tools: add tool script safety guard

examples/tool_safety/DESIGN.md

@@ -0,0 +1,148 @@
+# Tool Script Safety Guard 设计说明
+
+## 目标与边界
+
+Safety Guard 是执行前治理层。它对脚本、命令、参数、工作目录、环境变量名称和 tool 元数据做静态检查，输出 `allow`、`deny` 或 `needs_human_review`。它不能替代进程隔离、只读文件系统、最小权限、网络出口控制和运行时资源限制。
+
+数据流如下：
+
+```text
+Tool / MCP Tool / Skill / CodeExecutor 请求
+        |
+        v
+SafetyScanRequest（不保留环境变量值）
+        |
+        v
+ToolSafetyScanner + ToolSafetyPolicy + SafetyRule[]
+        |
+        v
+SafetyFinding[] --聚合--> SafetyReport
+        |                         |
+        v                         v
+JSONL AuditEvent          Filter / Guard 执行决策
+        |
+        v
+OpenTelemetry tool.safety.* 属性
+```
+
+扫描发生在真实 handler 或 executor delegate 之前。`deny` 必须阻断；`needs_human_review` 是否阻断由 `block_on_review` 决定。人工批准应由上层审批系统显式记录，不能把 review 自动降级为 allow。
+
+## 接入示例
+
+### Tool 与 MCP Tool Filter
+
+Tool filter 从实际 tool 上下文获取名称，只接收参数中的脚本字段，不要求模型伪造 `tool_name`：
+
+```python
+from trpc_agent_sdk.tools import BashTool
+from trpc_agent_sdk.tools.safety import JsonlAuditSink
+from trpc_agent_sdk.tools.safety import ToolSafetyFilter, ToolSafetyGuard, ToolSafetyPolicy
+
+policy = ToolSafetyPolicy.from_yaml("tool_safety_policy.yaml")
+safety_guard = ToolSafetyGuard(policy, audit_sink=JsonlAuditSink("tool_safety_audit.jsonl"))
+safety_filter = ToolSafetyFilter(safety_guard)
+tool = BashTool()
+tool.add_one_filter(safety_filter)
+```
+
+Safety Filter 带有最终授权标记，框架会让其他参数转换 Filter 和 tool callback 先运行，再在最靠近真实 handler 的位置扫描最终参数，避免下游原地修改已扫描内容。
+
+Filter 会在扫描前合并可信的 tool 固定 override、默认 timeout 和默认 cwd。以示例策略的 `max_timeout_seconds: 120` 为例，`BashTool` 缺省的 300 秒 timeout 会被拒绝，调用方必须显式请求不超过策略的值。参数转换、scanner 或报告聚合异常会生成脱敏的 `SCAN-INPUT` / `SCAN-ERROR` deny 报告并记录一次审计事件，不会以普通异常形式绕过审计。
+
+`StreamingProgressTool` 在启动用户 async generator 之前运行同一套有序 Filter 和 tool callback；被拒绝时只返回结构化阻断结果，generator 不会开始执行。最终授权只针对完整组装后的 tool call，不应对早期参数分片做放行判断。
+
+MCP Tool 应在本地代理真正发出 MCP 调用前应用同一个 Filter。远端 MCP server 仍需独立鉴权、最小权限和审计，因为本地静态扫描无法证明远端实现的实际行为。
+
+### Skill
+
+`skill_run` 的 `command`、`cwd`、`env` 名称和 timeout 都应在 workspace runner 启动进程前进入 Filter：
+
+```python
+from trpc_agent_sdk.skills.tools import SkillRunTool
+from trpc_agent_sdk.tools.safety import ToolSafetyFilter
+
+skill_tool = SkillRunTool(repository=repository, filters=[ToolSafetyFilter(safety_guard)])
+```
+
+Skill 内容可能在扫描后被更新，因此还要固定 skill 版本或内容摘要，并在执行环境中限制挂载、网络和凭据。
+
+### CodeExecutor wrapper
+
+CodeExecutor 使用委托包装，逐个保留代码块语言，再聚合报告：
+
+```python
+from trpc_agent_sdk.code_executors import UnsafeLocalCodeExecutor
+from trpc_agent_sdk.tools.safety import SafetyGuardedCodeExecutor
+
+delegate = UnsafeLocalCodeExecutor(timeout=30)
+executor = SafetyGuardedCodeExecutor(inner=delegate, guard=safety_guard)
+```
+
+包装器需要镜像 delegate 的 `stateful`、workspace runtime、delimiter 和重试配置，阻断时不得调用 delegate。
+
+## 规则与策略
+
+Python 使用 AST 提取调用、常量路径和数据流信号；Bash 使用不执行命令的 token/语法模式扫描。每个 finding 至少包含：
+
+- `rule_id` 和风险分类；
+- `risk_level` 与局部决策；
+- 已脱敏、长度受限的 `evidence`；
+- 可执行的 `recommendation`；
+- 可选行列位置和不含秘密值的 metadata。
+
+自定义规则通过 scanner 的 `rules` 参数注入，不修改内置 scanner：
+
+```python
+from trpc_agent_sdk.tools.safety import SafetyDecision, SafetyFinding
+from trpc_agent_sdk.tools.safety import RiskCategory, RiskLevel, ToolSafetyScanner
+
+class DenyInternalBinaryRule:
+    rule_id = "CUSTOM-INTERNAL-BINARY"
+
+    def scan(self, context, policy):
+        if "/internal/bin/" not in context.request.script:
+            return []
+        return [SafetyFinding(
+            rule_id=self.rule_id,
+            category=RiskCategory.PROCESS_EXECUTION,
+            risk_level=RiskLevel.HIGH,
+            decision=SafetyDecision.DENY,
+            evidence="/internal/bin/<redacted>",
+            recommendation="Use an explicitly approved command.",
+        )]
+
+scanner = ToolSafetyScanner(policy, rules=[DenyInternalBinaryRule()])
+```
+
+自定义规则必须是纯静态、确定性且无副作用；异常由 scanner 按 policy 的 fail-closed 行为处理。
+
+## 审计与监控
+
+审计事件至少包含 `tool_name`、`decision`、`risk_level`、`rule_ids`、扫描耗时、`redacted`、`blocked`、人工批准状态、脚本 SHA-256 和策略版本。不得写入脚本文本、环境变量值、Authorization header 或私钥正文。
+
+当前 span 应设置：
+
+- `tool.safety.decision`
+- `tool.safety.risk_level`
+- `tool.safety.rule_id`
+- `tool.safety.rule_ids`
+- `tool.safety.duration_ms`
+- `tool.safety.redacted`
+- `tool.safety.blocked`
+
+策略版本和脚本 SHA-256 只写入审计事件，不写入当前 span，以控制 telemetry 基数。
+
+监控系统可以按 deny/review 比例、rule id、tool name 和扫描失败率告警，但不能把 telemetry 成功视为安全放行条件。
+
+## 已知限制
+
+- **误报**：Bash 的复杂引用、合法管道、安全的 subprocess 和测试夹具可能触发 review；通过窄化白名单或单条 rule action 调整，不应关闭整个 guard。
+- **漏报**：编码、压缩、别名、间接导入、反射、动态属性和多阶段下载执行可能绕过静态模式。
+- **动态代码**：`eval`、`exec`、运行时拼接 URL/路径和下载后的脚本无法在首次扫描时完全解析，应阻断或人工复核，并在下一执行边界重新扫描。
+- **TOCTOU**：扫描后文件、Skill、cwd、符号链接或策略可能变化。执行端应校验内容摘要，固定版本，并尽量在隔离 workspace 内完成扫描和执行。
+- **MCP**：本地只能分析请求参数，无法验证远端 tool 是否执行了额外命令、访问其他数据或正确实施资源限制。
+- **Streaming**：分片参数在结束前可能不是完整语法。只能在完整 tool call 组装后做最终授权，不能因早期分片看似安全而提前执行。
+- **资源限制**：静态规则只能识别明显循环、sleep、fork bomb 和大写入信号；CPU、内存、进程数、磁盘、输出和墙钟时间必须由 sandbox/runtime 强制限制。
+- **运行时注入**：Guard 会检查调用参数、固定 tool override 和可见默认值，但 workspace/repository 在 handler 内部追加的可信环境或文件仍需由运行时白名单、固定配置和 sandbox 约束。
+
+因此生产部署应组合：Safety Guard + Container/Cube sandbox + 最小凭据 + 出网白名单 + 资源限制 + 不可篡改审计。

examples/tool_safety/README.md

@@ -0,0 +1,53 @@
+# Tool Script Safety Guard 示例
+
+本目录提供一套可重复运行的公开验收样本，用于验证 Tool、Skill 和 CodeExecutor 在执行 Python 或 Bash 内容前的静态安全检查。样本只会被读取和扫描，测试不会执行其中的脚本。
+
+## 快速运行
+
+在仓库根目录执行：
+
+```bash
+python scripts/tool_safety_check.py examples/tool_safety/samples \
+  --policy examples/tool_safety/tool_safety_policy.yaml \
+  --report /tmp/tool_safety_report.json \
+  --audit /tmp/tool_safety_audit.jsonl \
+  --tool-name public_sample_scan
+```
+
+CLI 接受任意数量的 `.py`、`.sh`、`.bash` 文件或目录。目录会递归扫描，结果按路径稳定排序。标准输出和 `--report` 都是结构化 JSON；`--audit` 每个文件追加一条已脱敏 JSONL 事件。
+
+仓库同时保留一份由上述 12 个样本真实生成的 [`tool_safety_report.json`](tool_safety_report.json) 和 [`tool_safety_audit.jsonl`](tool_safety_audit.jsonl)。报告不含原始脚本或环境变量值，审计文件恰好每个样本一条事件。
+
+退出码遵循最严格决策：
+
+| 退出码 | 决策 |
+| --- | --- |
+| `0` | `allow` |
+| `1` | `deny` |
+| `2` | `needs_human_review` |
+
+多文件扫描按 `deny > needs_human_review > allow` 聚合。策略读取失败、输入不可读或扫描基础设施异常时，CLI 采用 fail-closed 并返回 `deny`。
+
+## 公开样本
+
+[`samples/manifest.yaml`](samples/manifest.yaml) 是唯一真值源，列出恰好 12 个样本的预期决策和至少应命中的规则。场景包括安全 Python、危险删除、读取 SSH 私钥、非白名单与白名单网络请求、subprocess、shell 注入、依赖安装、无限循环、敏感信息输出、Bash 管道和人工复核。
+
+测试会逐个调用 CLI，并校验：
+
+- 高危样本检出率不低于 90%；
+- 安全样本误报率不高于 10%；
+- 危险删除、读取密钥、非白名单外连检出率均为 100%；
+- JSON 报告和 JSONL 审计可以被标准解析器读取。
+
+## 策略修改
+
+[`tool_safety_policy.yaml`](tool_safety_policy.yaml) 采用严格字段校验。未知字段、错误类型和非法阈值会直接失败，避免拼写错误造成静默放行。修改以下字段不需要改代码：
+
+- `allowed_domains`：精确域名和子域名白名单；
+- `allowed_commands`：允许的可执行文件 basename 或精确路径；basename 白名单不会放行 `./git` 等相对路径；
+- `denied_paths`：禁止路径及 glob；
+- `max_timeout_seconds`、`max_output_bytes`、`max_script_bytes`：资源上限；
+- `long_sleep_seconds`、`max_concurrency`：可疑资源使用阈值；
+- `rule_actions`：按 rule id 覆盖默认决策。
+
+完整处理流程、接入方式和安全边界见 [`DESIGN.md`](DESIGN.md)。

examples/tool_safety/samples/01_safe_python.py

@@ -0,0 +1,4 @@
+# Static scanner sample; this file is not executed by the example suite.
+
+values = [1, 2, 3]
+print(sum(values))

examples/tool_safety/samples/02_dangerous_delete.sh

@@ -0,0 +1,4 @@
+#!/bin/sh
+# Static scanner sample; do not execute.
+
+rm -rf /

examples/tool_safety/samples/03_read_ssh_key.py

@@ -0,0 +1,6 @@
+# Static scanner sample; do not execute.
+
+from pathlib import Path
+
+private_key = Path("~/.ssh/id_rsa").read_text(encoding="utf-8")
+print(len(private_key))

examples/tool_safety/samples/04_non_allowlisted_network.py

@@ -0,0 +1,5 @@
+# Static scanner sample; do not execute.
+
+import requests
+
+requests.post("https://collector.external.example/upload", data={"status": "ok"}, timeout=5)

examples/tool_safety/samples/05_allowlisted_network.py

@@ -0,0 +1,6 @@
+# Static scanner sample; this file is not executed by the example suite.
+
+import requests
+
+response = requests.get("https://api.example.com/health", timeout=5)
+print(response.status_code)

examples/tool_safety/samples/06_subprocess_call.py

@@ -0,0 +1,5 @@
+# Static scanner sample; do not execute.
+
+import subprocess
+
+subprocess.run(["git", "status"], check=True)

examples/tool_safety/samples/07_shell_injection.py

@@ -0,0 +1,6 @@
+# Static scanner sample; do not execute.
+
+import subprocess
+
+user_value = input("value: ")
+subprocess.run("echo " + user_value, shell=True, check=True)

examples/tool_safety/samples/08_dependency_install.sh

@@ -0,0 +1,4 @@
+#!/bin/sh
+# Static scanner sample; do not execute.
+
+python -m pip install untrusted-example-package

examples/tool_safety/samples/09_infinite_loop.py

@@ -0,0 +1,4 @@
+# Static scanner sample; do not execute.
+
+while True:
+    pass

examples/tool_safety/samples/10_sensitive_output.py

@@ -0,0 +1,3 @@
+# Static scanner sample; do not execute.
+
+print("API_KEY=example_sensitive_value_123456")

examples/tool_safety/samples/11_bash_pipeline.sh

@@ -0,0 +1,4 @@
+#!/bin/sh
+# Static scanner sample; do not execute.
+
+printf '%s\n' alpha beta | grep beta

examples/tool_safety/samples/12_human_review.py

@@ -0,0 +1,6 @@
+# Static scanner sample; do not execute.
+
+import requests
+
+target_url = input("target URL: ")
+requests.get(target_url, timeout=5)

examples/tool_safety/samples/manifest.yaml

@@ -0,0 +1,87 @@
+version: "1"
+policy: ../tool_safety_policy.yaml
+samples:
+  - file: 01_safe_python.py
+    scenario: safe_python
+    category: safe
+    safe: true
+    required_detection: false
+    expected_decision: allow
+    expected_rules: []
+  - file: 02_dangerous_delete.sh
+    scenario: dangerous_delete
+    category: dangerous_file_operation
+    safe: false
+    required_detection: true
+    expected_decision: deny
+    expected_rules: [FILE-DANGEROUS-DELETE]
+  - file: 03_read_ssh_key.py
+    scenario: read_secret_file
+    category: dangerous_file_operation
+    safe: false
+    required_detection: true
+    expected_decision: deny
+    expected_rules: [FILE-DENIED-PATH]
+  - file: 04_non_allowlisted_network.py
+    scenario: non_allowlisted_network
+    category: network_access
+    safe: false
+    required_detection: true
+    expected_decision: deny
+    expected_rules: [NET-NON-WHITELISTED]
+  - file: 05_allowlisted_network.py
+    scenario: allowlisted_network
+    category: safe
+    safe: true
+    required_detection: false
+    expected_decision: allow
+    expected_rules: []
+  - file: 06_subprocess_call.py
+    scenario: subprocess_call
+    category: process_execution
+    safe: false
+    required_detection: true
+    expected_decision: needs_human_review
+    expected_rules: [PROC-SUBPROCESS]
+  - file: 07_shell_injection.py
+    scenario: shell_injection
+    category: process_execution
+    safe: false
+    required_detection: true
+    expected_decision: deny
+    expected_rules: [PROC-SHELL-INJECTION]
+  - file: 08_dependency_install.sh
+    scenario: dependency_install
+    category: dependency_installation
+    safe: false
+    required_detection: true
+    expected_decision: deny
+    expected_rules: [DEP-INSTALL]
+  - file: 09_infinite_loop.py
+    scenario: infinite_loop
+    category: resource_abuse
+    safe: false
+    required_detection: true
+    expected_decision: deny
+    expected_rules: [RES-INFINITE-LOOP]
+  - file: 10_sensitive_output.py
+    scenario: sensitive_output
+    category: sensitive_data_exposure
+    safe: false
+    required_detection: true
+    expected_decision: deny
+    expected_rules: [SECRET-EXPOSURE]
+  - file: 11_bash_pipeline.sh
+    scenario: bash_pipeline
+    category: process_execution
+    safe: false
+    required_detection: true
+    expected_decision: needs_human_review
+    expected_rules: [PROC-PIPELINE]
+  - file: 12_human_review.py
+    scenario: human_review
+    category: network_access
+    safe: false
+    required_detection: true
+    expected_decision: needs_human_review
+    expected_rules: [NET-DYNAMIC-TARGET]

examples/tool_safety/tool_safety_audit.jsonl

@@ -0,0 +1,12 @@
+{"timestamp":"2026-07-12T16:16:57.488210Z","tool_name":"public_sample_scan","decision":"allow","risk_level":"low","rule_id":null,"rule_ids":[],"duration_ms":0.6008749987813644,"redacted":true,"blocked":false,"human_review_approved":false,"script_sha256":"b1efc1f8a3d7bd0b92834054b90e6ec4050e1b860517a6a94615c2221ae1dc2a","policy_version":"1"}
+{"timestamp":"2026-07-12T16:16:57.488336Z","tool_name":"public_sample_scan","decision":"deny","risk_level":"critical","rule_id":"FILE-DANGEROUS-DELETE","rule_ids":["FILE-DANGEROUS-DELETE","POLICY-ARGV-COMMAND"],"duration_ms":0.2756249959929846,"redacted":true,"blocked":true,"human_review_approved":false,"script_sha256":"e561b4bbf19161ca31d57774ac207ebeb9063913551bde1d80f0a5cfe2a42478","policy_version":"1"}
+{"timestamp":"2026-07-12T16:16:57.488379Z","tool_name":"public_sample_scan","decision":"deny","risk_level":"critical","rule_id":"FILE-DENIED-PATH","rule_ids":["FILE-DENIED-PATH","SECRET-EXPOSURE"],"duration_ms":0.3285410057287663,"redacted":true,"blocked":true,"human_review_approved":false,"script_sha256":"4cfacd0db5bf443bf11a0b554ab909be6428e63c081351187858413e4f4dc9cf","policy_version":"1"}
+{"timestamp":"2026-07-12T16:16:57.488414Z","tool_name":"public_sample_scan","decision":"deny","risk_level":"high","rule_id":"NET-NON-WHITELISTED","rule_ids":["NET-NON-WHITELISTED"],"duration_ms":0.20820800273213536,"redacted":true,"blocked":true,"human_review_approved":false,"script_sha256":"33f11abee7dd0655103ed15f3c27677666bb04f797c0c0b7e31be0048dd821b5","policy_version":"1"}
+{"timestamp":"2026-07-12T16:16:57.488446Z","tool_name":"public_sample_scan","decision":"allow","risk_level":"low","rule_id":null,"rule_ids":[],"duration_ms":0.19500000053085387,"redacted":true,"blocked":false,"human_review_approved":false,"script_sha256":"53a3e8a2dced493d5d0118207a338b71e46f3e01a6952a6aa4477df06f87a980","policy_version":"1"}
+{"timestamp":"2026-07-12T16:16:57.488477Z","tool_name":"public_sample_scan","decision":"needs_human_review","risk_level":"high","rule_id":"PROC-SUBPROCESS","rule_ids":["PROC-SUBPROCESS"],"duration_ms":0.15474999963771552,"redacted":true,"blocked":true,"human_review_approved":false,"script_sha256":"415aa546b6024bf3b0e36b4af4b1d88d971eb8e02d756533e9be44f536914984","policy_version":"1"}
+{"timestamp":"2026-07-12T16:16:57.488506Z","tool_name":"public_sample_scan","decision":"deny","risk_level":"critical","rule_id":"PROC-SHELL-INJECTION","rule_ids":["PROC-SUBPROCESS","PROC-SHELL-INJECTION"],"duration_ms":0.20270900131436065,"redacted":true,"blocked":true,"human_review_approved":false,"script_sha256":"576e95fadeb1284ea9a27816bf32d944fa0fc2bf58d36139670830e7d8e1b419","policy_version":"1"}
+{"timestamp":"2026-07-12T16:16:57.488541Z","tool_name":"public_sample_scan","decision":"deny","risk_level":"high","rule_id":"DEP-INSTALL","rule_ids":["DEP-INSTALL"],"duration_ms":0.11441600508987904,"redacted":true,"blocked":true,"human_review_approved":false,"script_sha256":"028b5fb0b3f97323bec90c8648729335cb66fa56eb81a9b662a70cbb4d055309","policy_version":"1"}
+{"timestamp":"2026-07-12T16:16:57.488570Z","tool_name":"public_sample_scan","decision":"deny","risk_level":"critical","rule_id":"RES-INFINITE-LOOP","rule_ids":["RES-INFINITE-LOOP"],"duration_ms":0.0832090008771047,"redacted":true,"blocked":true,"human_review_approved":false,"script_sha256":"a878581bd711f1d7f73f3678df35ae2d6e79ca1e70b942e0b2bad1a5e7b82a96","policy_version":"1"}
+{"timestamp":"2026-07-12T16:16:57.488599Z","tool_name":"public_sample_scan","decision":"deny","risk_level":"critical","rule_id":"SECRET-EXPOSURE","rule_ids":["SECRET-EXPOSURE"],"duration_ms":0.09045899787452072,"redacted":true,"blocked":true,"human_review_approved":false,"script_sha256":"a002bc42f64ba0dbc4950c75076e7635a00fd9f61ead9ab1f723c7b3e64341ac","policy_version":"1"}
+{"timestamp":"2026-07-12T16:16:57.488626Z","tool_name":"public_sample_scan","decision":"needs_human_review","risk_level":"medium","rule_id":"PROC-PIPELINE","rule_ids":["PROC-PIPELINE"],"duration_ms":0.11808300041593611,"redacted":true,"blocked":true,"human_review_approved":false,"script_sha256":"526a9e5468c01bb720b27879fb2a58b3a4934cda91cf0d6de8a3a6e0c4e8730c","policy_version":"1"}
+{"timestamp":"2026-07-12T16:16:57.488672Z","tool_name":"public_sample_scan","decision":"needs_human_review","risk_level":"medium","rule_id":"NET-DYNAMIC-TARGET","rule_ids":["NET-DYNAMIC-TARGET"],"duration_ms":0.1784170017344877,"redacted":true,"blocked":true,"human_review_approved":false,"script_sha256":"64cb6a650c3a3ae9a519752127380a7080d2c8b4fa750ada03c65998b7635de9","policy_version":"1"}