[BUG] ultrareview: CLI declares cloud review failed at 30 min while it is still running; free attempt consumed, findings never delivered to CLI

## Summary

`/code-review ultra` (ultrareview) launched from the CLI was declared **failed** by the client after 30 minutes ("cloud session exceeded 30 minutes"), and the CLI suggested retrying — while the cloud session was in fact **still running and progressing normally** (visible at claude.ai/code), eventually producing verified findings. The free ultrareview attempt was consumed, and the findings were never delivered back to the CLI session.

## Environment

- Claude Code **2.1.238**, macOS (darwin 25.5.0), interactive terminal session
- Model: claude-fable-5
- Review target: local branch → main, 23 files changed, 7385 insertions (large diff)

## Timeline

1. Launched ultrareview from the CLI: `Free ultrareview 1 of 3. Ultrareview launched for → main (~5-10 min, runs in the cloud).`
2. ~30+ minutes later the CLI received a task notification: `status: failed`, `Cloud review failed: cloud session exceeded 30 minutes`, with guidance "Cloud review did not produce output … retry /code-review ultra, or use plain /code-review".
3. Re-running `/code-review ultra` showed the launch dialog with **"Free ultrareview 2 of 3"** — i.e. the failed attempt was counted (cancelled at this point, so only one attempt was actually spent).
4. Opening the cloud session in the browser (claude.ai/code) showed the review **still actively running**: Setup ✓, Find ✓ (15 candidates), Verify ✓ (12 confirmed · 3 refuted), Dedupe in progress, with a live "Stop review" button. So the session had not failed at all — it just ran longer than the client-side 30-minute cap.
5. The task record was gone from the CLI (`TaskOutput` → "No task found"), so the findings could not be retrieved from the terminal session even after the cloud run finished.

## Expected behavior

- A cloud review that is still running should not be reported to the user as failed with "did not produce output"; the client-side cap should either wait longer, detach-and-poll, or clearly say "still running in the cloud, check the session link".
- If the harness gives up tracking, the findings should still be deliverable back to the originating CLI session once the cloud run completes (or at minimum the notification should point to the live session instead of advising a retry).
- Advising the user to retry while the original run is still executing invites double-spending the limited free attempts.
- A run that is timed out by the client (not actually failed in the cloud) arguably shouldn't consume a free-tier attempt — same concern as raised in #49905 and #54812 (both closed as stale).

## Actual behavior

- CLI reported failure and suggested a retry; free attempt consumed; findings only visible in the web session, unreachable from the CLI.

## Debug references (for Anthropic)

- Cloud session: `session_01C9j2e56sCzQuuwLv2QTLHq`
- Remote task id: `rimsx3dfb`
- Date/time: 2026-08-21, ~19:45–20:50 UTC

🤖 Filed with [Claude Code](https://claude.com/claude-code) on the user's behalf.

Start with the `/code-review ultra` entry point and the `TaskOutput` handling described in the report. Reproduce a large-diff review and compare the CLI's 30-minute timeout with the live cloud session status. Done means an in-progress cloud review is not reported as failed, and its results or live session link remain available without encouraging a duplicate retry.

Do mô hình lập chỉ mục viết ra từ nội dung của issue.