fix(fleet): verify via pane-change diff + non-resizing watch

Blocker fix: send --verify now captures a BEFORE snapshot immediately
before the send and an AFTER snapshot after the delay, then uses
classifySendResult(before, after) to classify. A wedged pane showing
stale non-empty content is no longer falsely reported as 'accepted' —
BEFORE==AFTER maps to 'unverifiable' (exit 1, "no pane change after
send"). Blank AFTER still fails closed as 'unverifiable'. Only
AFTER != BEFORE without a draft suffix counts as 'accepted' (exit 0).

Should-fix: agent watch now uses a GROUPED VIEWER SESSION instead of a
bare 'tmux attach -r' against the agent session. A bare attach lets the
viewer terminal shrink the agent's window; a grouped session has
independent sizing so the agent's window is never affected.
Sequence: new-session -d -t '=<agent>' -s '<agent>-watch-<pid>' (runner),
attach -r to viewer session (interactiveRunner), kill-session on detach
(runner). New builder functions exported: buildAgentWatchCreateViewerCommand,
buildAgentWatchAttachCommand, buildAgentWatchKillViewerCommand,
buildViewerSessionName. buildAgentWatchCommand kept but deprecated.

New exports: classifySendResult(before, after) — the testable classifier.

Tests added:
- classifySendResult unit suite (6 cases): accepted/draft/unverifiable/
  stale-pane/both-blank/before-blank-after-response
- send --verify regression: stale (before==after non-empty) => exit 1
- send --verify regression: blank AFTER => exit 1
- send --verify regression: draft after pane change => exit 1
- send --verify regression: changed non-draft => exit 0
- send --verify: 3-call sequence assertion (before-capture, send, after-capture)
- watch dispatch: grouped viewer session created/attached/killed; no bare
  attach against agent session; viewer name matches <agent>-watch-<pid>

PRD Known-limitations updated: pane-change check rationale, Phase-3
heartbeat-ack requirement, grouped-session watch design.

All gates pass: pnpm typecheck, pnpm lint, pnpm --filter @mosaicstack/mosaic test
(382 tests, 74 fleet), prettier --check.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RMoEx7hfdFGjUiCHuN1RRi

docs/fleet/PRD.md

@@ -80,13 +80,25 @@ observability and no safe way to watch a session.
 - **Verify heuristic is best-effort:** `agent send --verify` uses a `>` -prefix draft
   heuristic that is specific to pi/claude TUIs. Draft detection for codex and opencode
   TUIs is best-effort only; those runtimes may not use the same input-line indicator.
-- **Blank capture fails closed:** Full-screen TUIs (claude, codex, opencode, pi) render
+- **Pane-change check is the best Phase-2 signal:** `agent send --verify` compares a
-  blank for `tmux capture-pane`. When the captured output is empty, `send --verify`
+  BEFORE snapshot (captured immediately before the send) to an AFTER snapshot (captured
+  after the send delay). A pane that changed and does not end in a draft line is reported
+  as 'accepted'. A pane that did not change — including a wedged pane showing stale
+  non-empty content — is reported 'unverifiable' (exit 1, "no pane change after send").
+  Definitive acceptance ultimately requires a runtime acknowledgement (Phase-3
+  heartbeat-ack); the pane-change check is the best signal available against an opaque
+  TUI for Phase-2.
+- **Blank AFTER capture fails closed:** Full-screen TUIs (claude, codex, opencode, pi)
+  render blank for `tmux capture-pane`. When the AFTER snapshot is empty, `send --verify`
   returns non-zero with an "unverifiable" message rather than silently succeeding. This
   is an intentional fail-closed design (FR-5).
-- **`agent watch` requires TTY passthrough:** `tmux attach` is interactive and must be
+- **`agent watch` uses a grouped viewer session:** `tmux attach -r` directly against the
-  run with inherited stdio. It cannot be captured through a pipe. Tests inject a fake
+  agent session lets the viewer terminal shrink the agent's window. `agent watch` instead
-  `interactiveRunner`; the real implementation spawns with `stdio: 'inherit'`.
+  creates a throwaway grouped session (`tmux new-session -d -t '=<agent>' -s
+'<agent>-watch-<pid>'`), attaches read-only to that session, and kills it on detach.
+  The grouped session shares the agent's windows but has independent sizing, so the
+  agent's window is never affected. `tmux attach` is still interactive and requires
+  inherited stdio; the `interactiveRunner` handles TTY passthrough.
 
 ## Surfaces & parity (MVP-X1)
 

packages/mosaic/src/commands/fleet.spec.ts

@@ -5,11 +5,15 @@ import { Command } from 'commander';
 import { afterEach, describe, expect, it, vi } from 'vitest';
 import {
   buildAgentSendCommand,
+  buildAgentWatchAttachCommand,
   buildAgentWatchCommand,
+  buildAgentWatchCreateViewerCommand,
+  buildAgentWatchKillViewerCommand,
   buildAgentVerifyAcceptedCommand,
   buildFleetServiceCommand,
   buildSystemdShowCommand,
   buildTmuxListPanesCommand,
+  classifySendResult,
   detectDrift,
   generateAgentEnv,
   getDefaultOperatorSourceLabel,
@@ -1040,25 +1044,52 @@ describe('fleet ps — command sequences issued', () => {
 });
 
 describe('agent watch', () => {
-  it('builds exact read-only tmux attach command', () => {
+  it('builds exact grouped-viewer creation command', () => {
-    expect(buildAgentWatchCommand('canary-pi', 'mosaic-factory')).toEqual([
+    expect(
+      buildAgentWatchCreateViewerCommand('canary-pi', 'canary-pi-watch-123', 'mosaic-factory'),
+    ).toEqual([
+      'tmux',
+      '-L',
+      'mosaic-factory',
+      'new-session',
+      '-d',
+      '-t',
+      '=canary-pi',
+      '-s',
+      'canary-pi-watch-123',
+    ]);
+  });
+
+  it('builds exact viewer attach command (read-only)', () => {
+    expect(buildAgentWatchAttachCommand('canary-pi-watch-123', 'mosaic-factory')).toEqual([
       'tmux',
       '-L',
       'mosaic-factory',
       'attach',
       '-r',
       '-t',
-      '=canary-pi',
+      'canary-pi-watch-123',
     ]);
   });
 
-  it('uses DEFAULT_SOCKET_NAME when socket is omitted', () => {
+  it('builds exact viewer kill command', () => {
+    expect(buildAgentWatchKillViewerCommand('canary-pi-watch-123', 'mosaic-factory')).toEqual([
+      'tmux',
+      '-L',
+      'mosaic-factory',
+      'kill-session',
+      '-t',
+      'canary-pi-watch-123',
+    ]);
+  });
+
+  it('buildAgentWatchCommand (deprecated) still uses DEFAULT_SOCKET_NAME when socket is omitted', () => {
     const cmd = buildAgentWatchCommand('canary-pi');
     expect(cmd[2]).toBe('mosaic-factory');
     expect(cmd).toContain('-r');
   });
 
-  it('dispatches the read-only attach command through the interactiveRunner, NOT the capturing runner', async () => {
+  it('dispatch: creates grouped viewer session (runner) then attaches -r to viewer session (interactiveRunner), NOT a bare attach to the agent session', async () => {
     const home = await mkdtemp(join(tmpdir(), 'mosaic-fleet-'));
     await mkdir(join(home, 'fleet'), { recursive: true });
     await writeFile(
@@ -1086,11 +1117,28 @@ describe('agent watch', () => {
 
     try {
       await program.parseAsync(['node', 'mosaic', 'agent', 'watch', 'coder0']);
-      // Must go through interactiveRunner, not the capturing runner
+
-      expect(capturingCalls).toHaveLength(0);
+      // The capturing runner must be used for grouped-session creation and cleanup.
-      expect(interactiveCalls).toEqual([
+      // It must NOT be used for the interactive attach.
-        ['tmux', '-L', 'mosaic-factory', 'attach', '-r', '-t', '=coder0'],
+      expect(capturingCalls).toHaveLength(2); // new-session + kill-session
-      ]);
+      expect(capturingCalls[0]).toEqual(
+        expect.arrayContaining(['new-session', '-d', '-t', '=coder0']),
+      );
+      // The new-session command must include a viewer session name derived from agent name.
+      expect(capturingCalls[0]!.join(' ')).toMatch(/coder0-watch-\d+/);
+      // Kill-session must target the same viewer session, not the agent session.
+      expect(capturingCalls[1]).toEqual(expect.arrayContaining(['kill-session', '-t']));
+      expect(capturingCalls[1]!.join(' ')).toMatch(/coder0-watch-\d+/);
+      // The agent session itself must NOT be the attach target.
+      expect(capturingCalls[1]!.join(' ')).not.toContain('=coder0');
+
+      // The interactiveRunner must attach -r to the VIEWER session, not the agent session.
+      expect(interactiveCalls).toHaveLength(1);
+      expect(interactiveCalls[0]).toEqual(expect.arrayContaining(['attach', '-r', '-t']));
+      // Target must be the viewer session name (not "=coder0").
+      const attachTarget = interactiveCalls[0]![interactiveCalls[0]!.indexOf('-t') + 1]!;
+      expect(attachTarget).toMatch(/coder0-watch-\d+/);
+      expect(attachTarget).not.toBe('=coder0');
     } finally {
       await rm(home, { recursive: true, force: true });
     }
@@ -1150,7 +1198,44 @@ describe('agent send --verify', () => {
     expect(isSendAccepted('   \n  \n')).toBe('unverifiable');
   });
 
-  it('issues send then verify capture via injected runner when --verify is passed', async () => {
+  // ---------------------------------------------------------------------------
+  // classifySendResult — BEFORE/AFTER pane-diff classifier (regression suite)
+  // ---------------------------------------------------------------------------
+
+  describe('classifySendResult (BEFORE/AFTER pane-diff classifier)', () => {
+    it('returns "accepted" when AFTER differs from BEFORE and AFTER has no draft line', () => {
+      const before = 'Old content from prior interaction\n';
+      const after = 'Old content from prior interaction\nAgent response: task complete.\n';
+      expect(classifySendResult(before, after)).toBe('accepted');
+    });
+
+    it('returns "draft" when AFTER differs from BEFORE and AFTER ends in a draft line', () => {
+      const before = 'Previous output\n';
+      const after = 'Previous output\n> unsent message\n';
+      expect(classifySendResult(before, after)).toBe('draft');
+    });
+
+    it('returns "unverifiable" when AFTER is blank/empty (full-screen TUI blank render)', () => {
+      const before = 'Some previous content\n';
+      expect(classifySendResult(before, '')).toBe('unverifiable');
+      expect(classifySendResult(before, '   \n  \n')).toBe('unverifiable');
+    });
+
+    it('returns "unverifiable" when AFTER == BEFORE (stale/wedged pane — no change after send)', () => {
+      const staleContent = 'Old non-empty content that never changed\n';
+      expect(classifySendResult(staleContent, staleContent)).toBe('unverifiable');
+    });
+
+    it('returns "unverifiable" when both BEFORE and AFTER are blank (both blank => no change)', () => {
+      expect(classifySendResult('', '')).toBe('unverifiable');
+    });
+
+    it('returns "accepted" when BEFORE is blank and AFTER has non-draft content (pane woke up)', () => {
+      expect(classifySendResult('', 'Agent is now responding.\n')).toBe('accepted');
+    });
+  });
+
+  it('issues BEFORE-capture then send then AFTER-capture (3 calls) when --verify is passed', async () => {
     const home = await mkdtemp(join(tmpdir(), 'mosaic-fleet-'));
     await mkdir(join(home, 'fleet'), { recursive: true });
     await writeFile(
@@ -1160,11 +1245,17 @@ describe('agent send --verify', () => {
       ),
     );
 
+    let callIndex = 0;
     const calls: string[][] = [];
     const runner: CommandRunner = async (command, args) => {
       calls.push([command, ...args]);
-      // For agent-send.sh: success; for capture-pane: return accepted output
+      const idx = callIndex++;
-      return { stdout: 'Response from agent\n', stderr: '', exitCode: 0 };
+      if ([command, ...args].join(' ').includes('agent-send.sh')) {
+        return { stdout: '', stderr: '', exitCode: 0 };
+      }
+      // BEFORE capture: return old content; AFTER capture: return new content
+      const stdout = idx === 0 ? 'Old pane content\n' : 'New response from agent\n';
+      return { stdout, stderr: '', exitCode: 0 };
     };
 
     const program = new Command();
@@ -1183,11 +1274,11 @@ describe('agent send --verify', () => {
         '--verify',
       ]);
 
-      // First call should be agent-send.sh, second call should be capture-pane for verify
+      // 3 calls: BEFORE-capture, send, AFTER-capture
-      expect(calls).toHaveLength(2);
+      expect(calls).toHaveLength(3);
-      expect(calls[0]![0]).toContain('agent-send.sh');
+      expect(calls[0]).toEqual(buildAgentVerifyAcceptedCommand('coder0', 'mosaic-factory', 5));
-      const captureCall = calls[1]!;
+      expect(calls[1]![0]).toContain('agent-send.sh');
-      expect(captureCall).toEqual(buildAgentVerifyAcceptedCommand('coder0', 'mosaic-factory', 5));
+      expect(calls[2]).toEqual(buildAgentVerifyAcceptedCommand('coder0', 'mosaic-factory', 5));
     } finally {
       await rm(home, { recursive: true, force: true });
     }
@@ -1231,7 +1322,7 @@ describe('agent send --verify', () => {
     }
   });
 
-  it('send --verify: blank capture sets process.exitCode=1 (unverifiable, fails closed)', async () => {
+  it('send --verify: AFTER==BEFORE (stale/wedged pane) sets process.exitCode=1 (unverifiable)', async () => {
     const originalExitCode = process.exitCode;
     const stderrMessages: string[] = [];
     const stderrSpy = vi.spyOn(process.stderr, 'write').mockImplementation((msg) => {
@@ -1251,8 +1342,8 @@ describe('agent send --verify', () => {
     const runner: CommandRunner = async (command, args) => {
       const full = [command, ...args].join(' ');
       if (full.includes('agent-send.sh')) return { stdout: '', stderr: '', exitCode: 0 };
-      // capture-pane returns blank (full-screen TUI)
+      // BEFORE and AFTER are identical non-empty stale content — simulates a wedged pane
-      return { stdout: '', stderr: '', exitCode: 0 };
+      return { stdout: 'Stale old content that never changed\n', stderr: '', exitCode: 0 };
     };
 
     const program = new Command();
@@ -1271,7 +1362,8 @@ describe('agent send --verify', () => {
         '--verify',
       ]);
       expect(process.exitCode).toBe(1);
-      expect(stderrMessages.join('')).toMatch(/could not verify delivery.*blank/i);
+      // Must mention "no pane change" to distinguish from blank-capture case
+      expect(stderrMessages.join('')).toMatch(/no pane change after send/i);
     } finally {
       process.exitCode = originalExitCode;
       stderrSpy.mockRestore();
@@ -1279,7 +1371,7 @@ describe('agent send --verify', () => {
     }
   }, 10_000);
 
-  it('send --verify: draft line sets process.exitCode=1 with distinct wording', async () => {
+  it('send --verify: blank AFTER capture sets process.exitCode=1 (unverifiable, fails closed)', async () => {
     const originalExitCode = process.exitCode;
     const stderrMessages: string[] = [];
     const stderrSpy = vi.spyOn(process.stderr, 'write').mockImplementation((msg) => {
@@ -1296,11 +1388,65 @@ describe('agent send --verify', () => {
       ),
     );
 
+    let captureCallCount = 0;
     const runner: CommandRunner = async (command, args) => {
       const full = [command, ...args].join(' ');
       if (full.includes('agent-send.sh')) return { stdout: '', stderr: '', exitCode: 0 };
-      // capture-pane returns a draft line ("> unsent message")
+      captureCallCount++;
-      return { stdout: '> unsent message\n', stderr: '', exitCode: 0 };
+      // BEFORE: some content; AFTER: blank (full-screen TUI renders blank after send)
+      const stdout = captureCallCount === 1 ? 'Previous content\n' : '';
+      return { stdout, stderr: '', exitCode: 0 };
+    };
+
+    const program = new Command();
+    program.exitOverride();
+    registerAgentCommand(program, { runner, mosaicHome: home });
+
+    try {
+      await program.parseAsync([
+        'node',
+        'mosaic',
+        'agent',
+        'send',
+        'coder0',
+        '--message',
+        'hello',
+        '--verify',
+      ]);
+      expect(process.exitCode).toBe(1);
+      expect(stderrMessages.join('')).toMatch(/could not verify delivery/i);
+    } finally {
+      process.exitCode = originalExitCode;
+      stderrSpy.mockRestore();
+      await rm(home, { recursive: true, force: true });
+    }
+  }, 10_000);
+
+  it('send --verify: AFTER differs from BEFORE with draft line sets process.exitCode=1', async () => {
+    const originalExitCode = process.exitCode;
+    const stderrMessages: string[] = [];
+    const stderrSpy = vi.spyOn(process.stderr, 'write').mockImplementation((msg) => {
+      stderrMessages.push(String(msg));
+      return true;
+    });
+
+    const home = await mkdtemp(join(tmpdir(), 'mosaic-fleet-'));
+    await mkdir(join(home, 'fleet'), { recursive: true });
+    await writeFile(
+      join(home, 'fleet', 'roster.yaml'),
+      ['version: 1', 'transport: tmux', 'agents:', '  - name: coder0', '    runtime: codex'].join(
+        '\n',
+      ),
+    );
+
+    let captureCallCount = 0;
+    const runner: CommandRunner = async (command, args) => {
+      const full = [command, ...args].join(' ');
+      if (full.includes('agent-send.sh')) return { stdout: '', stderr: '', exitCode: 0 };
+      captureCallCount++;
+      // BEFORE: old content; AFTER: message appeared but ended as a draft line
+      const stdout = captureCallCount === 1 ? 'Previous output\n' : '> unsent message\n';
+      return { stdout, stderr: '', exitCode: 0 };
     };
 
     const program = new Command();
@@ -1327,7 +1473,7 @@ describe('agent send --verify', () => {
     }
   }, 10_000);
 
-  it('send --verify: real response content sets exitCode=0 (accepted)', async () => {
+  it('send --verify: AFTER differs from BEFORE with real response content sets exitCode=0 (accepted)', async () => {
     const originalExitCode = process.exitCode;
 
     const home = await mkdtemp(join(tmpdir(), 'mosaic-fleet-'));
@@ -1339,11 +1485,17 @@ describe('agent send --verify', () => {
       ),
     );
 
+    let captureCallCount = 0;
     const runner: CommandRunner = async (command, args) => {
       const full = [command, ...args].join(' ');
       if (full.includes('agent-send.sh')) return { stdout: '', stderr: '', exitCode: 0 };
-      // capture-pane returns real response content
+      captureCallCount++;
-      return { stdout: 'Agent response: task completed.\n', stderr: '', exitCode: 0 };
+      // BEFORE: old content; AFTER: new response content (pane changed)
+      const stdout =
+        captureCallCount === 1
+          ? 'Old pane content\n'
+          : 'Old pane content\nAgent response: task completed.\n';
+      return { stdout, stderr: '', exitCode: 0 };
     };
 
     const program = new Command();

packages/mosaic/src/commands/fleet.ts

@@ -456,8 +456,72 @@ export function getDefaultTenantAndHost(): { tenant_id: string; host: string } {
 }
 
 /**
- * Builds the `agent watch` command: read-only tmux attach.
+ * Builds the command to create a grouped viewer session targeting an agent session.
- * Uses `-r` flag to prevent keystrokes and `=<name>` exact-match session target.
+ * A grouped session shares the same windows as the target but gets INDEPENDENT sizing,
+ * so attaching the viewer never resizes the agent's window.
+ *
+ * The viewer session name is derived from the agent name and a unique suffix (typically
+ * the caller's PID) so multiple concurrent watchers don't collide.
+ *
+ * Usage sequence:
+ *   1. Run buildAgentWatchCreateViewerCommand → create grouped session (via capturing runner).
+ *   2. Run buildAgentWatchAttachCommand → attach -r to the viewer session (via interactiveRunner).
+ *   3. Run buildAgentWatchKillViewerCommand → kill the viewer session on detach (via capturing runner).
+ */
+export function buildAgentWatchCreateViewerCommand(
+  agentName: string,
+  viewerSessionName: string,
+  socketName = DEFAULT_SOCKET_NAME,
+): string[] {
+  return [
+    'tmux',
+    '-L',
+    socketName,
+    'new-session',
+    '-d',
+    '-t',
+    `=${agentName}`,
+    '-s',
+    viewerSessionName,
+  ];
+}
+
+/**
+ * Builds the interactive attach command for a viewer session (read-only).
+ * Must be run via interactiveRunner (stdio: 'inherit').
+ */
+export function buildAgentWatchAttachCommand(
+  viewerSessionName: string,
+  socketName = DEFAULT_SOCKET_NAME,
+): string[] {
+  return ['tmux', '-L', socketName, 'attach', '-r', '-t', viewerSessionName];
+}
+
+/**
+ * Builds the kill-session command to clean up a viewer session after detach.
+ * Keeps the agent session intact.
+ */
+export function buildAgentWatchKillViewerCommand(
+  viewerSessionName: string,
+  socketName = DEFAULT_SOCKET_NAME,
+): string[] {
+  return ['tmux', '-L', socketName, 'kill-session', '-t', viewerSessionName];
+}
+
+/**
+ * Returns a unique viewer session name for a given agent.
+ * Uses process.pid so concurrent watchers produce distinct names.
+ */
+export function buildViewerSessionName(agentName: string): string {
+  return `${agentName}-watch-${process.pid}`;
+}
+
+/**
+ * @deprecated Use buildAgentWatchCreateViewerCommand + buildAgentWatchAttachCommand +
+ * buildAgentWatchKillViewerCommand instead. This bare attach targets the agent session
+ * directly and can resize it when the viewer terminal is smaller than the agent's window.
+ *
+ * Kept for backward compatibility only.
  */
 export function buildAgentWatchCommand(
   agentName: string,
@@ -493,27 +557,73 @@ export function buildAgentVerifyAcceptedCommand(
  * Result of a send-verify check.
  * - 'accepted': positive evidence that the message was accepted (response content present).
  * - 'draft': last non-empty line matches the draft heuristic (unsubmitted input).
- * - 'unverifiable': blank/empty capture — full-screen TUIs (claude, codex, opencode, pi)
+ * - 'unverifiable': pane did not change after send (stale or blank) — we cannot determine
- *   render blank for capture-pane, so we cannot determine acceptance; fails closed per FR-5.
+ *   acceptance; fails closed per FR-5.
  */
 export type SendVerifyResult = 'accepted' | 'draft' | 'unverifiable';
 
 /**
- * Check whether a send was accepted (not left as draft).
+ * Classify the result of a send-verify check by comparing BEFORE and AFTER pane snapshots.
+ *
+ * This is the primary classifier for `send --verify`. It addresses the stale-pane
+ * false-success problem: if the pane content did not change after the send, the new
+ * message never registered in the TUI (wedged pane, send dropped, etc.).
+ *
+ * Classification logic:
+ *   'unverifiable' — AFTER is blank/empty OR AFTER == BEFORE (no pane change after send).
+ *   'draft'        — AFTER differs from BEFORE AND the last non-empty line of AFTER starts
+ *                    with the draft pattern ("> "); message was typed but not submitted.
+ *   'accepted'     — AFTER differs from BEFORE AND AFTER does not end in a draft line;
+ *                    positive evidence that the TUI accepted the message.
+ *
+ * NOTE on blank AFTER: Full-screen TUIs (claude, codex, opencode, pi) render blank for
+ * `tmux capture-pane`. A blank AFTER is indistinguishable from a wedged pane, so it
+ * is always classified 'unverifiable' (fail-closed).
+ *
+ * NOTE on definitive acceptance: Phase-2 can only observe the pane surface — there is no
+ * runtime acknowledgement (heartbeat-ack) at this phase. The pane-change check is the best
+ * signal available against an opaque TUI. Definitive acceptance ultimately requires a
+ * runtime acknowledgement (Phase-3 heartbeat-ack).
+ *
+ * Draft heuristic: a last non-empty line (after stripping ANSI escapes) that starts
+ * with "> " is treated as an unsubmitted input line. This pattern is specific to
+ * pi/claude TUIs; draft detection for codex/opencode TUIs is best-effort only.
+ *
+ * FR-5 requires `send --verify` to return non-zero when delivery cannot be verified.
+ *
+ * @param before  Pane snapshot captured BEFORE the send command.
+ * @param after   Pane snapshot captured AFTER the send command (after the delay).
+ */
+export function classifySendResult(before: string, after: string): SendVerifyResult {
+  const afterLines = after.split('\n').filter((l) => l.trim().length > 0);
+  // Blank/empty AFTER => full-screen TUI rendered blank, or pane is wedged => unverifiable.
+  if (afterLines.length === 0) return 'unverifiable';
+  // No change => message didn't register in the TUI (stale/wedged pane) => unverifiable.
+  if (after === before) return 'unverifiable';
+  // AFTER differs from BEFORE — check whether the pane is now showing a draft line.
+  const lastLine = afterLines[afterLines.length - 1]!;
+  const stripped = lastLine.replace(/\x1b\[[0-9;]*m/g, '').trim();
+  // Heuristic: if stripped last line starts with "> " — that's the common draft pattern
+  // in pi/claude TUIs for showing user input before submission.
+  // NOTE: this heuristic is pi/claude-specific; draft detection for codex/opencode
+  // TUIs is best-effort only and may miss other unsubmitted-input indicators.
+  if (/^>\s/.test(stripped)) return 'draft';
+  return 'accepted';
+}
+
+/**
+ * Check whether a send was accepted (not left as draft), using only the AFTER snapshot.
+ *
+ * @deprecated Prefer classifySendResult(before, after) which guards against stale-pane
+ * false-successes. This single-snapshot variant cannot detect a wedged pane that still
+ * shows old non-empty content — it will incorrectly return 'accepted' in that case.
+ *
+ * Retained for unit-test compatibility with single-snapshot assertions.
  *
  * Returns:
  *   'unverifiable' — blank/empty capture (full-screen TUIs render blank; we cannot tell).
  *   'draft'        — last non-empty line matches the draft heuristic.
- *   'accepted'     — positive evidence of response content; not blank and not draft.
+ *   'accepted'     — non-blank and not a draft line (but may be stale — see above).
- *
- * Draft heuristic: a last non-empty line (after stripping ANSI escapes) that starts
- * with "> " is treated as an unsubmitted input line. This pattern is specific to
- * pi/claude TUIs and may miss drafts in codex/opencode TUIs — draft detection for
- * those runtimes is best-effort only.
- *
- * FR-5 requires `send --verify` to return non-zero when delivery cannot be verified.
- * Blank capture (full-screen TUI case) is the known-unverifiable case; it is treated
- * as FAILURE (not success) so the caller fails closed rather than silently succeeding.
  */
 export function isSendAccepted(capturedOutput: string): SendVerifyResult {
   const lines = capturedOutput.split('\n').filter((l) => l.trim().length > 0);
@@ -834,22 +944,37 @@ export function registerFleetAgentCommands(
           resolveMosaicHomeFromCommand(agentCommand, deps.mosaicHome),
         );
         const sourceLabel = opts.sourceLabel ?? opts.source ?? getDefaultOperatorSourceLabel();
+        if (opts.verify) {
+          // Capture BEFORE snapshot so we can detect stale-pane false-successes.
+          // A wedged pane that still shows old non-empty content must not be reported
+          // as 'accepted' — we compare BEFORE vs AFTER to guard against that case.
+          const beforeResult = await runner(
+            ...splitCommand(buildAgentVerifyAcceptedCommand(agent, roster.tmux.socketName)),
+          );
+          if (beforeResult.exitCode !== 0) {
+            throw new Error(
+              `send --verify: could not capture pane output before send (tmux exited ${beforeResult.exitCode}).`,
+            );
+          }
+          const beforeSnapshot = beforeResult.stdout;
+
           await runChecked(
             runner,
             buildAgentSendCommand(paths, agent, opts.message, roster.tmux.socketName, sourceLabel),
           );
-        if (opts.verify) {
+
-          // Brief pause to allow the TUI to process the send before capturing
+          // Brief pause to allow the TUI to process the send before capturing the AFTER snapshot.
           await new Promise<void>((res) => setTimeout(res, 300));
-          const captureResult = await runner(
+          const afterResult = await runner(
             ...splitCommand(buildAgentVerifyAcceptedCommand(agent, roster.tmux.socketName)),
           );
-          if (captureResult.exitCode !== 0) {
+          if (afterResult.exitCode !== 0) {
             throw new Error(
-              `send --verify: could not capture pane output to verify acceptance (tmux exited ${captureResult.exitCode}).`,
+              `send --verify: could not capture pane output to verify acceptance (tmux exited ${afterResult.exitCode}).`,
             );
           }
-          const verifyResult = isSendAccepted(captureResult.stdout);
+          // Classify using BEFORE/AFTER comparison to guard against stale-pane false-successes.
+          const verifyResult = classifySendResult(beforeSnapshot, afterResult.stdout);
           if (verifyResult === 'draft') {
             process.exitCode = 1;
             process.stderr.write(
@@ -858,9 +983,14 @@ export function registerFleetAgentCommands(
           } else if (verifyResult === 'unverifiable') {
             process.exitCode = 1;
             process.stderr.write(
-              `send --verify: could not verify delivery (blank/no response captured) for agent "${agent}".\n`,
+              `send --verify: could not verify delivery (no pane change after send) for agent "${agent}".\n`,
             );
           }
+        } else {
+          await runChecked(
+            runner,
+            buildAgentSendCommand(paths, agent, opts.message, roster.tmux.socketName, sourceLabel),
+          );
         }
       },
     );
@@ -871,11 +1001,31 @@ export function registerFleetAgentCommands(
     .action(async (agent: string) => {
       const roster = await loadRosterFromAgentCommand(agentCommand, deps.mosaicHome);
       getRosterAgent(roster, agent);
-      // `tmux attach` is interactive and requires inherited TTY/stdin/stdout.
+
-      // Route through the interactiveRunner (stdio: 'inherit') instead of the
+      // Use a GROUPED VIEWER SESSION to prevent the observer from resizing the agent's
-      // capturing runner, which would hang or fail for full-screen TUI commands.
+      // window. A bare `tmux attach -r` against the agent session itself still lets the
-      const [bin, args] = splitCommand(buildAgentWatchCommand(agent, roster.tmux.socketName));
+      // client shrink the session to its terminal size; a grouped session gets INDEPENDENT
+      // sizing so the agent's window is never affected by the viewer's terminal dimensions.
+      //
+      // Sequence:
+      //   1. Create a throwaway grouped session targeting the agent (capturing runner).
+      //   2. Attach -r (read-only) to the viewer session (interactiveRunner / TTY).
+      //   3. Kill the viewer session on detach so stale sessions don't accumulate.
+      const viewerName = buildViewerSessionName(agent);
+      const socketName = roster.tmux.socketName;
+
+      await runChecked(runner, buildAgentWatchCreateViewerCommand(agent, viewerName, socketName));
+
+      const [bin, args] = splitCommand(buildAgentWatchAttachCommand(viewerName, socketName));
       const exitCode = await iRunner(bin, args);
+
+      // Best-effort cleanup of the viewer session regardless of how the user detached.
+      // Errors here are intentionally suppressed — the agent session is unaffected.
+      const killResult = await runner(
+        ...splitCommand(buildAgentWatchKillViewerCommand(viewerName, socketName)),
+      );
+      void killResult; // result is intentionally ignored
+
       if (exitCode !== 0) {
         process.exitCode = exitCode;
       }