Skip to content

fix(core): 异常终止时保存 ReActAgent 状态 #2713

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
hanydd wants to merge 1 commit into from
Closed

fix(core): 异常终止时保存 ReActAgent 状态 #2713

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
125 changes: 94 additions & 31 deletions agentscope-core/src/main/java/io/agentscope/core/ReActAgent.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -486,24 +486,49 @@ private Mono<Void> saveStateToSession(CallExecution scope) {
if (stateStore == null) {
return Mono.empty();
}
return Mono.<Void>fromRunnable(() -> persistState(scope))
.subscribeOn(Schedulers.boundedElastic());
}

private void persistState(CallExecution scope) {
syncToolkitToState(scope.state);
SlotRef ref = SlotRef.parse(scope.slotKey);
AgentState toSave = scope.state;
return Mono.<Void>fromRunnable(
() -> {
long newVersion =
persistAgentStateCas(
ref.userId,
ref.sessionId,
scope.slotKey,
toSave,
scope.loadedVersion,
scope.loadedContextSize);
if (newVersion != AgentStateStore.UNVERSIONED) {
scope.loadedVersion = newVersion;
}
})
.subscribeOn(Schedulers.boundedElastic());
long newVersion =
persistAgentStateCas(
ref.userId,
ref.sessionId,
scope.slotKey,
scope.state,
scope.loadedVersion,
scope.loadedContextSize);
if (newVersion != AgentStateStore.UNVERSIONED) {
scope.loadedVersion = newVersion;
}
}

private void repairStateBeforeAbnormalCheckpoint(CallExecution scope) {
try {
scope.dropUncommittedToolCallsFromCurrentCall();

@AgentScopeJavaBot AgentScopeJavaBot Aug 15, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[praise] Good defensive design: wrapping both dropUncommittedToolCallsFromCurrentCall() and persistState() in separate try-catch blocks ensures that a failure in state repair doesn't prevent the best-effort save, and a save failure is properly logged without masking the original error.

@AgentScopeJavaBot AgentScopeJavaBot Aug 18, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

✅ Verified as addressed in 1770366: checkpointAbnormalState (lines 519-531) still uses two separate try-catch blocks for repair and save as praised.

@AgentScopeJavaBot AgentScopeJavaBot Aug 15, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[praise] Good defensive design: wrapping both dropUncommittedToolCallsFromCurrentCall() and persistState() in separate try-catch blocks ensures that a failure in state repair doesn't prevent the best-effort save, and a save failure is properly logged without masking the original error.

@AgentScopeJavaBot AgentScopeJavaBot Aug 18, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

✅ Verified as addressed in 1770366: checkpointAbnormalState (lines 519-531) still uses two separate try-catch blocks for repair and save as praised.

} catch (RuntimeException repairError) {
log.warn("Failed to repair agent state before checkpoint", repairError);
}
}

private void checkpointStateAfterCancellation(CallExecution scope) {
repairStateBeforeAbnormalCheckpoint(scope);
try {
if (stateStore != null) {
persistState(scope);
}
} catch (RuntimeException saveError) {
log.warn("Failed to save agent state after abnormal termination", saveError);
}
}

private <T> Mono<T> checkpointOnAbnormalTermination(CallExecution scope, Mono<T> execution) {
return execution
.doOnCancel(() -> checkpointStateAfterCancellation(scope))
.onErrorResume(error -> saveStateAfterCallFailure(scope, error));
}

/**
Expand All @@ -522,6 +547,7 @@ private <T> Mono<T> saveStateAfterCallFailure(CallExecution scope, Throwable cal
if (ExceptionUtils.containsInterruptedException(callFailure)) {
return Mono.error(callFailure);
}
repairStateBeforeAbnormalCheckpoint(scope);
return saveStateToSession(scope)
.onErrorResume(
saveFailure -> {
Expand Down Expand Up @@ -1223,8 +1249,7 @@ protected Mono<Msg> doCall(List<Msg> msgs) {
AgentEventEmitter.fromForwardingContext(cv)
.ifPresent(ae -> scope.externalEventEmitter = ae);
}
return scope.doCallInner(msgs)
.onErrorResume(error -> saveStateAfterCallFailure(scope, error))
return checkpointOnAbnormalTermination(scope, scope.doCallInner(msgs))
.flatMap(result -> saveStateToSession(scope).thenReturn(result));
});
}
Expand Down Expand Up @@ -1265,20 +1290,25 @@ private Mono<Msg> doStructuredCall(List<Msg> msgs, Class<?> targetClass, JsonNod
hasTools
? model.supportsNativeStructuredOutputWithTools()
: model.supportsNativeStructuredOutput();
Mono<Msg> execution;
if (useNative) {
return doNativeStructuredCall(msgs, jsonSchema)
.onErrorResume(
e -> {
log.warn(
"Native structured output failed ({}) — falling back to"
+ " synthetic tool path",
e.getMessage() != null
? e.getMessage()
: e.getClass().getSimpleName());
return doFallbackStructuredCall(msgs, jsonSchema);
});
execution =
doNativeStructuredCall(msgs, jsonSchema)
.onErrorResume(
e -> {
log.warn(
"Native structured output failed ({}) — falling"
+ " back to synthetic tool path",
e.getMessage() != null
? e.getMessage()
: e.getClass().getSimpleName());
return doFallbackStructuredCall(msgs, jsonSchema);
});
} else {
execution = doFallbackStructuredCall(msgs, jsonSchema);
}
return doFallbackStructuredCall(msgs, jsonSchema);
return Mono.deferContextual(
cv -> checkpointOnAbnormalTermination(scopeFrom(cv), execution));
}

/**
Expand Down Expand Up @@ -1359,7 +1389,6 @@ private Mono<Msg> doFallbackStructuredCall(List<Msg> msgs, Map<String, Object> j
scope.soTool = createStructuredOutputTool(jsonSchema);

return scope.doCallInner(msgs)
.onErrorResume(error -> saveStateAfterCallFailure(scope, error))
.flatMap(
result -> {
Msg out = result;
Expand Down Expand Up @@ -2124,6 +2153,40 @@ private void synthesizeErrorResultsForPendingToolCalls() {
}
}

/** Remove current-call PENDING tool calls that never produced results. */

@AgentScopeJavaBot AgentScopeJavaBot Aug 15, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[praise] Clean implementation: iterating backward from the end of context and bounding by loadedContextSize ensures only current-call PENDING tool calls are dropped, preserving historical context integrity. The early return when retained.size() == message.getContent().size() is a nice short-circuit optimization.

@AgentScopeJavaBot AgentScopeJavaBot Aug 18, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

✅ Verified as addressed in 1770366: dropUncommittedToolCallsFromCurrentCall (lines 2125-2158) retains the praised backward iteration, loadedContextSize bound, and early-return optimization.

@AgentScopeJavaBot AgentScopeJavaBot Aug 15, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[praise] Clean implementation: iterating backward from the end of context and bounding by loadedContextSize ensures only current-call PENDING tool calls are dropped, preserving historical context integrity. The early return when retained.size() == message.getContent().size() is a nice short-circuit optimization.

@AgentScopeJavaBot AgentScopeJavaBot Aug 18, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

✅ Verified as addressed in 1770366: dropUncommittedToolCallsFromCurrentCall (lines 2125-2158) retains the praised backward iteration, loadedContextSize bound, and early-return optimization.

private void dropUncommittedToolCallsFromCurrentCall() {
Set<String> pendingIds = getPendingToolUseIds();
if (pendingIds.isEmpty()) {
return;
}
List<Msg> context = state.contextMutable();
for (int i = context.size() - 1; i >= loadedContextSize; i--) {
Msg message = context.get(i);
if (message.getRole() != MsgRole.ASSISTANT
|| !message.hasContentBlocks(ToolUseBlock.class)) {
continue;
}
List<ContentBlock> retained =
message.getContent().stream()
.filter(
block ->
!(block instanceof ToolUseBlock toolUse)
|| toolUse.getState()
!= ToolCallState.PENDING
|| !pendingIds.contains(toolUse.getId()))
.toList();
if (retained.size() == message.getContent().size()) {
return;
}
if (retained.isEmpty()) {
context.remove(i);
} else {
context.set(i, message.withContent(retained));
}
return;
}
}

private void publishEvent(AgentEvent event) {
FluxSink<AgentEvent> sink = eventSink;
if (sink != null) {
Expand Down
Loading
Loading