AutoZero/llm/outputlimit_test.go

package llm

import (
	"context"
	"strings"
	"testing"
)

// 输出上限切断的识别：三家的叫法不同，归一成 Message.OutputLimited 一个信号。
// 每个 fixture 都带正文——"什么都没有"是空回复那条路（ErrEmptyCompletion），
// 这里要抓的是"合法、有内容、但没写完"。
func TestProvidersDetectOutputLimit(t *testing.T) {
	cases := []struct {
		name string
		sse  string
		call func(url string) (Provider, error)
		want bool
	}{
		{
			name: "chat_completions：finish_reason=length",
			sse: "data: {\"choices\":[{\"delta\":{\"content\":\"第一段写完了，第二段\"}}]}\n\n" +
				"data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"length\"}],\"usage\":{\"prompt_tokens\":5,\"completion_tokens\":7}}\n\n" +
				"data: [DONE]\n\n",
			call: func(u string) (Provider, error) { return &OpenAIProvider{BaseURL: u, APIKey: "k", Model: "m"}, nil },
			want: true,
		},
		{
			name: "chat_completions：finish_reason=stop 是正常收尾",
			sse: "data: {\"choices\":[{\"delta\":{\"content\":\"说完了\"}}]}\n\n" +
				"data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":5,\"completion_tokens\":3}}\n\n" +
				"data: [DONE]\n\n",
			call: func(u string) (Provider, error) { return &OpenAIProvider{BaseURL: u, APIKey: "k", Model: "m"}, nil },
			want: false,
		},
		{
			// 上下文满 ≠ 输出被切断：续写只会连着撞同一堵墙，该走压缩。
			name: "chat_completions：context window 不算输出切断",
			sse: "data: {\"choices\":[{\"delta\":{\"content\":\"半截\"}}]}\n\n" +
				"data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"model_context_window_exceeded\"}],\"usage\":{\"prompt_tokens\":5,\"completion_tokens\":2}}\n\n" +
				"data: [DONE]\n\n",
			call: func(u string) (Provider, error) { return &OpenAIProvider{BaseURL: u, APIKey: "k", Model: "m"}, nil },
			want: false,
		},
		{
			name: "anthropic：message_delta 的 stop_reason=max_tokens",
			sse: "data: {\"type\":\"content_block_start\",\"index\":0,\"content_block\":{\"type\":\"text\",\"text\":\"\"}}\n\n" +
				"data: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"text_delta\",\"text\":\"还在写\"}}\n\n" +
				"data: {\"type\":\"message_delta\",\"delta\":{\"stop_reason\":\"max_tokens\"},\"usage\":{\"output_tokens\":64}}\n\n" +
				"data: {\"type\":\"message_stop\"}\n\n",
			call: func(u string) (Provider, error) {
				return &AnthropicProvider{BaseURL: u, APIKey: "k", Model: "m", MaxTokens: 128}, nil
			},
			want: true,
		},
		{
			name: "responses：incomplete + max_output_tokens",
			sse: "data: {\"type\":\"response.output_text.delta\",\"delta\":\"写了一半\"}\n\n" +
				"data: {\"type\":\"response.incomplete\",\"response\":{\"status\":\"incomplete\",\"output\":[],\"incomplete_details\":{\"reason\":\"max_output_tokens\"},\"usage\":{\"input_tokens\":4,\"output_tokens\":9}}}\n\n",
			call: func(u string) (Provider, error) { return &ResponsesProvider{BaseURL: u, APIKey: "k", Model: "m"}, nil },
			want: true,
		},
		{
			// 审核拦下的 incomplete 续写无意义（模型不是没写完，是不肯写完）。
			name: "responses：incomplete 但原因是内容审核",
			sse: "data: {\"type\":\"response.output_text.delta\",\"delta\":\"抱歉\"}\n\n" +
				"data: {\"type\":\"response.incomplete\",\"response\":{\"status\":\"incomplete\",\"output\":[],\"incomplete_details\":{\"reason\":\"content_filter\"},\"usage\":{\"input_tokens\":4,\"output_tokens\":2}}}\n\n",
			call: func(u string) (Provider, error) { return &ResponsesProvider{BaseURL: u, APIKey: "k", Model: "m"}, nil },
			want: false,
		},
	}
	for _, c := range cases {
		t.Run(c.name, func(t *testing.T) {
			var body map[string]any
			srv := captureServer(&body, c.sse)
			defer srv.Close()
			p, err := c.call(srv.URL)
			if err != nil {
				t.Fatal(err)
			}
			m, err := p.Stream(context.Background(), "sys",
				[]Message{{Role: RoleUser, Content: "hi"}}, nil, StreamOptions{}, nil)
			if err != nil {
				t.Fatalf("这一轮有正文，不该报错：%v", err)
			}
			if m.OutputLimited != c.want {
				t.Errorf("OutputLimited = %v，want %v（正文 %q）", m.OutputLimited, c.want, m.Content)
			}
			if strings.TrimSpace(m.Content) == "" {
				t.Error("fixture 没把正文收进来，用例本身失效")
			}
		})
	}
}

// TestToolCallAtOutputLimitIsNotMarkedCut：带工具调用的响应不算"被切断的最终回答"。
// 真是 length 也说明参数 JSON 已经残缺，续写救不回来——标了反而让上层去续一个
// 不该续的东西。
func TestToolCallAtOutputLimitIsNotMarkedCut(t *testing.T) {
	sse := "data: {\"choices\":[{\"delta\":{\"tool_calls\":[{\"index\":0,\"id\":\"c1\",\"type\":\"function\"," +
		"\"function\":{\"name\":\"bash\",\"arguments\":\"{\\\"command\\\":\\\"ls\\\"}\"}}]}}]}\n\n" +
		"data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"length\"}],\"usage\":{\"prompt_tokens\":3,\"completion_tokens\":4}}\n\n" +
		"data: [DONE]\n\n"
	var body map[string]any
	srv := captureServer(&body, sse)
	defer srv.Close()
	p := &OpenAIProvider{BaseURL: srv.URL, APIKey: "k", Model: "m"}
	m, err := p.Stream(context.Background(), "sys", []Message{{Role: RoleUser, Content: "hi"}}, nil, StreamOptions{}, nil)
	if err != nil {
		t.Fatal(err)
	}
	if len(m.ToolCalls) != 1 {
		t.Fatalf("工具调用没收到：%+v", m.ToolCalls)
	}
	if m.OutputLimited {
		t.Error("带 tool_calls 的响应被标成了输出切断")
	}
}

func TestMarksOutputLimit(t *testing.T) {
	for r, want := range map[string]bool{
		"length": true, "LENGTH": true, " max_tokens": true, "max_output_tokens": true,
		"stop": false, "tool_calls": false, "": false, "model_context_window_exceeded": false,
		"content_filter": false,
	} {
		if got := marksOutputLimit(r); got != want {
			t.Errorf("marksOutputLimit(%q) = %v，want %v", r, got, want)
		}
	}
}