restore
+/// something that was deliberately turned off.
+export function clearActivePill(slug: string): void {
+  if (!slug) return;
+  try {
+    unlinkSync(pillPath(slug));
+  } catch {
+    /* file may not exist — that's fine */
+  }
+}
diff --git a/cli/src/slack/bridge.ts b/cli/src/slack/bridge.ts
index 80e594e..d19615b 100644
--- a/cli/src/slack/bridge.ts
+++ b/cli/src/slack/bridge.ts
@@ -10,6 +10,7 @@ import { Runtime } from "../agents/runtime.js";
 import { recordAnnounceSuppress } from "./announce-suppress.js";
 import { WORKING_PILL_LABEL } from "./announce.js";
 import { writeThreadRoot, readThreadRoot } from "./thread-root.js";
+import { writeActivePill, readActivePill, clearActivePill } from "./active-pill.js";
 import { downloadSlackFile, formatPromptWithAttachments, DownloadedFile, sweepInbox, DEFAULT_RETENTION_DAYS } from "./inbox.js";
 import { parsePicker, Picker } from "./picker.js";
 import { findSessionJsonl, findCodexRollout, pasteTmuxPrompt, captureTmuxPane, sendTmuxKey } from "../data.js";
@@ -166,6 +167,42 @@ export async function runSlackBridge(opts: BridgeOptions): Promise<void> {
   interface ActivePill { channelId: string; threadTs: string; label: string; lastSetMs: number; }
   const active = new Map<string /* slug */, ActivePill>();
 
+  /// Restore pills from disk on startup so a bridge restart (config-sync
+  /// triggers one on every bundle apply, sometimes several per hour while
+  /// we iterate) doesn't drop the pill until the agent's next text post.
+  /// We only restore pills that were set recently — `MAX_RESTORE_AGE_MS`
+  /// older and the agent has likely already finished its turn, Slack's
+  /// own idle TTL would have cleared the visual pill, and re-lighting
+  /// would falsely advertise active work.
+  const MAX_RESTORE_AGE_MS = 10 * 60_000;
+  for (const a of agents) {
+    const pill = readActivePill(a.slug);
+    if (!pill) continue;
+    if (Date.now() - pill.lastSetMs > MAX_RESTORE_AGE_MS) {
+      clearActivePill(a.slug);
+      continue;
+    }
+    active.set(a.slug, pill);
+    // Re-light immediately rather than waiting for the next refresh tick,
+    // so the channel shows "is working…" within seconds of bridge start
+    // (the gap that prompted this whole change). Best-effort; the
+    // refresh loop will retry every cycle anyway.
+    try {
+      await client.setStatus(pill.channelId, pill.threadTs, pill.label);
+    } catch (e) {
+      console.error(`setStatus restore for ${a.slug} failed:`, e);
+    }
+  }
+
+  function setActivePill(slug: string, pill: ActivePill): void {
+    active.set(slug, pill);
+    writeActivePill(slug, pill);
+  }
+  function dropActivePill(slug: string): void {
+    active.delete(slug);
+    clearActivePill(slug);
+  }
+
   // Per-agent buffer of tool/thinking posts that have NOT yet been sent. We
   // hold them until the next text post arrives (or a picker pops), then drain
   // them all into the previous text's thread as a single message. This keeps
@@ -240,7 +277,7 @@ export async function runSlackBridge(opts: BridgeOptions): Promise<void> {
             }
             const ts = await client.post(t.channelId, post.text);
             if (ts) writeThreadRoot(t.slug, ts);
-            active.delete(t.slug);
+            dropActivePill(t.slug);
             // `terminal: true` posts (currently only the codex out-of-credits
             // sentinel) are the final word of the turn — no more work coming.
             // Skip the WORKING pill entirely so the channel doesn't show a
@@ -249,7 +286,7 @@ export async function runSlackBridge(opts: BridgeOptions): Promise<void> {
             if (ts && !post.terminal) {
               try {
                 await client.setStatus(t.channelId, ts, WORKING_LABEL);
-                active.set(t.slug, { channelId: t.channelId, threadTs: ts, label: WORKING_LABEL, lastSetMs: Date.now() });
+                setActivePill(t.slug, { channelId: t.channelId, threadTs: ts, label: WORKING_LABEL, lastSetMs: Date.now() });
               } catch (e) {
                 console.error(`setStatus (text) for ${t.slug} failed:`, e);
               }
@@ -279,6 +316,9 @@ export async function runSlackBridge(opts: BridgeOptions): Promise<void> {
       try {
         await client.setStatus(pill.channelId, pill.threadTs, pill.label);
         pill.lastSetMs = now;
+        // Persist the refresh so a restart right after this tick keeps
+        // the pill fresh (and within the MAX_RESTORE_AGE_MS window).
+        writeActivePill(slug, pill);
       } catch (e) {
         console.error(`setStatus refresh for ${slug} failed:`, e);
       }