The entrypoint field in a Claude Code transcript tells a Stop hook if anyone is watching

Exit 2 is the only enforcement lever a Stop hook has. In a terminal it stops the turn and puts the warning where the operator sees it ^{[2] [3]}. Under launchd, through the Agent SDK CLI, the same non-zero status becomes the exit status of the run ^{[4] [5]}. Nineteen times that night the guardrail did exactly what it was written to do, logged to a file no one read ^[6].

Nineteen nags from two launches averages 9.5 per run ^[18]. A cap of two nags per session had been added the previous day, after a performance audit flagged that a per-turn nag turns the audit into ritual ^[14]. Two sessions under that cap allow four ^[17]. The counter is a file in /tmp named for the session id, and the hook exits 0 once the count reaches two ^[15]. The script comment describes sdk-cli runs as single-shot with no next turn ^[12]. A run that opens a new session per invocation gets a new counter file and a fresh pair of strikes every time, which is one explanation the source does not confirm.

The repair reads provenance instead of lowering the bar. The hook already receives transcript_path on stdin, so it can open the file Claude Code wrote for the session that just ended ^[10]. The first 15 lines carry an entrypoint field ^[11]. A human-started session reports cli; an Agent SDK run reports sdk-cli ^[9]. Interactive keeps exit 2; headless gets a silent skip ^[12]. That is the right shape, because the alternative repair is to stop blocking, and a self-audit check that does not block is a log line in the one context where it currently works ^{[2] [5]}.

For this to transfer, one thing has to be true of your setup: the test reads launch mode, not attendance. The entrypoint value records how the session was started ^[9]. A scheduler that drives the interactive client rather than the SDK CLI still reports cli, and the nag still blocks against an empty chair. Attendance detection by entrypoint is accurate exactly as far as your automation's launch path differs from your own.

One part of the case remains an assertion. The author writes that a Stop hook returning exit 2 inside a CI-like pipeline changes how downstream steps are handled, and that mixing human-facing nags into unattended runs degraded execution quality rather than merely adding noise ^[8]. No downstream step is named and no failed image generation is shown. What the logs support is 19 nags and a path that exits non-zero without a reader; the pipeline damage is the author's read of his own system.

The generalizable claim in the piece is that a single Stop primitive covering both humans and machines makes this failure mode sharp, and that the same leak shows up wherever human-facing notification logic sits in an unattended execution path ^[19]. Stop is one of three documented hook events, alongside PreToolUse before a tool call and PostToolUse after ^[16]. Stop is the one that fires when the model finishes speaking, which is precisely when a hook author is tempted to say something to a person.