Bound parallel tool dispatch

cmd/olliesrv/internal/agent/loop.go

@@ -15,7 +15,10 @@ import (
 	"ollie/util"
 )
 
-const maxTransientRetries = 3
+const (
+	maxTransientRetries  = 3
+	maxParallelToolCalls = 16
+)
 
 // retryBaseDelay is the base delay for rate-limit retries (5<<attempt seconds).
 // Tests can override this to speed up retry tests.
@@ -162,31 +165,31 @@ func (ag *Agent) run(ctx context.Context) error {
 			}
 		}
 
-			msg := backend.Message{ID: responseID, Role: "assistant", Content: content, Reasoning: reasoning, ToolCalls: toolCalls}
+		msg := backend.Message{ID: responseID, Role: "assistant", Content: content, Reasoning: reasoning, ToolCalls: toolCalls}
-			results, interrupted := ag.execToolCalls(ctx, toolCalls)
+		results, interrupted := ag.execToolCalls(ctx, toolCalls)
 
-			h.update(msg, results)
+		h.update(msg, results)
-			ag.save()
+		ag.save()
 
-			if interrupted {
+		if interrupted {
-				return ctx.Err()
+			return ctx.Err()
-			}
-			if len(toolCalls) == 0 {
-				break
-			}
-
-			if abort := ag.trackErrors(ctx, results, &errs, step); abort != nil {
-				return abort
-			}
-
-			_, compactErr := ag.autoCompact(ctx)
-			if compactErr != nil {
-				ag.emit(Event{Role: "error", Content: compactErr.Error()})
-				break
-			}
-
-			step++
 		}
+		if len(toolCalls) == 0 {
+			break
+		}
+
+		if abort := ag.trackErrors(ctx, results, &errs, step); abort != nil {
+			return abort
+		}
+
+		_, compactErr := ag.autoCompact(ctx)
+		if compactErr != nil {
+			ag.emit(Event{Role: "error", Content: compactErr.Error()})
+			break
+		}
+
+		step++
+	}
 
 	return nil
 }
@@ -467,6 +470,7 @@ func (ag *Agent) execBatch(ctx context.Context, batch []backend.ToolCall) ([]too
 	uniqueResults := make([]inflightResult, len(batch))
 
 	var wg sync.WaitGroup
+	sem := make(chan struct{}, maxParallelToolCalls)
 	for k, call := range batch {
 		key := call.Name + "\x00" + string(call.Arguments)
 		if _, dup := inflight[key]; dup {
@@ -476,6 +480,13 @@ func (ag *Agent) execBatch(ctx context.Context, batch []backend.ToolCall) ([]too
 		wg.Add(1)
 		go func(k int, call backend.ToolCall) {
 			defer wg.Done()
+			select {
+			case sem <- struct{}{}:
+			case <-ctx.Done():
+				uniqueResults[k].wasInt = true
+				return
+			}
+			defer func() { <-sem }()
 			uniqueResults[k].tr, uniqueResults[k].wasInt = ag.execOne(ctx, call)
 		}(k, call)
 	}