package vision

import (
	"context"
	"encoding/json"
	"io"
	"net/http"
	"net/http/httptest"
	"sync"
	"testing"

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

// modelRecorder is a /chat/completions endpoint that records the requested model,
// so a test can pin which model each direction (Describe / Extract) asked for.
type modelRecorder struct {
	mu     sync.Mutex
	models []string
}

func (r *modelRecorder) seen() []string {
	r.mu.Lock()
	defer r.mu.Unlock()
	return append([]string(nil), r.models...)
}

func newModelRecorderServer(t *testing.T) (*httptest.Server, *modelRecorder) {
	t.Helper()
	rec := &modelRecorder{}
	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		var body struct {
			Model string `json:"model"`
		}
		_ = json.NewDecoder(r.Body).Decode(&body)
		rec.mu.Lock()
		rec.models = append(rec.models, body.Model)
		rec.mu.Unlock()
		w.Header().Set("Content-Type", "application/json")
		_, _ = io.WriteString(w, `{"choices":[{"message":{"content":"ok"}}]}`)
	}))
	t.Cleanup(srv.Close)
	return srv, rec
}

// TestOpenAICompatibleAdapter_OCRUsesOwnModel pins the independent OCR model
// (裁决 14.4.1 的「独立视觉模型 + OCR 模型」): when LLM_OCR_MODEL is set the OCR
// direction sends it, while Describe keeps the vision model.
func TestOpenAICompatibleAdapter_OCRUsesOwnModel(t *testing.T) {
	t.Parallel()
	srv, rec := newModelRecorderServer(t)
	ad := NewOpenAICompatible(config.LLMConfig{
		BaseURL: srv.URL, Model: "vision-main", OCRModel: "ocr-special",
	})

	if _, err := ad.Describe(context.Background(), ImageRef{Source: SourceBase64, Data: "aGVsbG8="}, ""); err != nil {
		t.Fatalf("Describe: %v", err)
	}
	if _, err := ad.Extract(context.Background(), ImageRef{Source: SourceBase64, Data: "aGVsbG8="}); err != nil {
		t.Fatalf("Extract: %v", err)
	}
	seen := rec.seen()
	if len(seen) != 2 || seen[0] != "vision-main" || seen[1] != "ocr-special" {
		t.Fatalf("models seen = %v, want [vision-main ocr-special]", seen)
	}
}

// TestOpenAICompatibleAdapter_OCRWithoutOwnModelFallsBack pins the compatibility
// edge: without LLM_OCR_MODEL both directions use the vision model.
func TestOpenAICompatibleAdapter_OCRWithoutOwnModelFallsBack(t *testing.T) {
	t.Parallel()
	srv, rec := newModelRecorderServer(t)
	ad := NewOpenAICompatible(config.LLMConfig{BaseURL: srv.URL, Model: "vision-main"})

	if _, err := ad.Extract(context.Background(), ImageRef{Source: SourceBase64, Data: "aGVsbG8="}); err != nil {
		t.Fatalf("Extract: %v", err)
	}
	if seen := rec.seen(); len(seen) != 1 || seen[0] != "vision-main" {
		t.Fatalf("models seen = %v, want [vision-main]", seen)
	}
	// WithOCRModel overrides cfg.OCRModel, so an explicit option still wins.
	ad2 := NewOpenAICompatible(config.LLMConfig{BaseURL: srv.URL, Model: "vision-main", OCRModel: "cfg-ocr"}, WithOCRModel("opt-ocr"))
	if _, err := ad2.Extract(context.Background(), ImageRef{Source: SourceBase64, Data: "aGVsbG8="}); err != nil {
		t.Fatalf("Extract: %v", err)
	}
	if seen := rec.seen(); len(seen) != 2 || seen[1] != "opt-ocr" {
		t.Fatalf("models seen = %v, want the explicit option opt-ocr", seen)
	}
}