Fix: when agentic chat fail, show the real error (#21964)

internal/agentic_rag/agent.go

@@ -553,6 +553,9 @@ func Run(ctx context.Context, in Input) (string, error) {
 		inputMessageIDs[adk.GetMessageID(msg)] = true
 	}
 	final, evidence, runErr := consumeAgentEvents(ctx, iter, in.OnDelta, in.ToolCallCounts, in.ToolCallErrors, in.ToolErrorSamples)
+	if runErr != nil {
+		runErr = fmt.Errorf("agentic research failed: %w", withCancellationCause(ctx, runErr))
+	}
 
 	// Recovery, cheapest and most faithful first. A run that ended on a
 	// narration tail or on a bare tool call still holds the model's own answer
@@ -586,7 +589,7 @@ func Run(ctx context.Context, in Input) (string, error) {
 		// would reintroduce the failure this call exists to survive. The
 		// unbounded-work risk that WithoutCancel normally carries is covered
 		// by the bounded budget attached here.
-		synthCtx, synthCancel := context.WithTimeout(context.WithoutCancel(ctx), finalizeTimeout)
+		synthCtx, synthCancel := context.WithTimeoutCause(context.WithoutCancel(ctx), finalizeTimeout, fmt.Errorf("final answer synthesis timed out after %s", finalizeTimeout))
 		defer synthCancel()
 		synthCtx = tokenizer.WithUsagePhase(synthCtx, "synth")
 		synthMessages := in.Messages
@@ -594,15 +597,17 @@ func Run(ctx context.Context, in Input) (string, error) {
 			synthMessages = append([]*schema.Message{schema.SystemMessage(tmpl.Content)}, synthMessages...)
 		}
 		synth, synthErr := finalizeAnswer(synthCtx, synthModelFor(in), synthMessages, final, evidence)
+		if synthErr == nil && strings.TrimSpace(synth) == "" {
+			synthErr = errors.New("empty answer")
+		}
 		if synthErr != nil {
+			synthErr = fmt.Errorf("final answer synthesis failed: %w", withCancellationCause(synthCtx, synthErr))
 			// Without this an empty final is completely silent: the run looks
 			// healthy in every other log line and the user simply gets nothing.
 			common.WarnCtx(ctx, "agentic_rag: finalizeAnswer failed", zap.Error(synthErr))
-		} else if strings.TrimSpace(synth) == "" {
-			common.WarnCtx(ctx, "agentic_rag: finalizeAnswer returned an empty answer")
 		}
 		switch {
-		case synthErr == nil && strings.TrimSpace(synth) != "":
+		case synthErr == nil:
 			// Replace rather than append: appending leaves the reader with a
 			// narration paragraph followed by the real answer. The single
 			// answer emission at the end of Run carries it.
@@ -611,6 +616,8 @@ func Run(ctx context.Context, in Input) (string, error) {
 		case runErr == nil:
 			// Normal termination but synthesis failed — surface the reason.
 			runErr = synthErr
+		default:
+			runErr = fmt.Errorf("%w; %w", runErr, synthErr)
 		}
 	}
 
@@ -622,6 +629,16 @@ func Run(ctx context.Context, in Input) (string, error) {
 	return final, runErr
 }
 
+// Eino wraps ctx.Err(), which omits cancellation causes such as token exhaustion.
+func withCancellationCause(ctx context.Context, err error) error {
+	if errors.Is(err, context.Canceled) || errors.Is(err, context.DeadlineExceeded) {
+		if cause := context.Cause(ctx); cause != nil && !errors.Is(err, cause) {
+			return fmt.Errorf("%w: %w", cause, err)
+		}
+	}
+	return err
+}
+
 var modelFailureRe = regexp.MustCompile(`exceeds max retries|failed on every model|short-circuited by failover cooldown|overloaded_error|status 529|insufficient balance`)
 
 // consumeAgentEvents drains the agent event iterator, streaming assistant

internal/service/agentic_pipeline.go

@@ -107,7 +107,7 @@ func (s *ChatPipelineService) agenticRag(
 	go func() {
 		defer close(out)
 
-		runCtx, cancel := context.WithTimeout(ctx, agenticRunTimeout)
+		runCtx, cancel := context.WithTimeoutCause(ctx, agenticRunTimeout, fmt.Errorf("agentic run timed out after %s", agenticRunTimeout))
 		defer cancel()
 		// The run's evidence registry: every tool result stamps its passages with
 		// the [ID:n] handle the model cites, and buildAgenticReference resolves
@@ -319,7 +319,7 @@ func (s *ChatPipelineService) agenticRag(
 				zap.Any("tool_errors", toolErrors),
 				zap.String("error", runErr.Error()),
 			}, usageFields...)...)
-			emitResult(AsyncChatResult{Answer: fmt.Sprintf("**ERROR**: %s", runErr.Error()), Final: true})
+			emitResult(AsyncChatResult{Answer: fmt.Sprintf("**ERROR**: %s", runtime.MarkdownSafeErrorText(runErr)), Final: true})
 			return
 		}
 		common.InfoCtx(ctx, "agentic turn done", append([]zap.Field{