iop/apps/node/internal/adapters/openai_compat/openai_compat_test.go

package openai_compat

import (
	"bytes"
	"context"
	"encoding/json"
	"errors"
	"fmt"
	"net/http"
	"net/http/httptest"
	"strings"
	"sync"
	"testing"
	"time"

	"go.uber.org/zap"

	"iop/apps/node/internal/runtime"
	"iop/packages/go/config"
)

type fakeSink struct {
	mu     sync.Mutex
	events []runtime.RuntimeEvent
}

func (s *fakeSink) Emit(_ context.Context, event runtime.RuntimeEvent) error {
	s.mu.Lock()
	defer s.mu.Unlock()
	s.events = append(s.events, event)
	return nil
}

func (s *fakeSink) all() []runtime.RuntimeEvent {
	s.mu.Lock()
	defer s.mu.Unlock()
	return append([]runtime.RuntimeEvent(nil), s.events...)
}

func TestOpenAICompatCapabilitiesQueryModels(t *testing.T) {
	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		if r.URL.Path != "/v1/models" {
			t.Fatalf("unexpected path %s", r.URL.Path)
		}
		_, _ = w.Write([]byte(`{"object":"list","data":[{"id":"model-a"},{"id":"model-b"}]}`))
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{
		Provider:         "lemonade",
		Endpoint:         server.URL,
		Capacity:         10,
		MaxQueue:         20,
		QueueTimeoutMS:   5000,
		RequestTimeoutMS: 10000,
	}, zap.NewNop())
	caps, err := adapter.Capabilities(context.Background())
	if err != nil {
		t.Fatalf("Capabilities failed: %v", err)
	}
	if caps.AdapterName != Name {
		t.Fatalf("AdapterName: got %q want %q", caps.AdapterName, Name)
	}
	if got := strings.Join(caps.Targets, ","); got != "model-a,model-b" {
		t.Fatalf("targets: got %q", got)
	}
	if caps.MaxConcurrency != 10 {
		t.Fatalf("expected MaxConcurrency 10, got %d", caps.MaxConcurrency)
	}
	if caps.MaxQueue != 20 || caps.QueueTimeoutMS != 5000 || caps.RequestTimeoutMS != 10000 {
		t.Fatalf("queue config: got %+v", caps)
	}
	if caps.ProviderStatus != runtime.ProviderStatusAvailable {
		t.Fatalf("expected available, got %s", caps.ProviderStatus)
	}

	adapterDefault := New(config.OpenAICompatConf{Endpoint: server.URL}, zap.NewNop())
	capsDefault, err := adapterDefault.Capabilities(context.Background())
	if err != nil {
		t.Fatalf("Capabilities failed: %v", err)
	}
	if capsDefault.MaxConcurrency != 8 {
		t.Fatalf("expected default MaxConcurrency 8, got %d", capsDefault.MaxConcurrency)
	}
}

func TestOpenAICompatProbeProviderAvailability(t *testing.T) {
	t.Run("200_ok_and_target_hit", func(t *testing.T) {
		server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
			if r.URL.Path != "/v1/models" {
				t.Fatalf("unexpected path %s", r.URL.Path)
			}
			_, _ = w.Write([]byte(`{"object":"list","data":[{"id":"model-a"},{"id":"model-b"}]}`))
		}))
		defer server.Close()

		adapter := New(config.OpenAICompatConf{Provider: "lemonade", Endpoint: server.URL}, zap.NewNop())
		res, err := adapter.ProbeProvider(context.Background(), "model-a")
		if err != nil {
			t.Fatalf("ProbeProvider failed: %v", err)
		}
		if res.Status != runtime.ProviderStatusAvailable {
			t.Errorf("expected available, got %s", res.Status)
		}
		if len(res.Targets) != 2 || res.Targets[0] != "model-a" {
			t.Errorf("unexpected Targets: %+v", res.Targets)
		}
		if res.Metadata["provider"] != "lemonade" {
			t.Errorf("expected provider metadata, got %+v", res.Metadata)
		}
	})

	t.Run("200_ok_and_target_miss", func(t *testing.T) {
		server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
			_, _ = w.Write([]byte(`{"object":"list","data":[{"id":"model-a"}]}`))
		}))
		defer server.Close()

		adapter := New(config.OpenAICompatConf{Endpoint: server.URL}, zap.NewNop())
		res, err := adapter.ProbeProvider(context.Background(), "model-b")
		if err != nil {
			t.Fatalf("ProbeProvider failed: %v", err)
		}
		if res.Status != runtime.ProviderStatusUnavailable {
			t.Errorf("expected unavailable, got %s", res.Status)
		}
		if !strings.Contains(res.Detail, "not found") {
			t.Errorf("expected 'not found' in detail, got %s", res.Detail)
		}
	})

	t.Run("500_internal_error", func(t *testing.T) {
		server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
			w.WriteHeader(http.StatusInternalServerError)
		}))
		defer server.Close()

		adapter := New(config.OpenAICompatConf{Endpoint: server.URL}, zap.NewNop())
		res, err := adapter.ProbeProvider(context.Background(), "model-a")
		if err != nil {
			t.Fatalf("ProbeProvider failed: %v", err)
		}
		if res.Status != runtime.ProviderStatusUnavailable {
			t.Errorf("expected unavailable, got %s", res.Status)
		}
	})

	t.Run("empty_target", func(t *testing.T) {
		server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
			_, _ = w.Write([]byte(`{"object":"list","data":[{"id":"model-a"}]}`))
		}))
		defer server.Close()

		adapter := New(config.OpenAICompatConf{Endpoint: server.URL}, zap.NewNop())
		res, err := adapter.ProbeProvider(context.Background(), "")
		if err != nil {
			t.Fatalf("ProbeProvider failed: %v", err)
		}
		if res.Status != runtime.ProviderStatusAvailable {
			t.Errorf("expected available, got %s", res.Status)
		}
	})
}
func TestOpenAICompatExecuteStreamsDeltasAndFinishReason(t *testing.T) {
	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		if r.URL.Path != "/v1/chat/completions" {
			t.Fatalf("unexpected path %s", r.URL.Path)
		}
		w.Header().Set("Content-Type", "text/event-stream")
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"reasoning_content":"think "}}]}`)
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"content":"hello "}}]}`)
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"content":"world"},"finish_reason":"stop"}],"usage":{"prompt_tokens":5,"completion_tokens":2}}`)
		_, _ = fmt.Fprintf(w, "data: [DONE]\n\n")
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{Endpoint: server.URL}, zap.NewNop())
	sink := &fakeSink{}
	err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
		RunID:  "run-1",
		Target: "lemonade-model",
		Input:  map[string]any{"prompt": "say hello"},
	}, sink)
	if err != nil {
		t.Fatalf("Execute failed: %v", err)
	}

	events := sink.all()
	// start + reasoning_delta + delta + delta + complete
	if len(events) != 5 {
		t.Fatalf("expected 5 events, got %d: %+v", len(events), events)
	}
	if events[0].Type != runtime.EventTypeStart {
		t.Fatalf("expected start, got %+v", events[0])
	}
	if events[1].Type != runtime.EventTypeReasoningDelta || events[1].Delta != "think " {
		t.Fatalf("expected reasoning delta, got %+v", events[1])
	}
	if events[2].Delta+events[3].Delta != "hello world" {
		t.Fatalf("unexpected deltas: %q + %q", events[2].Delta, events[3].Delta)
	}
	complete := events[4]
	if complete.Type != runtime.EventTypeComplete {
		t.Fatalf("expected complete, got %+v", complete)
	}
	if complete.Metadata["finish_reason"] != "stop" {
		t.Fatalf("expected finish_reason stop, got %+v", complete.Metadata)
	}
	if complete.Usage == nil || complete.Usage.OutputTokens != 2 || complete.Usage.InputTokens != 5 {
		t.Fatalf("expected usage from chunk, got %+v", complete.Usage)
	}
}

func TestOpenAICompatExecuteParsesReasoningAndCachedInputTokens(t *testing.T) {
	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
		w.Header().Set("Content-Type", "text/event-stream")
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"content":"hi"},"finish_reason":"stop"}],"usage":{"prompt_tokens":12,"completion_tokens":7,"prompt_tokens_details":{"cached_tokens":4},"completion_tokens_details":{"reasoning_tokens":3}}}`)
		_, _ = fmt.Fprintf(w, "data: [DONE]\n\n")
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{Endpoint: server.URL}, zap.NewNop())
	sink := &fakeSink{}
	if err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
		RunID:  "run-1",
		Target: "qwen-model",
		Input:  map[string]any{"prompt": "hi"},
	}, sink); err != nil {
		t.Fatalf("Execute failed: %v", err)
	}

	events := sink.all()
	complete := events[len(events)-1]
	if complete.Type != runtime.EventTypeComplete || complete.Usage == nil {
		t.Fatalf("expected complete event with usage, got %+v", complete)
	}
	u := complete.Usage
	if u.InputTokens != 12 || u.OutputTokens != 7 {
		t.Fatalf("input/output tokens: got %d/%d, want 12/7", u.InputTokens, u.OutputTokens)
	}
	if u.CachedInputTokens != 4 {
		t.Fatalf("cached_input_tokens: got %d, want 4", u.CachedInputTokens)
	}
	if u.ReasoningTokens != 3 {
		t.Fatalf("reasoning_tokens: got %d, want 3", u.ReasoningTokens)
	}
}

func TestOpenAICompatExecuteStreamsReasoningAlias(t *testing.T) {
	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		if r.URL.Path != "/v1/chat/completions" {
			t.Fatalf("unexpected path %s", r.URL.Path)
		}
		w.Header().Set("Content-Type", "text/event-stream")
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"reasoning":"think alias "}}]}`)
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"content":"answer"}}]}`)
		_, _ = fmt.Fprintf(w, "data: [DONE]\n\n")
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{Endpoint: server.URL}, zap.NewNop())
	sink := &fakeSink{}
	err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
		RunID:  "run-1",
		Target: "qwen-model",
		Input:  map[string]any{"prompt": "say hello"},
	}, sink)
	if err != nil {
		t.Fatalf("Execute failed: %v", err)
	}

	events := sink.all()
	if len(events) != 4 {
		t.Fatalf("expected 4 events, got %d: %+v", len(events), events)
	}
	if events[1].Type != runtime.EventTypeReasoningDelta || events[1].Delta != "think alias " {
		t.Fatalf("expected reasoning alias delta, got %+v", events[1])
	}
	if events[2].Type != runtime.EventTypeDelta || events[2].Delta != "answer" {
		t.Fatalf("expected answer delta, got %+v", events[2])
	}
}

func TestOpenAICompatExecuteSendsHeaders(t *testing.T) {
	var gotAuth, gotContentType string
	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		gotAuth = r.Header.Get("Authorization")
		gotContentType = r.Header.Get("Content-Type")
		w.Header().Set("Content-Type", "text/event-stream")
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"content":"ok"}}]}`)
		_, _ = fmt.Fprintf(w, "data: [DONE]\n\n")
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{
		Endpoint: server.URL,
		Headers: map[string]string{
			"Authorization": "Bearer secret-token",
			// A user-provided Content-Type must not override the adapter-owned one.
			"Content-Type": "text/plain",
		},
	}, zap.NewNop())
	sink := &fakeSink{}
	if err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
		RunID:  "run-h",
		Target: "lemonade-model",
		Input:  map[string]any{"prompt": "hi"},
	}, sink); err != nil {
		t.Fatalf("Execute failed: %v", err)
	}
	if gotAuth != "Bearer secret-token" {
		t.Fatalf("expected Authorization header, got %q", gotAuth)
	}
	if gotContentType != "application/json" {
		t.Fatalf("expected adapter-owned content type, got %q", gotContentType)
	}
}

func TestOpenAICompatExecutePassesOptionsAsTopLevelFields(t *testing.T) {
	var body map[string]any
	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
			t.Fatalf("decode: %v", err)
		}
		w.Header().Set("Content-Type", "text/event-stream")
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"content":"ok"}}]}`)
		_, _ = fmt.Fprintf(w, "data: [DONE]\n\n")
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{Endpoint: server.URL}, zap.NewNop())
	sink := &fakeSink{}
	if err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
		RunID:  "run-o",
		Target: "lemonade-model",
		Input: map[string]any{
			"prompt": "hi",
			"options": map[string]any{
				"temperature": 0.2,
				"max_tokens":  float64(8),
				// adapter-owned fields must not be overridden by options
				"model":  "evil",
				"stream": false,
			},
		},
	}, sink); err != nil {
		t.Fatalf("Execute failed: %v", err)
	}
	if body["temperature"] != 0.2 {
		t.Fatalf("expected temperature top-level, got %v", body["temperature"])
	}
	if body["max_tokens"] != float64(8) {
		t.Fatalf("expected max_tokens top-level, got %v", body["max_tokens"])
	}
	if body["model"] != "lemonade-model" {
		t.Fatalf("options must not override model, got %v", body["model"])
	}
	if body["stream"] != true {
		t.Fatalf("options must not override stream, got %v", body["stream"])
	}
}

func TestOpenAICompatExecutePassesToolsAndToolChoice(t *testing.T) {
	var body map[string]any
	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
			t.Fatalf("decode: %v", err)
		}
		w.Header().Set("Content-Type", "text/event-stream")
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"content":"ok"}}]}`)
		_, _ = fmt.Fprintf(w, "data: [DONE]\n\n")
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{Endpoint: server.URL}, zap.NewNop())
	sink := &fakeSink{}
	if err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
		RunID:  "run-t",
		Target: "qwen3.6:35b",
		Input: map[string]any{
			"prompt":      "hi",
			"tools":       []any{map[string]any{"type": "function"}},
			"tool_choice": "none",
		},
	}, sink); err != nil {
		t.Fatalf("Execute failed: %v", err)
	}
	if tools, ok := body["tools"].([]any); !ok || len(tools) != 1 {
		t.Fatalf("tools not passed: %+v", body["tools"])
	}
	if body["tool_choice"] != "none" {
		t.Fatalf("tool_choice not passed: %+v", body["tool_choice"])
	}
}

func TestOpenAICompatExecuteRetriesSingleToolWhenAutoUnsupported(t *testing.T) {
	attempts := 0
	var retryBody map[string]any
	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		attempts++
		if attempts == 1 {
			w.WriteHeader(http.StatusBadRequest)
			_, _ = w.Write([]byte(`{"error":{"message":"\"auto\" tool choice requires --enable-auto-tool-choice and --tool-call-parser to be set"}}`))
			return
		}
		if err := json.NewDecoder(r.Body).Decode(&retryBody); err != nil {
			t.Fatalf("decode retry: %v", err)
		}
		w.Header().Set("Content-Type", "text/event-stream")
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"content":"ok"}}]}`)
		_, _ = fmt.Fprintf(w, "data: [DONE]\n\n")
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{Endpoint: server.URL}, zap.NewNop())
	sink := &fakeSink{}
	if err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
		RunID:  "run-retry",
		Target: "qwen3.6:35b",
		Input: map[string]any{
			"prompt": "status",
			"tools": []any{map[string]any{
				"type": "function",
				"function": map[string]any{
					"name": "run_commands",
				},
			}},
		},
	}, sink); err != nil {
		t.Fatalf("Execute failed: %v", err)
	}
	if attempts != 2 {
		t.Fatalf("attempts: got %d, want 2", attempts)
	}
	choice, ok := retryBody["tool_choice"].(map[string]any)
	if !ok {
		t.Fatalf("retry tool_choice missing: %+v", retryBody)
	}
	fn := choice["function"].(map[string]any)
	if choice["type"] != "function" || fn["name"] != "run_commands" {
		t.Fatalf("retry tool_choice: %+v", choice)
	}
}

func TestOpenAICompatExecuteFallsBackToTextToolsWhenNativeToolsUnsupported(t *testing.T) {
	attempts := 0
	var fallbackBody map[string]any
	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		attempts++
		switch attempts {
		case 1:
			w.WriteHeader(http.StatusBadRequest)
			_, _ = w.Write([]byte(`{"error":{"message":"\"auto\" tool choice requires --enable-auto-tool-choice and --tool-call-parser to be set"}}`))
			return
		case 2:
			w.WriteHeader(http.StatusBadRequest)
			_, _ = w.Write([]byte(`{"error":{"message":"tool_choice=\"function=ChatCompletionNamedFunction(name='run_commands') type='function'\" requires --tool-call-parser to be set"}}`))
			return
		}
		if err := json.NewDecoder(r.Body).Decode(&fallbackBody); err != nil {
			t.Fatalf("decode fallback: %v", err)
		}
		w.Header().Set("Content-Type", "text/event-stream")
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"content":"<tool_call>\n<function=run_commands>\n<parameter=commands>[\"git status\"]</parameter>\n</function>\n</tool_call>"}}]}`)
		_, _ = fmt.Fprintf(w, "data: [DONE]\n\n")
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{Endpoint: server.URL}, zap.NewNop())
	sink := &fakeSink{}
	if err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
		RunID:  "run-text-fallback",
		Target: "qwen3.6:35b",
		Input: map[string]any{
			"messages": []any{
				map[string]any{"role": "system", "content": "Existing Cline system prompt."},
				map[string]any{"role": "user", "content": "status"},
			},
			"tools": []any{map[string]any{
				"type": "function",
				"function": map[string]any{
					"name": "run_commands",
					"parameters": map[string]any{
						"type": "object",
						"properties": map[string]any{
							"commands": map[string]any{"type": "array", "items": map[string]any{"type": "string"}},
						},
					},
				},
			}},
		},
	}, sink); err != nil {
		t.Fatalf("Execute failed: %v", err)
	}
	if attempts != 3 {
		t.Fatalf("attempts: got %d, want 3", attempts)
	}
	if _, ok := fallbackBody["tools"]; ok {
		t.Fatalf("fallback must omit tools: %+v", fallbackBody["tools"])
	}
	if _, ok := fallbackBody["tool_choice"]; ok {
		t.Fatalf("fallback must omit tool_choice: %+v", fallbackBody["tool_choice"])
	}
	messages := fallbackBody["messages"].([]any)
	first := messages[0].(map[string]any)
	if first["role"] != "system" || !strings.Contains(first["content"].(string), "<tool_call>") || !strings.Contains(first["content"].(string), "run_commands") {
		t.Fatalf("fallback system instruction missing tool format/name: %+v", first)
	}
	if !strings.Contains(first["content"].(string), "client workspace root") || !strings.Contains(first["content"].(string), "Do not prepend cd") {
		t.Fatalf("fallback system instruction missing workspace-root guidance: %+v", first)
	}
	if !strings.Contains(first["content"].(string), "Existing Cline system prompt.") {
		t.Fatalf("fallback system instruction did not preserve existing system content: %+v", first)
	}
	if len(messages) != 2 || messages[1].(map[string]any)["role"] != "user" {
		t.Fatalf("fallback should keep a single leading system message: %+v", messages)
	}
	events := sink.all()
	complete := events[len(events)-1]
	if complete.Metadata[runtimeMetadataOpenAITextToolFallback] != "true" {
		t.Fatalf("fallback metadata missing: %+v", complete.Metadata)
	}
}

func TestOpenAICompatExecutePreservesToolCallMessages(t *testing.T) {
	var body map[string]any
	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
			t.Fatalf("decode: %v", err)
		}
		w.Header().Set("Content-Type", "text/event-stream")
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"content":"ok"}}]}`)
		_, _ = fmt.Fprintf(w, "data: [DONE]\n\n")
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{Endpoint: server.URL}, zap.NewNop())
	sink := &fakeSink{}
	if err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
		RunID:  "run-tool-turn",
		Target: "qwen3.6:35b",
		Input: map[string]any{
			"messages": []any{
				map[string]any{
					"role":    "assistant",
					"content": "",
					"tool_calls": []any{map[string]any{
						"id":   "call_iop_1",
						"type": "function",
						"function": map[string]any{
							"name":      "run_commands",
							"arguments": `{"commands":["git status"]}`,
						},
					}},
				},
				map[string]any{
					"role":         "tool",
					"content":      "clean",
					"tool_call_id": "call_iop_1",
				},
			},
		},
	}, sink); err != nil {
		t.Fatalf("Execute failed: %v", err)
	}
	messages, ok := body["messages"].([]any)
	if !ok || len(messages) != 2 {
		t.Fatalf("messages not preserved: %+v", body["messages"])
	}
	assistantMsg := messages[0].(map[string]any)
	if calls, ok := assistantMsg["tool_calls"].([]any); !ok || len(calls) != 1 {
		t.Fatalf("assistant tool_calls not preserved: %+v", assistantMsg)
	}
	toolMsg := messages[1].(map[string]any)
	if toolMsg["tool_call_id"] != "call_iop_1" {
		t.Fatalf("tool_call_id not preserved: %+v", toolMsg)
	}
}

func TestOpenAICompatExecutePreservesNativeToolCalls(t *testing.T) {
	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
		w.Header().Set("Content-Type", "text/event-stream")
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"tool_calls":[{"index":0,"id":"call_1","type":"function","function":{"name":"run_commands","arguments":"{\"commands\":"}}]}}]}`)
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"tool_calls":[{"index":0,"function":{"arguments":"[\"git status\"]}"}}]},"finish_reason":"tool_calls"}]}`)
		_, _ = fmt.Fprintf(w, "data: [DONE]\n\n")
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{Endpoint: server.URL}, zap.NewNop())
	sink := &fakeSink{}
	if err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
		RunID:  "run-native-tools",
		Target: "qwen3.6:35b",
		Input:  map[string]any{"prompt": "status"},
	}, sink); err != nil {
		t.Fatalf("Execute failed: %v", err)
	}
	events := sink.all()
	complete := events[len(events)-1]
	if complete.Metadata["finish_reason"] != "tool_calls" {
		t.Fatalf("finish_reason: %+v", complete.Metadata)
	}
	var toolCalls []map[string]any
	if err := json.Unmarshal([]byte(complete.Metadata[runtimeMetadataOpenAIToolCalls]), &toolCalls); err != nil {
		t.Fatalf("tool_calls metadata JSON: %v", err)
	}
	if len(toolCalls) != 1 || toolCalls[0]["id"] != "call_1" {
		t.Fatalf("tool_calls: %+v", toolCalls)
	}
	fn := toolCalls[0]["function"].(map[string]any)
	if fn["name"] != "run_commands" || fn["arguments"] != `{"commands":["git status"]}` {
		t.Fatalf("function: %+v", fn)
	}
}

func TestOpenAICompatExecuteRejectsEmptyEndpointOrModel(t *testing.T) {
	t.Run("empty_endpoint", func(t *testing.T) {
		adapter := New(config.OpenAICompatConf{}, zap.NewNop())
		err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
			RunID: "r", Target: "m", Input: map[string]any{"prompt": "hi"},
		}, &fakeSink{})
		if err == nil || !strings.Contains(err.Error(), "endpoint is required") {
			t.Fatalf("expected endpoint error, got %v", err)
		}
	})

	t.Run("empty_model", func(t *testing.T) {
		adapter := New(config.OpenAICompatConf{Endpoint: "http://localhost:8000"}, zap.NewNop())
		err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
			RunID: "r", Input: map[string]any{"prompt": "hi"},
		}, &fakeSink{})
		if err == nil || !strings.Contains(err.Error(), "target/model is required") {
			t.Fatalf("expected model error, got %v", err)
		}
	})
}

func TestOpenAICompatJoinOpenAIPathNoDuplicateV1(t *testing.T) {
	cases := []struct {
		endpoint string
		path     string
		wantTail string
	}{
		{"http://host:8000", "/v1/models", "/v1/models"},
		{"http://host:8000/", "/v1/chat/completions", "/v1/chat/completions"},
		{"http://host:8000/v1", "/v1/models", "/v1/models"},
		{"http://host:8000/v1/", "/v1/chat/completions", "/v1/chat/completions"},
	}
	for _, c := range cases {
		got := joinOpenAIPath(strings.TrimRight(c.endpoint, "/"), c.path)
		if !strings.HasSuffix(got, c.wantTail) || strings.Contains(got, "/v1/v1") {
			t.Errorf("joinOpenAIPath(%q,%q)=%q want suffix %q without /v1/v1", c.endpoint, c.path, got, c.wantTail)
		}
	}
}

func TestOpenAICompatExecuteMapsThinkFalseForVLLMChatTemplateKwargs(t *testing.T) {
	var body map[string]any
	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
			t.Fatalf("decode: %v", err)
		}
		w.Header().Set("Content-Type", "text/event-stream")
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"content":"ok"}}]}`)
		_, _ = fmt.Fprintf(w, "data: [DONE]\n\n")
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{
		Provider: "vllm",
		Endpoint: server.URL,
	}, zap.NewNop())

	sink := &fakeSink{}
	err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
		RunID:  "run-vllm",
		Target: "qwen-model",
		Input: map[string]any{
			"prompt": "hello",
			"think":  false,
		},
	}, sink)
	if err != nil {
		t.Fatalf("Execute failed: %v", err)
	}

	ctk, ok := body["chat_template_kwargs"].(map[string]any)
	if !ok {
		t.Fatalf("expected chat_template_kwargs to be a map, got %T", body["chat_template_kwargs"])
	}
	if ctk["enable_thinking"] != false {
		t.Fatalf("expected enable_thinking to be false, got %v", ctk["enable_thinking"])
	}
}

func TestOpenAICompatExecuteVLLMMLXUnsupportedThinkDisable(t *testing.T) {
	// The vLLM-MLX runtime cannot produce a reasoning-free streaming response,
	// so think=false / reasoning_effort=none must fail loudly instead of being
	// silently accepted as a reasoning stream.
	for _, disable := range []map[string]any{
		{"prompt": "hello", "think": false},
		{"prompt": "hello", "reasoning_effort": "none"},
	} {
		adapter := New(config.OpenAICompatConf{
			Provider: "vllm-mlx",
			Endpoint: "http://localhost:8000",
		}, zap.NewNop())

		sink := &fakeSink{}
		err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
			RunID:  "run-vllm-mlx-disable",
			Target: "qwen-model",
			Input:  disable,
		}, sink)
		if err == nil {
			t.Fatalf("expected error for vllm-mlx think disable %v, got nil", disable)
		}
		if !strings.Contains(err.Error(), "unsupported think control") {
			t.Fatalf("expected error message to contain 'unsupported think control' for %v, got: %v", disable, err)
		}
	}
}

func TestOpenAICompatExecuteMapsThinkFalseForLemonadeChatTemplateKwargs(t *testing.T) {
	var body map[string]any
	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
			t.Fatalf("decode: %v", err)
		}
		w.Header().Set("Content-Type", "text/event-stream")
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"content":"ok"}}]}`)
		_, _ = fmt.Fprintf(w, "data: [DONE]\n\n")
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{
		Provider: "lemonade",
		Endpoint: server.URL,
	}, zap.NewNop())

	sink := &fakeSink{}
	err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
		RunID:  "run-lemonade",
		Target: "lemonade-model",
		Input: map[string]any{
			"prompt": "hello",
			"think":  false,
		},
	}, sink)
	if err != nil {
		t.Fatalf("Execute failed: %v", err)
	}

	if _, ok := body["think"]; ok {
		t.Fatalf("expected top-level think to be omitted for lemonade, got %v", body["think"])
	}
	ctk, ok := body["chat_template_kwargs"].(map[string]any)
	if !ok {
		t.Fatalf("expected chat_template_kwargs to be a map, got %T", body["chat_template_kwargs"])
	}
	if ctk["enable_thinking"] != false {
		t.Fatalf("expected enable_thinking to be false, got %v", ctk["enable_thinking"])
	}
}

func TestOpenAICompatExecuteOmitsThinkControlWhenRequestOmitsIt(t *testing.T) {
	for _, provider := range []string{"vllm", "vllm-mlx", "lemonade", ""} {
		t.Run(provider, func(t *testing.T) {
			var body map[string]any
			server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
				if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
					t.Fatalf("decode: %v", err)
				}
				w.Header().Set("Content-Type", "text/event-stream")
				_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"content":"ok"}}]}`)
				_, _ = fmt.Fprintf(w, "data: [DONE]\n\n")
			}))
			defer server.Close()

			adapter := New(config.OpenAICompatConf{
				Provider: provider,
				Endpoint: server.URL,
			}, zap.NewNop())

			sink := &fakeSink{}
			err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
				RunID:  "run-omit",
				Target: "qwen-model",
				Input: map[string]any{
					"prompt": "hello",
				},
			}, sink)
			if err != nil {
				t.Fatalf("Execute failed: %v", err)
			}

			for _, key := range []string{"think", "reasoning_effort", "thinking_token_budget", "chat_template_kwargs"} {
				if _, ok := body[key]; ok {
					t.Fatalf("expected key %q to be omitted for provider %q", key, provider)
				}
			}
		})
	}
}

func TestOpenAICompatExecuteVLLMUnsupportedReasoningEffort(t *testing.T) {
	adapter := New(config.OpenAICompatConf{
		Provider: "vllm",
		Endpoint: "http://localhost:8000",
	}, zap.NewNop())

	sink := &fakeSink{}
	err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
		RunID:  "run-unsupported",
		Target: "qwen-model",
		Input: map[string]any{
			"prompt":           "hello",
			"reasoning_effort": "low",
		},
	}, sink)
	if err == nil {
		t.Fatalf("expected error for unsupported reasoning_effort, got nil")
	}
	if !strings.Contains(err.Error(), "unsupported think control") {
		t.Fatalf("expected error message to contain 'unsupported think control', got: %v", err)
	}
}

func TestOpenAICompatExecuteVLLMUnsupportedNegativeBudget(t *testing.T) {
	adapter := New(config.OpenAICompatConf{
		Provider: "vllm",
		Endpoint: "http://localhost:8000",
	}, zap.NewNop())

	sink := &fakeSink{}
	err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
		RunID:  "run-unsupported",
		Target: "qwen-model",
		Input: map[string]any{
			"prompt":                "hello",
			"thinking_token_budget": -10,
		},
	}, sink)
	if err == nil {
		t.Fatalf("expected error for negative budget, got nil")
	}
	if !strings.Contains(err.Error(), "unsupported think control") {
		t.Fatalf("expected error message to contain 'unsupported think control', got: %v", err)
	}
}

func TestOpenAICompatExecuteMapsReasoningEffortNone(t *testing.T) {
	t.Run("vllm", func(t *testing.T) {
		var body map[string]any
		server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
			if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
				t.Fatalf("decode: %v", err)
			}
			w.Header().Set("Content-Type", "text/event-stream")
			_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"content":"ok"}}]}`)
			_, _ = fmt.Fprintf(w, "data: [DONE]\n\n")
		}))
		defer server.Close()

		adapter := New(config.OpenAICompatConf{
			Provider: "vllm",
			Endpoint: server.URL,
		}, zap.NewNop())

		sink := &fakeSink{}
		err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
			RunID:  "run-vllm-none",
			Target: "qwen-model",
			Input: map[string]any{
				"prompt":           "hello",
				"reasoning_effort": "none",
			},
		}, sink)
		if err != nil {
			t.Fatalf("Execute failed: %v", err)
		}

		ctk, ok := body["chat_template_kwargs"].(map[string]any)
		if !ok {
			t.Fatalf("expected chat_template_kwargs to be a map, got %T", body["chat_template_kwargs"])
		}
		if ctk["enable_thinking"] != false {
			t.Fatalf("expected enable_thinking to be false, got %v", ctk["enable_thinking"])
		}
	})

	t.Run("lemonade", func(t *testing.T) {
		var body map[string]any
		server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
			if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
				t.Fatalf("decode: %v", err)
			}
			w.Header().Set("Content-Type", "text/event-stream")
			_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"content":"ok"}}]}`)
			_, _ = fmt.Fprintf(w, "data: [DONE]\n\n")
		}))
		defer server.Close()

		adapter := New(config.OpenAICompatConf{
			Provider: "lemonade",
			Endpoint: server.URL,
		}, zap.NewNop())

		sink := &fakeSink{}
		err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
			RunID:  "run-lemonade-none",
			Target: "lemonade-model",
			Input: map[string]any{
				"prompt":           "hello",
				"reasoning_effort": "none",
			},
		}, sink)
		if err != nil {
			t.Fatalf("Execute failed: %v", err)
		}

		if _, ok := body["reasoning_effort"]; ok {
			t.Fatalf("expected reasoning_effort to be omitted, got %v", body["reasoning_effort"])
		}
		ctk, ok := body["chat_template_kwargs"].(map[string]any)
		if !ok {
			t.Fatalf("expected chat_template_kwargs to be a map, got %T", body["chat_template_kwargs"])
		}
		if ctk["enable_thinking"] != false {
			t.Fatalf("expected enable_thinking to be false, got %v", ctk["enable_thinking"])
		}
	})
}

func TestOpenAICompatExecuteMapsThinkTrueForVLLMChatTemplateKwargs(t *testing.T) {
	var body map[string]any
	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
			t.Fatalf("decode: %v", err)
		}
		w.Header().Set("Content-Type", "text/event-stream")
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"content":"ok"}}]}`)
		_, _ = fmt.Fprintf(w, "data: [DONE]\n\n")
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{
		Provider: "vllm",
		Endpoint: server.URL,
	}, zap.NewNop())

	sink := &fakeSink{}
	err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
		RunID:  "run-vllm",
		Target: "qwen-model",
		Input: map[string]any{
			"prompt": "hello",
			"think":  true,
		},
	}, sink)
	if err != nil {
		t.Fatalf("Execute failed: %v", err)
	}

	ctk, ok := body["chat_template_kwargs"].(map[string]any)
	if !ok {
		t.Fatalf("expected chat_template_kwargs to be a map, got %T", body["chat_template_kwargs"])
	}
	if ctk["enable_thinking"] != true {
		t.Fatalf("expected enable_thinking to be true, got %v", ctk["enable_thinking"])
	}
}

func TestOpenAICompatExecuteMapsThinkTrueForVLLMMLXChatTemplateKwargs(t *testing.T) {
	var body map[string]any
	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
			t.Fatalf("decode: %v", err)
		}
		w.Header().Set("Content-Type", "text/event-stream")
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"content":"ok"}}]}`)
		_, _ = fmt.Fprintf(w, "data: [DONE]\n\n")
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{
		Provider: "vllm-mlx",
		Endpoint: server.URL,
	}, zap.NewNop())

	sink := &fakeSink{}
	err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
		RunID:  "run-vllm-mlx",
		Target: "qwen-model",
		Input: map[string]any{
			"prompt": "hello",
			"think":  true,
		},
	}, sink)
	if err != nil {
		t.Fatalf("Execute failed: %v", err)
	}

	ctk, ok := body["chat_template_kwargs"].(map[string]any)
	if !ok {
		t.Fatalf("expected chat_template_kwargs to be a map, got %T", body["chat_template_kwargs"])
	}
	if ctk["enable_thinking"] != true {
		t.Fatalf("expected enable_thinking to be true, got %v", ctk["enable_thinking"])
	}
}

func TestOpenAICompatExecuteMapsThinkTrueForLemonadeChatTemplateKwargs(t *testing.T) {
	var body map[string]any
	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
			t.Fatalf("decode: %v", err)
		}
		w.Header().Set("Content-Type", "text/event-stream")
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"content":"ok"}}]}`)
		_, _ = fmt.Fprintf(w, "data: [DONE]\n\n")
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{
		Provider: "lemonade",
		Endpoint: server.URL,
	}, zap.NewNop())

	sink := &fakeSink{}
	err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
		RunID:  "run-lemonade",
		Target: "lemonade-model",
		Input: map[string]any{
			"prompt":                "hello",
			"think":                 true,
			"thinking_token_budget": 1024,
		},
	}, sink)
	if err != nil {
		t.Fatalf("Execute failed: %v", err)
	}

	if _, ok := body["think"]; ok {
		t.Fatalf("expected top-level think to be omitted for lemonade, got %v", body["think"])
	}
	if _, ok := body["thinking_token_budget"]; ok {
		t.Fatalf("expected top-level thinking_token_budget to be omitted for lemonade, got %v", body["thinking_token_budget"])
	}
	ctk, ok := body["chat_template_kwargs"].(map[string]any)
	if !ok {
		t.Fatalf("expected chat_template_kwargs to be a map, got %T", body["chat_template_kwargs"])
	}
	if ctk["enable_thinking"] != true {
		t.Fatalf("expected enable_thinking to be true, got %v", ctk["enable_thinking"])
	}
	if ctk["thinking_token_budget"] != float64(1024) {
		t.Fatalf("expected thinking_token_budget to be float64 1024, got %v", ctk["thinking_token_budget"])
	}
}

func TestOpenAICompatExecuteMapsThinkingTokenBudgetForVLLMChatTemplateKwargs(t *testing.T) {
	var body map[string]any
	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
			t.Fatalf("decode: %v", err)
		}
		w.Header().Set("Content-Type", "text/event-stream")
		_, _ = fmt.Fprintf(w, "data: %s\n\n", `{"choices":[{"delta":{"content":"ok"}}]}`)
		_, _ = fmt.Fprintf(w, "data: [DONE]\n\n")
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{
		Provider: "vllm",
		Endpoint: server.URL,
	}, zap.NewNop())

	sink := &fakeSink{}
	err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
		RunID:  "run-vllm-budget",
		Target: "qwen-model",
		Input: map[string]any{
			"prompt":                "hello",
			"thinking_token_budget": 2048,
		},
	}, sink)
	if err != nil {
		t.Fatalf("Execute failed: %v", err)
	}

	if _, ok := body["thinking_token_budget"]; ok {
		t.Fatalf("expected top-level thinking_token_budget to be omitted for vllm, got %v", body["thinking_token_budget"])
	}

	ctk, ok := body["chat_template_kwargs"].(map[string]any)
	if !ok {
		t.Fatalf("expected chat_template_kwargs to be a map, got %T", body["chat_template_kwargs"])
	}
	if ctk["enable_thinking"] != true {
		t.Fatalf("expected enable_thinking to be true for budget-only request, got %v", ctk["enable_thinking"])
	}
	if ctk["thinking_token_budget"] != float64(2048) {
		t.Fatalf("expected thinking_token_budget to be 2048, got %v", ctk["thinking_token_budget"])
	}
}

func TestOpenAICompatExecuteVLLMUnsupportedInvalidBudget(t *testing.T) {
	adapter := New(config.OpenAICompatConf{
		Provider: "vllm",
		Endpoint: "http://localhost:8000",
	}, zap.NewNop())

	sink := &fakeSink{}
	invalidBudgets := []any{"invalid", -5, 12.34}
	for _, val := range invalidBudgets {
		err := adapter.Execute(context.Background(), runtime.ExecutionSpec{
			RunID:  "run-vllm-invalid-budget",
			Target: "qwen-model",
			Input: map[string]any{
				"prompt":                "hello",
				"thinking_token_budget": val,
			},
		}, sink)
		if err == nil {
			t.Fatalf("expected error for budget val %v, got nil", val)
		}
		if !strings.Contains(err.Error(), "unsupported think control") {
			t.Fatalf("expected error message to contain 'unsupported think control' for val %v, got: %v", val, err)
		}
	}
}

type fakeTunnelSink struct {
	mu     sync.Mutex
	frames []runtime.ProviderTunnelFrame
}

func (s *fakeTunnelSink) EmitTunnelFrame(_ context.Context, frame runtime.ProviderTunnelFrame) error {
	s.mu.Lock()
	defer s.mu.Unlock()
	s.frames = append(s.frames, frame)
	return nil
}

func (s *fakeTunnelSink) all() []runtime.ProviderTunnelFrame {
	s.mu.Lock()
	defer s.mu.Unlock()
	return append([]runtime.ProviderTunnelFrame(nil), s.frames...)
}

// failingTunnelSink emulates the Edge socket write failing after failAfter
// successful frames, as when the Edge relay aborts an abandoned tunnel.
type failingTunnelSink struct {
	mu        sync.Mutex
	frames    []runtime.ProviderTunnelFrame
	failAfter int
}

func (s *failingTunnelSink) EmitTunnelFrame(_ context.Context, frame runtime.ProviderTunnelFrame) error {
	s.mu.Lock()
	defer s.mu.Unlock()
	if len(s.frames) >= s.failAfter {
		return errors.New("edge socket write failed")
	}
	s.frames = append(s.frames, frame)
	return nil
}

func (s *failingTunnelSink) all() []runtime.ProviderTunnelFrame {
	s.mu.Lock()
	defer s.mu.Unlock()
	return append([]runtime.ProviderTunnelFrame(nil), s.frames...)
}

// assertOrderedTunnelFrames verifies S09 frame semantics: contiguous sequences
// starting at 0 (no drop/reorder) and exactly one terminal END/ERROR frame,
// which must be the last frame emitted.
func assertOrderedTunnelFrames(t *testing.T, frames []runtime.ProviderTunnelFrame) {
	t.Helper()
	if len(frames) == 0 {
		t.Fatal("no tunnel frames emitted")
	}
	terminals := 0
	for i, f := range frames {
		if f.Sequence != int64(i) {
			t.Errorf("frame %d: sequence %d, want %d (ordered, lossless)", i, f.Sequence, i)
		}
		switch f.Kind {
		case runtime.ProviderTunnelFrameKindEnd, runtime.ProviderTunnelFrameKindError:
			terminals++
			if i != len(frames)-1 {
				t.Errorf("terminal %s frame at index %d must be the last of %d frames", f.Kind, i, len(frames))
			}
		}
	}
	if terminals != 1 {
		t.Errorf("expected exactly one terminal END/ERROR frame, got %d", terminals)
	}
}

func TestOpenAICompatTunnelProvider(t *testing.T) {
	expectedBody := "data: {\"choices\":[{\"delta\":{\"reasoning_content\":\"think Qwen\"}}]}\n\ndata: [DONE]\n\n"

	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		if r.Method != http.MethodPost {
			t.Errorf("expected POST method, got %s", r.Method)
		}
		if r.URL.Path != "/v1/chat/completions" {
			t.Errorf("expected path /v1/chat/completions, got %s", r.URL.Path)
		}
		if r.Header.Get("Authorization") != "Bearer test-key" {
			t.Errorf("expected Authorization header, got %s", r.Header.Get("Authorization"))
		}

		w.Header().Set("Content-Type", "text/event-stream")
		w.Header().Set("X-Custom-Header", "custom-value")
		w.WriteHeader(http.StatusOK)
		_, _ = w.Write([]byte(expectedBody))
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{
		Endpoint: server.URL,
	}, zap.NewNop())

	sink := &fakeTunnelSink{}
	req := runtime.ProviderTunnelRequest{
		RunID:    "run-1",
		TunnelID: "tunnel-1",
		Method:   "POST",
		Path:     "/v1/chat/completions",
		Headers:  map[string]string{"Authorization": "Bearer test-key"},
		Body:     []byte(`{"model":"qwen3.6:35b","prompt":"hi"}`),
	}

	err := adapter.TunnelProvider(context.Background(), req, sink)
	if err != nil {
		t.Fatalf("TunnelProvider failed: %v", err)
	}

	frames := sink.all()
	if len(frames) < 3 {
		t.Fatalf("expected at least 3 frames, got %d", len(frames))
	}

	startFrame := frames[0]
	if startFrame.Kind != runtime.ProviderTunnelFrameKindResponseStart {
		t.Errorf("expected RESPONSE_START, got %s", startFrame.Kind)
	}
	if startFrame.StatusCode != http.StatusOK {
		t.Errorf("expected 200 OK, got %d", startFrame.StatusCode)
	}
	if startFrame.Headers["X-Custom-Header"] != "custom-value" {
		t.Errorf("expected Custom Header, got %v", startFrame.Headers)
	}

	var bodyBuffer bytes.Buffer
	var lastSeq int64 = 0
	for _, f := range frames[1 : len(frames)-1] {
		if f.Kind != runtime.ProviderTunnelFrameKindBody {
			t.Errorf("expected BODY kind, got %s", f.Kind)
		}
		if f.Sequence != lastSeq+1 {
			t.Errorf("expected seq %d, got %d", lastSeq+1, f.Sequence)
		}
		bodyBuffer.Write(f.Body)
		lastSeq = f.Sequence
	}
	if bodyBuffer.String() != expectedBody {
		t.Errorf("body mismatch: got %q, want %q", bodyBuffer.String(), expectedBody)
	}

	endFrame := frames[len(frames)-1]
	if endFrame.Kind != runtime.ProviderTunnelFrameKindEnd {
		t.Errorf("expected END, got %s", endFrame.Kind)
	}
	if !endFrame.End {
		t.Errorf("expected End true, got %v", endFrame.End)
	}
	if endFrame.Sequence != lastSeq+1 {
		t.Errorf("expected seq %d, got %d", lastSeq+1, endFrame.Sequence)
	}
	assertOrderedTunnelFrames(t, frames)
}

func TestOpenAICompatTunnelProvider_RelaysProviderHTTPError(t *testing.T) {
	expectedBody := `{"error":{"message":"unsupported field","type":"invalid_request_error","param":"custom_provider_options"}}`

	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		if r.Method != http.MethodPost {
			t.Errorf("expected POST method, got %s", r.Method)
		}
		if r.URL.Path != "/v1/chat/completions" {
			t.Errorf("expected path /v1/chat/completions, got %s", r.URL.Path)
		}

		w.Header().Set("Content-Type", "application/json")
		w.WriteHeader(http.StatusUnprocessableEntity)
		_, _ = w.Write([]byte(expectedBody))
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{
		Endpoint: server.URL,
	}, zap.NewNop())

	sink := &fakeTunnelSink{}
	req := runtime.ProviderTunnelRequest{
		RunID:    "run-1",
		TunnelID: "tunnel-1",
		Method:   "POST",
		Path:     "/v1/chat/completions",
		Body:     []byte(`{"model":"qwen3.6:35b","custom_provider_options":{"reject":true}}`),
	}

	err := adapter.TunnelProvider(context.Background(), req, sink)
	if err != nil {
		t.Fatalf("TunnelProvider must relay provider HTTP errors as response frames, got error: %v", err)
	}

	frames := sink.all()
	if len(frames) < 3 {
		t.Fatalf("expected at least 3 frames, got %d", len(frames))
	}
	if frames[0].Kind != runtime.ProviderTunnelFrameKindResponseStart {
		t.Fatalf("expected RESPONSE_START, got %s", frames[0].Kind)
	}
	if frames[0].StatusCode != http.StatusUnprocessableEntity {
		t.Fatalf("expected provider status 422, got %d", frames[0].StatusCode)
	}
	if got := frames[0].Headers["Content-Type"]; got != "application/json" {
		t.Fatalf("expected provider Content-Type header, got %q", got)
	}
	var bodyBuffer bytes.Buffer
	for _, f := range frames[1 : len(frames)-1] {
		if f.Kind != runtime.ProviderTunnelFrameKindBody {
			t.Errorf("expected BODY kind, got %s", f.Kind)
		}
		bodyBuffer.Write(f.Body)
	}
	if bodyBuffer.String() != expectedBody {
		t.Fatalf("provider error body mismatch:\n got: %q\nwant: %q", bodyBuffer.String(), expectedBody)
	}
	if frames[len(frames)-1].Kind != runtime.ProviderTunnelFrameKindEnd {
		t.Fatalf("expected END, got %s", frames[len(frames)-1].Kind)
	}
	assertOrderedTunnelFrames(t, frames)
}

func TestOpenAICompatTunnelProvider_ResponsesPath(t *testing.T) {
	requestBody := `{"model":"served-model","input":"hi","max_output_tokens":123,"store":false}`
	expectedBody := "data: {\"type\":\"response.output_text.delta\",\"delta\":\"hi\"}\n\ndata: [DONE]\n\n"

	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		if r.Method != http.MethodPost {
			t.Errorf("expected POST method, got %s", r.Method)
		}
		if r.URL.Path != "/v1/responses" {
			t.Errorf("expected path /v1/responses, got %s", r.URL.Path)
		}
		if r.Header.Get("Authorization") != "Bearer test-key" {
			t.Errorf("expected Authorization header, got %s", r.Header.Get("Authorization"))
		}
		var got bytes.Buffer
		if _, err := got.ReadFrom(r.Body); err != nil {
			t.Fatalf("read request body: %v", err)
		}
		if got.String() != requestBody {
			t.Errorf("request body mismatch: got %q want %q", got.String(), requestBody)
		}

		w.Header().Set("Content-Type", "text/event-stream")
		w.WriteHeader(http.StatusOK)
		_, _ = w.Write([]byte(expectedBody))
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{
		Endpoint: server.URL,
	}, zap.NewNop())

	sink := &fakeTunnelSink{}
	req := runtime.ProviderTunnelRequest{
		RunID:    "run-1",
		TunnelID: "tunnel-1",
		Method:   "POST",
		Path:     "/v1/responses",
		Headers:  map[string]string{"Authorization": "Bearer test-key"},
		Body:     []byte(requestBody),
		Stream:   true,
	}

	err := adapter.TunnelProvider(context.Background(), req, sink)
	if err != nil {
		t.Fatalf("TunnelProvider failed: %v", err)
	}

	frames := sink.all()
	if len(frames) < 3 {
		t.Fatalf("expected at least 3 frames, got %d", len(frames))
	}
	if frames[0].Kind != runtime.ProviderTunnelFrameKindResponseStart {
		t.Errorf("expected RESPONSE_START, got %s", frames[0].Kind)
	}
	if frames[0].StatusCode != http.StatusOK {
		t.Errorf("expected 200 OK, got %d", frames[0].StatusCode)
	}
	var bodyBuffer bytes.Buffer
	for _, f := range frames[1 : len(frames)-1] {
		if f.Kind != runtime.ProviderTunnelFrameKindBody {
			t.Errorf("expected BODY kind, got %s", f.Kind)
		}
		bodyBuffer.Write(f.Body)
	}
	if bodyBuffer.String() != expectedBody {
		t.Errorf("body mismatch: got %q, want %q", bodyBuffer.String(), expectedBody)
	}
	if frames[len(frames)-1].Kind != runtime.ProviderTunnelFrameKindEnd {
		t.Errorf("expected END, got %s", frames[len(frames)-1].Kind)
	}
	assertOrderedTunnelFrames(t, frames)
}

func TestOpenAICompatTunnelProvider_Cancel(t *testing.T) {
	ctx, cancel := context.WithCancel(context.Background())
	handlerObservedCancel := make(chan struct{})

	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		w.Header().Set("Content-Type", "text/event-stream")
		w.WriteHeader(http.StatusOK)

		flusher, ok := w.(http.Flusher)
		for {
			_, err := w.Write([]byte("data: chunk\n\n"))
			if err != nil {
				close(handlerObservedCancel)
				return
			}
			if ok {
				flusher.Flush()
			}
			select {
			case <-r.Context().Done():
				close(handlerObservedCancel)
				return
			case <-time.After(10 * time.Millisecond):
			}
		}
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{
		Endpoint: server.URL,
	}, zap.NewNop())

	sink := &fakeTunnelSink{}
	req := runtime.ProviderTunnelRequest{
		RunID:    "run-1",
		TunnelID: "tunnel-1",
		Method:   "POST",
		Path:     "/v1/chat/completions",
		Body:     []byte(`{"prompt":"hi"}`),
	}

	errCh := make(chan error, 1)
	go func() {
		errCh <- adapter.TunnelProvider(ctx, req, sink)
	}()

	time.Sleep(100 * time.Millisecond)
	cancel()

	select {
	case err := <-errCh:
		if err == nil {
			t.Error("expected error from TunnelProvider on cancellation, got nil")
		}
	case <-time.After(2 * time.Second):
		t.Fatal("timeout waiting for TunnelProvider to return")
	}

	select {
	case <-handlerObservedCancel:
	case <-time.After(2 * time.Second):
		t.Fatal("handler did not observe context cancel")
	}

	frames := sink.all()
	if len(frames) < 2 {
		t.Fatalf("expected at least 2 frames, got %d: %+v", len(frames), frames)
	}
	assertOrderedTunnelFrames(t, frames)

	if frames[0].Kind != runtime.ProviderTunnelFrameKindResponseStart {
		t.Errorf("expected first frame RESPONSE_START, got %s", frames[0].Kind)
	}
	for i, f := range frames[1 : len(frames)-1] {
		if f.Kind != runtime.ProviderTunnelFrameKindBody {
			t.Errorf("mid frame %d: expected BODY before the terminal ERROR, got %s", i+1, f.Kind)
		}
	}
	lastFrame := frames[len(frames)-1]
	if lastFrame.Kind != runtime.ProviderTunnelFrameKindError {
		t.Errorf("expected last frame to be ERROR, got %s", lastFrame.Kind)
	}
	if lastFrame.Error == "" {
		t.Error("expected non-empty error message in frame")
	}
	if !strings.Contains(lastFrame.Error, "context canceled") {
		t.Errorf("expected cancellation error in terminal frame, got %q", lastFrame.Error)
	}
}

func TestOpenAICompatTunnelProvider_BodyReadError(t *testing.T) {
	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		hj, ok := w.(http.Hijacker)
		if !ok {
			t.Fatal("webserver doesn't support hijacking")
		}
		conn, _, err := hj.Hijack()
		if err != nil {
			t.Fatalf("hijack failed: %v", err)
		}
		defer conn.Close()

		_, _ = conn.Write([]byte("HTTP/1.1 200 OK\r\nContent-Type: text/event-stream\r\nContent-Length: 100\r\n\r\n"))
		_, _ = conn.Write([]byte("data: chunk1\n\n"))
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{
		Endpoint: server.URL,
	}, zap.NewNop())

	sink := &fakeTunnelSink{}
	req := runtime.ProviderTunnelRequest{
		RunID:    "run-1",
		TunnelID: "tunnel-1",
		Method:   "POST",
		Path:     "/v1/chat/completions",
		Body:     []byte(`{"prompt":"hi"}`),
	}

	err := adapter.TunnelProvider(context.Background(), req, sink)
	if err == nil {
		t.Fatal("expected error from body read failure, got nil")
	}

	frames := sink.all()
	if len(frames) < 2 {
		t.Fatalf("expected at least 2 frames, got %d: %+v", len(frames), frames)
	}
	assertOrderedTunnelFrames(t, frames)

	lastFrame := frames[len(frames)-1]
	if lastFrame.Kind != runtime.ProviderTunnelFrameKindError {
		t.Errorf("expected last frame to be ERROR, got %s", lastFrame.Kind)
	}
	if !strings.Contains(lastFrame.Error, "read response body") && !strings.Contains(lastFrame.Error, "EOF") {
		t.Errorf("expected body read error msg, got %q", lastFrame.Error)
	}
}

// TestOpenAICompatTunnelProvider_SinkWriteFailureStopsUpstreamRelay covers the
// Edge-write-failure leg of S09: when EmitTunnelFrame fails mid-stream the
// adapter stops relaying immediately, returns the emit error without appending
// further frames, and the upstream provider request is torn down.
func TestOpenAICompatTunnelProvider_SinkWriteFailureStopsUpstreamRelay(t *testing.T) {
	handlerObservedTeardown := make(chan struct{})

	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		w.Header().Set("Content-Type", "text/event-stream")
		w.WriteHeader(http.StatusOK)

		flusher, ok := w.(http.Flusher)
		for {
			_, err := w.Write([]byte("data: chunk\n\n"))
			if err != nil {
				close(handlerObservedTeardown)
				return
			}
			if ok {
				flusher.Flush()
			}
			select {
			case <-r.Context().Done():
				close(handlerObservedTeardown)
				return
			case <-time.After(10 * time.Millisecond):
			}
		}
	}))
	defer server.Close()

	adapter := New(config.OpenAICompatConf{
		Endpoint: server.URL,
	}, zap.NewNop())

	// RESPONSE_START and the first BODY frame succeed; the next emit fails as
	// if the Edge socket write broke after the caller went away.
	sink := &failingTunnelSink{failAfter: 2}
	req := runtime.ProviderTunnelRequest{
		RunID:    "run-1",
		TunnelID: "tunnel-1",
		Method:   "POST",
		Path:     "/v1/chat/completions",
		Body:     []byte(`{"prompt":"hi"}`),
	}

	errCh := make(chan error, 1)
	go func() {
		errCh <- adapter.TunnelProvider(context.Background(), req, sink)
	}()

	var tunnelErr error
	select {
	case tunnelErr = <-errCh:
	case <-time.After(2 * time.Second):
		t.Fatal("timeout waiting for TunnelProvider to return after sink write failure")
	}
	if tunnelErr == nil || !strings.Contains(tunnelErr.Error(), "edge socket write failed") {
		t.Fatalf("expected the sink emit error to propagate, got %v", tunnelErr)
	}

	select {
	case <-handlerObservedTeardown:
	case <-time.After(2 * time.Second):
		t.Fatal("provider handler did not observe upstream request teardown")
	}

	frames := sink.all()
	if len(frames) != 2 {
		t.Fatalf("expected relay to stop at 2 frames after emit failure, got %d: %+v", len(frames), frames)
	}
	if frames[0].Kind != runtime.ProviderTunnelFrameKindResponseStart {
		t.Errorf("expected first frame RESPONSE_START, got %s", frames[0].Kind)
	}
	if frames[1].Kind != runtime.ProviderTunnelFrameKindBody {
		t.Errorf("expected second frame BODY, got %s", frames[1].Kind)
	}
	for i, f := range frames {
		if f.Sequence != int64(i) {
			t.Errorf("frame %d: sequence %d, want %d", i, f.Sequence, i)
		}
	}
}