@takuma-hirai/hirai-method 0.1.0

package/.claude/hooks/gateguard.sh

@@ -0,0 +1,209 @@
+#!/usr/bin/env bash
+# gateguard.sh — Fact-forcing PreToolUse gate (ECC GateGuard)
+#
+# Args: $1 = tool name (Edit|Write|Bash) used as fallback if JSON tool_name absent
+# Stdin: PreToolUse JSON
+#
+# Bypass:
+#   ECC_GATEGUARD=off          # disable globally
+#   /gate-clear                # reset state
+#   /gate-bypass <file>        # pre-clear one file
+#
+# Config keys consumed (.claude/harness-config.yml):
+#   gateguard_state_dir        # state file root (cleared markers)
+#   agent_marker_dir           # for subagent passthrough detection
+
+set -u
+
+# config 読み込み (HC_* 変数 export)
+# shellcheck source=lib/config-loader.sh
+source "$(dirname "$0")/lib/config-loader.sh"
+
+# Feature toggle 参照 (task-45 Phase 2)
+if command -v is_feature_enabled >/dev/null 2>&1 && ! is_feature_enabled gateguard; then
+  echo '{}'
+  exit 0   # feature OFF で no-op
+fi
+
+# Phase β: F1 disable for SWE-bench grid evaluation (fail-open)
+# ECC_F{1,2,3}_OFF env vars allow runner.py to bypass gates per-task
+# while measuring gate effectiveness. Production usage MUST NOT set these.
+if [ "${ECC_F1_OFF:-0}" = "1" ] || [ "${ECC_F1_OFF:-}" = "true" ]; then
+  if command -v jq >/dev/null 2>&1; then
+    jq -n '{decision:"approve", reason:"F1 (gateguard) disabled via ECC_F1_OFF"}'
+  else
+    printf '{"decision":"approve","reason":"F1 (gateguard) disabled via ECC_F1_OFF"}\n'
+  fi
+  exit 0
+fi
+
+# Bypass
+if [ "${ECC_GATEGUARD:-}" = "off" ]; then
+  echo '{}'
+  exit 0
+fi
+
+input=$(cat)
+
+# Need jq
+if ! command -v jq >/dev/null 2>&1; then
+  echo '{}'
+  exit 0
+fi
+
+# Detect tool from JSON, fallback to $1
+tool=$(printf '%s' "$input" | jq -r '.tool_name // empty' 2>/dev/null)
+if [ -z "$tool" ] || [ "$tool" = "null" ]; then
+  tool="${1:-}"
+fi
+
+# Subagent passthrough (same multi-stage detection as delegation-guard)
+is_subagent="false"
+if [ "${CLAUDE_HARNESS_ROLE:-}" = "subagent" ]; then
+  is_subagent="true"
+fi
+if [ "$is_subagent" = "false" ]; then
+  for field in agent_type subagent_type parent_tool_use_id agent_id; do
+    v=$(printf '%s' "$input" | jq -r ".${field} // empty" 2>/dev/null)
+    if [ -n "$v" ] && [ "$v" != "null" ]; then
+      is_subagent="true"
+      break
+    fi
+  done
+fi
+if [ "$is_subagent" = "false" ] && [ -d "$HC_AGENT_MARKER_DIR" ]; then
+  if ls "$HC_AGENT_MARKER_DIR"/*.lock >/dev/null 2>&1; then
+    is_subagent="true"
+  fi
+fi
+if [ "$is_subagent" = "true" ]; then
+  echo '{}'
+  exit 0
+fi
+
+state_dir="$HC_GATEGUARD_STATE_DIR"
+mkdir -p "$state_dir" 2>/dev/null
+
+hash_str() {
+  printf '%s' "$1" | shasum -a 256 2>/dev/null | cut -c1-12
+}
+
+emit_block() {
+  jq -n --arg r "$1" '{decision:"block", reason:$r}'
+}
+
+# === Edit gate ===
+if [ "$tool" = "Edit" ]; then
+  file=$(printf '%s' "$input" | jq -r '.tool_input.file_path // .tool_input.filePath // empty')
+  if [ -z "$file" ]; then echo '{}'; exit 0; fi
+
+  # Excluded paths
+  case "$file" in
+    */CLAUDE.md|*/README.md|*/.gitignore|*/.dockerignore|*/.env.example|*/.gitmessage|*/.claude/.agent-markers/*|*/.claude/.gateguard-state/*|*.lock)
+      echo '{}'; exit 0;;
+  esac
+
+  fhash=$(hash_str "$file")
+  state="$state_dir/edit-$fhash.cleared"
+
+  if [ -f "$state" ]; then
+    echo '{}'; exit 0
+  fi
+
+  touch "$state" 2>/dev/null
+
+  msg="[GateGuard] First Edit to: $file"
+  msg="$msg"$'\n\n'"Before retrying, gather and present these facts:"
+  msg="$msg"$'\n\n'"  1. Run Grep to list ALL files importing/requiring this file"
+  msg="$msg"$'\n'"     -> which downstream code is affected by this edit?"
+  msg="$msg"$'\n'"  2. Identify public functions / classes / exports affected"
+  msg="$msg"$'\n'"     -> API surface that callers depend on"
+  msg="$msg"$'\n'"  3. Describe data structure (field names, types, date formats)"
+  msg="$msg"$'\n'"     -> values may be redacted"
+  msg="$msg"$'\n'"  4. Quote the user's verbatim instruction authorizing this change"
+  msg="$msg"$'\n\n'"After presenting these facts, retry the Edit — the gate is now cleared"
+  msg="$msg"$'\n'"for this file in this session ($state created)."
+  msg="$msg"$'\n\n'"Bypass: ECC_GATEGUARD=off  /  /gate-clear  /  /gate-bypass $file"
+
+  emit_block "$msg"
+  exit 0
+fi
+
+# === Write gate ===
+if [ "$tool" = "Write" ]; then
+  file=$(printf '%s' "$input" | jq -r '.tool_input.file_path // .tool_input.filePath // empty')
+  if [ -z "$file" ]; then echo '{}'; exit 0; fi
+
+  case "$file" in
+    */CLAUDE.md|*/README.md|*/.gitignore|*/.dockerignore|*/.env.example|*/.gitmessage|*/.claude/.agent-markers/*|*/.claude/.gateguard-state/*|*.lock)
+      echo '{}'; exit 0;;
+  esac
+
+  fhash=$(hash_str "$file")
+  state="$state_dir/write-$fhash.cleared"
+
+  if [ -f "$state" ]; then
+    echo '{}'; exit 0
+  fi
+
+  touch "$state" 2>/dev/null
+
+  msg="[GateGuard] First Write of new file: $file"
+  msg="$msg"$'\n\n'"Before retrying, gather and present these facts:"
+  msg="$msg"$'\n\n'"  1. Which files / lines will call or import this new file?"
+  msg="$msg"$'\n'"     -> if none yet, name the caller you'll add next"
+  msg="$msg"$'\n'"  2. Confirm via Glob that no existing file duplicates this purpose"
+  msg="$msg"$'\n'"  3. Describe data structure (if applicable)"
+  msg="$msg"$'\n'"  4. Quote the user's verbatim instruction authorizing creation"
+  msg="$msg"$'\n\n'"After presenting these facts, retry the Write — the gate is now cleared"
+  msg="$msg"$'\n'"for this file in this session ($state created)."
+  msg="$msg"$'\n\n'"Bypass: ECC_GATEGUARD=off  /  /gate-clear  /  /gate-bypass $file"
+
+  emit_block "$msg"
+  exit 0
+fi
+
+# === Bash gate ===
+if [ "$tool" = "Bash" ]; then
+  cmd=$(printf '%s' "$input" | jq -r '.tool_input.command // empty')
+  if [ -z "$cmd" ]; then echo '{}'; exit 0; fi
+
+  is_destructive="false"
+  if printf '%s' "$cmd" | grep -qE '(rm -rf|rm -fr|git reset --hard|git push --force|git push -f|git branch -D|drop table|DROP TABLE|truncate table|TRUNCATE TABLE|supabase db reset)'; then
+    is_destructive="true"
+  fi
+
+  if [ "$is_destructive" = "false" ]; then
+    echo '{}'; exit 0
+  fi
+
+  chash=$(hash_str "$cmd")
+  state="$state_dir/bash-$chash.cleared"
+
+  if [ -f "$state" ]; then
+    echo '{}'; exit 0
+  fi
+
+  touch "$state" 2>/dev/null
+
+  cmd_short=$(printf '%s' "$cmd" | head -c 200)
+
+  msg="[GateGuard] First destructive Bash command:"
+  msg="$msg"$'\n'"  \$ $cmd_short"
+  msg="$msg"$'\n\n'"Before retrying, gather and present these facts:"
+  msg="$msg"$'\n\n'"  1. List ALL files / data this command modifies or deletes"
+  msg="$msg"$'\n'"     -> enumerate explicitly, not 'various files'"
+  msg="$msg"$'\n'"  2. Provide a one-line rollback procedure"
+  msg="$msg"$'\n'"     -> e.g. git reflog / DB snapshot restore"
+  msg="$msg"$'\n'"  3. Quote the user's verbatim instruction authorizing this destructive action"
+  msg="$msg"$'\n\n'"After presenting these facts, retry — the gate is cleared for this exact"
+  msg="$msg"$'\n'"command in this session ($state created)."
+  msg="$msg"$'\n\n'"Bypass: ECC_GATEGUARD=off  /  /gate-clear"
+
+  emit_block "$msg"
+  exit 0
+fi
+
+# Default: allow
+echo '{}'
+exit 0

package/.claude/hooks/improvement-proposal.sh

@@ -0,0 +1,112 @@
+#!/usr/bin/env bash
+# improvement-proposal.sh — SessionStart hook that surfaces self-improvement hints.
+#
+# Orchestrator: env load + kill switch + cache check + Python aggregate dispatch + output flush.
+# 集計 / cache / heuristics logic は lib/improvement-proposal/ に分割。
+#
+# 設計原則:
+#   - fail-open: jq / python3 が無くても、または読み取り失敗でも exit 0
+#   - noisy にしない: 提案 0 件なら何も出さない、dedup で同提案を 24h 以内に再表示しない
+#   - data 駆動: heuristics は Python 側に集約 (lib/improvement-proposal/aggregate.py)
+#   - cache: 集計結果を TTL 1h で cache し SessionStart 時間を短縮 (lib/improvement-proposal/cache.sh)
+#   - kill switch:
+#       ECC_IMPROVEMENT_PROPOSAL=off          → 全 skip
+#       HC_IMPROVEMENT_PROPOSAL_ENABLED=false → 全 skip
+#       HC_IMPROVEMENT_PROPOSAL_CACHE_ENABLED=false → cache skip (常に再集計)
+#
+# Stdin:  SessionStart hook JSON（読むが現状未使用）
+# Stdout: 空
+# Stderr: 提案行 + フッター 1 行（提案 1+ 件のときのみ）
+# Exit:   常に 0
+
+set -uo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+# shellcheck source=lib/config-loader.sh
+. "$SCRIPT_DIR/lib/config-loader.sh"
+
+# Feature toggle 参照 (task-45 Phase 2)
+if command -v is_feature_enabled >/dev/null 2>&1 && ! is_feature_enabled improvement_proposal; then
+  exit 0   # feature OFF で no-op
+fi
+
+# shellcheck source=lib/improvement-proposal/cache.sh
+. "$SCRIPT_DIR/lib/improvement-proposal/cache.sh"
+
+# stdin を捨てる
+cat >/dev/null 2>&1 || true
+
+# kill switch (env > config)
+if [ "${ECC_IMPROVEMENT_PROPOSAL:-}" = "off" ]; then
+  exit 0
+fi
+case "${HC_IMPROVEMENT_PROPOSAL_ENABLED:-true}" in
+  false|False|FALSE|0|no|off) exit 0 ;;
+esac
+
+if ! command -v python3 >/dev/null 2>&1; then
+  exit 0
+fi
+
+# project root を決定（CLAUDE_PROJECT_DIR > git toplevel > pwd）
+PROJ_ROOT="${CLAUDE_PROJECT_DIR:-}"
+if [ -z "$PROJ_ROOT" ] || [ ! -d "$PROJ_ROOT" ]; then
+  PROJ_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
+fi
+
+# state dir 解決（config 相対パスを project root に基づき絶対化）
+STATE_DIR="${HC_IMPROVEMENT_PROPOSAL_STATE_DIR:-.claude/.improvement-proposal-state}"
+case "$STATE_DIR" in
+  /*) ABS_STATE_DIR="$STATE_DIR" ;;
+  *)  ABS_STATE_DIR="$PROJ_ROOT/$STATE_DIR" ;;
+esac
+
+mkdir -p "$ABS_STATE_DIR" 2>/dev/null || true
+
+# === cache layer (task-22 W5) =================================================
+# cache TTL は秒指定 (default 3600 = 1h)。テスト用に env で短縮可能。
+CACHE_TTL_SECONDS="${HC_IMPROVEMENT_PROPOSAL_CACHE_TTL:-3600}"
+# 整数値検証 (非数値なら default 3600 に fallback)
+case "$CACHE_TTL_SECONDS" in
+  ''|*[!0-9]*) CACHE_TTL_SECONDS=3600 ;;
+esac
+CACHE_FILE="$ABS_STATE_DIR/cache.json"
+CACHE_ENABLED="${HC_IMPROVEMENT_PROPOSAL_CACHE_ENABLED:-true}"
+
+# cache hit → payload を stderr に流して exit 0 (集計 skip)
+if _cache_hit; then
+  jq -r '.payload' "$CACHE_FILE" >&2 2>/dev/null || true
+  exit 0
+fi
+
+# === 集計実行 (cache miss or expired or corrupt) ==============================
+# Python 側に集計を委譲（標準ライブラリのみ使用、lib/improvement-proposal/aggregate.py）
+# 環境変数経由で設定を渡す
+export HC_IMPROVEMENT_PROPOSAL_LOOKBACK_DAYS
+export HC_IMPROVEMENT_PROPOSAL_MAX_COUNT
+export HC_IMPROVEMENT_PROPOSAL_DEDUP_HOURS
+export HC_GATEGUARD_STATE_DIR
+export HC_TASKGUARD_STATE_DIR
+export HC_FAILURE_WINDOW_DIR
+export HC_CONFIDENCE_STATE_DIR
+export HC_HOMUNCULUS_ROOT
+export ECC_IMPROVEMENT_PROPOSAL_TEST_NOW
+export ABS_STATE_DIR
+export PROJ_ROOT
+
+# Python の stderr を一旦 tmp に捕捉、終了後に cache 書き込み + stderr に流す。
+CACHE_TMP="$(mktemp 2>/dev/null || printf '%s' "$ABS_STATE_DIR/.cache.tmp.$$")"
+# observations count を Python 側から拾うための tmp path (cache metadata 用)
+export HC_IMPROVEMENT_PROPOSAL_OBS_COUNT_FILE="$ABS_STATE_DIR/.obs-count.tmp.$$"
+
+python3 "$SCRIPT_DIR/lib/improvement-proposal/aggregate.py" 2>"$CACHE_TMP"
+
+_write_cache
+
+# 集計結果 stderr を実際に流す (cache hit 時と同じ挙動を保証)
+cat "$CACHE_TMP" >&2 2>/dev/null || true
+
+# クリーンアップ
+rm -f "$CACHE_TMP" "$HC_IMPROVEMENT_PROPOSAL_OBS_COUNT_FILE" 2>/dev/null || true
+
+exit 0

package/.claude/hooks/init-tasks-on-start.sh

@@ -0,0 +1,34 @@
+#!/usr/bin/env bash
+# init-tasks-on-start.sh — SessionStart hook
+#
+# 役割:
+#   セッション開始時に docs/tasks/ docs/draft/ が未整備なら自動でテンプレ展開する。
+#   既存ファイルは触らない。失敗してもセッションをブロックしない。
+#
+# 設計:
+#   - quiet モードで init-tasks.sh を呼ぶ
+#   - exit 0 を強制（fail-open）
+
+set -u
+
+# config 読み込み (HC_* 変数 + is_feature_enabled 関数 export、task-45 Phase 2)
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+# shellcheck source=lib/config-loader.sh
+. "$SCRIPT_DIR/lib/config-loader.sh" 2>/dev/null || true
+
+# Feature toggle 参照 (task-45 Phase 2)
+if command -v is_feature_enabled >/dev/null 2>&1 && ! is_feature_enabled init_tasks_on_start; then
+  exit 0   # feature OFF で no-op
+fi
+
+# プロジェクト直下で実行
+if [ -n "${CLAUDE_PROJECT_DIR:-}" ] && [ -d "$CLAUDE_PROJECT_DIR" ]; then
+  cd "$CLAUDE_PROJECT_DIR" 2>/dev/null || true
+fi
+
+# init-tasks.sh が見つかれば実行
+if [ -f .claude/scripts/init-tasks.sh ]; then
+  bash .claude/scripts/init-tasks.sh --quiet 2>/dev/null || true
+fi
+
+exit 0

package/.claude/hooks/lib/bypass-logger.sh

@@ -0,0 +1,82 @@
+#!/usr/bin/env bash
+# .claude/hooks/lib/bypass-logger.sh
+# Centralized bypass logging — 全 hook の env-var bypass を統一フォーマットで
+# .claude/.workflow-state/bypass.log に記録する共通ライブラリ。
+#
+# 役割:
+#   各 hook が env-var (例: ECC_WORKFLOW_GUARD_OFF / HC_WORKFLOW_GUARD_ENABLED=false)
+#   で skip された際に audit log を 1 行 append する。
+#   形式は .claude/.workflow-state/bypass.log.template と一致:
+#     <ISO-8601> | <session_id> | <hook_name> | <env_var> | <reason>
+#
+# 設計:
+#   - file-top に set を書かない (CLAUDE.md Critical Lessons HIGH 準拠、caller への
+#     shell flags leak 防止。pipefail leak → SIGPIPE → exit 141 サイレント終了の根絶)
+#   - set -e は使わない (mode-loader.sh の CB-verify 教訓 - 5846925)
+#   - 各関数 body 内で set -uo pipefail を subshell `( ... )` 形式で局所化
+#   - 失敗しても hook 本体を止めない (mkdir / printf を best-effort)
+#   - reason 未指定時は "(not provided)"
+#   - session_id 未取得時は "unknown"
+#
+# Security (task-40 Step 7 iter3 HIGH-1):
+#   - 全フィールド (hook_name / env_var / reason / session_id) に対し改行 -> 空白、
+#     pipe -> カンマの sanitize を実施。Claude tool_input から取得した file_path 等が
+#     reason に含まれる場合に改行/pipe 注入で bypass.log の audit trail (format:
+#     <ts> | <session> | <hook> | <env> | <reason>) が破壊・偽造されることを防止。
+#   - defense in depth: timestamp は date(1) 出力 (control 化済) だが念のため対象外。
+#
+# 使い方:
+#   source "${CLAUDE_PROJECT_DIR:-$(pwd)}/.claude/hooks/lib/bypass-logger.sh"
+#   log_bypass "workflow-guard" "ECC_WORKFLOW_GUARD_OFF" "$ECC_BYPASS_REASON"
+#
+# Args:
+#   $1 = hook_name  (例: "workflow-guard")
+#   $2 = env_var    (例: "ECC_WORKFLOW_GUARD_OFF")
+#   $3 = reason     (任意 — 未指定時は "(not provided)")
+
+# Note: file-top に `set -uo pipefail` を書かない。caller hook (autonomous-action-guard /
+# confidence-gate / parallel-subagent-reminder 等) が source した瞬間に pipefail flag が
+# leak し、`cmd | head -1` 等の SIGPIPE で exit 141 サイレント終了する事故を防ぐ
+# (CLAUDE.md Critical Operational Lessons HIGH 準拠)。各関数は subshell `( ... )`
+# 形式で定義し、内部で `set -uo pipefail` を局所適用する。
+
+# log_bypass <hook_name> <env_var> <reason>
+# 副作用: .claude/.workflow-state/bypass.log に 1 行 append
+# subshell 関数化 ( ... ): 内部の set -uo pipefail は caller の shell flags に leak しない
+log_bypass() (
+  set -uo pipefail
+  local hook_name="${1:-unknown}"
+  local env_var="${2:-unknown}"
+  local reason="${3:-(not provided)}"
+  local session_id="${CLAUDE_SESSION_ID:-unknown}"
+  local timestamp
+  timestamp=$(date -u +"%Y-%m-%dT%H:%M:%SZ" 2>/dev/null)
+
+  # ----------------------------------------------------------------------
+  # HIGH-1 (task-40 Step 7 iter3): audit trail 改行/pipe 注入 sanitize
+  # reason は Claude tool_input.file_path 等の untrusted source を含み得る。
+  # 改行 (\n) を空白に置換 -> 後続行偽造防止。
+  # pipe (|) をカンマに置換 -> 5-field format (<ts> | <s> | <h> | <e> | <r>) 破壊防止。
+  # 全フィールドに適用 (defense in depth、env_var / hook_name は通常静的だが念のため)。
+  # CR (\r) も改行扱いで sanitize (Windows-style line ending 経由の偽造防止)。
+  # ----------------------------------------------------------------------
+  hook_name="${hook_name//$'\n'/ }"
+  hook_name="${hook_name//$'\r'/ }"
+  hook_name="${hook_name//|/,}"
+  env_var="${env_var//$'\n'/ }"
+  env_var="${env_var//$'\r'/ }"
+  env_var="${env_var//|/,}"
+  reason="${reason//$'\n'/ }"
+  reason="${reason//$'\r'/ }"
+  reason="${reason//|/,}"
+  session_id="${session_id//$'\n'/ }"
+  session_id="${session_id//$'\r'/ }"
+  session_id="${session_id//|/,}"
+
+  local log_dir="${CLAUDE_PROJECT_DIR:-$(pwd)}/.claude/.workflow-state"
+  local log_file="${log_dir}/bypass.log"
+  mkdir -p "$log_dir" 2>/dev/null
+  printf '%s | %s | %s | %s | %s\n' \
+    "$timestamp" "$session_id" "$hook_name" "$env_var" "$reason" \
+    >> "$log_file" 2>/dev/null
+)

package/.claude/hooks/lib/confidence-gate/bypass.sh

@@ -0,0 +1,48 @@
+#!/usr/bin/env bash
+# bypass.sh — confidence-gate の bypass marker 判定 + log_failure helper
+#
+# 提供関数:
+#   init_bypass_paths             — state_dir / bypass_marker / bypass_log を設定
+#   log_failure <phase> [detail]  — 失敗理由を bypass.log に構造化記録
+#   handle_bypass_marker          — bypass.cleared 存在時に通過 + 1 回 re-arm
+#
+# caller (orchestrator) が事前に export 必須:
+#   HC_CONFIDENCE_STATE_DIR (config-loader 経由で setup 済)
+#
+# 副作用:
+#   handle_bypass_marker は bypass marker 検出時に jq '{}' echo + exit 0
+
+init_bypass_paths() {
+  state_dir="${HC_CONFIDENCE_STATE_DIR:-.claude/.confidence-gate-state}"
+  mkdir -p "$state_dir" 2>/dev/null
+  bypass_marker="${state_dir}/bypass.cleared"
+  bypass_log="${state_dir}/bypass.log"
+}
+
+# bypass.log 拡張: 失敗理由を構造化ログ。
+#   引数: $1 = phase (regex_no_match / no_transcript / file_not_found / extract_failed
+#                     / below_threshold / below_threshold_via_tool_response)
+#         $2 = detail (任意 free text、200 char 上限、改行/タブは空白に置換)
+log_failure() {
+  local phase="$1"
+  local detail="${2:-}"
+  local ts
+  ts=$(date -u +%Y-%m-%dT%H:%M:%SZ)
+  detail=$(printf '%s' "$detail" | tr '\n\t' '  ' | cut -c1-200)
+  printf '%s\tfailed: %s\t%s\n' "$ts" "$phase" "$detail" >> "$bypass_log" 2>/dev/null
+}
+
+handle_bypass_marker() {
+  if [ -f "$bypass_marker" ]; then
+    local bypass_reason=""
+    if [ -s "$bypass_marker" ]; then
+      bypass_reason=$(cat "$bypass_marker" 2>/dev/null | head -c 200)
+    fi
+    local ts
+    ts=$(date -u +%Y-%m-%dT%H:%M:%SZ)
+    printf '%s\tbypassed: %s\n' "$ts" "${bypass_reason:-(no reason given)}" >> "$bypass_log" 2>/dev/null
+    rm -f "$bypass_marker" 2>/dev/null
+    echo '{}'
+    exit 0
+  fi
+}

package/.claude/hooks/lib/confidence-gate/extract.sh

@@ -0,0 +1,99 @@
+#!/usr/bin/env bash
+# extract.sh — confidence-gate の text 抽出 helper
+#
+# 提供関数:
+#   extract_confidence <text>           — text から `confidence: 0.X` を抽出 (variant 対応)
+#   extract_tool_response_text <input>  — SubagentStop JSON の tool_response から最終 reply text 抽出
+#   resolve_subagent_transcript <transcript> <agent_id>
+#                                       — subagent transcript path 解決 (agent_id 付きなら subagents/ 優先)
+#   extract_final_assistant_text <transcript>
+#                                       — transcript 末尾 500 行から assistant message text を join
+
+# === 抽出ヘルパー: confidence 数値を文字列から抜く ===
+# 対応 variant:
+#   confidence: 0.85
+#   Confidence: 0.85
+#   confidence_score: 0.85
+#   confidence score: 0.85
+#   信頼度: 0.85
+#   信頼度 0.85
+# 数値レンジ: 0(.0..) / 1(.0..)
+# 区切り: 半角 ":" / 全角 "："（記号のみ env-portable な POSIX class で）
+extract_confidence() {
+  local text="$1"
+  local m
+  # 英語 variant — confidence / confidence_score / confidence score
+  m=$(printf '%s' "$text" | grep -ioE '(confidence([_[:space:]]*score)?)[[:space:]]*[:：][[:space:]]*(0(\.[0-9]+)?|1(\.0+)?)' | tail -n 1)
+  if [ -z "$m" ]; then
+    # 日本語 variant — 信頼度: 0.X / 信頼度 0.X
+    m=$(printf '%s' "$text" | grep -ioE '信頼度[[:space:]]*[:：]?[[:space:]]*(0(\.[0-9]+)?|1(\.0+)?)' | tail -n 1)
+  fi
+  printf '%s' "$m"
+}
+
+# === fallback: hook input の tool_response から最終 reply text を構築 ===
+extract_tool_response_text() {
+  printf '%s' "$1" | jq -r '
+    (.tool_response // {})
+    | if type == "string" then .
+      elif type == "object" then
+        (.content // .text // "")
+        | if type == "array" then
+            [ .[] | (.text // (if type == "string" then . else "" end)) ] | join("\n")
+          elif type == "string" then .
+          else "" end
+      else "" end
+  ' 2>/dev/null
+}
+
+# === subagent transcript 解決 ===
+# SubagentStop hook に渡される transcript_path は親セッションの JSONL を指すことがある。
+# 実際のサブエージェント応答は <parent_stem>/subagents/agent-<id>.jsonl に書かれる。
+# agent_id が SubagentStop JSON に含まれる場合は subagent ファイルを優先して読む。
+# stdout に解決後 transcript path を返す
+resolve_subagent_transcript() {
+  local transcript="$1"
+  local agent_id="$2"
+  if [ -n "$agent_id" ] && [ -n "$transcript" ] && [ "$transcript" != "null" ]; then
+    local stem subagent_file
+    stem="${transcript%.jsonl}"
+    subagent_file="${stem}/subagents/agent-${agent_id}.jsonl"
+    if [ -f "$subagent_file" ]; then
+      printf '%s' "$subagent_file"
+      return
+    fi
+  fi
+  printf '%s' "$transcript"
+}
+
+# === 最終 assistant text を抽出 ===
+# transcript は JSONL。末尾 500 行から assistant role のメッセージを集める
+# (元 200 行では subagent の長い summary が境界をまたぎ confidence を見落とす
+# ケースが頻発したため拡大)。
+# Claude Code transcript schema 新旧両対応:
+#   旧: {"type":"assistant","message":{"content":[{"type":"text","text":"..."}]}}
+#   新: {"role":"assistant","content":[{"type":"text","text":"..."}]} 等
+extract_final_assistant_text() {
+  local transcript="$1"
+  tail -n 500 "$transcript" 2>/dev/null | jq -rs '
+    [
+      .[]
+      | select(
+          ((.type // "") == "assistant")
+          or ((.role // "") == "assistant")
+          or (((.message // {}).role // "") == "assistant")
+        )
+      | (
+          ((.message // .).content)
+          | if type == "array" then
+              [ .[] | (.text // "") ] | join("\n")
+            elif type == "string" then
+              .
+            else
+              ""
+            end
+        )
+    ]
+    | join("\n")
+  ' 2>/dev/null
+}

package/.claude/hooks/lib/confidence-gate/major-agent-filter.sh

@@ -0,0 +1,59 @@
+#!/usr/bin/env bash
+# major-agent-filter.sh — F3 major subagent only block の判定
+#
+# 役割:
+#   SubagentStop hook は agent_type を問わず全 stop event で fire するため、
+#   軽量 sidechain (Task tool query / tool-use only) で 96 件の regex_no_match
+#   が累積する事故があった (/harness-audit 2026-05-13 観察)。
+#   major subagent (general-purpose 等の allowlist or path_subagents) 以外は
+#   fail-open で通過させ、major subagent only に confidence 自己評価を強制する。
+#
+# bypass: HC_CONFIDENCE_MAJOR_AGENT_ONLY=false で従来動作 (全 stop event で block 判定)。
+#
+# 提供関数:
+#   classify_sidechain <transcript>     — transcript path / record から is_sidechain 判定
+#                                          stdout: "path_subagents" / "record_isSidechain" / "no" / "unknown"
+#   is_major_subagent <agent_type> <is_sidechain>
+#                                       — major subagent 判定
+#                                          stdout: "true" / "false"
+
+# transcript path から sidechain 判定
+classify_sidechain() {
+  local transcript="$1"
+  case "$transcript" in
+    */subagents/*) printf 'path_subagents'; return ;;
+  esac
+  printf 'unknown'
+}
+
+# transcript file 内の record から sidechain 判定 (path 判定で unknown のとき)
+refine_sidechain_from_records() {
+  local current="$1"
+  local transcript="$2"
+  if [ "$current" = "unknown" ]; then
+    if grep -q '"isSidechain"[[:space:]]*:[[:space:]]*true' "$transcript" 2>/dev/null; then
+      printf 'record_isSidechain'
+    else
+      printf 'no'
+    fi
+    return
+  fi
+  printf '%s' "$current"
+}
+
+# === major subagent 判定 (W1, task #9) ===
+# allowlist match (agent_type が general-purpose / Explore / Task) または
+# is_sidechain==path_subagents なら major subagent (confidence 必須)。
+# それ以外 (軽量 sidechain / Task tool query etc.) は extract_confidence 失敗時に fail-open。
+is_major_subagent() {
+  local agent_type="$1"
+  local is_sidechain="$2"
+  local result="false"
+  case "$agent_type" in
+    general-purpose|Explore|Task) result="true" ;;
+  esac
+  if [ "$is_sidechain" = "path_subagents" ]; then
+    result="true"
+  fi
+  printf '%s' "$result"
+}