full tests

Author: impself

docs/mkdocs/en/tool_safety_guard.md

@@ -393,7 +393,7 @@ tool/
     _telemetry.py               # OTel span attrs + metrics (no-op safe)
     _tool_adapter.py            # ToolInputAdapter + built-ins
     _filter.py                  # ToolScriptSafetyFilter (terminal)
-  wrapper.py                    # SafetyWrappedCallable, SafetyCheckedExecutor
+    wrapper.py                  # SafetyWrappedCallable, SafetyCheckedExecutor
 scripts/
   tool_safety_check.py          # CLI
 tests/tool_safety/              # safety guard tests
@@ -402,7 +402,7 @@ trpc_agent_sdk/tools/safety/examples/
   samples/                      # 14 public samples + manifest
   tool_safety_report.json       # generated report
   tool_safety_audit.jsonl       # generated audit log
-  manifest_run.json             # manifest execution summary
+  manifest_run.json             # 14 full reports plus expectation checks
 docs/
   tool_safety_guard.md          # this document
   tool_safety_guard.zh_CN.md    # Chinese version

docs/mkdocs/zh/tool_safety_guard.zh_CN.md

@@ -345,13 +345,17 @@ python -m pytest tests/tool_safety/ -v
 ## 文件布局
 
 ```text
-tool/
-  safety/                       # Guard、策略、规则、扫描器、审计和 Telemetry
+trpc_agent_sdk/tools/safety/    # Guard、策略、规则、扫描器、审计和 Telemetry
   wrapper.py                    # SafetyWrappedCallable、SafetyCheckedExecutor
+  examples/
+    tool_safety_policy.yaml     # 示例策略
+    samples/                    # 14 个公开样例及 manifest
+    tool_safety_report.json     # 单次扫描完整报告
+    tool_safety_audit.jsonl     # JSONL 审计日志
+    manifest_run.json           # 14 份完整报告及预期结果核对
 scripts/
   tool_safety_check.py          # CLI
 tests/tool_safety/              # 安全检查器测试
-trpc_agent_sdk/tools/safety/examples/           # 策略、14 个样例、报告和审计样例
 docs/
   tool_safety_guard.md          # English version
   tool_safety_guard.zh_CN.md    # 本文

scripts/tool_safety_check.py

@@ -168,6 +168,7 @@ def _run_manifest(guard: ToolSafetyGuard,
             "risk_level": report.risk_level.label(),
             "duration_ms": report.scan_duration_ms,
             "matches_expected": _decision_matches(report, expected),
+            "report": report.model_dump(mode="json"),
         })
         blocked = report.decision != SafetyDecision.ALLOW
         asyncio.run(_emit_audit(

tests/tool_safety/test_cli.py

@@ -92,6 +92,24 @@ def test_manifest_writes_output(tmp_path):
     assert output.exists()
     payload = json.loads(output.read_text(encoding="utf-8"))
     assert len(payload) == 14
+    for item in payload:
+        report = item["report"]
+        assert {
+            "decision",
+            "risk_level",
+            "rule_ids",
+            "findings",
+            "recommendation",
+        } <= report.keys()
+        if report["decision"] != "allow":
+            assert report["findings"]
+        for finding in report["findings"]:
+            assert {
+                "category",
+                "rule_id",
+                "evidence",
+                "recommendation",
+            } <= finding.keys()
     # Audit file has one line per sample.
     lines = [ln for ln in audit.read_text(encoding="utf-8").splitlines()
              if ln.strip()]

trpc_agent_sdk/tools/safety/examples/README.md

@@ -25,7 +25,7 @@ and the generated report / audit outputs.
 | `samples/manifest.yaml` | Expected decisions for all 14 samples |
 | `tool_safety_report.json` | Single-sample report (regenerated by CLI) |
 | `tool_safety_audit.jsonl` | Audit events (regenerated by CLI) |
-| `manifest_run.json` | Manifest execution summary |
+| `manifest_run.json` | Expected-vs-actual results and full structured reports for all samples |
 
 ## Run
 
@@ -54,8 +54,8 @@ python ../../../../scripts/tool_safety_check.py \
   hash.
 * `tool_safety_audit.jsonl` shows one line per scan. Note that no raw
   script text, env values, or unredacted arguments ever appear.
-* `manifest_run.json` shows expected-vs-actual decisions across all 14
-  samples.
+* `manifest_run.json` shows expected-vs-actual decisions and the full
+  structured report for each of the 14 samples.
 
 See the full design document in
 [English](../../../../docs/mkdocs/en/tool_safety_guard.md) or