feat(session,llm): restore session runtime on resume; align Responses API details

pkg/llm/responses_test.go

@@ -59,12 +59,15 @@ func TestBuildResponsesRequest(t *testing.T) {
 	if raw["model"] != "gpt-test" || raw["stream"] != true {
 		t.Fatalf("model/stream = %#v / %#v", raw["model"], raw["stream"])
 	}
+	if raw["instructions"] != "be concise" {
+		t.Fatalf("instructions = %#v", raw["instructions"])
+	}
 	if raw["tool_choice"] != "auto" {
 		t.Fatalf("tool_choice = %#v", raw["tool_choice"])
 	}
 
 	input, ok := raw["input"].([]any)
-	if !ok || len(input) != 6 {
+	if !ok || len(input) != 5 {
 		t.Fatalf("input = %#v", raw["input"])
 	}
 	checkInputItem := func(index int, wantType, wantKey, wantValue string) {
@@ -80,16 +83,15 @@ func TestBuildResponsesRequest(t *testing.T) {
 			t.Fatalf("input[%d].type = %#v, want %q", index, item["type"], wantType)
 		}
 	}
-	checkInputItem(0, "", "role", "system")
-	checkInputItem(1, "", "role", "user")
-	checkInputItem(2, "", "role", "assistant")
-	checkInputItem(3, "function_call", "call_id", "call_1")
-	checkInputItem(4, "function_call_output", "call_id", "call_1")
-	checkInputItem(5, "", "role", "assistant")
-	if item := input[3].(map[string]any); item["name"] != "file_read" || item["arguments"] != `{"path":"README.md"}` {
+	checkInputItem(0, "", "role", "user")
+	checkInputItem(1, "", "role", "assistant")
+	checkInputItem(2, "function_call", "call_id", "call_1")
+	checkInputItem(3, "function_call_output", "call_id", "call_1")
+	checkInputItem(4, "", "role", "assistant")
+	if item := input[2].(map[string]any); item["name"] != "file_read" || item["arguments"] != `{"path":"README.md"}` {
 		t.Fatalf("function_call item = %#v", item)
 	}
-	if item := input[4].(map[string]any); item["output"] != "contents" {
+	if item := input[3].(map[string]any); item["output"] != "contents" {
 		t.Fatalf("function_call_output item = %#v", item)
 	}
 
@@ -117,6 +119,36 @@ func TestBuildResponsesRequest(t *testing.T) {
 	}
 }
 
+func TestResponsesInstructionsCombinesSystemMessages(t *testing.T) {
+	payload := buildResponsesRequest(ChatRequest{
+		Model: "gpt-test",
+		Messages: []Message{
+			{Role: RoleSystem, Content: "first instruction"},
+			{Role: RoleUser, Content: "hi"},
+			{Role: RoleSystem, Content: "second instruction"},
+		},
+	})
+	data, err := json.Marshal(payload)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var raw map[string]any
+	if err := json.Unmarshal(data, &raw); err != nil {
+		t.Fatal(err)
+	}
+	if raw["instructions"] != "first instruction\n\nsecond instruction" {
+		t.Fatalf("instructions = %#v", raw["instructions"])
+	}
+	input, ok := raw["input"].([]any)
+	if !ok || len(input) != 1 {
+		t.Fatalf("input = %#v", raw["input"])
+	}
+	item, ok := input[0].(map[string]any)
+	if !ok || item["role"] != "user" {
+		t.Fatalf("input[0] = %#v", input[0])
+	}
+}
+
 func TestOpenAIResponsesClientStreamsContentAndToolCalls(t *testing.T) {
 	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		if r.URL.Path != "/v1/responses" {