feat(agent): add executor evidence output contract

This commit is contained in:
zhuyongxin
2026-07-07 19:02:02 +08:00
parent 7f2e47ca38
commit 04eb50e2b4
19 changed files with 986 additions and 25 deletions
@@ -101,6 +101,8 @@ public class DiagnosisTraceEvaluator {
failedChecks.add("reject output does not use degraded template");
}
failedChecks.addAll(validateExecutorStructuredOutput(trace));
Integer toolCallCount = trace.getToolInvocations() == null ? 0 : trace.getToolInvocations().size();
Integer durationMs = trace.getSession() == null ? null : trace.getSession().getTotalDurationMs();
@@ -174,6 +176,32 @@ public class DiagnosisTraceEvaluator {
return value == null ? null : String.valueOf(value);
}
private List<String> validateExecutorStructuredOutput(DiagnosisTraceResponse trace) {
Object structuredOutput = nestedValue(trace, "verifier_evaluation", "executor_structured_output");
if (!(structuredOutput instanceof Map<?, ?> output)) {
return List.of();
}
Object claims = output.get("claims");
if (!(claims instanceof List<?> claimList)) {
return List.of("executor structured output missing claims array");
}
List<String> failedChecks = new ArrayList<>();
for (Object item : claimList) {
if (!(item instanceof Map<?, ?> claim)) {
failedChecks.add("executor structured claim is not an object");
continue;
}
Object claimIdValue = claim.get("claim_id");
String claimId = claimIdValue == null ? "unknown" : String.valueOf(claimIdValue);
Object bindings = claim.get("evidence_bindings");
if (!(bindings instanceof List<?> bindingList) || bindingList.isEmpty()) {
failedChecks.add("executor confirmed claim missing evidence bindings: " + claimId);
}
}
return failedChecks;
}
private Object nestedValue(DiagnosisTraceResponse trace, String firstKey, String secondKey) {
if (trace.getSession() == null || trace.getSession().getSelfEvaluation() == null) {
return null;
@@ -5,6 +5,8 @@ import com.alibaba.cloud.ai.graph.agent.hook.HookPosition;
import com.alibaba.cloud.ai.graph.agent.hook.HookPositions;
import com.alibaba.cloud.ai.graph.agent.hook.messages.AgentCommand;
import com.alibaba.cloud.ai.graph.agent.hook.messages.MessagesModelHook;
import com.fasterxml.jackson.core.type.TypeReference;
import com.fasterxml.jackson.databind.JsonNode;
import com.fasterxml.jackson.databind.ObjectMapper;
import com.superbiz.agent.service.ToolTraceSummaryService;
import com.superbiz.agent.util.SessionContextHolder;
@@ -27,6 +29,8 @@ public class VerifierInputHook extends MessagesModelHook {
private final ToolTraceSummaryService toolTraceSummaryService;
private final ObjectMapper objectMapper = new ObjectMapper();
private static final TypeReference<Map<String, Object>> MAP_TYPE = new TypeReference<>() {
};
public VerifierInputHook(ToolTraceSummaryService toolTraceSummaryService) {
this.toolTraceSummaryService = toolTraceSummaryService;
@@ -48,6 +52,10 @@ public class VerifierInputHook extends MessagesModelHook {
executorFinalAnswer = extractLastAssistantText(previousMessages);
}
ExecutorOutputParseResult parseResult = parseExecutorOutput(executorFinalAnswer);
VerifierContextHolder.setExecutorStructuredOutput(parseResult.structuredOutput());
VerifierContextHolder.setExecutorOutputParseStatus(parseResult.status());
List<Map<String, Object>> toolTraceSummary =
toolTraceSummaryService.buildVerifierTraceSummary(sessionId, executorFinalAnswer);
VerifierContextHolder.setToolTraceSummary(toolTraceSummary);
@@ -55,6 +63,8 @@ public class VerifierInputHook extends MessagesModelHook {
Map<String, Object> verifierInput = new LinkedHashMap<>();
verifierInput.put("original_query", VerifierContextHolder.getOriginalQuery());
verifierInput.put("executor_final_answer", executorFinalAnswer);
verifierInput.put("executor_structured_output", parseResult.structuredOutput());
verifierInput.put("executor_output_parse_status", parseResult.status());
verifierInput.put("tool_trace_summary", toolTraceSummary);
verifierInput.put("retry_context", VerifierContextHolder.getRetryContext());
@@ -66,6 +76,59 @@ public class VerifierInputHook extends MessagesModelHook {
}
}
private ExecutorOutputParseResult parseExecutorOutput(String executorFinalAnswer) {
if (executorFinalAnswer == null || executorFinalAnswer.isBlank()) {
return new ExecutorOutputParseResult(null, status("missing", "executor_final_answer is blank"));
}
String sanitized = sanitizeJsonPayload(executorFinalAnswer);
if (!looksJsonLike(sanitized)) {
return new ExecutorOutputParseResult(null, status("missing", "executor output is not JSON"));
}
try {
JsonNode root = objectMapper.readTree(sanitized);
if (!root.isObject() || !root.path("claims").isArray()) {
return new ExecutorOutputParseResult(null, status("malformed",
"executor output JSON does not match evidence-attribution contract"));
}
Map<String, Object> structuredOutput = objectMapper.convertValue(root, MAP_TYPE);
return new ExecutorOutputParseResult(structuredOutput, status("valid", "parsed executor evidence contract"));
} catch (Exception e) {
log.debug("Failed to parse executor structured output", e);
return new ExecutorOutputParseResult(null, status("malformed", e.getMessage()));
}
}
private String sanitizeJsonPayload(String raw) {
String trimmed = raw.trim();
int fenceStart = trimmed.indexOf("```");
if (fenceStart >= 0) {
int firstNewline = trimmed.indexOf('\n', fenceStart);
int lastFence = trimmed.indexOf("```", firstNewline + 1);
if (firstNewline >= 0 && lastFence > firstNewline) {
return trimmed.substring(firstNewline + 1, lastFence).trim();
}
}
int objectStart = trimmed.indexOf('{');
int objectEnd = trimmed.lastIndexOf('}');
if (objectStart >= 0 && objectEnd > objectStart) {
return trimmed.substring(objectStart, objectEnd + 1).trim();
}
return trimmed;
}
private boolean looksJsonLike(String text) {
return text.startsWith("{") && text.endsWith("}");
}
private Map<String, Object> status(String status, String detail) {
Map<String, Object> result = new LinkedHashMap<>();
result.put("status", status);
result.put("detail", detail == null ? "" : detail);
return result;
}
private String extractLastAssistantText(List<Message> previousMessages) {
for (int i = previousMessages.size() - 1; i >= 0; i--) {
if (previousMessages.get(i) instanceof AssistantMessage assistantMessage) {
@@ -102,4 +165,10 @@ public class VerifierInputHook extends MessagesModelHook {
}
return message.toString();
}
private record ExecutorOutputParseResult(
Map<String, Object> structuredOutput,
Map<String, Object> status
) {
}
}
@@ -444,7 +444,8 @@ public class ChatService {
}
if ("PASS".equals(finalDecision.verdict())) {
answer = answer == null || answer.isBlank() ? "抱歉,多 Agent 分析未能生成有效结论。" : answer;
answer = extractUserFacingAnswer(answer)
.orElse(answer == null || answer.isBlank() ? "抱歉,多 Agent 分析未能生成有效结论。" : answer);
persistVerifierEvaluation(session, finalDecision, round);
break;
}
@@ -744,6 +745,10 @@ public class ChatService {
verifierEvaluation.put("rationale", decision.rationale());
verifierEvaluation.put("round", round);
verifierEvaluation.put("traceability_version", "v1");
verifierEvaluation.put("executor_output_parse_status",
Optional.ofNullable(VerifierContextHolder.getExecutorOutputParseStatus())
.orElse(Map.of("status", "missing", "detail", "executor parse status unavailable")));
verifierEvaluation.put("executor_structured_output", VerifierContextHolder.getExecutorStructuredOutput());
verifierEvaluation.put("tool_trace_summary",
Optional.ofNullable(VerifierContextHolder.getToolTraceSummary()).orElse(List.of()));
@@ -795,6 +800,22 @@ public class ChatService {
return output.toString();
}
private Optional<String> extractUserFacingAnswer(String executorAnswer) {
if (executorAnswer == null || executorAnswer.isBlank()) {
return Optional.empty();
}
try {
JsonNode root = objectMapper.readTree(sanitizeJsonPayload(executorAnswer));
String userFacingAnswer = root.path("user_facing_answer").asText("");
if (!userFacingAnswer.isBlank()) {
return Optional.of(userFacingAnswer);
}
} catch (Exception e) {
logger.debug("Executor answer is not structured JSON, keep raw answer");
}
return Optional.empty();
}
private String buildDegradedOutput(VerifierDecision decision) {
StringBuilder output = new StringBuilder(DEGRADED_PREFIX);
@@ -11,6 +11,8 @@ public final class VerifierContextHolder {
private static final ThreadLocal<String> ORIGINAL_QUERY = new ThreadLocal<>();
private static final ThreadLocal<String> RETRY_CONTEXT = new ThreadLocal<>();
private static final ThreadLocal<String> EXECUTOR_FINAL_ANSWER = new ThreadLocal<>();
private static final ThreadLocal<Map<String, Object>> EXECUTOR_STRUCTURED_OUTPUT = new ThreadLocal<>();
private static final ThreadLocal<Map<String, Object>> EXECUTOR_OUTPUT_PARSE_STATUS = new ThreadLocal<>();
private static final ThreadLocal<List<Map<String, Object>>> TOOL_TRACE_SUMMARY = new ThreadLocal<>();
private VerifierContextHolder() {
@@ -40,6 +42,22 @@ public final class VerifierContextHolder {
return EXECUTOR_FINAL_ANSWER.get();
}
public static void setExecutorStructuredOutput(Map<String, Object> executorStructuredOutput) {
EXECUTOR_STRUCTURED_OUTPUT.set(executorStructuredOutput);
}
public static Map<String, Object> getExecutorStructuredOutput() {
return EXECUTOR_STRUCTURED_OUTPUT.get();
}
public static void setExecutorOutputParseStatus(Map<String, Object> executorOutputParseStatus) {
EXECUTOR_OUTPUT_PARSE_STATUS.set(executorOutputParseStatus);
}
public static Map<String, Object> getExecutorOutputParseStatus() {
return EXECUTOR_OUTPUT_PARSE_STATUS.get();
}
public static void setToolTraceSummary(List<Map<String, Object>> toolTraceSummary) {
TOOL_TRACE_SUMMARY.set(toolTraceSummary);
}
@@ -52,6 +70,8 @@ public final class VerifierContextHolder {
ORIGINAL_QUERY.remove();
RETRY_CONTEXT.remove();
EXECUTOR_FINAL_ANSWER.remove();
EXECUTOR_STRUCTURED_OUTPUT.remove();
EXECUTOR_OUTPUT_PARSE_STATUS.remove();
TOOL_TRACE_SUMMARY.remove();
}
}