KeiSeiKit-1.0/hooks/chat-numeric-prewarn.sh
Parfii-bot 4afc85ca30 fix(hooks): post-audit hook chain hardening + 4 new defensive hooks
Hook chain repairs (Group A):
- alignment-check.sh: read .prompt (was .user_prompt) — hook was dead
- block-dangerous.sh: jq instead of inline interpreter (RULE 0.2 + fail-open fix)
- destructive-guard.sh: explicit INPUT=cat + jq guard + exit 0 — was silent no-op
- numeric-claims-guard.sh: exit 1 -> exit 2 (Claude Code spec — was non-blocking)
                          comments updated 0.17 -> 0.18 (env var name kept)
- no-downgrade.sh: removed (?i) PCRE syntax — POSIX ERE matched literal text
- task-timer.sh: jq -nc instead of bare printf — JSON injection on quotes/backslashes
                 in description was corrupting RULE 0.18 evidence journal
- check-error-patterns.sh: replaced with no-op stub — had hardcoded /Users/denis/...
                            PATH LEAK in public kit, plus inline interpreter use
- post-commit-audit.sh: added trailing exit 0 — grep return code was hook exit code
- citation-verify.sh: ALLOW_REGEX accepts HOOK-BYPASS marker — bypass was documented
                       but never matched
- settings-snippet.json: agent-stub-scan moved PreToolUse:Agent -> PostToolUse:Agent
                          (RULE 0.16 enforcement was firing before transcript existed)
- check-error-patterns hook removed from settings-snippet.json

New defensive hooks (Group H):
- no-github-push.sh: PreToolUse:Bash hard deny on github.com push/create/sync/remote-add
                      (RULE 0.1 — patent IP protection; was missing from public kit)
- secrets-pre-guard.sh: PreToolUse:Edit|Write — token-pattern scan with allowlist (RULE 0.8)
- chat-numeric-prewarn.sh: UserPromptSubmit reminder when prompt mentions time/cost
                            (RULE 0.18 chat extension)
- chat-numeric-postflag.sh: Stop event scans last assistant message for naked numerics
                             without REAL/FROM-JOURNAL/ESTIMATE-HTC markers

Source: full Sonnet test-retest audit 2026-05-02 (3 parallel waves of 6 agents each)
identified hook chain bugs as HIGH severity in all 3 runs independently.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-02 21:38:47 +08:00

48 lines
2 KiB
Bash
Executable file

#!/bin/sh
# chat-numeric-prewarn.sh — UserPromptSubmit remind (RULE 0.18 chat-output)
#
# Detects time/cost/effort keywords in the user's prompt and injects an
# additionalContext reminder asking the assistant to attach RULE 0.18
# evidence markers before emitting any numeric claim in its response.
#
# Severity: remind — always exits 0, never blocks.
#
# Bypass: set RULE_018_CHAT_BYPASS=1 in the calling environment.
set -u
if [ "${RULE_018_CHAT_BYPASS:-0}" = "1" ]; then
exit 0
fi
if ! command -v jq > /dev/null 2>&1; then
exit 0
fi
INPUT=$(cat)
PROMPT=$(printf '%s' "$INPUT" | jq -r '.prompt // empty' 2>/dev/null)
[ -z "$PROMPT" ] && exit 0
PROMPT_LC=$(printf '%s' "$PROMPT" | tr '[:upper:]' '[:lower:]')
# Keywords that imply the user is asking for a time/cost/effort estimate
MATCH=0
if printf '%s' "$PROMPT_LC" | grep -qE \
'сколько|как долго|estimate|how long|how much|duration|time|effort|займёт|сколько стоит|cost|стоимость|за сколько|за (сколько|это)'; then
MATCH=1
fi
[ "$MATCH" -eq 0 ] && exit 0
# Emit additionalContext JSON to stdout (Claude Code hook protocol)
cat <<'EOF'
{
"hookSpecificOutput": {
"hookEventName": "UserPromptSubmit",
"additionalContext": "<rule-018-chat-prewarn>\nRULE 0.18 REMINDER — user prompt contains time/cost/effort keywords.\n\nBefore emitting ANY duration, count, cost, size, or percentage claim in your response, attach one of these evidence markers inline:\n\n [REAL: <source — file:line, commit SHA, or timestamp>]\n [FROM-JOURNAL: ~/.claude/memory/time-metrics/<file>.jsonl#<id>]\n [ESTIMATE-HTC: <one sentence: why this cannot be measured precisely>]\n\nNaked numbers are forbidden by RULE 0.18 (lock 2026-04-29).\nIf you do not have a journal entry for the task, use [ESTIMATE-HTC:] and state the reason.\nDo NOT fabricate a number from latent space — refusal to estimate is preferred over a false estimate.\n\nSee: ~/.claude/rules/chat-numeric-pre-output.md\n</rule-018-chat-prewarn>"
}
}
EOF
exit 0