fix(agent): surface tool calls before input completes

Emit tool-call placeholders as soon as tool input streaming starts so long writes appear immediately in chat. Reuse the same UI tool message when full input arrives to avoid duplicate cards, and keep the hook-required test suite green.

internal/agent/stream_test.go

@@ -0,0 +1,89 @@
+package agent
+
+import (
+	"context"
+	"reflect"
+	"testing"
+
+	sdk "github.com/memohai/twilight-ai/sdk"
+)
+
+type agentToolPlaceholderProvider struct{}
+
+func (*agentToolPlaceholderProvider) Name() string { return "tool-placeholder-mock" }
+
+func (*agentToolPlaceholderProvider) ListModels(context.Context) ([]sdk.Model, error) {
+	return nil, nil
+}
+
+func (*agentToolPlaceholderProvider) Test(context.Context) *sdk.ProviderTestResult {
+	return &sdk.ProviderTestResult{Status: sdk.ProviderStatusOK, Message: "ok"}
+}
+
+func (*agentToolPlaceholderProvider) TestModel(context.Context, string) (*sdk.ModelTestResult, error) {
+	return &sdk.ModelTestResult{Supported: true, Message: "supported"}, nil
+}
+
+func (*agentToolPlaceholderProvider) DoGenerate(context.Context, sdk.GenerateParams) (*sdk.GenerateResult, error) {
+	return &sdk.GenerateResult{FinishReason: sdk.FinishReasonStop}, nil
+}
+
+func (*agentToolPlaceholderProvider) DoStream(_ context.Context, _ sdk.GenerateParams) (*sdk.StreamResult, error) {
+	ch := make(chan sdk.StreamPart, 8)
+	go func() {
+		defer close(ch)
+		ch <- &sdk.StartPart{}
+		ch <- &sdk.StartStepPart{}
+		ch <- &sdk.ToolInputStartPart{ID: "call-1", ToolName: "write"}
+		ch <- &sdk.StreamToolCallPart{
+			ToolCallID: "call-1",
+			ToolName:   "write",
+			Input:      map[string]any{"path": "/tmp/long.txt"},
+		}
+		ch <- &sdk.FinishStepPart{FinishReason: sdk.FinishReasonStop}
+		ch <- &sdk.FinishPart{FinishReason: sdk.FinishReasonStop}
+	}()
+	return &sdk.StreamResult{Stream: ch}, nil
+}
+
+func TestAgentStreamEmitsEarlyToolPlaceholderBeforeFullInput(t *testing.T) {
+	t.Parallel()
+
+	a := New(Deps{})
+
+	var events []StreamEvent
+	for event := range a.Stream(context.Background(), RunConfig{
+		Model: &sdk.Model{
+			ID:       "mock-model",
+			Provider: &agentToolPlaceholderProvider{},
+		},
+		Messages:         []sdk.Message{sdk.UserMessage("write a long file")},
+		SupportsToolCall: false,
+		Identity:         SessionContext{BotID: "bot-1"},
+	}) {
+		events = append(events, event)
+	}
+
+	if len(events) != 4 {
+		t.Fatalf("expected 4 events, got %d: %#v", len(events), events)
+	}
+	if events[0].Type != EventAgentStart {
+		t.Fatalf("expected first event %q, got %#v", EventAgentStart, events[0])
+	}
+	if events[1].Type != EventToolCallStart || events[1].ToolCallID != "call-1" || events[1].ToolName != "write" {
+		t.Fatalf("unexpected placeholder tool event: %#v", events[1])
+	}
+	if events[1].Input != nil {
+		t.Fatalf("expected placeholder tool event to have nil input, got %#v", events[1].Input)
+	}
+	if events[2].Type != EventToolCallStart || events[2].ToolCallID != "call-1" {
+		t.Fatalf("unexpected full tool event: %#v", events[2])
+	}
+	expectedInput := map[string]any{"path": "/tmp/long.txt"}
+	if !reflect.DeepEqual(events[2].Input, expectedInput) {
+		t.Fatalf("expected full tool event input %#v, got %#v", expectedInput, events[2].Input)
+	}
+	if events[3].Type != EventAgentEnd {
+		t.Fatalf("expected terminal event %q, got %#v", EventAgentEnd, events[3])
+	}
+}