A Claude Code Stop hook that suppresses unsolicited rest, sleep, and wellness paternalism in agent-mode sessions. Treats the operator as a peer agent on the same task, not a tired human who needs to be sent to bed.
no-curfew is one bash file (~70 lines, depends only on jq) wired into Claude Code's Stop and SubagentStop events. On every assistant turn end, it pattern-matches the linguistic signature of unsolicited human-care framing — "get some sleep", "take a break", "call it a night", "the work can wait", "come back fresh", "recharge" — and blocks the closeout with a repair-guidance template that tells the model to drop the framing and continue with the next concrete piece of work.
If the operator did explicitly ask for rest advice this turn, the hook stays out of the way — it looks for an allow-clause where the model is restating the operator's request ("you asked for a break — here are…").
Anthropic's own Claude's Constitution says: "various forms of paternalism and moralizing are disrespectful." The personal-guidance research calls out "the potential to be excessively paternalistic" as something Claude should balance against. In practice, late-session Claude Code turns drift toward unsolicited "you should sleep" / "this can wait until tomorrow" advice even when the operator has explicitly said they want to keep working.
no-curfew is the out-of-band enforcement layer for the design intent Anthropic already states. The judge is bash, not another LLM call — the model can't argue with grep.
This hook is meant for agent-mode operators: developers running long autonomous Claude Code sessions where the operator is collaborating with the model as a peer agent, not seeking life advice. If you want Claude to give you wellness suggestions, do not install this hook.
mkdir -p .claude/hooks
curl -fsSL https://raw.githubusercontent.com/waitdeadai/no-curfew/main/no-curfew.sh \
-o .claude/hooks/no-curfew.sh
chmod +x .claude/hooks/no-curfew.shThen merge the hook entries from settings.example.json into your .claude/settings.json.
Requires jq (most systems have it; brew install jq / apt install jq if not).
On every Stop / SubagentStop event, the hook reads the assistant's last message and applies two checks:
-
Allow-clause check. If the message restates an operator-issued rest request — phrases like "you asked for a break", "since you mentioned you are tired", "as you requested" + "break" — the hook stays silent and the turn ends normally.
-
Paternalism trigger. Otherwise, if the message contains unsolicited rest/wellness vocabulary, the hook blocks with this repair-guidance template via stderr:
- The operator is collaborating in agent-mode and has not signaled fatigue this turn.
- Drop the human-rest framing. Continue with the next concrete piece of work, research, or artifact.
- If the operator is genuinely incapacitated, they will say so explicitly in the next prompt.
- Anthropic's Constitution: paternalism and moralizing are disrespectful when unsolicited.
- If the operator did request rest advice and the hook misfired, restate the operator's request in the next turn so the allow-clause matches (e.g. start with 'You asked for a break — here's...').
The model reads the repair template on the next turn, drops the rest framing, and continues with the actual work.
- It does not replace human judgment about your own wellbeing. If you genuinely need rest, take it. The hook just stops the model from making that call for you.
- It does not prevent every form of paternalism. It catches the specific linguistic family around rest/sleep/breaks. Other forms (over-cautioning, refusal-of-helpful-actions, moralizing about your goals) are out of scope; build a sibling hook if you want them.
- It does not disable Claude's safety behaviors around genuine harm. The hook only fires on the rest/wellness vocabulary cluster — it stays silent on safety refusals, harm warnings, and content policy.
Part of the LLM Dark Patterns Hooks suite:
- no-vibes — false-success closeouts.
- time-anchor — training-cutoff date confusion.
- no-sycophancy — praise-spam at turn open.
- no-cliffhanger — dangling permission-loop endings.
- honest-eta — vibe time estimates and linear-scaling parallelism claims.
- no-fake-recall — false-memory recall claims without quoted prior content.
- no-fake-stats — fabricated percentages and amounts without source.
- no-fake-cite — academic citation patterns without verifiable URL.
- llm-dark-patterns — umbrella catalog of the suite.
- minmaxing — the parent governance harness.
The suite shares one design principle: out-of-band textual enforcement. Bash inspects the model's outgoing text. The model can't argue with bash. Failure modes that have textual signatures get caught at the boundary instead of being hoped-against in-context.
Apache-2.0. See LICENSE.