feat(agent): harden verifier evidence references

This commit is contained in:
aruo
2026-07-08 16:12:56 +08:00
parent a08672b31e
commit 7b8c75e571
29 changed files with 3426 additions and 94 deletions
@@ -18,14 +18,18 @@ class ExecutorGatekeeperServiceTest {
void validatePassesForExecutorEvidenceV2WithMatchingInvocation() {
ToolInvocationRepository repository = mock(ToolInvocationRepository.class);
when(repository.findBySessionIdOrderByIdAsc("session-1")).thenReturn(List.of(
ToolInvocation.builder().id(101L).sessionId("session-1").toolName("query_metrics").build()
invocation(101L, "query_metrics", "$.alerts[0]",
"HighCPUUsage firing, service=payment-service, current=92%, duration=25m")
));
ExecutorGatekeeperService service = new ExecutorGatekeeperService(repository);
Map<String, Object> result = service.validate("session-1", validOutput(101L, "query_metrics"),
Map<String, Object> result = service.validate("session-1",
validOutput(101L, "query_metrics", "$.alerts[0]",
"HighCPUUsage firing, service=payment-service, current=92%"),
Map.of("status", "valid"));
assertEquals("pass", result.get("status"));
assertEquals("none", result.get("severity"));
assertTrue(((List<?>) result.get("failed_rules")).isEmpty());
}
@@ -34,12 +38,13 @@ class ExecutorGatekeeperServiceTest {
ToolInvocationRepository repository = mock(ToolInvocationRepository.class);
when(repository.findBySessionIdOrderByIdAsc("session-1")).thenReturn(List.of());
ExecutorGatekeeperService service = new ExecutorGatekeeperService(repository);
Map<String, Object> output = validOutput(101L, "query_metrics");
Map<String, Object> output = validOutput(101L, "query_metrics", "$.alerts[0]", "cpu=92");
output.put("user_facing_answer", "旧版最终答案");
Map<String, Object> result = service.validate("session-1", output, Map.of("status", "valid"));
assertEquals("fail", result.get("status"));
assertEquals("reject", result.get("severity"));
assertTrue(((List<?>) result.get("failed_rules")).contains("schema.executor_v2"));
}
@@ -47,14 +52,16 @@ class ExecutorGatekeeperServiceTest {
void validateFailsForFabricatedInvocationId() {
ToolInvocationRepository repository = mock(ToolInvocationRepository.class);
when(repository.findBySessionIdOrderByIdAsc("session-1")).thenReturn(List.of(
ToolInvocation.builder().id(101L).sessionId("session-1").toolName("query_metrics").build()
invocation(101L, "query_metrics", "$.alerts[0]", "cpu=92")
));
ExecutorGatekeeperService service = new ExecutorGatekeeperService(repository);
Map<String, Object> result = service.validate("session-1", validOutput(999L, "query_metrics"),
Map<String, Object> result = service.validate("session-1",
validOutput(999L, "query_metrics", "$.alerts[0]", "cpu=92"),
Map.of("status", "valid"));
assertEquals("fail", result.get("status"));
assertEquals("reject", result.get("severity"));
assertTrue(((List<?>) result.get("failed_rules")).contains("evidence.invocation_ref"));
}
@@ -62,19 +69,137 @@ class ExecutorGatekeeperServiceTest {
void validateFailsForToolNameMismatch() {
ToolInvocationRepository repository = mock(ToolInvocationRepository.class);
when(repository.findBySessionIdOrderByIdAsc("session-1")).thenReturn(List.of(
ToolInvocation.builder().id(101L).sessionId("session-1").toolName("query_logs").build()
invocation(101L, "query_logs", "$.logs[0]", "cpu=92")
));
ExecutorGatekeeperService service = new ExecutorGatekeeperService(repository);
Map<String, Object> result = service.validate("session-1", validOutput(101L, "query_metrics"),
Map<String, Object> result = service.validate("session-1",
validOutput(101L, "query_metrics", "$.alerts[0]", "cpu=92"),
Map.of("status", "valid"));
assertEquals("fail", result.get("status"));
assertEquals("reject", result.get("severity"));
assertTrue(((List<?>) result.get("failed_rules")).contains("evidence.invocation_ref"));
}
@SuppressWarnings("unchecked")
private Map<String, Object> validOutput(Long invocationId, String toolName) {
@Test
void validateDowngradesMissingRawPathToLowConfidence() {
ToolInvocationRepository repository = mock(ToolInvocationRepository.class);
when(repository.findBySessionIdOrderByIdAsc("session-1")).thenReturn(List.of(
invocation(101L, "query_metrics", "$.alerts[0]", "cpu=92")
));
ExecutorGatekeeperService service = new ExecutorGatekeeperService(repository);
Map<String, Object> output = validOutput(101L, "query_metrics", null, "cpu=92");
Map<String, Object> result = service.validate("session-1", output, Map.of("status", "valid"));
assertEquals("fail", result.get("status"));
assertEquals("low_confid", result.get("severity"));
assertTrue(((List<?>) result.get("failed_rules")).contains("evidence.raw_path"));
}
@Test
void validateRejectsUnknownRawPath() {
ToolInvocationRepository repository = mock(ToolInvocationRepository.class);
when(repository.findBySessionIdOrderByIdAsc("session-1")).thenReturn(List.of(
invocation(101L, "query_metrics", "$.alerts[0]", "cpu=92")
));
ExecutorGatekeeperService service = new ExecutorGatekeeperService(repository);
Map<String, Object> result = service.validate("session-1",
validOutput(101L, "query_metrics", "$.alerts[99]", "cpu=92"),
Map.of("status", "valid"));
assertEquals("fail", result.get("status"));
assertEquals("reject", result.get("severity"));
assertTrue(((List<?>) result.get("failed_rules")).contains("evidence.raw_path"));
}
@Test
void validateDowngradesOldInvocationWithoutEvidenceRefsToLowConfidence() {
ToolInvocationRepository repository = mock(ToolInvocationRepository.class);
when(repository.findBySessionIdOrderByIdAsc("session-1")).thenReturn(List.of(
ToolInvocation.builder().id(101L).sessionId("session-1").toolName("query_metrics")
.retrievalDetails("{\"evidence_status\":\"supported\"}")
.build()
));
ExecutorGatekeeperService service = new ExecutorGatekeeperService(repository);
Map<String, Object> result = service.validate("session-1",
validOutput(101L, "query_metrics", "$.alerts[0]", "cpu=92"),
Map.of("status", "valid"));
assertEquals("fail", result.get("status"));
assertEquals("low_confid", result.get("severity"));
assertTrue(((List<?>) result.get("failed_rules")).contains("evidence.missing"));
}
@Test
void validateRejectsMismatchedExcerpt() {
ToolInvocationRepository repository = mock(ToolInvocationRepository.class);
when(repository.findBySessionIdOrderByIdAsc("session-1")).thenReturn(List.of(
invocation(101L, "query_metrics", "$.alerts[0]", "HighCPUUsage firing service payment-service current 92")
));
ExecutorGatekeeperService service = new ExecutorGatekeeperService(repository);
Map<String, Object> result = service.validate("session-1",
validOutput(101L, "query_metrics", "$.alerts[0]", "HikariCP active=50/50 waiting=32"),
Map.of("status", "valid"));
assertEquals("fail", result.get("status"));
assertEquals("reject", result.get("severity"));
assertTrue(((List<?>) result.get("failed_rules")).contains("evidence.excerpt_mismatch"));
}
@Test
void validateFailsForRecommendedActionFabricatedInvocationId() {
ToolInvocationRepository repository = mock(ToolInvocationRepository.class);
when(repository.findBySessionIdOrderByIdAsc("session-1")).thenReturn(List.of(
invocation(101L, "query_metrics", "$.alerts[0]", "cpu=92")
));
ExecutorGatekeeperService service = new ExecutorGatekeeperService(repository);
Map<String, Object> output = validOutput(101L, "query_metrics", "$.alerts[0]", "cpu=92");
output.put("recommended_actions", List.of(Map.of(
"action_text", "restart service",
"reason", "alert is firing",
"evidence_bindings", List.of(Map.of(
"source_type", "tool_trace",
"source_id", "trace-1",
"tool_name", "query_metrics",
"source_invocation_id", 999L,
"raw_path", "$.alerts[0]",
"evidence_excerpt", "cpu=92"
))
)));
Map<String, Object> result = service.validate("session-1", output, Map.of("status", "valid"));
assertEquals("fail", result.get("status"));
assertEquals("reject", result.get("severity"));
assertTrue(((List<?>) result.get("failed_rules")).contains("evidence.invocation_ref"));
}
private ToolInvocation invocation(Long id, String toolName, String rawPath, String text) {
return ToolInvocation.builder()
.id(id)
.sessionId("session-1")
.toolName(toolName)
.retrievalDetails("{\"evidence_refs\":[{\"raw_path\":\"" + rawPath
+ "\",\"text\":\"" + text + "\"}]}")
.build();
}
private Map<String, Object> validOutput(Long invocationId, String toolName, String rawPath, String excerpt) {
Map<String, Object> binding = new java.util.LinkedHashMap<>();
binding.put("source_type", "tool_trace");
binding.put("source_id", "trace-1");
binding.put("tool_name", toolName);
binding.put("source_invocation_id", invocationId);
if (rawPath != null) {
binding.put("raw_path", rawPath);
}
binding.put("evidence_excerpt", excerpt);
return new java.util.LinkedHashMap<>(Map.of(
"answer_version", "executor_evidence_v2",
"claims", List.of(Map.of(
@@ -82,13 +207,7 @@ class ExecutorGatekeeperServiceTest {
"claim_type", "symptom",
"claim_text", "连接池 active 达到上限",
"support_level", "direct",
"evidence_bindings", List.of(Map.of(
"source_type", "tool_trace",
"source_id", "trace-1",
"tool_name", toolName,
"source_invocation_ids", List.of(invocationId),
"evidence_excerpt", "active=50 max=50"
))
"evidence_bindings", List.of(binding)
)),
"hypotheses", List.of(),
"recommended_actions", List.of(),