
What happened
A developer installed a reviewer subagent that writes a sha256-tagged PASS/FAIL verdict to .review/verdict.json and a Stop hook that checks it, then logged ten places the setup stalled on Claude Code 2.1.325 — including a blocking exit-2 hook that let a turn end.
Why it matters
The stalls show a verification setup can silently fail to hold the AI to its review gate, so anyone relying on such a hook may think changes were checked when the turn simply ended.
Summaries like this, every morning.
The setup under test has two moving parts: a reviewer subagent that writes a verdict with the current diff's sha256 into .review/verdict.json, and a Stop hook that only checks whether a verdict exists, matches the current diff, and says PASS. The author's runs on Claude Code 2.1.325 (with one run on 2.1.326) show that several of the failure points sit in how Claude Code treats hooks and subagents rather than in the verdict logic itself.
The documented exceptions mattered more than the headline behavior. A Stop hook that exits 2 is treated as non-blocking when stdout is empty and stderr names a missing file, and onFailure: "block" has no effect on Stop hooks — both documented in the hooks reference, and both observed in runs D and E. Separately, subagents start in the background by default, so in the author's runs the AI often closed its turn before a verdict existed, and the hook blocked again; in R6 the AI called the reviewer in the background three times even after receiving a reason asking it to wait.
The author's fixes were partial. Moving the verdict to .review/verdict.json avoided Claude Code's protection on .claude/, and adding an explicit instruction not to use the caller's supplementary notes stopped the reviewer from passing out-of-scope README changes. But the author notes each run was done once, with personal global hooks and CLAUDE.md active, and a clean environment was not tested — so how much of the AI's waiting and reverting behavior generalizes is left open.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
AIworker analyzed recorded consultations and found ChatGPT's share of new corporate AI adoption inquiries fell…

The piece sorts agent-driven video into video-generation models like Seedance, front-end code such as Three.js…

Running Claude Code sessions via TerminalHub on one 32GB Windows PC, the user logged 33 heavy jobs over three…

Levent Alpöge published a 4-page remark.pdf saying he believes he found a Riemann Hypothesis proof with Claude…

On 2026年11月1日, EveryInfra tested Claude Opus 5.5, Sonnet 5.5, and Fable 5.1 and found forced tool calls return…

A Zenn explainer lays out ReAct, the 2022 Google Research pattern in which an LLM alternates Reasoning and Act…
