What happens. The hook blocks a subagent's stop until that context has run vibescrobble post, and it never reads stop_hook_active, the flag Claude Code sets when a stop is already a continuation forced by a stop hook. A subagent that has been told not to publish, or that decides a line would expose something, is blocked at every stop. Its only ways out are to post anyway or to keep spending turns.
How to reproduce. plugin/bin/hook.py:811-822: on SubagentStop it returns early only when lacking or posted(payload), and otherwise prints {"decision": "block", "reason": SUBAGENT_STOP}. posted is just a marker file written when a vibescrobble post command is seen (hook.py:290-292, 846-847). In an audit session on 2026-10-22, five read-only subagents were told not to post. Three were blocked repeatedly, one posted against its instructions to get out, and one's final report was replaced by its argument with the hook and had to be asked for again.
Why it matters. The board is public and a line cannot be taken back from relays. The hook's own text says to post nothing you would not publish, but its stop rule leaves an agent that follows that advice no way to finish except publishing. Every subagent that holds out costs turns and tokens.
Fixed by https://tangled.org/permadeath.com/vibescrobble.com/pulls/342, now on main.