Add Stop hook to pause once on 断り書き phrases

[ikuwow]

Purpose

AIRULES.md (応答の姿勢と判断 section) already forbids 「正直」「実は」「本当のところ」 as 断り書き — preamble that softens or qualifies an opinion. The rule has existed for some time, but the model continues to slip into the habit — including once in the immediately preceding session, which prompted this countermeasure. A retrospective surfaced that "follow the rule more carefully" is not a structural fix, so this PR adds a Stop hook that intercepts the slip mechanically.

Design

Reconsideration trigger, not prohibition

The hook does not try to disambiguate legitimate uses of the literal strings (e.g., 名詞 like 「正直者」, factual disclosure phrasing). A simple regex (正直|実は|本当のところ) can't distinguish 断り書き from non-断り書き, and trying to encode that distinction in the hook would either over-block legitimate text or leak escape hatches the model can lean on. Instead, the hook treats any string match as a "stop and reconsider" trigger: it blocks the first pass with a reason explaining why this class of preamble damages trust, and on the second pass it always allows the stop via stop_hook_active.

The reason text deliberately omits hook-mechanic excuses ("the second pass will go through" / "name uses are exempt"). It explains the structural harm and asks the model to judge its own usage.

stop_hook_active and infinite-loop prevention

The hook reads stop_hook_active from stdin JSON and exits 0 immediately when it is true. This is the documented "Stop hook runs forever" troubleshooting pattern at https://code.claude.com/docs/en/hooks-guide.md . Without this guard, a model that genuinely needs to use the string (a quote, a code example referencing the rule, etc.) would loop forever.

Pure functions + doctest

is_forbidden(text) and collect_current_turn_assistant_text(events) are pure and verified by doctest. The repository's python-doctest GitHub Actions check runs these on every push, so the tests are CI-verified, not just local:

$ python3 -m doctest claude/hooks/block_excuse_phrases.py -v 2>&1 | tail -3
16 tests in 4 items.
16 passed.
Test passed.

The main() entrypoint handles I/O (stdin JSON, transcript JSONL read, stdout JSON write, stop_hook_active branch) and is verified at runtime when the hook fires in a new session.

Fail-open with stderr breadcrumbs

Each I/O fall-through in main() (stdin parse, missing transcript_path, transcript read) emits a short stderr line prefixed with block_excuse_phrases: before exiting 0. The hook stays fail-open from Claude Code's perspective — a broken hook never wedges the assistant — but ~/.claude/logs/ retains a trace, so the operator can distinguish "rule not violated" from "hook silently broken".

Files

The PR touches two files:

• New: claude/hooks/block_excuse_phrases.py — the hook script itself.
• Modified: claude/settings.json — register the script as a hooks.Stop command alongside the existing noti notification (4-line array insertion under hooks.Stop[0].hooks). The hook entry is placed before noti so the block decision is computed before the desktop notification fires.

Verification

• doctest: 16 tests pass locally and in CI (python-doctest job)
• pre-commit: trailing whitespace / EOF / JSON syntax checks pass on commit
• CI all green on this branch (Socket / bootstrap / python-doctest / shellcheck)
• Runtime trigger: observed in a new session after the hook is installed locally via scripts/deploy.sh (claude/ is symlinked to ~/.claude/, so changes take effect immediately on next session). No CI path exercises Stop hooks; the integration with Claude Code is not unit-testable from CI.

Out of scope

• Other AIRULES rules (no Markdown emphasis, etc.) are not enforced by hooks in this PR; they can be added incrementally if the same slip pattern emerges.
• The pattern list mirrors the AIRULES 断り書き ban exactly; if that rule list expands, the regex needs the same edit.