@takuma-hirai/hirai-method 0.1.0

package/.claude/hooks/tests/test-confidence-gate.sh

@@ -0,0 +1,170 @@
+#!/usr/bin/env bash
+# test-confidence-gate.sh — smoke tests for confidence-gate.sh transcript extraction
+#
+# Verifies the improvements documented in feat/harness-improvement-2026-05-04:
+#   1. tail size 200 -> 500: confidence near transcript boundary still detected
+#   2. sidechain detection: transcript with isSidechain:true reads directly (no parent traversal)
+#   3. regex variants: "Confidence: 0.X" / "信頼度: 0.X" / "confidence_score: 0.X" all match
+#   4. tool_response fallback: when transcript missing, hook input tool_response provides text
+#   5. no-match detailed log: bypass.log records failure phase + detail
+#
+# 既存 test-gate-disable.sh と同居して run-tests.sh から自動で動く。
+
+set -u
+
+HOOKS_DIR="$(cd "$(dirname "$0")/.." && pwd)"
+HOOK="$HOOKS_DIR/confidence-gate.sh"
+PASS=0
+FAIL=0
+
+# Sandbox state dir
+TMP=$(mktemp -d)
+trap 'rm -rf "$TMP"' EXIT
+export HC_CONFIDENCE_STATE_DIR="$TMP/conf"
+mkdir -p "$HC_CONFIDENCE_STATE_DIR"
+
+# Force gate ON (env > YAML override) so we exercise the real extraction path
+export HC_CONFIDENCE_REQUIRED=true
+export HC_CONFIDENCE_THRESHOLD=0.6
+unset ECC_CONFIDENCE_GATE
+unset ECC_F3_OFF
+
+assert_pass() {
+  local name="$1"
+  local out="$2"
+  # pass = empty {} JSON (or any output without "decision":"block")
+  local compact
+  compact=$(printf '%s' "$out" | tr -d '[:space:]')
+  if [[ "$compact" != *'"decision":"block"'* ]]; then
+    printf "PASS: %s\n" "$name"
+    PASS=$((PASS + 1))
+  else
+    printf "FAIL: %s\n  expected pass, got block: %s\n" "$name" "$out"
+    FAIL=$((FAIL + 1))
+  fi
+}
+
+assert_block() {
+  local name="$1"
+  local out="$2"
+  local compact
+  compact=$(printf '%s' "$out" | tr -d '[:space:]')
+  if [[ "$compact" == *'"decision":"block"'* ]]; then
+    printf "PASS: %s\n" "$name"
+    PASS=$((PASS + 1))
+  else
+    printf "FAIL: %s\n  expected block, got: %s\n" "$name" "$out"
+    FAIL=$((FAIL + 1))
+  fi
+}
+
+# Helper: build a transcript JSONL with N filler lines + tail entries
+make_transcript() {
+  local path="$1"
+  local filler_count="$2"
+  shift 2
+  : > "$path"
+  local i=0
+  while [ "$i" -lt "$filler_count" ]; do
+    printf '{"type":"user","message":{"role":"user","content":"filler %d"}}\n' "$i" >> "$path"
+    i=$((i + 1))
+  done
+  for line in "$@"; do
+    printf '%s\n' "$line" >> "$path"
+  done
+}
+
+# === Case 1: tail size 200 -> 500. ===
+# Place an assistant "confidence: 0.85" so it survives the 500-line tail.
+# 350 filler lines push old 200-tail past the assistant message; 500-tail catches it.
+T1="$TMP/t1.jsonl"
+ASSISTANT_LINE='{"type":"assistant","message":{"role":"assistant","content":[{"type":"text","text":"Implementation complete. confidence: 0.85"}]}}'
+TAIL_FILLER_COUNT=300
+{
+  printf '%s\n' "$ASSISTANT_LINE"
+  i=0
+  while [ "$i" -lt "$TAIL_FILLER_COUNT" ]; do
+    printf '{"type":"user","message":{"role":"user","content":"trailing %d"}}\n' "$i"
+    i=$((i + 1))
+  done
+} > "$T1"
+INPUT="{\"transcript_path\":\"$T1\"}"
+out=$(printf '%s' "$INPUT" | bash "$HOOK" 2>/dev/null)
+assert_pass "Case1: confidence detected within 500-line tail (~300 trailing filler)" "$out"
+
+# === Case 2: sidechain detection - path contains /subagents/ ===
+# Place transcript inside a /subagents/ subdir; assistant text has confidence; expect pass.
+SIDE_DIR="$TMP/projects/abc/subagents"
+mkdir -p "$SIDE_DIR"
+T2="$SIDE_DIR/agent-side.jsonl"
+{
+  printf '{"isSidechain":true,"type":"user","message":{"role":"user","content":"task"}}\n'
+  printf '%s\n' '{"type":"assistant","message":{"role":"assistant","content":[{"type":"text","text":"Done. confidence: 0.9"}]}}'
+} > "$T2"
+INPUT="{\"transcript_path\":\"$T2\"}"
+out=$(printf '%s' "$INPUT" | bash "$HOOK" 2>/dev/null)
+assert_pass "Case2: sidechain transcript (/subagents/ path) read directly" "$out"
+
+# === Case 3: regex variants ===
+T3a="$TMP/t3a.jsonl"
+printf '%s\n' '{"type":"assistant","message":{"role":"assistant","content":[{"type":"text","text":"Done. Confidence: 0.75"}]}}' > "$T3a"
+out=$(printf '%s' "{\"transcript_path\":\"$T3a\"}" | bash "$HOOK" 2>/dev/null)
+assert_pass "Case3a: 'Confidence:' (capitalized) variant" "$out"
+
+T3b="$TMP/t3b.jsonl"
+printf '%s\n' '{"type":"assistant","message":{"role":"assistant","content":[{"type":"text","text":"Done. confidence_score: 0.8"}]}}' > "$T3b"
+out=$(printf '%s' "{\"transcript_path\":\"$T3b\"}" | bash "$HOOK" 2>/dev/null)
+assert_pass "Case3b: 'confidence_score:' variant" "$out"
+
+T3c="$TMP/t3c.jsonl"
+printf '%s\n' '{"type":"assistant","message":{"role":"assistant","content":[{"type":"text","text":"完了。信頼度: 0.95"}]}}' > "$T3c"
+out=$(printf '%s' "{\"transcript_path\":\"$T3c\"}" | bash "$HOOK" 2>/dev/null)
+assert_pass "Case3c: '信頼度:' Japanese variant" "$out"
+
+# === Case 4: tool_response fallback ===
+# transcript_path missing entirely; tool_response.content holds confidence.
+INPUT='{"tool_response":{"content":[{"type":"text","text":"Result. confidence: 0.7"}]}}'
+out=$(printf '%s' "$INPUT" | bash "$HOOK" 2>/dev/null)
+assert_pass "Case4a: tool_response fallback (array of text blocks)" "$out"
+
+# tool_response as a plain string
+INPUT='{"tool_response":"Done with high confidence. Confidence: 0.88"}'
+out=$(printf '%s' "$INPUT" | bash "$HOOK" 2>/dev/null)
+assert_pass "Case4b: tool_response fallback (string payload)" "$out"
+
+# tool_response present but below threshold -> block
+INPUT='{"tool_response":{"content":[{"type":"text","text":"shaky. confidence: 0.3"}]}}'
+out=$(printf '%s' "$INPUT" | bash "$HOOK" 2>/dev/null)
+assert_block "Case4c: tool_response fallback but below threshold blocks" "$out"
+
+# === Case 5: no-match detailed log ===
+# Provide transcript with assistant text that has NO confidence marker -> regex_no_match.
+# Then bypass.log should have a "failed: regex_no_match" line.
+T5="$TMP/t5.jsonl"
+printf '%s\n' '{"type":"assistant","message":{"role":"assistant","content":[{"type":"text","text":"All done, looks great"}]}}' > "$T5"
+INPUT="{\"transcript_path\":\"$T5\"}"
+out=$(printf '%s' "$INPUT" | bash "$HOOK" 2>/dev/null)
+assert_block "Case5a: no confidence variant -> block" "$out"
+if grep -q "failed: regex_no_match" "$HC_CONFIDENCE_STATE_DIR/bypass.log" 2>/dev/null; then
+  printf "PASS: Case5b: bypass.log records 'failed: regex_no_match'\n"
+  PASS=$((PASS + 1))
+else
+  printf "FAIL: Case5b: bypass.log missing 'failed: regex_no_match'\n  log: %s\n" \
+    "$(cat "$HC_CONFIDENCE_STATE_DIR/bypass.log" 2>/dev/null)"
+  FAIL=$((FAIL + 1))
+fi
+
+# Case 5c: file_not_found logged when transcript path absent on disk
+INPUT='{"transcript_path":"/no/such/path.jsonl"}'
+out=$(printf '%s' "$INPUT" | bash "$HOOK" 2>/dev/null)
+if grep -q "failed: file_not_found" "$HC_CONFIDENCE_STATE_DIR/bypass.log" 2>/dev/null; then
+  printf "PASS: Case5c: bypass.log records 'failed: file_not_found'\n"
+  PASS=$((PASS + 1))
+else
+  printf "FAIL: Case5c: bypass.log missing 'failed: file_not_found'\n  log: %s\n" \
+    "$(cat "$HC_CONFIDENCE_STATE_DIR/bypass.log" 2>/dev/null)"
+  FAIL=$((FAIL + 1))
+fi
+
+printf "\n--- test-confidence-gate.sh: %d pass / %d fail ---\n" "$PASS" "$FAIL"
+[ "$FAIL" -eq 0 ]

package/.claude/hooks/tests/test-config-env-override.sh

@@ -0,0 +1,220 @@
+#!/usr/bin/env bash
+# test-config-env-override.sh — env > YAML > defaults precedence for config-loader.sh
+#
+# config-loader.sh の汎用 env override 機能を検証する。
+# 検証する仕様:
+#   1. env 未設定 → YAML 値が読まれる
+#   2. env 設定 → YAML 値より env 優先
+#   3. env 設定 + YAML 不在 → env 値で動作（warning は出るが defaults を上書き）
+#   4. YAML / env / defaults の三重指定 → env が最優先
+#   5. 配列キー (HC_PROTECTED_PATHS) も env で上書き可能
+#   6. tilde 含む path (HC_HOMUNCULUS_ROOT=~/foo) は env でも展開される
+#   7. confidence-gate (HC_CONFIDENCE_REQUIRED=false) が env override で動作
+#
+# 既存 test-confidence-gate.sh / test-gate-disable.sh と同居して
+# run-tests.sh から自動で動く。
+
+set -u
+
+HOOKS_DIR="$(cd "$(dirname "$0")/.." && pwd)"
+LOADER="$HOOKS_DIR/lib/config-loader.sh"
+PASS=0
+FAIL=0
+
+TMP=$(mktemp -d)
+trap 'rm -rf "$TMP"' EXIT
+
+assert_eq() {
+  local name="$1"
+  local expected="$2"
+  local actual="$3"
+  if [ "$expected" = "$actual" ]; then
+    printf "PASS: %s\n" "$name"
+    PASS=$((PASS + 1))
+  else
+    printf "FAIL: %s\n  expected: [%s]\n  actual:   [%s]\n" "$name" "$expected" "$actual"
+    FAIL=$((FAIL + 1))
+  fi
+}
+
+# Helper: run a sub-shell with a fresh env and source config-loader,
+# then echo the requested HC_* variable.
+# Args: <yaml-path-or-empty> <var-name> [<env-spec>...]
+#   env-spec: NAME=VALUE (will be exported in the sub-shell before sourcing)
+load_and_echo() {
+  local yaml="$1"
+  local var="$2"
+  shift 2
+  local env_setup=""
+  for spec in "$@"; do
+    # Use printf %q-equivalent quoting via shell escape
+    env_setup="$env_setup export $spec;"
+  done
+  local config_setup=""
+  if [ -n "$yaml" ]; then
+    config_setup="export HC_CONFIG_PATH=\"$yaml\";"
+  else
+    # /dev/null path -> trigger "config not found" branch + suppress WARN
+    config_setup="export HC_CONFIG_PATH=$TMP/nonexistent.yml;"
+  fi
+  bash -c "
+    set +u
+    # Strip any inherited HC_* vars so each test starts clean
+    for v in \$(env | grep -oE '^HC_[A-Z_]+' || true); do unset \"\$v\"; done
+    $config_setup
+    $env_setup
+    source '$LOADER' 2>/dev/null
+    eval \"echo \\\"\\\$$var\\\"\"
+  "
+}
+
+# === Case 1: env 未設定 → YAML 値 ===
+cat > "$TMP/case1.yml" <<'YAML'
+task_dir: docs/yaml-tasks
+draft_dir: docs/yaml-draft
+homunculus_root: ~/.claude/yaml-root
+YAML
+out=$(load_and_echo "$TMP/case1.yml" HC_TASK_DIR)
+assert_eq "Case1a: no env -> YAML task_dir" "docs/yaml-tasks" "$out"
+out=$(load_and_echo "$TMP/case1.yml" HC_DRAFT_DIR)
+assert_eq "Case1b: no env -> YAML draft_dir" "docs/yaml-draft" "$out"
+out=$(load_and_echo "$TMP/case1.yml" HC_HOMUNCULUS_ROOT)
+assert_eq "Case1c: no env -> YAML homunculus_root tilde-expanded" "$HOME/.claude/yaml-root" "$out"
+
+# === Case 2: env 設定 → env > YAML ===
+out=$(load_and_echo "$TMP/case1.yml" HC_TASK_DIR "HC_TASK_DIR=docs/env-tasks")
+assert_eq "Case2a: env task_dir overrides YAML" "docs/env-tasks" "$out"
+out=$(load_and_echo "$TMP/case1.yml" HC_DRAFT_DIR "HC_DRAFT_DIR=docs/env-draft")
+assert_eq "Case2b: env draft_dir overrides YAML" "docs/env-draft" "$out"
+
+# === Case 3: env + YAML 不在 → env 値 (defaults を上書き) ===
+out=$(load_and_echo "" HC_TASK_DIR "HC_TASK_DIR=docs/env-only")
+assert_eq "Case3a: env wins when YAML missing" "docs/env-only" "$out"
+# YAML 不在 + env 不在 → defaults
+out=$(load_and_echo "" HC_TASK_DIR)
+assert_eq "Case3b: defaults when both YAML and env absent" "docs/tasks" "$out"
+
+# === Case 4: 三重指定 → env 最優先 ===
+# YAML が defaults と異なる値を持ち、env が更に異なる値を持つ。
+cat > "$TMP/case4.yml" <<'YAML'
+task_dir: docs/yaml-value
+YAML
+out=$(load_and_echo "$TMP/case4.yml" HC_TASK_DIR "HC_TASK_DIR=docs/env-value")
+assert_eq "Case4: env > YAML > defaults — env wins" "docs/env-value" "$out"
+
+# === Case 5: 配列キー env 上書き ===
+# bash の改行リスト形式を env で渡すと PROTECTED_DISPLAY 等の派生値も
+# 正しく再生成されるか。
+cat > "$TMP/case5.yml" <<'YAML'
+protected_paths: [src, tests, scripts]
+YAML
+# YAML 値が読まれるか確認 (control)
+out=$(load_and_echo "$TMP/case5.yml" HC_PROTECTED_DISPLAY)
+assert_eq "Case5a: YAML protected_paths -> display" "src/ tests/ scripts/" "$out"
+# env で上書き (改行区切り — bash の $'a\nb\nc' リテラル)
+out=$(bash -c "
+  set +u
+  for v in \$(env | grep -oE '^HC_[A-Z_]+' || true); do unset \"\$v\"; done
+  export HC_CONFIG_PATH='$TMP/case5.yml'
+  export HC_PROTECTED_PATHS=\$'app\nlib\nmiddleware.ts'
+  source '$LOADER' 2>/dev/null
+  echo \"\$HC_PROTECTED_DISPLAY\"
+")
+assert_eq "Case5b: env protected_paths overrides YAML" "app/ lib/ middleware.ts/" "$out"
+# 派生 glob も正しく再生成されているか
+out=$(bash -c "
+  set +u
+  for v in \$(env | grep -oE '^HC_[A-Z_]+' || true); do unset \"\$v\"; done
+  export HC_CONFIG_PATH='$TMP/case5.yml'
+  export HC_PROTECTED_PATHS=\$'app\nlib'
+  source '$LOADER' 2>/dev/null
+  echo \"\$HC_PROTECTED_GLOB_FILE\"
+")
+assert_eq "Case5c: env protected_paths regenerates GLOB_FILE" "*/app/*|*/lib/*" "$out"
+
+# === Case 6: tilde 含む path env 上書き ===
+out=$(load_and_echo "$TMP/case1.yml" HC_HOMUNCULUS_ROOT "HC_HOMUNCULUS_ROOT=~/foo-from-env")
+assert_eq "Case6a: env tilde expanded (~/foo)" "$HOME/foo-from-env" "$out"
+# tilde 単体 (`~`)
+out=$(load_and_echo "$TMP/case1.yml" HC_HOMUNCULUS_ROOT "HC_HOMUNCULUS_ROOT=~")
+assert_eq "Case6b: env tilde alone (~)" "$HOME" "$out"
+# 絶対 path (展開不要)
+out=$(load_and_echo "$TMP/case1.yml" HC_HOMUNCULUS_ROOT "HC_HOMUNCULUS_ROOT=/absolute/path")
+assert_eq "Case6c: env absolute path unchanged" "/absolute/path" "$out"
+
+# === Case 7: confidence-gate 個別対応 (HC_CONFIDENCE_REQUIRED=false) ===
+# YAML が confidence_required: true、env が false → 即 pass を返すこと
+cat > "$TMP/case7.yml" <<'YAML'
+confidence_required: true
+confidence_threshold: 0.95
+YAML
+# control: env なし、YAML true → block (confidence 未記載)
+INPUT='{"transcript_path":"/nonexistent.jsonl"}'
+out=$(bash -c "
+  set +u
+  for v in \$(env | grep -oE '^HC_[A-Z_]+' || true); do unset \"\$v\"; done
+  unset ECC_CONFIDENCE_GATE ECC_F3_OFF
+  export HC_CONFIG_PATH='$TMP/case7.yml'
+  export HC_CONFIDENCE_STATE_DIR='$TMP/conf7-control'
+  mkdir -p \"\$HC_CONFIDENCE_STATE_DIR\"
+  printf '%s' '$INPUT' | bash '$HOOKS_DIR/confidence-gate.sh' 2>/dev/null
+")
+# fallback path (transcript missing) returns {} (fail-open) - so this proves env wasn't broken.
+# We instead test the early-OFF branch directly: HC_CONFIDENCE_REQUIRED=false should yield {}
+out=$(bash -c "
+  set +u
+  for v in \$(env | grep -oE '^HC_[A-Z_]+' || true); do unset \"\$v\"; done
+  unset ECC_CONFIDENCE_GATE ECC_F3_OFF
+  export HC_CONFIG_PATH='$TMP/case7.yml'
+  export HC_CONFIDENCE_REQUIRED=false
+  export HC_CONFIDENCE_STATE_DIR='$TMP/conf7-env'
+  mkdir -p \"\$HC_CONFIDENCE_STATE_DIR\"
+  printf '%s' '$INPUT' | bash '$HOOKS_DIR/confidence-gate.sh' 2>/dev/null
+")
+out_compact=$(printf '%s' "$out" | tr -d '[:space:]')
+assert_eq "Case7a: HC_CONFIDENCE_REQUIRED=false (env) overrides YAML true -> pass" "{}" "$out_compact"
+
+# Threshold env override: with required=true, threshold=0.0, any confidence passes.
+INPUT2='{"tool_response":{"content":[{"type":"text","text":"shaky. confidence: 0.1"}]}}'
+out=$(bash -c "
+  set +u
+  for v in \$(env | grep -oE '^HC_[A-Z_]+' || true); do unset \"\$v\"; done
+  unset ECC_CONFIDENCE_GATE ECC_F3_OFF
+  export HC_CONFIG_PATH='$TMP/case7.yml'
+  export HC_CONFIDENCE_REQUIRED=true
+  export HC_CONFIDENCE_THRESHOLD=0.0
+  export HC_CONFIDENCE_STATE_DIR='$TMP/conf7-thresh'
+  mkdir -p \"\$HC_CONFIDENCE_STATE_DIR\"
+  printf '%s' '$INPUT2' | bash '$HOOKS_DIR/confidence-gate.sh' 2>/dev/null
+")
+out_compact=$(printf '%s' "$out" | tr -d '[:space:]')
+# Should pass (0.1 >= 0.0) — env threshold beats YAML threshold (0.95)
+if [ "$out_compact" = "{}" ]; then
+  printf "PASS: Case7b: env HC_CONFIDENCE_THRESHOLD=0.0 overrides YAML 0.95 -> pass at 0.1\n"
+  PASS=$((PASS + 1))
+else
+  printf "FAIL: Case7b: expected {} got: %s\n" "$out"
+  FAIL=$((FAIL + 1))
+fi
+
+# === Case 8: 内部変数 leak チェック (env override 機能の副作用) ===
+# config-loader source 後に _HC_PRESET_KEYS / _hc_k 等が caller scope に
+# 残らないことを確認する (regression: caller env 汚染).
+out=$(bash -c "
+  set +u
+  for v in \$(env | grep -oE '^HC_[A-Z_]+' || true); do unset \"\$v\"; done
+  export HC_CONFIG_PATH='$TMP/case1.yml'
+  export HC_TASK_DIR=docs/probe
+  source '$LOADER' 2>/dev/null
+  printf 'PRESET_KEYS=[%s] PRESET_TASK_DIR=[%s] hc_k=[%s] hc_v=[%s] hc_set=[%s]\n' \
+    \"\${_HC_PRESET_KEYS-UNSET}\" \\
+    \"\${_HC_PRESET_TASK_DIR-UNSET}\" \\
+    \"\${_hc_k-UNSET}\" \\
+    \"\${_hc_v-UNSET}\" \\
+    \"\${_hc_set-UNSET}\"
+")
+expected="PRESET_KEYS=[UNSET] PRESET_TASK_DIR=[UNSET] hc_k=[UNSET] hc_v=[UNSET] hc_set=[UNSET]"
+assert_eq "Case8: internal vars are unset after source" "$expected" "$out"
+
+printf "\n--- test-config-env-override.sh: %d pass / %d fail ---\n" "$PASS" "$FAIL"
+[ "$FAIL" -eq 0 ]

package/.claude/hooks/tests/test-gate-disable.sh

@@ -0,0 +1,118 @@
+#!/usr/bin/env bash
+# test-gate-disable.sh — smoke test for ECC_F{1,2,3}_OFF env var bypass
+#
+# Verifies that:
+#   - F1 (gateguard.sh):        ECC_F1_OFF=1 → approve, unset → existing behavior
+#   - F2 (task-rule-guard.sh):  ECC_F2_OFF=1 → approve, unset → existing behavior
+#   - F3 (confidence-gate.sh):  ECC_F3_OFF=1 → approve, unset → existing behavior
+#
+# Test design:
+#   - "OFF": invoke hook with ECC_F{n}_OFF=1, expect output containing
+#            decision":"approve" (or {"decision":"approve",...}). No file
+#            system mutation needed.
+#   - "ON" (envvar unset): invoke hook with an input that the existing
+#            logic would block. Expect output NOT containing the F{n}_OFF
+#            marker reason — proves the early-exit branch is not hit and
+#            existing logic still runs.
+
+set -u
+
+HOOKS_DIR="$(cd "$(dirname "$0")/.." && pwd)"
+PASS=0
+FAIL=0
+
+run_test() {
+  local name="$1"
+  local expected_substr="$2"
+  local must_not_contain="$3"
+  local actual="$4"
+
+  # Strip whitespace so that pretty-printed jq output ({"decision": "approve"})
+  # matches the same substring patterns as compact output ({"decision":"approve"}).
+  local actual_compact
+  actual_compact=$(printf '%s' "$actual" | tr -d '[:space:]')
+
+  if [[ "$actual_compact" == *"$expected_substr"* ]]; then
+    if [ -n "$must_not_contain" ] && [[ "$actual_compact" == *"$must_not_contain"* ]]; then
+      printf "FAIL: %s\n  expected '%s', forbidden '%s' both present\n  got: %s\n" \
+        "$name" "$expected_substr" "$must_not_contain" "$actual"
+      FAIL=$((FAIL + 1))
+      return
+    fi
+    printf "PASS: %s\n" "$name"
+    PASS=$((PASS + 1))
+  else
+    printf "FAIL: %s\n  expected substring: %s\n  got: %s\n" \
+      "$name" "$expected_substr" "$actual"
+    FAIL=$((FAIL + 1))
+  fi
+}
+
+# --- Sandbox: isolate state dirs so existing markers don't leak in ---
+TMP_STATE=$(mktemp -d)
+trap 'rm -rf "$TMP_STATE"' EXIT
+export HC_GATEGUARD_STATE_DIR="$TMP_STATE/gateguard"
+export HC_TASKGUARD_STATE_DIR="$TMP_STATE/taskguard"
+export HC_CONFIDENCE_STATE_DIR="$TMP_STATE/confidence"
+mkdir -p "$HC_GATEGUARD_STATE_DIR" "$HC_TASKGUARD_STATE_DIR" "$HC_CONFIDENCE_STATE_DIR"
+
+# Make sure no stale agent markers / subagent env interferes.
+unset CLAUDE_HARNESS_ROLE
+export HC_AGENT_MARKER_DIR="$TMP_STATE/agent-markers"
+mkdir -p "$HC_AGENT_MARKER_DIR"
+
+# === F1 (gateguard) ===
+F1_INPUT_BLOCK='{"tool_name":"Edit","tool_input":{"file_path":"/tmp/some/protected/foo.ts"}}'
+
+# Case 1: ECC_F1_OFF=1 → approve
+out=$(printf '%s' "$F1_INPUT_BLOCK" | ECC_F1_OFF=1 bash "$HOOKS_DIR/gateguard.sh" Edit 2>/dev/null)
+run_test "F1: ECC_F1_OFF=1 yields approve" \
+  '"decision":"approve"' "" "$out"
+
+# Case 2: ECC_F1_OFF unset → existing logic runs (block path triggers because
+# this is a fresh state dir / first edit). We assert it did NOT take the
+# OFF early-exit branch by checking the OFF reason string is absent.
+unset ECC_F1_OFF
+out=$(printf '%s' "$F1_INPUT_BLOCK" | bash "$HOOKS_DIR/gateguard.sh" Edit 2>/dev/null)
+run_test "F1: unset preserves existing logic" \
+  '' "ECC_F1_OFF" "$out"
+# Also assert it actually emitted SOMETHING (block JSON or {} per existing logic):
+if [ -z "$out" ]; then
+  printf "FAIL: F1 unset case produced empty output\n"
+  FAIL=$((FAIL + 1))
+  PASS=$((PASS - 1))
+fi
+
+# === F2 (task-rule-guard) ===
+F2_INPUT='{"tool_name":"Write","tool_input":{"file_path":"/tmp/some/repo/docs/tasks/task-99-nodraft.md"}}'
+
+# Case 3: ECC_F2_OFF=1 → approve
+out=$(printf '%s' "$F2_INPUT" | ECC_F2_OFF=1 bash "$HOOKS_DIR/task-rule-guard.sh" Write 2>/dev/null)
+run_test "F2: ECC_F2_OFF=1 yields approve" \
+  '"decision":"approve"' "" "$out"
+
+# Case 4: ECC_F2_OFF unset → OFF marker absent
+unset ECC_F2_OFF
+out=$(printf '%s' "$F2_INPUT" | bash "$HOOKS_DIR/task-rule-guard.sh" Write 2>/dev/null)
+run_test "F2: unset preserves existing logic" \
+  '' "ECC_F2_OFF" "$out"
+
+# === F3 (confidence-gate) ===
+# Use a transcript_path pointing at a non-existent file → existing logic
+# fail-opens with `{}`. With ECC_F3_OFF=1 we should get an approve JSON instead.
+F3_INPUT='{"transcript_path":"/nonexistent/file.jsonl"}'
+
+# Case 5: ECC_F3_OFF=1 → approve
+out=$(printf '%s' "$F3_INPUT" | ECC_F3_OFF=1 bash "$HOOKS_DIR/confidence-gate.sh" 2>/dev/null)
+run_test "F3: ECC_F3_OFF=1 yields approve" \
+  '"decision":"approve"' "" "$out"
+
+# Case 6: ECC_F3_OFF unset → OFF marker absent (logic falls through to fail-open `{}`)
+unset ECC_F3_OFF
+out=$(printf '%s' "$F3_INPUT" | bash "$HOOKS_DIR/confidence-gate.sh" 2>/dev/null)
+run_test "F3: unset preserves existing logic" \
+  '' "ECC_F3_OFF" "$out"
+
+# --- summary ---
+printf "\n--- test-gate-disable.sh: %d pass / %d fail ---\n" "$PASS" "$FAIL"
+[ "$FAIL" -eq 0 ]