Agent exits success after 5-10 turns without completing the review (silent no-op runs) — ~50% of runs since ~v1.0.190

## Summary

Since 2026-08-11 we see runs where the action exits **success** but the agent silently dies mid-task: the execution envelope reports `"subtype": "success", "is_error": false` with an anomalously low `num_turns` (5-10), the sticky comment is left frozen partway through its own todo list (`- [ ] Run code review…`), zero findings/inline comments are posted, and the run bills in full. Roughly **50% of attempts** are affected. Healthy runs of the same workload complete in 6-15 min with 30+ turns.

## Setup

- `anthropics/claude-code-action@v1` (floating tag), tag mode on `pull_request`
- plugin `code-review@claude-code-plugins` (marketplace `https://github.com/anthropics/claude-code.git`), prompt `/code-review --comment`, `use_sticky_comment: true`, `track_progress: true`, OAuth token auth
- Also reproduces in **`issue_comment` (@claude) mode** on a workflow file untouched since 2026-06-19 with near-default config — so it is not specific to the review plugin or our workflow config.

## Evidence (17 attempts observed live, 2026-08-11 → 2026-08-13)

| date | run / attempt | outcome | envelope |
|---|---|---|---|
| 08-11 | 31434572614 | silent | success, 6 turns, $8.84 |
| 08-11 | 31505493057 a1 | silent | success, 5 turns, $5.84 |
| 08-11 | 31507765842 a1 | silent | success, 10 turns, $9.69 |
| 08-11 | 31507765842 a2 (rerun) | silent | success, 7 turns, $6.15 |
| 08-12 | 31654155016 a1+a2 | silent ×2 (same commit) | success, low turns |
| 08-13 | 31704228111 a1+a2 | silent ×2 (same commit) | success, low turns |
| 08-13 | 31705575564 (@claude mode) | silent | frozen at 2nd todo item |
| various | 7 attempts | completed normally | 15-38 turns, real findings |

Same commit can produce both a silent run and a complete review across attempts — nondeterministic, not diff-dependent. Two commits went silent twice consecutively, so retry is not a reliable remedy.

## Version correlation

- Reviews were consistently reliable through **v1.0.183** (heavy observed load 2026-07-27: four overlapping reviews, all completed).
- The silent cluster began the evening of 2026-08-11 — **v1.0.191 published 16 minutes before** the worst burst; one earlier same-day silent run puts **v1.0.190** in scope too. We did not observe v1.0.185-189.
- After pinning back to the v1.0.183 SHA we will follow up with whether the rate drops (will update this issue).

## Impact

A silent run is indistinguishable from "reviewed, found nothing" on the PR page — teams merge unreviewed code behind a green check, at full billing. We now run a post-step that fails the job red when the run's sticky comment still contains an unchecked todo, which is how these were caught.

Repo is private; happy to share full job logs / envelopes with maintainers on request.