Cyrene_v2/internal/runtime/model/reasoning_shape_test.go

package model

import (
	"context"
	"net/http"
	"strings"
	"testing"
	"time"

	"git.yeij.top/AskaEth/Cyrene_v2/internal/config"
	"git.yeij.top/AskaEth/Cyrene_v2/internal/config/runtimeconfig"
)

// 本文件钉住「按提供商独立适配推理开关」（裁决 03.4.2 / 08.4.1）：声明了形状的端点
// 必须真的下发**自己形状**的开关（而不是只 WARN），且绝不发别人的形状——发错形状会
// 400，比不发更糟。自动模式（未声明形状）保持保守：形状有歧义时不下发。

// recvWithin 在超时内从记录通道取一次请求。路由/下发出错时（请求打到了另一个端点
// 或根本没发），用例必须**立刻失败**并指明哪个端点没收到，而不是挂在通道上把整个包
// 拖到测试超时——挂死的红是让人看不懂的红。
func recvWithin[T any](t *testing.T, ch <-chan T, endpoint string) T {
	t.Helper()
	select {
	case got := <-ch:
		return got
	case <-time.After(2 * time.Second):
		var zero T
		t.Fatalf("%s did not receive the request within 2s", endpoint)
		return zero
	}
}

// newKindAdapter 构造一个 OpenAI 兼容适配器，其配置声明了给定的推理开关形状。
// 真实端点由 stub server 顶替，请求体被记录供断言。
func newKindAdapter(t *testing.T, kind string, ch chan<- capturedRequest, opts ...OpenAIOption) *OpenAIChatAdapter {
	t.Helper()
	srv := newCompletionServer(t, http.StatusOK, "application/json", okCompletion, ch)
	return NewOpenAIChat(config.LLMConfig{BaseURL: srv.URL, Model: "endpoint-default", ReasoningKind: kind}, opts...)
}

// declaredShapeCases 是「声明形状 → 该形状的开关字段」的对照表。model 刻意取一个
// **会误导自动识别**的名字：显式形状必须压过模型名/BaseURL 的厂商推断。
var declaredShapeCases = []struct {
	name  string
	kind  string
	model string
	field string
}{
	{"deepseek", runtimeconfig.ReasoningKindDeepSeek, "qwen3-max", "thinking"},
	{"vllm", runtimeconfig.ReasoningKindVLLM, "deepseek-flash", "chat_template_kwargs"},
	{"dashscope", runtimeconfig.ReasoningKindDashScope, "deepseek-flash", "enable_thinking"},
}

// allSwitchFields 是全部形状的开关字段，用于断言「只发自己那一个」。
var allSwitchFields = []string{"thinking", "chat_template_kwargs", "enable_thinking"}

// switchBool 把一个开关字段读成布尔值（thinking 的开关读嵌套的 type）。
func switchBool(body map[string]any, field string) (bool, bool) {
	value, present := body[field]
	if !present {
		return false, false
	}
	switch v := value.(type) {
	case bool:
		return v, true
	case map[string]any:
		if field == "thinking" {
			return v["type"] == thinkingTypeEnabled, true
		}
		if nested, ok := v["enable_thinking"].(bool); ok {
			return nested, true
		}
	}
	return false, true
}

// TestOpenAIChat_DeclaredShapeSendsItsOwnSwitch 是本次裁决的主用例：每个 provider
// 声明自己的形状后，关闭思考必须真的下发到该形状的字段上，且请求体不含任何别的形状。
func TestOpenAIChat_DeclaredShapeSendsItsOwnSwitch(t *testing.T) {
	t.Parallel()
	for _, tc := range declaredShapeCases {
		t.Run(tc.name, func(t *testing.T) {
			t.Parallel()
			ch := make(chan capturedRequest, 1)
			ad := newKindAdapter(t, tc.kind, ch)
			if _, err := ad.Chat(context.Background(), Request{
				Model:    tc.model,
				Messages: []Message{TextMessage(RoleUser, "hi")},
				Thinking: &Thinking{Enable: boolPtr(false)},
			}, Options{}); err != nil {
				t.Fatalf("Chat error: %v", err)
			}
			body := recvWithin(t, ch, "the declared endpoint").body
			got, present := switchBool(body, tc.field)
			if !present {
				t.Fatalf("body[%q] missing: the declared shape %q was not forwarded: %v", tc.field, tc.kind, body)
			}
			if got {
				t.Fatalf("body[%q] = thinking on, want off", tc.field)
			}
			for _, other := range allSwitchFields {
				if other == tc.field {
					continue
				}
				if _, present := body[other]; present {
					t.Errorf("body[%q] = %v, want no other shape's field (shape %q)", other, body[other], tc.kind)
				}
			}
		})
	}
}

// TestOpenAIChat_DeclaredShapeEnablesThinking pins the on direction as well, so the
// switch cannot be hard-wired off by a refactor.
func TestOpenAIChat_DeclaredShapeEnablesThinking(t *testing.T) {
	t.Parallel()
	for _, tc := range declaredShapeCases {
		t.Run(tc.name, func(t *testing.T) {
			t.Parallel()
			ch := make(chan capturedRequest, 1)
			ad := newKindAdapter(t, tc.kind, ch)
			if _, err := ad.Chat(context.Background(), Request{
				Model:    tc.model,
				Messages: []Message{TextMessage(RoleUser, "hi")},
				Thinking: &Thinking{Enable: boolPtr(true)},
			}, Options{}); err != nil {
				t.Fatalf("Chat error: %v", err)
			}
			got, present := switchBool(recvWithin(t, ch, "the declared endpoint").body, tc.field)
			if !present || !got {
				t.Fatalf("body[%q] present=%v value=%v, want an explicit on", tc.field, present, got)
			}
		})
	}
}

// TestOpenAIChat_DeclaredSwitchShapeDropsEffort 是「该形状没有 reasoning_effort」
// 的红线用例：强度档被丢弃并报告，绝不改写成近似档（改档会反转调用方意图）。
func TestOpenAIChat_DeclaredSwitchShapeDropsEffort(t *testing.T) {
	t.Parallel()
	ch := make(chan capturedRequest, 1)
	log, buf := captureLog()
	ad := newKindAdapter(t, runtimeconfig.ReasoningKindVLLM, ch, WithOpenAILogger(log))

	if _, err := ad.Chat(context.Background(), Request{
		Model:           "qwen3-max",
		Messages:        []Message{TextMessage(RoleUser, "hi")},
		ReasoningEffort: "high",
	}, Options{}); err != nil {
		t.Fatalf("Chat error: %v", err)
	}
	body := recvWithin(t, ch, "the declared endpoint").body
	if _, present := body["reasoning_effort"]; present {
		t.Fatalf("body.reasoning_effort = %v, want the field dropped", body["reasoning_effort"])
	}
	if _, present := body["chat_template_kwargs"]; present {
		t.Fatalf("body.chat_template_kwargs = %v, want no switch invented from an effort", body["chat_template_kwargs"])
	}
	rec := buf.String()
	for _, want := range []string{"level=WARN", "reasoning_effort", "high", "vllm", "chat_template_kwargs.enable_thinking"} {
		if !strings.Contains(rec, want) {
			t.Errorf("log %q missing %q", rec, want)
		}
	}
}

// TestOpenAIChat_EffortNoneBecomesExplicitSwitchOff 是红线用例：`reasoning_effort:
// none` 的语义是**关思考**。在只有布尔开关的端点上它必须落成显式关闭，绝不能被
// 翻成任何强度档（那正是旧实现把「关」变成「开」的语义反转）。
func TestOpenAIChat_EffortNoneBecomesExplicitSwitchOff(t *testing.T) {
	t.Parallel()
	for _, tc := range declaredShapeCases {
		if tc.kind == runtimeconfig.ReasoningKindDeepSeek {
			continue // DeepSeek 有 reasoning_effort 字段，none 原样下发（另有用例钉住）
		}
		t.Run(tc.kind, func(t *testing.T) {
			t.Parallel()
			ch := make(chan capturedRequest, 1)
			log, buf := captureLog()
			ad := newKindAdapter(t, tc.kind, ch, WithOpenAILogger(log))
			if _, err := ad.Chat(context.Background(), Request{
				Model:           tc.model,
				Messages:        []Message{TextMessage(RoleUser, "hi")},
				ReasoningEffort: effortNone,
			}, Options{}); err != nil {
				t.Fatalf("Chat error: %v", err)
			}
			body := recvWithin(t, ch, "the declared endpoint").body
			got, present := switchBool(body, tc.field)
			if !present || got {
				t.Fatalf("body[%q] present=%v value=%v, want an explicit off", tc.field, present, got)
			}
			if _, present := body["reasoning_effort"]; present {
				t.Fatalf("body.reasoning_effort = %v, want the field dropped", body["reasoning_effort"])
			}
			if rec := buf.String(); !strings.Contains(rec, effortNone) || !strings.Contains(rec, "level=WARN") {
				t.Errorf("log %q, want a WARN explaining the none→off translation", rec)
			}
		})
	}
}

// TestOpenAIChat_ReasoningKindOpenAIKeepsOpenAITiers 用一个 DeepSeek 风格的模型名
// 证明显式形状压过厂商推断：`medium` 在 OpenAI 白名单里是合法档，必须原样下发，
// 不能按 DeepSeek 别名表被改写成 high。
func TestOpenAIChat_ReasoningKindOpenAIKeepsOpenAITiers(t *testing.T) {
	t.Parallel()
	ch := make(chan capturedRequest, 1)
	ad := newKindAdapter(t, runtimeconfig.ReasoningKindOpenAI, ch)

	if _, err := ad.Chat(context.Background(), Request{
		Model:           "deepseek-flash",
		Messages:        []Message{TextMessage(RoleUser, "hi")},
		ReasoningEffort: "medium",
	}, Options{}); err != nil {
		t.Fatalf("Chat error: %v", err)
	}
	body := recvWithin(t, ch, "the declared endpoint").body
	if got := body["reasoning_effort"]; got != "medium" {
		t.Fatalf("body.reasoning_effort = %v, want medium (an OpenAI tier, not the DeepSeek alias high)", got)
	}
	for _, other := range allSwitchFields {
		if _, present := body[other]; present {
			t.Errorf("body[%q] = %v, want no switch shape on an effort-only endpoint", other, body[other])
		}
	}
}

// TestOpenAIChat_ReasoningKindOpenAIReportsBooleanSwitch asserts the effort-only
// shape does not silently swallow a boolean thinking switch: it reports it, and
// points at the documented way to disable reasoning.
func TestOpenAIChat_ReasoningKindOpenAIReportsBooleanSwitch(t *testing.T) {
	t.Parallel()
	ch := make(chan capturedRequest, 1)
	log, buf := captureLog()
	ad := newKindAdapter(t, runtimeconfig.ReasoningKindOpenAI, ch, WithOpenAILogger(log))

	if _, err := ad.Chat(context.Background(), Request{
		Model:    "gpt-4o",
		Messages: []Message{TextMessage(RoleUser, "hi")},
		Thinking: &Thinking{Enable: boolPtr(false)},
	}, Options{}); err != nil {
		t.Fatalf("Chat error: %v", err)
	}
	body := recvWithin(t, ch, "the declared endpoint").body
	for _, other := range allSwitchFields {
		if _, present := body[other]; present {
			t.Fatalf("body[%q] present, want nothing sent for a boolean switch on an effort-only endpoint", other)
		}
	}
	rec := buf.String()
	for _, want := range []string{"level=WARN", "thinking.enable", "openai", "reasoning_effort"} {
		if !strings.Contains(rec, want) {
			t.Errorf("log %q missing %q", rec, want)
		}
	}
}

// TestOpenAIChat_UnknownReasoningKindSendsNothing asserts a shape this build does
// not implement is reported and sends nothing — even when the model name would
// otherwise look like a vendor we do understand. Guessing here is exactly the
// "fake adaptation" the ruling forbids.
func TestOpenAIChat_UnknownReasoningKindSendsNothing(t *testing.T) {
	t.Parallel()
	ch := make(chan capturedRequest, 1)
	log, buf := captureLog()
	ad := newKindAdapter(t, "gemini", ch, WithOpenAILogger(log))

	if _, err := ad.Chat(context.Background(), Request{
		Model:           "gpt-4o",
		Messages:        []Message{TextMessage(RoleUser, "hi")},
		Thinking:        &Thinking{Enable: boolPtr(false)},
		ReasoningEffort: "low",
	}, Options{}); err != nil {
		t.Fatalf("Chat error: %v", err)
	}
	body := recvWithin(t, ch, "the declared endpoint").body
	for _, field := range append(append([]string{}, allSwitchFields...), "reasoning_effort") {
		if _, present := body[field]; present {
			t.Errorf("body[%q] = %v, want nothing sent for an unknown shape", field, body[field])
		}
	}
	rec := buf.String()
	for _, want := range []string{"level=WARN", "reasoning.kind", "gemini"} {
		if !strings.Contains(rec, want) {
			t.Errorf("log %q missing %q", rec, want)
		}
	}
}

// TestOpenAIChat_AutoQwenShapeStaysConservative pins the un-declared default: a
// Qwen endpoint may be vLLM or DashScope, whose field shapes differ, so auto mode
// sends nothing and asks for an explicit declaration instead of guessing one.
func TestOpenAIChat_AutoQwenShapeStaysConservative(t *testing.T) {
	t.Parallel()
	ch := make(chan capturedRequest, 1)
	log, buf := captureLog()
	ad := newKindAdapter(t, "", ch, WithOpenAILogger(log))

	if _, err := ad.Chat(context.Background(), Request{
		Model:    "qwen3-max",
		Messages: []Message{TextMessage(RoleUser, "hi")},
		Thinking: &Thinking{Enable: boolPtr(false)},
	}, Options{}); err != nil {
		t.Fatalf("Chat error: %v", err)
	}
	body := recvWithin(t, ch, "the declared endpoint").body
	for _, field := range allSwitchFields {
		if _, present := body[field]; present {
			t.Errorf("body[%q] = %v, want nothing sent while the shape is ambiguous", field, body[field])
		}
	}
	rec := buf.String()
	for _, want := range []string{"level=WARN", "enable_thinking", "qwen", "reasoning.kind"} {
		if !strings.Contains(rec, want) {
			t.Errorf("log %q missing %q", rec, want)
		}
	}
}

// TestNewChatChain_CarriesReasoningKind 钉住配置到适配器的贯通：链的每一步都从
// 同一个 group 配置构建，因此声明过的形状必须随配置抵达适配器（否则 models.json
// 里的声明会静默失效）。
func TestNewChatChain_CarriesReasoningKind(t *testing.T) {
	t.Parallel()
	chain, err := NewChatChain(config.LLMConfig{
		Provider: ProviderOpenAI, BaseURL: "http://127.0.0.1:1/v1", Model: "m", Fallback: "fb",
		ReasoningKind: runtimeconfig.ReasoningKindDashScope,
	})
	if err != nil {
		t.Fatalf("NewChatChain: %v", err)
	}
	if len(chain.steps) != 2 {
		t.Fatalf("chain steps = %d, want 2 (primary + fallback)", len(chain.steps))
	}
	for i, step := range chain.steps {
		ad, ok := step.Adapter.(*OpenAIChatAdapter)
		if !ok {
			t.Fatalf("step %d adapter is %T, want *OpenAIChatAdapter", i, step.Adapter)
		}
		if ad.reasoningKind != runtimeconfig.ReasoningKindDashScope {
			t.Errorf("step %d reasoningKind = %q, want %q", i, ad.reasoningKind, runtimeconfig.ReasoningKindDashScope)
		}
	}
}