// LocalLS provides a gRPC client for the local Windsurf LanguageServerService.
//
// The correct chat flow routes through the local LS binary (which handles
// all auth/session management internally) rather than calling the upstream
// ApiServerService directly. The Cascade flow:
//
//  1. InitializeCascadePanelState — one-shot per LS session
//  2. AddTrackedWorkspace          — one-shot per LS session
//  3. UpdateWorkspaceTrust         — one-shot per LS session
//  4. StartCascade                 → cascade_id
//  5. SendUserCascadeMessage       — send prompt + model config
//  6. GetCascadeTrajectorySteps    — poll until trajectory status is IDLE (2)
package windsurf

import (
	"bytes"
	"context"
	"crypto/sha256"
	"crypto/tls"
	"encoding/binary"
	"encoding/hex"
	"fmt"
	"io"
	"log/slog"
	"net"
	"net/http"
	"net/url"
	"strings"
	"sync"
	"time"

	"golang.org/x/net/http2"
	"google.golang.org/protobuf/proto"

	pb "github.com/Wei-Shaw/sub2api/internal/gen/language_server_pb"
)

const (
	StartCascadeRPC               = "/exa.language_server_pb.LanguageServerService/StartCascade"
	InitPanelStateRPC             = "/exa.language_server_pb.LanguageServerService/InitializeCascadePanelState"
	AddTrackedWorkspaceRPC        = "/exa.language_server_pb.LanguageServerService/AddTrackedWorkspace"
	UpdateWorkspaceTrustRPC       = "/exa.language_server_pb.LanguageServerService/UpdateWorkspaceTrust"
	SendUserCascadeMessageRPC     = "/exa.language_server_pb.LanguageServerService/SendUserCascadeMessage"
	GetCascadeTrajectoryStepsRPC  = "/exa.language_server_pb.LanguageServerService/GetCascadeTrajectorySteps"
	GetCascadeTrajectoryStatusRPC = "/exa.language_server_pb.LanguageServerService/GetCascadeTrajectory"
	GetCascadeModelConfigsRPC     = "/exa.language_server_pb.LanguageServerService/GetCascadeModelConfigs"
)

// cascadeModelCapsCacheEntry 是单个 API key 下模型能力的缓存条目。
type cascadeModelCapsCacheEntry struct {
	SupportsImages map[string]bool
	FetchedAt      time.Time
}

// cascadeModelCapsTTL 能力缓存 TTL（5 分钟）。
const cascadeModelCapsTTL = 5 * time.Minute

// LocalLSClient talks to the local Windsurf LanguageServerService via h2c (plain HTTP/2 over TCP).
type LocalLSClient struct {
	BaseURL   string
	CSRFToken string
	HTTP      *http.Client
	SessionID string
	Warmed    bool
	// TrackedWorkspace is optional. When empty, the LS is treated as having no
	// server-side repository context and relies on caller-provided tool results.
	TrackedWorkspace string
	mu               sync.Mutex

	// 模型能力缓存（per-API-key hash），供 Cascade 图像 gate 使用。
	modelCapsMu    sync.Mutex
	modelCapsCache map[string]cascadeModelCapsCacheEntry
}

// NewLocalLSClient builds a client for the local LS at the given port.
func NewLocalLSClient(port int, csrfToken string) *LocalLSClient {
	h2cTransport := &http2.Transport{
		AllowHTTP: true,
		DialTLSContext: func(ctx context.Context, network, addr string, _ *tls.Config) (net.Conn, error) {
			return (&net.Dialer{Timeout: 5 * time.Second}).DialContext(ctx, network, addr)
		},
	}
	return &LocalLSClient{
		BaseURL:          fmt.Sprintf("http://localhost:%d", port),
		CSRFToken:        csrfToken,
		SessionID:        generateUUID(),
		TrackedWorkspace: "",
		HTTP: &http.Client{
			Transport: h2cTransport,
			Timeout:   60 * time.Second,
		},
	}
}

// WarmupCascade runs the one-shot panel init sequence required before StartCascade.
// Idempotent — skip if already warmed.
func (l *LocalLSClient) WarmupCascade(ctx context.Context, token string) error {
	return l.warmupCascade(ctx, token, false)
}

// ForceWarmupCascade resets session state and re-runs warmup.
func (l *LocalLSClient) ForceWarmupCascade(ctx context.Context, token string) error {
	return l.warmupCascade(ctx, token, true)
}

func (l *LocalLSClient) warmupCascade(ctx context.Context, token string, force bool) error {
	l.mu.Lock()
	defer l.mu.Unlock()

	if force {
		l.Warmed = false
		l.SessionID = generateUUID()
	}
	if l.Warmed {
		return nil
	}
	if l.SessionID == "" {
		l.SessionID = generateUUID()
	}

	var firstErr error

	// InitializeCascadePanelState: F1=metadata, F3=workspace_trusted (bool, true)
	initReq := encodeBytesField(1, buildMetadata(token, l.SessionID))
	initReq = append(initReq, encodeVarintField(3, 1)...)
	if err := l.grpcUnary(ctx, InitPanelStateRPC, initReq); err != nil {
		firstErr = err
	}

	// AddTrackedWorkspace is optional. Default Windsurf mode should not pretend
	// to have a mounted repository when the server does not actually have one.
	if strings.TrimSpace(l.TrackedWorkspace) != "" {
		addWsReq := encodeStringField(1, l.TrackedWorkspace)
		_ = l.grpcUnary(ctx, AddTrackedWorkspaceRPC, addWsReq)
	}

	// UpdateWorkspaceTrust: F1=metadata, F2=workspace_trusted (bool, true)
	trustReq := encodeBytesField(1, buildMetadata(token, l.SessionID))
	trustReq = append(trustReq, encodeVarintField(2, 1)...)
	if err := l.grpcUnary(ctx, UpdateWorkspaceTrustRPC, trustReq); err != nil && firstErr == nil {
		firstErr = err
	}

	// Only mark warmed on success (unlike the old code which always set true)
	if firstErr == nil {
		l.Warmed = true
	}
	return firstErr
}

// StartCascade calls StartCascade and returns the cascade_id.
// Retries once on panel-state-not-found.
func (l *LocalLSClient) StartCascade(ctx context.Context, token string) (string, error) {
	doStart := func() (string, error) {
		body := encodeBytesField(1, buildMetadata(token, l.SessionID))
		resp, err := l.grpcUnaryRaw(ctx, StartCascadeRPC, body)
		if err != nil {
			return "", fmt.Errorf("StartCascade: %w", err)
		}
		cascadeID, err := parseStringField1(resp)
		if err != nil {
			return "", fmt.Errorf("StartCascade parse: %w", err)
		}
		if cascadeID == "" {
			return "", fmt.Errorf("StartCascade: empty cascade_id (hex=%x)", resp)
		}
		return cascadeID, nil
	}

	cascadeID, err := doStart()
	if err != nil && isPanelStateNotFound(err) {
		_ = l.ForceWarmupCascade(ctx, token)
		return doStart()
	}
	return cascadeID, err
}

// SendUserCascadeMessage sends a message into an existing cascade session.
// Returns the (possibly new) cascadeID — it changes if panel-state retry triggers a new StartCascade.
// toolPreamble, if non-empty, is injected into the tool_calling_section override.
// SendUserCascadeMessage sends a user chat message to Cascade.
// Returns the (possibly new) cascadeID — it changes if panel-state retry triggers a new StartCascade.
// toolPreamble, if non-empty, is injected into the tool_calling_section override.
// images（可选）作为 SendUserCascadeMessageRequest.images (field 6) 追加到 proto wire。
//
// allowRecreate 控制 panel-state-not-found 时是否内部静默重建 cascade：
//   - true：ForceWarmup + StartCascade + 用 SAME text 再发一次。调用方须保证
//     text 已包含完整历史（否则新 cascade 无状态 + text 无历史 = 上下文丢失）。
//   - false：直接返回错误，让调用方重建含完整历史的 text 后再调。
//
// 经验值：StreamCascadeChat 内当 reuseCascadeID 为空（本地 StartCascade 的流程，
// text 已是 full-history）时传 true；reuse 场景（text 可能仅含最后一条消息）传 false。
func (l *LocalLSClient) SendUserCascadeMessage(ctx context.Context, token, cascadeID, text, modelUID, toolPreamble string, modelEnumHint int, images []CascadeImage, allowRecreate bool) (string, error) {
	modelEnum := resolveModelEnum(modelUID)
	if modelEnum == 0 && modelEnumHint > 0 {
		modelEnum = modelEnumHint
	}

	doSend := func(cid string) error {
		body := encodeStringField(1, cid)
		body = append(body, encodeBytesField(2, encodeStringField(1, text))...)
		body = append(body, encodeBytesField(3, buildMetadata(token, l.SessionID))...)
		body = append(body, encodeBytesField(5, buildCascadeConfig(modelUID, modelEnum, toolPreamble))...)
		// field 6: repeated CodeiumImage images（逆向自 Windsurf.app chat-client）
		body = appendSendUserCascadeImages(body, images)
		return l.grpcUnary(ctx, SendUserCascadeMessageRPC, body)
	}

	if err := doSend(cascadeID); err != nil {
		if !isPanelStateNotFound(err) {
			return "", err
		}
		if !allowRecreate {
			// reuse 场景：不要静默用 last-message-only 的 text 去灌新 cascade。
			// 返回错误，让 chatCascade 用 full-history text 重建整个调用。
			return "", err
		}
		_ = l.ForceWarmupCascade(ctx, token)
		newCascadeID, startErr := l.StartCascade(ctx, token)
		if startErr != nil {
			return "", startErr
		}
		if err := doSend(newCascadeID); err != nil {
			return "", err
		}
		return newCascadeID, nil
	}
	return cascadeID, nil
}

// buildMetadata builds the full Metadata proto for local LS calls, aligned with WindsurfAPI.
func buildMetadata(token, sessionID string) []byte {
	if sessionID == "" {
		sessionID = generateUUID()
	}
	var meta []byte
	meta = append(meta, encodeStringField(1, AppName)...)                        // ide_name
	meta = append(meta, encodeStringField(2, ExtensionVersion)...)               // extension_version
	meta = append(meta, encodeStringField(3, token)...)                          // api_key
	meta = append(meta, encodeStringField(4, "en")...)                           // locale
	meta = append(meta, encodeStringField(5, RuntimeOS())...)                    // os
	meta = append(meta, encodeStringField(7, IDEVersion)...)                     // ide_version
	meta = append(meta, encodeStringField(8, HardwareArch())...)                 // hardware
	meta = append(meta, encodeVarintField(9, uint64(time.Now().UnixMilli()))...) // request_id
	meta = append(meta, encodeStringField(10, sessionID)...)                     // session_id
	meta = append(meta, encodeStringField(12, AppName)...)                       // extension_name
	return meta
}

// buildSectionOverride builds a SectionOverrideConfig { mode=OVERRIDE(1), content=text }.
func buildSectionOverride(content string) []byte {
	var out []byte
	out = append(out, encodeVarintField(1, 1)...) // SECTION_OVERRIDE_MODE_OVERRIDE
	out = append(out, encodeStringField(2, content)...)
	return out
}

// buildCascadeConfig builds a CascadeConfig for the given model UID and enum.
// Uses NO_TOOL planner mode (3) with section overrides for pure conversational responses.
//
// Key insight (2026-04-12): NO_TOOL mode SUPPRESSES field 10 (tool_calling_section) —
// it is injected but never rendered to the model. Tool definitions MUST go into
// field 12 (additional_instructions_section) which IS rendered regardless of planner mode.
// Field 10 is kept as belt-and-suspenders.
func buildCascadeConfig(modelUID string, modelEnum int, toolPreamble string) []byte {
	var convParts []byte
	convParts = append(convParts, encodeVarintField(4, 3)...) // planner_mode=NO_TOOL(3)

	const toolReinforcement = "\n\nThe functions listed above are available and callable. " +
		"When the user's request can be answered by calling a function, emit a <tool_call> block as described. " +
		"Use this exact format: <tool_call>{\"name\":\"...\",\"arguments\":{...}}</tool_call>"

	if toolPreamble != "" {
		// Primary: field 12 (additional_instructions_section) — always rendered in NO_TOOL mode
		convParts = append(convParts, encodeBytesField(12, buildSectionOverride(toolPreamble+toolReinforcement))...)
		// Belt-and-suspenders: field 10 (tool_calling_section)
		convParts = append(convParts, encodeBytesField(10, buildSectionOverride(toolPreamble))...)
		// field 13 (communication_section)
		convParts = append(convParts, encodeBytesField(13, buildSectionOverride(
			"You are accessed via API. Respond in the same language as the user. "+
				"Use the functions above when relevant."))...)
	} else {
		// field 10: suppress built-in tool list
		convParts = append(convParts, encodeBytesField(10, buildSectionOverride("No tools are available."))...)
		// field 12: reinforce direct-answer mode
		convParts = append(convParts, encodeBytesField(12, buildSectionOverride(
			"You have no tools, no file access, and no command execution. "+
				"Answer all questions directly using your knowledge. "+
				"Never pretend to create files or check directories."))...)
		// field 11 (code_changes_section): suppress IDE-specific boilerplate
		convParts = append(convParts, encodeBytesField(11, buildSectionOverride(""))...)
		// field 13 (communication_section)
		convParts = append(convParts, encodeBytesField(13, buildSectionOverride(
			"You are accessed via API. Answer directly. "+
				"Respond in the same language as the user."))...)
	}

	// CortexPlannerConfig
	var plannerParts []byte
	plannerParts = append(plannerParts, encodeBytesField(2, convParts)...) // conversational=2

	if modelUID != "" {
		plannerParts = append(plannerParts, encodeStringField(35, modelUID)...)
		plannerParts = append(plannerParts, encodeStringField(34, modelUID)...)
	}
	if modelEnum > 0 {
		plannerParts = append(plannerParts, encodeBytesField(15, encodeVarintField(1, uint64(modelEnum)))...)
		plannerParts = append(plannerParts, encodeVarintField(1, uint64(modelEnum))...)
	}

	// max_output_tokens (field 6) = 32768 — prevents long response truncation
	plannerParts = append(plannerParts, encodeVarintField(6, 32768)...)

	// BrainConfig: F1=enabled=true, F6=update_strategy{dynamic_update{}}
	var brainParts []byte
	brainParts = append(brainParts, encodeVarintField(1, 1)...)
	brainParts = append(brainParts, encodeBytesField(6, encodeBytesField(6, nil))...)

	// memory_config (field 5): {enabled=false} — prevent LS injecting user's stored memories
	memoryConfig := encodeVarintField(1, 0) // bool enabled = false

	var cfg []byte
	cfg = append(cfg, encodeBytesField(1, plannerParts)...)
	cfg = append(cfg, encodeBytesField(5, memoryConfig)...)
	cfg = append(cfg, encodeBytesField(7, brainParts)...)
	return cfg
}

// isPanelStateNotFound detects "panel state not found" gRPC errors.
func isPanelStateNotFound(err error) bool {
	if err == nil {
		return false
	}
	s := strings.ToLower(err.Error())
	return strings.Contains(s, "panel state not found") ||
		strings.Contains(s, "not_found") && strings.Contains(s, "panel")
}

// NativeToolCall holds a structured tool call extracted from trajectory step metadata
// or from step oneof fields (tool_call_proposal, mcp_tool).
type NativeToolCall struct {
	ID            string
	Name          string
	ArgumentsJSON string
}

// TrajectoryStep holds the parsed content from a trajectory step.
type TrajectoryStep struct {
	Type         int
	Status       int
	Text         string // modifiedText || responseText (final preferred)
	ResponseText string // raw responseText (field 20/1) — monotonic during streaming
	Thinking     string // field 20/3
	ErrorText    string // field 24 or field 31
	Usage        *StepUsage
	ToolCall     *NativeToolCall // structured tool call from metadata/step oneof
}

// StepUsage holds server-reported token counts from step metadata.
type StepUsage struct {
	InputTokens      int
	OutputTokens     int
	CacheReadTokens  int
	CacheWriteTokens int
}

// GetTrajectoryStatus polls the trajectory status (field 2 varint).
// Returns 1 when the trajectory is IDLE (complete).
func (l *LocalLSClient) GetTrajectoryStatus(ctx context.Context, cascadeID string) (int, error) {
	body := encodeStringField(1, cascadeID)
	resp, err := l.grpcUnaryRaw(ctx, GetCascadeTrajectoryStatusRPC, body)
	if err != nil {
		return 0, err
	}
	status, _ := parseVarintField2(resp)
	return int(status), nil
}

// GetTrajectorySteps fetches trajectory steps starting at stepOffset.
func (l *LocalLSClient) GetTrajectorySteps(ctx context.Context, cascadeID string, stepOffset int) ([]TrajectoryStep, error) {
	body := encodeStringField(1, cascadeID)
	if stepOffset > 0 {
		body = append(body, encodeVarintField(2, uint64(stepOffset))...)
	}
	resp, err := l.grpcUnaryRaw(ctx, GetCascadeTrajectoryStepsRPC, body)
	if err != nil {
		return nil, err
	}
	return parseTrajectorySteps(resp), nil
}

// CascadeChatResult holds the full output from StreamCascadeChat.
type CascadeChatResult struct {
	Text        string
	Thinking    string
	Usage       *StepUsage // aggregated from all steps; nil if no server-reported data
	CascadeID   string
	FirstTextAt time.Time // when text first appeared (zero if no text)
	ToolCalls   []NativeToolCall
}

// CascadeModelError is raised when the trajectory contains an error step (type=17)
// or the planner stalls. Callers should retry with a different account.
type CascadeModelError struct {
	Msg string
}

func (e *CascadeModelError) Error() string { return e.Msg }

// StreamCascadeChat performs the full Cascade chat flow and returns accumulated text + thinking.
// Includes cold/warm stall detection, step error handling, and final sweep (aligned with JS v1.9).
// If reuseCascadeID is non-empty, skips StartCascade and reuses the existing cascade session.
// images 作为当前 user turn 的图像 sidecar 传递给 SendUserCascadeMessage（proto field 6）。
func (l *LocalLSClient) StreamCascadeChat(ctx context.Context, token, modelUID, userText, toolPreamble, reuseCascadeID string, modelEnumHint int, images []CascadeImage) (*CascadeChatResult, error) {
	if err := l.WarmupCascade(ctx, token); err != nil {
		return nil, fmt.Errorf("warmup: %w", err)
	}

	var cascadeID string
	var err error
	if reuseCascadeID != "" {
		cascadeID = reuseCascadeID
	} else {
		cascadeID, err = l.StartCascade(ctx, token)
		if err != nil {
			return nil, err
		}
	}

	// When reusing a cascade, capture the pre-existing step count so subsequent
	// polls only fetch new steps. Without this, Turn 2 would re-read Turn 1's
	// completed steps and append them again to the accumulated text, causing
	// the response to duplicate Turn 1's prefix (including prior <tool_call>).
	startStepIndex := 0
	if reuseCascadeID != "" {
		baselineSteps, berr := l.GetTrajectorySteps(ctx, cascadeID, 0)
		if berr == nil {
			startStepIndex = len(baselineSteps)
		}
	}

	// allowRecreate=true 仅对本流程内 StartCascade 出来的全新 cascade 安全：
	// 此时 userText 已是 full-history，内部遇到 panel-not-found 可静默重建再发。
	// reuse 场景（caller 传入 reuseCascadeID）下 userText 可能只含最后一条消息，
	// 静默重建会把空状态 cascade 当成有历史的 resume 用 → 上下文丢失，所以禁止。
	allowRecreate := reuseCascadeID == ""
	cascadeID, err = l.SendUserCascadeMessage(ctx, token, cascadeID, userText, modelUID, toolPreamble, modelEnumHint, images, allowRecreate)
	if err != nil {
		return nil, fmt.Errorf("SendUserCascadeMessage: %w", err)
	}

	const (
		maxWait          = 180 * time.Second
		idleGrace        = 8 * time.Second
		pollInterval     = 250 * time.Millisecond
		noGrowthStallMs  = 25000
		stallRetryMinLen = 300
	)

	textCursors := make(map[int]int)
	thinkCursors := make(map[int]int)
	var totalText, totalThinking int
	var accText, accThinking string
	var firstTextAt time.Time
	idleCount := 0
	sawActive := false
	sawText := false
	lastGrowthAt := time.Now()

	// Native tool call tracking
	seenToolCalls := make(map[string]bool)
	var nativeToolCalls []NativeToolCall
	lastStatus := 0
	startTime := time.Now()
	deadline := startTime.Add(maxWait)
	graceEnd := startTime.Add(idleGrace)
	inputChars := len(userText)

	// Aggregated step usage
	usageByStep := make(map[int]*StepUsage)

	for time.Now().Before(deadline) {
		select {
		case <-ctx.Done():
			return &CascadeChatResult{Text: SanitizePath(accText), Thinking: accThinking, CascadeID: cascadeID, FirstTextAt: firstTextAt, ToolCalls: nativeToolCalls}, ctx.Err()
		default:
		}

		time.Sleep(pollInterval)

		steps, err := l.GetTrajectorySteps(ctx, cascadeID, startStepIndex)
		if err != nil {
			continue
		}

		// Check for error steps (type=17)
		for _, s := range steps {
			if s.Type == 17 && s.ErrorText != "" {
				return nil, &CascadeModelError{Msg: s.ErrorText}
			}
		}

		// Cold stall: active but no text/thinking after threshold
		elapsed := time.Since(startTime)
		coldThreshold := 30*time.Second + time.Duration(inputChars/1500)*5*time.Second
		if coldThreshold > maxWait {
			coldThreshold = maxWait
		}
		if elapsed > coldThreshold && sawActive && !sawText && totalThinking == 0 {
			return nil, &CascadeModelError{Msg: fmt.Sprintf("Cascade planner stalled — no output after %ds", int(coldThreshold.Seconds()))}
		}

		for idx, s := range steps {
			// Usage
			if s.Usage != nil {
				usageByStep[idx] = s.Usage
			}

			// Thinking delta
			if s.Thinking != "" {
				prev := thinkCursors[idx]
				if len(s.Thinking) > prev {
					accThinking += s.Thinking[prev:]
					totalThinking += len(s.Thinking) - prev
					thinkCursors[idx] = len(s.Thinking)
					lastGrowthAt = time.Now()
				}
			}

			// Native tool call from structured step data
			if s.ToolCall != nil && s.ToolCall.Name != "" {
				key := s.ToolCall.Name + "|" + s.ToolCall.ID
				if !seenToolCalls[key] {
					seenToolCalls[key] = true
					nativeToolCalls = append(nativeToolCalls, *s.ToolCall)
					lastGrowthAt = time.Now()
					sawText = true
					if firstTextAt.IsZero() {
						firstTextAt = lastGrowthAt
					}
				}
			}

			// Text delta — use ResponseText during streaming for monotonic cursor
			liveText := s.ResponseText
			if liveText == "" {
				liveText = s.Text
			}
			if liveText == "" {
				continue
			}
			prev := textCursors[idx]
			if len(liveText) > prev {
				accText += liveText[prev:]
				totalText += len(liveText) - prev
				textCursors[idx] = len(liveText)
				lastGrowthAt = time.Now()
				if !sawText {
					firstTextAt = lastGrowthAt
				}
				sawText = true
			}
		}

		// Warm stall: text stopped growing for 25s while planner is active
		if sawText && lastStatus != 1 && time.Since(lastGrowthAt).Milliseconds() > noGrowthStallMs {
			if totalText < stallRetryMinLen {
				return nil, &CascadeModelError{Msg: "Cascade planner stalled after preamble — no progress for 25s"}
			}
			break // accept partial result
		}

		status, err := l.GetTrajectoryStatus(ctx, cascadeID)
		if err != nil {
			continue
		}
		lastStatus = status

		if status != 1 {
			sawActive = true
		}

		if status == 1 { // IDLE
			if !sawActive && time.Now().Before(graceEnd) {
				continue
			}
			idleCount++
			growthSettled := time.Since(lastGrowthAt) > pollInterval*2
			canBreak := false
			if sawText {
				canBreak = idleCount >= 2 && growthSettled
			} else {
				canBreak = idleCount >= 4
			}
			if canBreak {
				// Final sweep: fetch one more time to get modifiedText top-up
				finalSteps, err := l.GetTrajectorySteps(ctx, cascadeID, startStepIndex)
				if err == nil {
					for idx, s := range finalSteps {
						if s.Usage != nil {
							usageByStep[idx] = s.Usage
						}
						// Top up from responseText
						rt := s.ResponseText
						if rt == "" {
							rt = s.Text
						}
						prev := textCursors[idx]
						if len(rt) > prev {
							accText += rt[prev:]
							totalText += len(rt) - prev
							textCursors[idx] = len(rt)
						}
						// Modified-response top-up: only if it extends what we already emitted
						mt := s.Text // Text = modifiedText || responseText
						cursor := textCursors[idx]
						if len(mt) > cursor && strings.HasPrefix(mt, rt) {
							accText += mt[cursor:]
							totalText += len(mt) - cursor
							textCursors[idx] = len(mt)
						}
						// Thinking final sweep
						if s.Thinking != "" {
							prev := thinkCursors[idx]
							if len(s.Thinking) > prev {
								accThinking += s.Thinking[prev:]
								totalThinking += len(s.Thinking) - prev
								thinkCursors[idx] = len(s.Thinking)
							}
						}
					}
				}
				break
			}
		} else {
			idleCount = 0
		}
	}

	slog.Info("windsurf_cascade_poll_result",
		"cascade_id", cascadeID[:min(8, len(cascadeID))],
		"acc_text_len", len(accText),
		"acc_thinking_len", len(accThinking),
		"native_tool_calls", len(nativeToolCalls),
		"saw_active", sawActive,
		"saw_text", sawText,
		"steps_seen", len(textCursors),
		"idle_count", idleCount,
	)

	// Aggregate step usage
	var aggUsage *StepUsage
	for _, u := range usageByStep {
		if aggUsage == nil {
			aggUsage = &StepUsage{}
		}
		aggUsage.InputTokens += u.InputTokens
		aggUsage.OutputTokens += u.OutputTokens
		aggUsage.CacheReadTokens += u.CacheReadTokens
		aggUsage.CacheWriteTokens += u.CacheWriteTokens
	}

	return &CascadeChatResult{
		Text:        SanitizePath(accText),
		Thinking:    accThinking,
		Usage:       aggUsage,
		CascadeID:   cascadeID,
		FirstTextAt: firstTextAt,
		ToolCalls:   nativeToolCalls,
	}, nil
}

// ── gRPC helpers ───────────────────────────────────────────

func (l *LocalLSClient) grpcUnary(ctx context.Context, path string, body []byte) error {
	_, err := l.grpcUnaryRaw(ctx, path, body)
	return err
}

func (l *LocalLSClient) grpcUnaryRaw(ctx context.Context, path string, body []byte) ([]byte, error) {
	env := make([]byte, 5+len(body))
	env[0] = 0
	binary.BigEndian.PutUint32(env[1:5], uint32(len(body)))
	copy(env[5:], body)

	req, err := http.NewRequestWithContext(ctx, http.MethodPost, l.BaseURL+path, bytes.NewReader(env))
	if err != nil {
		return nil, fmt.Errorf("build request: %w", err)
	}
	req.Header.Set("Content-Type", "application/grpc")
	req.Header.Set("TE", "trailers")
	req.Header.Set("User-Agent", "grpc-go/1.64.0")
	if l.CSRFToken != "" {
		req.Header.Set("x-codeium-csrf-token", l.CSRFToken)
	}
	slog.Debug("windsurf_grpc_request", "url", l.BaseURL+path, "csrf_token", l.CSRFToken)

	resp, err := l.HTTP.Do(req)
	if err != nil {
		return nil, fmt.Errorf("roundtrip: %w", err)
	}
	defer resp.Body.Close()

	respBody, _ := io.ReadAll(resp.Body)

	if resp.StatusCode != 200 {
		return nil, fmt.Errorf("HTTP %d: %s", resp.StatusCode, truncate(string(respBody), 200))
	}

	grpcStatus := resp.Header.Get("grpc-status")
	grpcMsg := resp.Header.Get("grpc-message")
	if grpcStatus == "" {
		grpcStatus = resp.Trailer.Get("grpc-status")
		grpcMsg = resp.Trailer.Get("grpc-message")
	}
	if grpcStatus != "" && grpcStatus != "0" {
		slog.Warn("windsurf_grpc_error",
			"url", l.BaseURL+path,
			"grpc_status", grpcStatus,
			"grpc_msg", grpcMsg,
			"http_status", resp.StatusCode,
			"resp_headers", fmt.Sprintf("%v", resp.Header),
			"resp_trailers", fmt.Sprintf("%v", resp.Trailer),
			"body_len", len(respBody),
		)
		decoded, decErr := url.QueryUnescape(grpcMsg)
		if decErr == nil {
			grpcMsg = decoded
		}
		return nil, fmt.Errorf("gRPC status %s: %s", grpcStatus, grpcMsg)
	}

	return stripGRPCFrame(respBody), nil
}

func stripGRPCFrame(data []byte) []byte {
	if len(data) < 5 {
		return data
	}
	msgLen := binary.BigEndian.Uint32(data[1:5])
	if 5+int(msgLen) <= len(data) {
		return data[5 : 5+msgLen]
	}
	return data[5:]
}

// ── Model enum mapping ─────────────────────────────────────

// modelEnumByUID maps modelUid strings to their deprecated enum values.
// Only entries with enumValue > 0 are included. Sourced from WindsurfAPI models.js.
var modelEnumByUID = map[string]int{
	// Anthropic
	"MODEL_CLAUDE_4_SONNET":          281,
	"MODEL_CLAUDE_4_SONNET_THINKING": 282,
	"MODEL_CLAUDE_4_OPUS":            290,
	"MODEL_CLAUDE_4_OPUS_THINKING":   291,
	"MODEL_CLAUDE_4_1_OPUS":          328,
	"MODEL_CLAUDE_4_1_OPUS_THINKING": 329,
	"MODEL_PRIVATE_2":                353,
	"MODEL_PRIVATE_3":                354,
	"MODEL_CLAUDE_4_5_OPUS":          391,
	"MODEL_CLAUDE_4_5_OPUS_THINKING": 392,
	// OpenAI
	"MODEL_CHAT_GPT_4O_2024_08_06":  109,
	"MODEL_CHAT_GPT_4_1_2025_04_14": 259,
	"MODEL_PRIVATE_6":               340,
	"MODEL_CHAT_GPT_5_CODEX":        346,
	"MODEL_GPT_5_2_LOW":             400,
	"MODEL_GPT_5_2_MEDIUM":          401,
	"MODEL_GPT_5_2_HIGH":            402,
	"MODEL_GPT_5_2_XHIGH":           403,
	"MODEL_CHAT_O3":                 218,
	// Google
	"MODEL_GOOGLE_GEMINI_2_5_PRO":          246,
	"MODEL_GOOGLE_GEMINI_2_5_FLASH":        312,
	"MODEL_GOOGLE_GEMINI_3_0_PRO_LOW":      412,
	"MODEL_GOOGLE_GEMINI_3_0_FLASH_MEDIUM": 415,
	// Others
	"MODEL_XAI_GROK_3":   217,
	"MODEL_KIMI_K2":      323,
	"MODEL_GLM_4_7":      417,
	"MODEL_SWE_1_5_SLOW": 369,
	"MODEL_SWE_1_5":      359,
}

func resolveModelEnum(modelUID string) int {
	if v, ok := modelEnumByUID[modelUID]; ok {
		return v
	}
	return 0
}

// ── Proto parsers ──────────────────────────────────────────

func parseStringField1(data []byte) (string, error) {
	pos := 0
	for pos < len(data) {
		tag, np, ok := ReadVarint(data, pos)
		if !ok {
			break
		}
		pos = np
		fieldNum := tag >> 3
		wireType := tag & 7

		switch wireType {
		case 2:
			length, np2, ok := ReadVarint(data, pos)
			if !ok {
				return "", fmt.Errorf("parse length at pos %d", pos)
			}
			pos = np2
			if pos+int(length) > len(data) {
				return "", fmt.Errorf("field out of bounds")
			}
			field := data[pos : pos+int(length)]
			pos += int(length)
			if fieldNum == 1 {
				return string(field), nil
			}
		case 0:
			_, np2, ok := ReadVarint(data, pos)
			if !ok {
				break
			}
			pos = np2
		case 1:
			pos += 8
		case 5:
			pos += 4
		default:
			return "", fmt.Errorf("unknown wire type %d at pos %d", wireType, pos)
		}
	}
	return "", nil
}

func parseVarintField2(data []byte) (uint64, error) {
	pos := 0
	for pos < len(data) {
		tag, np, ok := ReadVarint(data, pos)
		if !ok {
			break
		}
		pos = np
		fieldNum := tag >> 3
		wireType := tag & 7

		switch wireType {
		case 0:
			val, np2, ok := ReadVarint(data, pos)
			if !ok {
				break
			}
			pos = np2
			if fieldNum == 2 {
				return val, nil
			}
		case 2:
			length, np2, ok := ReadVarint(data, pos)
			if !ok {
				break
			}
			pos = np2 + int(length)
		case 1:
			pos += 8
		case 5:
			pos += 4
		default:
			break
		}
	}
	return 0, nil
}

func parseTrajectorySteps(data []byte) []TrajectoryStep {
	var steps []TrajectoryStep
	pos := 0
	for pos < len(data) {
		tag, np, ok := ReadVarint(data, pos)
		if !ok {
			break
		}
		pos = np
		fieldNum := tag >> 3
		wireType := tag & 7

		if wireType == 2 {
			length, np2, ok := ReadVarint(data, pos)
			if !ok {
				break
			}
			pos = np2
			if pos+int(length) > len(data) {
				break
			}
			field := data[pos : pos+int(length)]
			pos += int(length)
			if fieldNum == 1 {
				steps = append(steps, parseOneTrajectoryStep(field))
			}
		} else if wireType == 0 {
			_, np2, ok := ReadVarint(data, pos)
			if !ok {
				break
			}
			pos = np2
		} else if wireType == 1 {
			pos += 8
		} else if wireType == 5 {
			pos += 4
		} else {
			break
		}
	}
	return steps
}

func parseOneTrajectoryStep(data []byte) TrajectoryStep {
	var s TrajectoryStep
	pos := 0
	for pos < len(data) {
		tag, np, ok := ReadVarint(data, pos)
		if !ok {
			break
		}
		pos = np
		fieldNum := tag >> 3
		wireType := tag & 7

		switch wireType {
		case 0:
			val, np2, ok := ReadVarint(data, pos)
			if !ok {
				break
			}
			pos = np2
			switch fieldNum {
			case 1:
				s.Type = int(val)
			case 4:
				s.Status = int(val)
			}
		case 2:
			length, np2, ok := ReadVarint(data, pos)
			if !ok {
				break
			}
			pos = np2
			if pos+int(length) > len(data) {
				break
			}
			field := data[pos : pos+int(length)]
			pos += int(length)
			switch fieldNum {
			case 5: // step metadata (CortexStepMetadata)
				s.Usage = parseStepUsage(field)
				if tc := parseMetadataToolCall(field); tc != nil {
					s.ToolCall = tc
				}
			case 47: // mcp_tool (CortexStepMcpTool) — tool_call is field 2
				if tc := parseChatToolCallFromContainer(field, 2); tc != nil {
					s.ToolCall = tc
				}
			case 49: // tool_call_proposal (CortexStepToolCallProposal) — tool_call is field 1
				if tc := parseChatToolCallFromContainer(field, 1); tc != nil {
					s.ToolCall = tc
				}
			case 20: // planner_response
				pr := parseFields2(field)
				var responseText, modifiedText, thinking string
				for _, pf := range pr {
					switch pf.fn {
					case 1:
						responseText = string(pf.val)
					case 3:
						thinking = string(pf.val)
					case 8:
						modifiedText = string(pf.val)
					}
				}
				if modifiedText != "" {
					s.Text = modifiedText
				} else {
					s.Text = responseText
				}
				s.ResponseText = responseText
				s.Thinking = thinking
			case 24: // error_message
				s.ErrorText = extractErrorText(field)
			case 31: // error (fallback)
				if s.ErrorText == "" {
					s.ErrorText = extractErrorText(field)
				}
			}
		case 1:
			pos += 8
		case 5:
			pos += 4
		default:
			pos = len(data)
		}
	}
	return s
}

// parseChatToolCall parses a ChatToolCall proto message:
//
//	field 1 (string) = id
//	field 2 (string) = name
//	field 3 (string) = arguments_json
func parseChatToolCall(data []byte) *NativeToolCall {
	var tc NativeToolCall
	pos := 0
	for pos < len(data) {
		tag, np, ok := ReadVarint(data, pos)
		if !ok {
			break
		}
		pos = np
		fieldNum := tag >> 3
		wireType := tag & 7
		if wireType == 2 {
			length, np2, ok := ReadVarint(data, pos)
			if !ok {
				break
			}
			pos = np2
			if pos+int(length) > len(data) {
				break
			}
			val := string(data[pos : pos+int(length)])
			pos += int(length)
			switch fieldNum {
			case 1:
				tc.ID = val
			case 2:
				tc.Name = val
			case 3:
				tc.ArgumentsJSON = val
			}
		} else if wireType == 0 {
			_, np2, ok := ReadVarint(data, pos)
			if !ok {
				break
			}
			pos = np2
		} else if wireType == 1 {
			pos += 8
		} else if wireType == 5 {
			pos += 4
		} else {
			break
		}
	}
	if tc.Name == "" {
		return nil
	}
	return &tc
}

// parseChatToolCallFromContainer extracts ChatToolCall from a container message
// where the ChatToolCall is at the given field number.
func parseChatToolCallFromContainer(data []byte, toolCallFieldNum uint64) *NativeToolCall {
	pos := 0
	for pos < len(data) {
		tag, np, ok := ReadVarint(data, pos)
		if !ok {
			break
		}
		pos = np
		fieldNum := tag >> 3
		wireType := tag & 7
		if wireType == 2 {
			length, np2, ok := ReadVarint(data, pos)
			if !ok {
				break
			}
			pos = np2
			if pos+int(length) > len(data) {
				break
			}
			field := data[pos : pos+int(length)]
			pos += int(length)
			if fieldNum == toolCallFieldNum {
				return parseChatToolCall(field)
			}
		} else if wireType == 0 {
			_, np2, ok := ReadVarint(data, pos)
			if !ok {
				break
			}
			pos = np2
		} else if wireType == 1 {
			pos += 8
		} else if wireType == 5 {
			pos += 4
		} else {
			break
		}
	}
	return nil
}

// parseMetadataToolCall extracts ChatToolCall from CortexStepMetadata (field 4 = tool_call).
func parseMetadataToolCall(metaData []byte) *NativeToolCall {
	return parseChatToolCallFromContainer(metaData, 4)
}

// parseStepUsage extracts token usage from CortexStepMetadata (field 5).
// CortexStepMetadata.model_usage = field 9 → ModelUsageStats {2=input, 3=output, 4=cacheWrite, 5=cacheRead}
func parseStepUsage(metaData []byte) *StepUsage {
	// Find field 9 (model_usage) in metadata
	usageData := extractLenDelimField(metaData, 9)
	if usageData == nil {
		return nil
	}
	var u StepUsage
	found := false
	pos := 0
	for pos < len(usageData) {
		tag, np, ok := ReadVarint(usageData, pos)
		if !ok {
			break
		}
		pos = np
		fn := tag >> 3
		wt := tag & 7
		if wt == 0 {
			val, np2, ok := ReadVarint(usageData, pos)
			if !ok {
				break
			}
			pos = np2
			switch fn {
			case 2:
				u.InputTokens = int(val)
				found = true
			case 3:
				u.OutputTokens = int(val)
				found = true
			case 4:
				u.CacheWriteTokens = int(val)
				found = true
			case 5:
				u.CacheReadTokens = int(val)
				found = true
			}
		} else if wt == 2 {
			length, np2, ok := ReadVarint(usageData, pos)
			if !ok {
				break
			}
			pos = np2 + int(length)
		} else if wt == 1 {
			pos += 8
		} else if wt == 5 {
			pos += 4
		} else {
			break
		}
	}
	if !found {
		return nil
	}
	return &u
}

// extractLenDelimField finds the first length-delimited field with the given number.
func extractLenDelimField(data []byte, targetField uint64) []byte {
	pos := 0
	for pos < len(data) {
		tag, np, ok := ReadVarint(data, pos)
		if !ok {
			break
		}
		pos = np
		fn := tag >> 3
		wt := tag & 7
		if wt == 2 {
			length, np2, ok := ReadVarint(data, pos)
			if !ok {
				break
			}
			pos = np2
			if pos+int(length) > len(data) {
				break
			}
			if fn == targetField {
				return data[pos : pos+int(length)]
			}
			pos += int(length)
		} else if wt == 0 {
			_, np2, ok := ReadVarint(data, pos)
			if !ok {
				break
			}
			pos = np2
		} else if wt == 1 {
			pos += 8
		} else if wt == 5 {
			pos += 4
		} else {
			break
		}
	}
	return nil
}

type protoField struct {
	fn  uint64
	val []byte
}

func parseFields2(data []byte) []protoField {
	var fields []protoField
	pos := 0
	for pos < len(data) {
		tag, np, ok := ReadVarint(data, pos)
		if !ok {
			break
		}
		pos = np
		fn := tag >> 3
		wt := tag & 7
		switch wt {
		case 2:
			length, np2, ok := ReadVarint(data, pos)
			if !ok {
				break
			}
			pos = np2
			if pos+int(length) > len(data) {
				break
			}
			fields = append(fields, protoField{fn, data[pos : pos+int(length)]})
			pos += int(length)
		case 0:
			_, np2, ok := ReadVarint(data, pos)
			if !ok {
				break
			}
			pos = np2
		case 1:
			pos += 8
		case 5:
			pos += 4
		default:
			pos = len(data)
		}
	}
	return fields
}

func extractErrorText(data []byte) string {
	return extractErrorTextDepth(data, 0)
}

func extractErrorTextDepth(data []byte, depth int) string {
	if depth > 3 {
		return ""
	}
	for _, pf := range parseFields2(data) {
		if pf.fn >= 1 && pf.fn <= 5 && len(pf.val) > 10 {
			txt := string(pf.val)
			for len(txt) > 0 && txt[0] < 0x20 {
				txt = txt[1:]
			}
			if len(txt) > 10 && !hasNonPrintable(txt[:10]) {
				return txt
			}
			if inner := extractErrorTextDepth(pf.val, depth+1); inner != "" {
				return inner
			}
		}
	}
	return ""
}

func hasNonPrintable(s string) bool {
	for _, c := range s {
		if c < 0x20 && c != '\n' && c != '\r' {
			return true
		}
	}
	return false
}

// GetCascadeModelConfigs 查询 LS 的 GetCascadeModelConfigs RPC，
// 返回 model_name -> supports_images 的映射。模型名按小写归一化。
func (l *LocalLSClient) GetCascadeModelConfigs(ctx context.Context, token string) (map[string]bool, error) {
	// 请求 body：只需 metadata，即 field 1 encode(Metadata)
	// 参考 package.json 提到的 proto；这里用 metadata-only encoding 与其他 RPC 一致。
	body := encodeBytesField(1, buildMetadata(token, l.SessionID))
	raw, err := l.grpcUnaryRaw(ctx, GetCascadeModelConfigsRPC, body)
	if err != nil {
		return nil, fmt.Errorf("get_cascade_model_configs: %w", err)
	}
	var resp pb.GetCascadeModelConfigsResponse
	if err := proto.Unmarshal(raw, &resp); err != nil {
		return nil, fmt.Errorf("unmarshal: %w", err)
	}
	out := make(map[string]bool, len(resp.GetModels()))
	for _, m := range resp.GetModels() {
		out[strings.ToLower(strings.TrimSpace(m.GetName()))] = m.GetSupportsImages()
	}
	return out, nil
}

// ModelSupportsImages 带缓存的图像能力查询。
// fail-open：RPC 失败且无缓存时返回 (false, false, nil)，由上层决定策略。
// 返回值：(found, supportsImages, error)
func (l *LocalLSClient) ModelSupportsImages(ctx context.Context, token, modelName string) (bool, bool, error) {
	key := apiKeyHash(token)

	l.modelCapsMu.Lock()
	if l.modelCapsCache == nil {
		l.modelCapsCache = make(map[string]cascadeModelCapsCacheEntry)
	}
	entry, ok := l.modelCapsCache[key]
	fresh := ok && time.Since(entry.FetchedAt) < cascadeModelCapsTTL
	l.modelCapsMu.Unlock()

	if fresh {
		v, found := entry.SupportsImages[strings.ToLower(strings.TrimSpace(modelName))]
		return found, v, nil
	}

	// 拉新：失败时保留 stale
	caps, err := l.GetCascadeModelConfigs(ctx, token)
	if err != nil {
		// stale fallback
		if ok {
			v, found := entry.SupportsImages[strings.ToLower(strings.TrimSpace(modelName))]
			return found, v, nil
		}
		return false, false, err
	}

	l.modelCapsMu.Lock()
	l.modelCapsCache[key] = cascadeModelCapsCacheEntry{
		SupportsImages: caps,
		FetchedAt:      time.Now(),
	}
	l.modelCapsMu.Unlock()

	v, found := caps[strings.ToLower(strings.TrimSpace(modelName))]
	return found, v, nil
}

func apiKeyHash(token string) string {
	sum := sha256.Sum256([]byte(token))
	return hex.EncodeToString(sum[:8]) // 16 hex chars 足够区分
}