Summary

When an unleashed session ends (Claude Code exits, user presses Ctrl+C, or the PTY dies), the wrapper just... stops. There's no summary of what happened during the session. The user is dropped back to their shell prompt with no indication of:

• How long the session ran
• How many permissions were auto-approved
• How many sentinel checks occurred (and their verdicts)
• Whether any commands were blocked
• How many friction prompts were logged
• The path to the session's mirror transcript and logs

Current Exit Behavior

[session runs for 45 minutes]
[Claude Code exits]
$                    ← user is back at shell, no context

Proposed Exit Behavior

[session runs for 45 minutes]
[Claude Code exits]

── unleashed session summary ──────────────────
  Duration:        45m 12s
  Approvals:       47
  Sentinel:        12 local-allow, 3 api-allow, 1 blocked
  Mirror:          logs/mirror-2026-02-14_14-30.txt (2.1 KB)
  Friction log:    logs/friction-2026-02-14_14-30.jsonl
───────────────────────────────────────────────
$

Implementation

The data is already available:

• Duration: time.time() - session_start (session_start already tracked)
• Approval count: increment a counter in do_approval()
• Sentinel stats: self.sentinel_gate.stats dict (already has local_allow, local_block, api_allow, api_block, api_error)
• Mirror path: self.mirror_path (already stored)
• Friction log: self.friction_logger.log_path if friction logger is active

Print the summary in the finally block of run(), right before cleanup:

finally:
    # Print session summary
    elapsed = time.time() - session_start
    mins, secs = divmod(int(elapsed), 60)
    print(f"\n── unleashed session summary {'─' * 40}", file=sys.stderr)
    print(f"  Duration:     {mins}m {secs}s", file=sys.stderr)
    print(f"  Approvals:    {self.approval_count}", file=sys.stderr)
    if self.sentinel_gate:
        s = self.sentinel_gate.stats
        print(f"  Sentinel:     {s['local_allow']} local-allow, {s['api_allow']} api-allow, "
              f"{s['local_block'] + s['api_block']} blocked, {s['api_error']} errors", file=sys.stderr)
    if self.mirror_path:
        size = os.path.getsize(self.mirror_path) if os.path.exists(self.mirror_path) else 0
        print(f"  Mirror:       {self.mirror_path} ({size // 1024:.1f} KB)", file=sys.stderr)
    print(f"{'─' * 55}", file=sys.stderr)
    # ... existing cleanup ...

Why This Matters

Unleashed sessions are autonomous by design — the user steps away and comes back to results. A session summary is the bridge between "I left it running" and "here's what happened." Without it, the user must:

1. Find the mirror transcript file
2. Open it and scroll through
3. Mentally reconstruct the session

A 5-line summary on exit gives instant context. It's especially valuable for sentinel adoption: seeing "Sentinel: 47 local-allow, 0 blocked" builds confidence that sentinel isn't interfering with productivity.

Affected Files

• src/unleashed-c-21.py — run() method's finally block

Context

This is the lowest-effort, highest-UX-value improvement on this list. All the data already exists in memory — it just needs to be printed. The sentinel stats integration (issue #12 Phase 3) already accumulates the numbers. This issue is about surfacing them to the user at the one moment they're most valuable: when the session ends and they're deciding "did that go well?"