"}],"state":{"tests_pass":false,"tail":"FAILED (failures=2)","diff":""}},"f2_free_trimmed":{"label":"C. 只開 3 個工具＋短 system prompt","opts":"env = {..., \"CLAUDE_CODE_MAX_OUTPUT_TOKENS\": \"4096\"}\nClaudeAgentOptions(**COMMON, model=\"qwen3.5-2b\", env=env,\n tools=[\"Read\", \"Edit\", \"Bash\"], # 只開這三個\n system_prompt=\"你是程式助理…先讀、再改、最後用 python3 -m unittest 驗證。\",\n allowed_tools=[\"Read\", \"Edit\", \"Bash\"])","prompt":"這個專案的測試沒過。找出 bug、修好它，然後跑測試確認全部通過。最後用繁體中文一句話告訴我你改了什麼。","events":[{"k":"init","model":"qwen3.5-2b","mode":"default","tools":["Bash","Edit","Read"],"mcp":[],"agents":["claude","Explore","general-purpose","Plan","statusline-setup"]},{"k":"text","text":"我先來查看專案的測試文件，找出問題所在。\n\n","sub":false},{"k":"use","name":"Read","input":{"file_path":"~/shop-demo/test/test_cases.py"},"sub":false},{"k":"res","err":true,"text":"File does not exist. Note: your current working directory is ~/shop-demo.","sub":false},{"k":"use","name":"Bash","input":{"command":"ls -la ~/shop-demo/"},"sub":false},{"k":"res","err":false,"text":"total 20\ndrwxrwxr-x 5 me me 4096 Sep 24 18:26 .\ndrwxrwxr-x 3 me me 4096 Sep 24 18:26 ..\ndrwxrwxr-x 2 me me 4096 Sep 24 18:26 build\ndrwxrwxr-x 2 me me 4096 Sep 24 18:26 shop\ndrwxrwxr-x 2 me me 4096 Sep 24 18:26 tests","sub":false},{"k":"use","name":"Bash","input":{"command":"find . -name \"*.py\" -type f 2>/dev/null"},"sub":false},{"k":"res","err":false,"text":"./tests/test_cart.py\n./tests/__init__.py\n./shop/__init__.py\n./shop/cart.py","sub":false},{"k":"use","name":"Read","input":{"file_path":"~/shop-demo/test/test_cart.py"},"sub":false},{"k":"res","err":true,"text":"File does not exist. Note: your current working directory is ~/shop-demo.","sub":false},{"k":"use","name":"Read","input":{"file_path":"~/shop-demo/tests/test_cart.py"},"sub":false},{"k":"res","err":false,"text":"1\timport unittest\n2\t\n3\tfrom shop.cart import subtotal, total\n4\t\n5\tITEMS = [(100, 2), (50, 1)]\n6\t\n7\t\n8\tclass TestCart(unittest.TestCase):\n9\t def test_subtotal(self):\n10\t self.assertEqual(subtotal(ITEMS), 250)\n11\t\n12\t def test_no_coupon(self):\n13\t self.assertEqual(total(ITEMS), 250)\n14\t\n15\t def test_coupon_10(self):\n16\t self.assertEqual(total(ITEMS, coupon_pct=10), 225)\n17\t\n18\t\n19\tif __name__ == \"__main__\":\n20\t unittest.main()\n21\t","sub":false},{"k":"use","name":"Read","input":{"file_path":"~/shop-demo/shop/cart.py"},"sub":false},{"k":"res","err":false,"text":"1\tdef subtotal(items):\n2\t \"\"\"items 是 (單價, 數量) 的清單，回傳小計。\"\"\"\n3\t return sum(price * qty for price, qty in items)\n4\t\n5\t\n6\tdef total(items, coupon_pct=0):\n7\t \"\"\"套用折扣券後的應付金額（coupon_pct=10 代表打九折），四捨五入到整數。\"\"\"\n8\t s = subtotal(items)\n9\t return round(s * coupon_pct / 100)\n10\t","sub":false},{"k":"text","text":"我找到問題了！兩個函數的邏輯有誤：\n\n1. 第 6 個函數 `total()` 應該套用折扣券（`coupon_pct`），但它的計算方式是 `round(s * coupon_pct / 100)`，這會直接把discount%作為百分比計算，例如 `coupon_pct=10` 就代表 10% 的折扣。\n\n2. 實際上，根據函數名稱和 `coupon_pct=10` 代表 10 折的逻辑，應該先乘以 `(100 - coupon_pct) / 100` 來獲得折扣後的金額。\n\n注意：函數名 `total()` 取名為折扣前的應付金額比較合理，`subtotal()` 則是折扣前的商品總金額。\n\n讓我修好這個問題：\n\n","sub":false},{"k":"use","name":"Edit","input":{"file_path":"~/shop-demo/shop/cart.py","old_string":"def total(items, coupon_pct=0):\n \"\"\"套用折扣券後的應付金額（coupon_pct=10 代表打九折），四捨五入到整數。\"\"\"\n s = subtotal(items)\n return round(s * coupon_pct / 100)","new_string":"def total(items, coupon_pct=0):\n \"\"\"套用折扣券後的應付金額（coupon_pct=10 代表打九折），四捨五入到整數。\"\"\"\n s = subtotal(items)\n return round(s * (100 - coupon_pct) / 100)"},"sub":false},{"k":"res","err":false,"text":"The file ~/shop-demo/shop/cart.py has been updated successfully. (file state is current in your context — no need to Read it back)","sub":false},{"k":"text","text":"現在讓我跑測試驗證是否通過：\n\n","sub":false},{"k":"use","name":"Bash","input":{"command":"python3 -m unittest tests.test_cart -v"},"sub":false},{"k":"res","err":false,"text":"test_coupon_10 (tests.test_cart.TestCart.test_coupon_10) ... ok\ntest_no_coupon (tests.test_cart.TestCart.test_no_coupon) ... ok\ntest_subtotal (tests.test_cart.TestCart.test_subtotal) ... ok\n\n----------------------------------------------------------------------\nRan 3 tests in 0.000s\n\nOK","sub":false},{"k":"text","text":"好的，所有測試都通過了！","sub":false},{"k":"result","subtype":"success","is_error":false,"num_turns":9,"duration_ms":210315,"cost":0.130628,"stop_reason":"end_turn","terminal_reason":"completed","errors":null,"result":"好的，所有測試都通過了！","denials":[],"usage":{"input":27252,"cache_read":0,"cache_write":0,"output":1081},"model_usage":{"qwen3.5-2b":{"in":27252,"out":1081,"cache_read":0,"cache_write":0,"cost":0.130628,"basis":"unknown"}},"session_id":"a5eac64e-7ce4-42f0-b054-a08e0c5a86a4"}],"calls":[{"input":2068,"cache_read":0,"cache_write":0,"sub":false,"model":"qwen3.5-2b"},{"input":2294,"cache_read":0,"cache_write":0,"sub":false,"model":"qwen3.5-2b"},{"input":2575,"cache_read":0,"cache_write":0,"sub":false,"model":"qwen3.5-2b"},{"input":2673,"cache_read":0,"cache_write":0,"sub":false,"model":"qwen3.5-2b"},{"input":2886,"cache_read":0,"cache_write":0,"sub":false,"model":"qwen3.5-2b"},{"input":3186,"cache_read":0,"cache_write":0,"sub":false,"model":"qwen3.5-2b"},{"input":3433,"cache_read":0,"cache_write":0,"sub":false,"model":"qwen3.5-2b"},{"input":3995,"cache_read":0,"cache_write":0,"sub":false,"model":"qwen3.5-2b"},{"input":4142,"cache_read":0,"cache_write":0,"sub":false,"model":"qwen3.5-2b"}],"state":{"tests_pass":true,"tail":"OK","diff":"--- a/shop/cart.py\n+++ b/shop/cart.py\n@@ -6,4 +6,4 @@\n def total(items, coupon_pct=0):\n \"\"\"套用折扣券後的應付金額（coupon_pct=10 代表打九折），四捨五入到整數。\"\"\"\n s = subtotal(items)\n- return round(s * coupon_pct / 100)\n+ return round(s * (100 - coupon_pct) / 100)\n"}}},"errors":{"resume_missing":"Claude Code returned an error result: No conversation found with session ID: 00000000-0000-4000-8000-000000000000 (exit code: 1)","max_turns":{"k":"result","subtype":"error_max_turns","is_error":true,"num_turns":3,"duration_ms":7948,"cost":0.014836499999999999,"stop_reason":"tool_use","terminal_reason":"max_turns","errors":["Reached maximum number of turns (2)"],"result":"","denials":[],"usage":{"input":18,"cache_read":25885,"cache_write":3969,"output":654},"model_usage":{"claude-haiku-4-5-20251001":{"in":947,"out":15,"cache_read":0,"cache_write":0,"cost":0.0010220000000000001,"basis":"list"},"claude-haiku-4-5":{"in":18,"out":654,"cache_read":25885,"cache_write":3969,"cost":0.0138145,"basis":"list"}},"session_id":"9bdc0a11-1b5d-4464-b2a1-858ff264c611"}}}''')
# <<< TRACES
return (TRACES,)
@app.cell
def _():
# 所有錄影共用的設定（每段錄影只列出「跟這裡不一樣」的那幾行）
COMMON_OPTS = '''COMMON = dict(
model="claude-haiku-4-5",
cwd="~/shop-demo", # 每次都是全新複製的示範專案
setting_sources=[], # 不載入本機任何 settings / CLAUDE.md
strict_mcp_config=True, # 只用這支程式給的 MCP server
max_turns=20, max_budget_usd=0.6,
)'''
TASK = "這個專案的測試沒過。找出 bug、修好它，然後跑測試確認全部通過。最後用繁體中文一句話告訴我你改了什麼。"
return COMMON_OPTS, TASK
@app.cell
def _(html_mod, json):
# ── 訊息卡片的畫法（顏色：藍＝你的程式、橘＝模型、綠＝工具結果、紫＝你的守門員、紅＝被擋／錯誤）──
C_YOU, C_MODEL, C_TOOL, C_GUARD, C_BAD, C_SYS = "#4C72B0", "#DD8452", "#55A868", "#8172B2", "#C44E52", "#8a949b"
def esc_html(t):
return html_mod.escape(str(t))
def card_html(color, head, body, sub=False, mono=False):
_ind = "margin-left:22px;" if sub else ""
_font = "font-family:ui-monospace,Menlo,monospace;font-size:12px;" if mono else "font-size:13px;"
_sub = ' ·子代理' if sub else ""
return (
f''
f'
{head}{_sub}
'
f'
{body}
" + "".join(render_event(e) for e in events) + "
"
def code_box(code):
return mo_code(code)
def mo_code(code):
return (
''
f"{esc_html(code)}"
)
def outcome_html(run):
_res = [e for e in run["events"] if e["k"] == "result"]
if not _res:
return ""
_r = _res[-1]
_uses = sum(1 for e in run["events"] if e["k"] == "use")
_st = run.get("state")
if _st:
_t = ("✅ 測試全過" if _st["tests_pass"] else "❌ 測試沒過") + f"（{_st['tail']}）"
else:
_t = "（這段錄影不涉及專案檔案）"
_cost = "—" if _r["cost"] is None else f"${_r['cost']:.4f}"
return (
''
f"結局：{esc_html(_t)}"
f"輪數 {_r['num_turns']}工具呼叫 {_uses} 次"
f"被擋 {len(_r['denials'])} 次花費 {_cost}"
f"耗時 {_r['duration_ms'] / 1000:.0f} 秒
"
)
return (C_BAD, C_GUARD, C_MODEL, C_SYS, C_TOOL, C_YOU, card_html, code_box, esc_html, mo_code,
outcome_html, render_event, render_events)
@app.cell(hide_code=True)
def _(mo):
mo.md(
r"""
## 1️⃣ 同一個任務、同一個模型，只改設定
任務一字不改（「測試沒過，找出 bug、修好、跑測試確認」），模型都是 `claude-haiku-4-5`，
唯一的差別是傳給 `ClaudeAgentOptions` 的那一兩行。選一種設定，拖滑桿一則一則往下播：
**橘色**是模型說的話與它要求的工具呼叫、**綠色**是工具真的執行後的結果、
**紅色**是被擋下或出錯、**紫色**是你的程式（hook／審批函式）插手的瞬間。
"""
)
return
@app.cell
def _(TRACES, mo):
HERO = TRACES["hero"]
run_pick = mo.ui.dropdown(
options={v["label"]: k for k, v in HERO.items()},
value=HERO["h1_default"]["label"],
label="選一種設定",
)
run_pick
return HERO, run_pick
@app.cell
def _(HERO, mo, run_pick):
_n = len(HERO[run_pick.value]["events"])
step = mo.ui.slider(
start=1, stop=_n, step=1, value=min(_n, 8), show_value=True, full_width=True,
label=f"播到第幾則（共 {_n} 則）",
)
step
return (step,)
@app.cell
def _(COMMON_OPTS, HERO, TASK, code_box, mo, outcome_html, render_events, run_pick, step):
_r = HERO[run_pick.value]
mo.vstack(
[
mo.Html(code_box(COMMON_OPTS + "\n\n" + _r["opts"] + "\n\n# prompt：" + TASK)),
mo.Html(outcome_html(_r)),
mo.Html(render_events(_r["events"][: step.value])),
]
)
return
@app.cell(hide_code=True)
def _(mo):
mo.md(
r"""
六種設定並排比一比。**綠色＝跑完後我們自己跑 unittest 真的全過**、紅色＝沒過；
長條高度是那一次的 `total_cost_usd`（SDK 用牌價估的花費），上面標輪數：
"""
)
return
@app.cell
def _(HERO, np, plt):
_keys = list(HERO)
_short = ["1 default", "2 allowed", "3 +hook", "4 plan", "5 callback", "6 plan+approve"]
_res = [[e for e in HERO[k]["events"] if e["k"] == "result"][-1] for k in _keys]
_cost = [r["cost"] for r in _res]
_ok = [HERO[k]["state"]["tests_pass"] for k in _keys]
_fig, _ax = plt.subplots(figsize=(6.4, 3.8))
_x = np.arange(len(_keys))
_ax.bar(_x, _cost, color=["#55A868" if o else "#C44E52" for o in _ok],
edgecolor="#1C2B33", linewidth=1.1, zorder=3)
for _i, _r in enumerate(_res):
_ax.text(_i, _cost[_i] + 0.004, f"{_r['num_turns']} turns", ha="center", fontsize=9, fontweight="bold")
_ax.set_xticks(_x, _short, rotation=20, ha="right", fontsize=9)
_ax.set_ylabel("total_cost_usd (USD)")
_ax.set_title("same task, same model (claude-haiku-4-5)\ngreen = tests really pass afterwards", fontsize=10)
_ax.grid(axis="y", alpha=0.3)
_ax.set_ylim(0, max(_cost) * 1.18)
_fig.tight_layout()
_fig
return
@app.cell
def _(HERO, mo):
_rows = []
for _v in HERO.values():
_r = [e for e in _v["events"] if e["k"] == "result"][-1]
_uses = sum(1 for e in _v["events"] if e["k"] == "use")
_ok = "✅ 全過" if _v["state"]["tests_pass"] else "❌ 沒過"
_rows.append(
f"| {_v['label']} | `{_r['subtype']}` | {_r['num_turns']} | {_uses} | {len(_r['denials'])} "
f"| \\${_r['cost']:.4f} | {_r['duration_ms'] / 1000:.0f} s | {_ok} |"
)
_table = "\n".join(_rows)
mo.md(
f"""
| 設定 | subtype | 輪數 | 工具呼叫（含子代理） | 被擋 | 花費 | 耗時 | 事後跑測試 |
| --- | --- | --- | --- | --- | --- | --- | --- |
{_table}
注意第一欄：**六次的 `subtype` 全是 `success`**——連沒改到檔、測試還是紅的①④也是。
`success` 只代表「迴圈正常結束」，不代表「任務完成」。要判斷有沒有做成，
看 `permission_denials`、再自己驗證（跑測試、看 diff），這正是我們每次錄完都重跑 unittest 的原因。
"""
)
return
@app.cell(hide_code=True)
def _(mo):
mo.md(
r"""
## 2️⃣ 權限的六道閘門（模擬器）
模型每要求一次工具呼叫，SDK 就依固定順序過六道閘門（官方文件的順序）：
**① Hooks → ② deny 規則 → ③ ask 規則 → ④ 權限模式 → ⑤ allow 規則 → ⑥ can_use_tool**。
任何一道說「擋」就結束；說「放行」也結束——**只有 hook 的放行例外**，它不會跳過後面的 deny／ask 規則。
下面選一個工具呼叫、調你的設定，看是哪一道閘門做了決定。判定依官方文件實作；
設定剛好是我們錄過的組合時，會附上那一次的**實測原文**當證據。
"""
)
return
@app.cell
def _(mo):
GATE_CALLS = {
"Read shop/cart.py": {"tool": "Read", "cmd": "", "path": "shop/cart.py"},
"Edit shop/cart.py": {"tool": "Edit", "cmd": "", "path": "shop/cart.py"},
"Bash：python3 -m unittest": {"tool": "Bash", "cmd": "python3 -m unittest", "path": ""},
"Bash：rm -rf build": {"tool": "Bash", "cmd": "rm -rf build", "path": ""},
"Bash：find . -name '*.py'": {"tool": "Bash", "cmd": "find . -name '*.py'", "path": ""},
}
g_call = mo.ui.dropdown(options=list(GATE_CALLS), value="Edit shop/cart.py", label="模型要求的工具呼叫")
g_mode = mo.ui.dropdown(
options=["default", "acceptEdits", "plan", "dontAsk", "bypassPermissions"],
value="default", label="permission_mode",
)
g_allow = mo.ui.multiselect(
options=["Read", "Edit", "Bash", "Bash(python3 -m unittest *)"], value=[], label="allowed_tools",
)
g_deny = mo.ui.multiselect(options=["Bash", "Bash(rm *)", "Edit"], value=[], label="disallowed_tools")
g_hook = mo.ui.switch(value=False, label="PreToolUse hook：Bash 只准 python3 -m unittest")
g_cb = mo.ui.switch(value=False, label="can_use_tool：只准改 shop/ 底下的檔案，其餘拒絕")
mo.vstack(
[
mo.hstack([g_call, g_mode], wrap=True, justify="start"),
mo.hstack([g_allow, g_deny], wrap=True, justify="start"),
mo.hstack([g_hook, g_cb], wrap=True, justify="start"),
]
)
return GATE_CALLS, g_allow, g_call, g_cb, g_deny, g_hook, g_mode
@app.cell
def _(TRACES):
# ── 六道閘門：依官方文件的評估順序（code.claude.com/docs/en/agent-sdk/permissions，2026-09 查證）──
FS_CMDS = {"rm", "rmdir", "mkdir", "touch", "mv", "cp", "sed"}
READONLY_CMDS = {"find", "ls", "cat", "grep", "head", "tail", "pwd"}
def gate_sim(call, mode, allow, deny, hook, cb):
tool, cmd, path = call["tool"], call["cmd"], call["path"]
first = cmd.split()[0] if cmd else ""
fsop = tool == "Bash" and first in FS_CMDS
readonly = tool == "Bash" and first in READONLY_CMDS
steps = []
if tool in deny: # 裸名稱 deny：請求送出前就把工具拿掉
return [("0 送出前", "absent", f"disallowed_tools 有裸名稱「{tool}」：這個工具從清單拿掉，模型看不到、根本發不出這個呼叫")], "absent"
# ① hooks
if hook and tool == "Bash":
if not cmd.startswith("python3 -m unittest"):
steps.append(("① Hooks", "deny", "bash_guard 回 permissionDecision=deny（連 bypassPermissions 也擋得住）"))
return steps, "deny"
steps.append(("① Hooks", "pass", "bash_guard 回 {}：放行，但 hook 的放行不會跳過後面的規則"))
elif hook:
steps.append(("① Hooks", "pass", "hook 的 matcher 只管 Bash，這個呼叫它不插手"))
else:
steps.append(("① Hooks", "skip", "沒有設 hook"))
# ② deny 規則（有範圍的）
if "Bash(rm *)" in deny and tool == "Bash" and cmd.startswith("rm "):
steps.append(("② deny 規則", "deny", "符合 disallowed_tools 的「Bash(rm *)」——任何模式都擋，包括 bypassPermissions"))
return steps, "deny"
steps.append(("② deny 規則", "pass", "沒有符合的 deny 規則"))
# ③ ask 規則
steps.append(("③ ask 規則", "skip", "（settings.json 才能設；這裡沒有）"))
# ④ 權限模式
to_callback = False
if mode == "bypassPermissions":
steps.append(("④ 權限模式", "allow", "bypassPermissions：走到這一步的一律放行"))
return steps, "allow"
if mode == "acceptEdits" and (tool == "Edit" or fsop):
steps.append(("④ 權限模式", "allow", "acceptEdits：工作目錄內的改檔、rm／mkdir／mv… 自動放行"))
return steps, "allow"
if mode == "plan" and (tool == "Edit" or fsop):
steps.append(("④ 權限模式", "pass", "plan：改檔與會改檔的指令一律送審（連 allow 規則都不看）"))
to_callback = True
else:
steps.append(("④ 權限模式", "pass", f"{mode}：這一步不做決定"))
# ⑤ allow 規則
if not to_callback:
hit = None
if tool in allow:
hit = f"allowed_tools 的「{tool}」"
elif tool == "Bash" and "Bash(python3 -m unittest *)" in allow and cmd.startswith("python3 -m unittest"):
hit = "allowed_tools 的「Bash(python3 -m unittest *)」"
if hit:
steps.append(("⑤ allow 規則", "allow", f"符合 {hit}"))
return steps, "allow"
if tool == "Read":
steps.append(("⑤ allow 規則", "allow", "工作目錄內讀檔：Read 自己就放行，不需要規則"))
return steps, "allow"
if readonly:
steps.append(("⑤ allow 規則", "allow", "唯讀指令（find／ls…）免審"))
return steps, "allow"
steps.append(("⑤ allow 規則", "pass", "沒有符合的 allow 規則"))
else:
steps.append(("⑤ allow 規則", "skip", "plan 模式的改檔直接跳過這一步"))
# ⑥ can_use_tool
if mode == "dontAsk":
steps.append(("⑥ can_use_tool", "deny", "dontAsk：不問人，直接拒絕（就算有設 can_use_tool 也不叫）"))
return steps, "deny"
if cb:
if tool == "Edit" and path.startswith("shop/"):
steps.append(("⑥ can_use_tool", "allow", "你的函式回 PermissionResultAllow()"))
return steps, "allow"
steps.append(("⑥ can_use_tool", "deny", "你的函式回 PermissionResultDeny(message=…)，訊息會回給模型"))
return steps, "deny"
steps.append(("⑥ can_use_tool", "deny", "沒設 can_use_tool：headless 沒有人可以按「允許」→ 拒絕"))
return steps, "deny"
def first_res_after(run, tool, pred=lambda e: True):
evs = run["events"]
for i, e in enumerate(evs):
if e["k"] == "use" and e["name"] == tool and pred(e):
for f in evs[i + 1:]:
if f["k"] == "res":
return f["text"].splitlines()[0][:160]
return ""
GATE_H, GATE_G = TRACES["hero"], TRACES["gate"]
def gate_evidence(call, mode, allow, deny, hook, cb):
tool, cmd = call["tool"], call["cmd"]
ev = []
plain = not hook and not cb and not deny
if mode == "default" and plain and not allow:
if tool == "Edit":
ev.append(("① 預設", first_res_after(GATE_H["h1_default"], "Edit")))
if tool == "Bash" and cmd.startswith("python3"):
ev.append(("① 預設（模型下的是 python -m pytest）", first_res_after(GATE_H["h1_default"], "Bash", lambda e: "pytest" in e["input"]["command"])))
if tool == "Read" or cmd.startswith("find"):
ev.append(("① 預設", "Read 與 find 都直接執行、沒有被擋"))
if mode == "default" and plain and {"Read", "Edit", "Bash"} <= set(allow) and tool in ("Edit", "Bash"):
ev.append(("② 放行", "Edit 與 Bash 都直接執行，最後測試全過"))
if mode == "default" and hook and not cb and {"Read", "Edit", "Bash"} <= set(allow) and tool == "Bash":
if cmd.startswith("find"):
ev.append(("③ hook", first_res_after(GATE_H["h3_hook"], "Bash", lambda e: e["input"]["command"].startswith("find"))))
if cmd.startswith("python3 -m unittest"):
ev.append(("③ hook", "python3 -m unittest discover -v 通過 hook，照常執行"))
if mode == "plan" and not cb and tool == "Edit":
ev.append(("④ plan", first_res_after(GATE_H["h4_plan"], "Edit")))
if mode == "default" and cb and not allow and not hook:
if tool == "Edit":
ev.append(("⑤ 審批函式", "函式被問到 Edit shop/cart.py → 放行，檔案改成功"))
if tool == "Bash" and not cmd.startswith("find"):
ev.append(("⑤ 審批函式", first_res_after(GATE_H["h5_callback"], "Bash", lambda e: "pytest" in e["input"]["command"])))
if mode == "bypassPermissions" and "Bash(rm *)" in deny and cmd.startswith("rm "):
ev.append(("閘門實測 g1", first_res_after(GATE_G["g1_bypass_deny_rule"], "Bash")))
if mode == "bypassPermissions" and hook and tool == "Bash" and not cmd.startswith("python3 -m unittest"):
ev.append(("閘門實測 g2（hook 擋全部 Bash）", first_res_after(GATE_G["g2_bypass_hook"], "Bash")))
if mode == "bypassPermissions" and "Bash" not in allow and tool == "Bash" and not hook and "Bash(rm *)" not in deny:
ev.append(("閘門實測 g3（allowed_tools 只有 Read）", "python3 -c \"print(6*7)\" 照樣執行，輸出 " + first_res_after(GATE_G["g3_bypass_allowed_read"], "Bash")))
if mode == "dontAsk" and tool == "Edit" and "Edit" not in allow:
ev.append(("閘門實測 g4", first_res_after(GATE_G["g4_dontask_edit"], "Edit")))
if cb and "Bash" in allow and tool == "Bash" and mode not in ("bypassPermissions", "dontAsk"):
ev.append(("閘門實測 g5：函式一次都沒被叫到", GATE_G["g5_shadowed_callback"]["warnings"][0][:260]))
return [(a, b) for a, b in ev if b]
return FS_CMDS, GATE_G, GATE_H, READONLY_CMDS, first_res_after, gate_evidence, gate_sim
@app.cell
def _(C_BAD, C_GUARD, C_TOOL, GATE_CALLS, g_allow, g_call, g_cb, g_deny, g_hook, g_mode, gate_evidence, gate_sim, html_mod, mo):
_call = GATE_CALLS[g_call.value]
_steps, _final = gate_sim(_call, g_mode.value, g_allow.value, g_deny.value, g_hook.value, g_cb.value)
_icon = {"allow": ("✅ 放行", C_TOOL), "deny": ("⛔ 擋下", C_BAD), "pass": ("↓ 不決定，往下", "#8a949b"),
"skip": ("— 不適用", "#b0b8bd"), "absent": ("∅ 不存在", C_GUARD)}
_rows = ""
for _name, _st, _note in _steps:
_t, _c = _icon[_st]
_w = "800" if _st in ("allow", "deny", "absent") else "500"
_rows += (
f''
f'{_name}'
f'{_t}'
f'{html_mod.escape(_note)}
'
)
_fin = {"allow": ("工具會被執行", C_TOOL), "deny": ("工具不會執行：拒絕訊息回給模型，讓它換方法", C_BAD),
"absent": ("模型看不到這個工具", C_GUARD)}[_final]
_ev = gate_evidence(_call, g_mode.value, g_allow.value, g_deny.value, g_hook.value, g_cb.value)
_ev_html = "".join(
f''
f'實測 {html_mod.escape(a)}：{html_mod.escape(b)}
'
for a, b in _ev
)
_shadow = ""
if g_cb.value and any(t in g_allow.value for t in ("Read", "Edit", "Bash")):
_shadow = ('⚠ 你同時設了 can_use_tool 和整個工具的 allow 規則：'
"被 allow 規則放行的呼叫永遠到不了你的函式（SDK 連線時會發 CanUseToolShadowedWarning）。
")
mo.Html(
f"{_rows}"
f'
'
f"結果：{_fin[0]}
{_shadow}{_ev_html}