Skip to content

Commit 4077074

Browse files
authored
Merge pull request #358 from XianReallyHot-ZZH/feat/issue-280-message-id-consistency
fix(agent): stream deltas and snapshots share per-turn message IDs (#280)
2 parents 8a6dc3c + ced4d20 commit 4077074

3 files changed

Lines changed: 722 additions & 27 deletions

File tree

‎argi-agent/argi-agent-core/src/main/java/io/github/agentic/ai/agent/react/ReActAgent.java‎

Lines changed: 50 additions & 22 deletions
Original file line numberDiff line numberDiff line change
@@ -345,16 +345,20 @@ private void run() {
345345
context.control().assertActive();
346346
ToolRegistry availableTools = resolveTools(state);
347347
AtomicReference<ToolRegistry> effectiveTools = new AtomicReference<>(availableTools);
348-
ChatResponse response = modelTurn(messages, state, usage, availableTools, effectiveTools);
349-
if (response.messages().isEmpty()) {
350-
throw new IllegalStateException("Model response did not contain an assistant message");
351-
}
352-
for (ChatMessage message : response.messages()) {
353-
messages.add(message);
354-
emit(new AgentEvent.Message("model-" + turn, message));
348+
ChatResponse response = modelTurn(messages, state, usage, availableTools, effectiveTools, turn);
349+
// The turn contract guarantees exactly one assistant message: direct
350+
// responses are validated by ModelResponse, and interceptor-supplied
351+
// streams are validated here, so the turn's complete snapshot and its
352+
// streamed deltas share exactly one invocation-local ID (issue #280).
353+
if (response.messages().size() != 1 || response.messages().get(0).role() != ChatRole.ASSISTANT) {
354+
throw new IllegalStateException("Model response must contain exactly one assistant message");
355355
}
356+
ChatMessage message = response.messages().get(0);
357+
String messageId = modelMessageId(turn);
358+
messages.add(message);
359+
emit(new AgentEvent.Message(messageId, message));
356360
if (response.toolCalls().isEmpty()) {
357-
finish(messages, state, usage.value(), response.messages().get(response.messages().size() - 1));
361+
finish(messages, state, usage.value(), message, messageId);
358362
return;
359363
}
360364
for (ToolCall call : response.toolCalls()) {
@@ -366,14 +370,26 @@ private void run() {
366370
return;
367371
}
368372
if (step.returnDirect()) {
369-
finish(messages, state, usage.value(), messages.get(messages.size() - 1));
373+
finish(messages, state, usage.value(), messages.get(messages.size() - 1),
374+
toolMessageId(call));
370375
return;
371376
}
372377
}
373378
}
374379
throw new IllegalStateException("ReAct agent exceeded the configured turn limit");
375380
}
376381

382+
// Message-ID scheme for the invocation-local AgentEvent contract (issue #280):
383+
// exactly one ID per logical message, shared by its streamed deltas and its
384+
// complete snapshot, centralized so the two can never drift apart again.
385+
private static String modelMessageId(int turn) {
386+
return "model-" + turn;
387+
}
388+
389+
private static String toolMessageId(ToolCall call) {
390+
return "tool-" + call.id();
391+
}
392+
377393
private ToolRegistry resolveTools(Map<String, Object> state) {
378394
if (dynamicToolProvider == null) {
379395
return tools;
@@ -395,8 +411,13 @@ private ToolRegistry resolveTools(Map<String, Object> state) {
395411

396412
private ChatResponse modelTurn(List<ChatMessage> messages, Map<String, Object> state,
397413
UsageAccumulator usage, ToolRegistry availableTools,
398-
AtomicReference<ToolRegistry> effectiveTools) {
414+
AtomicReference<ToolRegistry> effectiveTools, int turn) {
399415
context.control().assertActive();
416+
// Streamed deltas and the turn's complete snapshot share the invocation-local
417+
// model-<turn> ID (issue #280): a consumer assembles streamed text per
418+
// message ID and seals it with the same-ID snapshot, never with a fixed
419+
// stream marker shared by every turn.
420+
String streamMessageId = modelMessageId(turn);
400421
ModelRequest request = ModelRequest.builder().messages(messages).toolRegistry(availableTools)
401422
.trustedExecutionContext(context.trusted())
402423
.executionGuard(context.control())
@@ -406,15 +427,16 @@ private ChatResponse modelTurn(List<ChatMessage> messages, Map<String, Object> s
406427
// The effective tool set is only known at the end of the interceptor chain,
407428
// where selection semantics are resolved; execution in the same turn must use
408429
// exactly the set the model was shown (issue #268: declared and executed agree).
409-
ModelCallHandler handler = modelRequest -> invokeModel(modelRequest, usage, effectiveTools);
430+
ModelCallHandler handler = modelRequest -> invokeModel(modelRequest, usage, effectiveTools,
431+
streamMessageId);
410432
for (int index = interceptors.size() - 1; index >= 0; index--) {
411433
ModelInterceptor interceptor = interceptors.get(index);
412434
ModelCallHandler next = handler;
413435
handler = current -> interceptor.interceptModel(current, next);
414436
}
415437
ModelResponse response = Objects.requireNonNull(handler.call(request), "model interceptor returned no response");
416438
if (response.isStreaming()) {
417-
return collectModelStream(response.getStream(), usage);
439+
return collectModelStream(response.getStream(), usage, streamMessageId);
418440
}
419441
return response.getChatResponse();
420442
}
@@ -470,14 +492,14 @@ private List<ChatMessage> composeMessages(ModelRequest request) {
470492
}
471493

472494
private ModelResponse invokeModel(ModelRequest request, UsageAccumulator usage,
473-
AtomicReference<ToolRegistry> effectiveTools) {
495+
AtomicReference<ToolRegistry> effectiveTools, String streamMessageId) {
474496
ToolRegistry effective = effectiveToolSet(request);
475497
effectiveTools.set(effective);
476498
ChatClient selectedClient = request.getChatClient() != null ? request.getChatClient() : chatClient;
477499
ChatRequest chatRequest = new ChatRequest(composeMessages(request), effective.definitions(),
478500
request.getOptions(), request.getAttributes());
479501
if (streamModel) {
480-
return ModelResponse.of(collectModelStream(selectedClient, chatRequest, usage));
502+
return ModelResponse.of(collectModelStream(selectedClient, chatRequest, usage, streamMessageId));
481503
}
482504
CompletableFuture<ChatResponse> future = selectedClient.call(chatRequest).toCompletableFuture();
483505
toolOperation.set(future);
@@ -495,12 +517,15 @@ private ModelResponse invokeModel(ModelRequest request, UsageAccumulator usage,
495517
}
496518
}
497519

498-
private ChatResponse collectModelStream(ChatClient selectedClient, ChatRequest request, UsageAccumulator usage) {
520+
private ChatResponse collectModelStream(ChatClient selectedClient, ChatRequest request, UsageAccumulator usage,
521+
String messageId) {
499522
return collectModelStream(
500-
Objects.requireNonNull(selectedClient.stream(request), "chat client returned no stream"), usage);
523+
Objects.requireNonNull(selectedClient.stream(request), "chat client returned no stream"), usage,
524+
messageId);
501525
}
502526

503-
private ChatResponse collectModelStream(Flow.Publisher<ChatEvent> publisher, UsageAccumulator usage) {
527+
private ChatResponse collectModelStream(Flow.Publisher<ChatEvent> publisher, UsageAccumulator usage,
528+
String messageId) {
504529
CompletableFuture<ChatResponse> future = new CompletableFuture<>();
505530
try (AutoCloseable ignored = context.control().onLoss(() -> {
506531
Flow.Subscription subscription = modelSubscription.getAndSet(null);
@@ -534,7 +559,7 @@ public void onSubscribe(Flow.Subscription subscription) {
534559
public void onNext(ChatEvent event) {
535560
context.control().assertActive();
536561
switch (event.type()) {
537-
case TEXT_DELTA -> emit(new AgentEvent.MessageDelta("model-stream", event.text()));
562+
case TEXT_DELTA -> emit(new AgentEvent.MessageDelta(messageId, event.text()));
538563
case USAGE -> emit(new AgentEvent.UsageUpdate(usage.valueIncluding(event.usage())));
539564
case RESPONSE -> response = event.response();
540565
case TOOL_CALL -> emit(new AgentEvent.Progress("model.tool-call", Map.of("name",
@@ -632,7 +657,7 @@ private void appendToolResult(ToolCall call, ToolResult result, List<ChatMessage
632657
}
633658
ChatMessage message = result.toMessage(call);
634659
messages.add(message);
635-
emit(new AgentEvent.Message("tool-" + call.id(), message));
660+
emit(new AgentEvent.Message(toolMessageId(call), message));
636661
}
637662

638663
private void admitPatch(String toolName, ToolResult result, Map<String, Object> patch,
@@ -666,7 +691,7 @@ private void admitPatch(String toolName, ToolResult result, Map<String, Object>
666691
}
667692

668693
private void finish(List<ChatMessage> messages, Map<String, Object> state, Usage usage,
669-
ChatMessage terminal) {
694+
ChatMessage terminal, String terminalMessageId) {
670695
Map<String, Object> outputs = new LinkedHashMap<>(state);
671696
if (outputSchema != null && outputFinalizer != null) {
672697
AgentResult result = outputFinalizer.finalizeOutput(terminal, outputSchema, usage, context.control());
@@ -676,11 +701,14 @@ private void finish(List<ChatMessage> messages, Map<String, Object> state, Usage
676701
// A successful repair replaces the delivered terminal: the message that
677702
// actually passed validation becomes the final message while the failed
678703
// original stays in history, and stream consumers see it through the
679-
// complete-message event before the finished snapshot.
704+
// complete-message event before the finished snapshot. The repaired
705+
// message is emitted under the replaced terminal's ID (issue #280), so
706+
// an ID-keyed consumer view replaces the failed message and ends the
707+
// run consistent instead of accumulating a separate repair message.
680708
ChatMessage delivered = result.message();
681709
if (delivered != null && !Objects.equals(delivered.text(), terminal.text())) {
682710
messages.add(delivered);
683-
emit(new AgentEvent.Message("model-repair", delivered));
711+
emit(new AgentEvent.Message(terminalMessageId, delivered));
684712
}
685713
if (result.data() != null) {
686714
outputs.put("structuredOutput", result.data());

0 commit comments

Comments
 (0)