core/server/handlers/completion.go

package handlers

import (
	"context"
	"database/sql"
	"encoding/base64"
	"encoding/json"
	"fmt"
	"io"
	"log"
	"net/http"
	"strings"

	"github.com/gin-gonic/gin"

	capspkg "git.gobha.me/xcaliber/chat-switchboard/capabilities"
	"git.gobha.me/xcaliber/chat-switchboard/crypto"
	"git.gobha.me/xcaliber/chat-switchboard/database"
	"git.gobha.me/xcaliber/chat-switchboard/events"
	"git.gobha.me/xcaliber/chat-switchboard/models"
	"git.gobha.me/xcaliber/chat-switchboard/providers"
	"git.gobha.me/xcaliber/chat-switchboard/storage"
	"git.gobha.me/xcaliber/chat-switchboard/store"
	"git.gobha.me/xcaliber/chat-switchboard/tools"
)

// ── Request Types ───────────────────────────

type completionRequest struct {
	ChannelID     string   `json:"channel_id"` // preferred; validated manually below
	ChatID        string   `json:"chat_id"`    // deprecated alias — maps to channel_id
	Content       string   `json:"content" binding:"required"`
	Model         string   `json:"model,omitempty"`
	PresetID      string   `json:"preset_id,omitempty"` // if set, unwraps preset → base model + config
	APIConfigID   string   `json:"provider_config_id,omitempty"`
	MaxTokens     int      `json:"max_tokens,omitempty"`
	Temperature   *float64 `json:"temperature,omitempty"`
	TopP          *float64 `json:"top_p,omitempty"`
	Stream        *bool    `json:"stream,omitempty"`
	AttachmentIDs []string `json:"attachment_ids,omitempty"` // staged attachment UUIDs to include in request
	DisabledTools []string `json:"disabled_tools,omitempty"` // tool names to exclude from this request
}

// CompletionHandler proxies LLM requests through the backend.
type CompletionHandler struct {
	vault    *crypto.KeyResolver
	stores   store.Stores
	hub      *events.Hub          // WebSocket hub for browser tool bridge
	objStore storage.ObjectStore  // file storage for attachment content (nil = disabled)
}

// NewCompletionHandler creates a new handler.
func NewCompletionHandler(vault *crypto.KeyResolver, stores store.Stores, hub *events.Hub, objStore storage.ObjectStore) *CompletionHandler {
	return &CompletionHandler{vault: vault, stores: stores, hub: hub, objStore: objStore}
}

// ── Chat Completion ─────────────────────────
// POST /api/v1/chat/completions
//
// Flow:
//   1. Validate request, verify chat ownership
//   2. Resolve api_config (from request, chat, or user default)
//   3. Load conversation history from messages
//   4. Persist user message
//   5. Attach tool definitions if model supports tools
//   6. Call provider (stream or non-stream)
//   7. Tool loop: if LLM requests tool calls, execute them and re-call
//   8. Stream SSE to client / return JSON
//   9. Persist assistant message with token counts

func (h *CompletionHandler) Complete(c *gin.Context) {
	var req completionRequest
	if err := c.ShouldBindJSON(&req); err != nil {
		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
		return
	}

	// Support chat_id as alias during frontend transition
	channelID := req.ChannelID
	if channelID == "" {
		channelID = req.ChatID
	}
	if channelID == "" {
		c.JSON(http.StatusBadRequest, gin.H{"error": "channel_id is required"})
		return
	}

	userID := getUserID(c)

	// Verify channel ownership
	if !userOwnsChannel(c, channelID, userID) {
		return
	}

	// ── Preset unwrap: preset overrides defaults, explicit request fields win ──
	var presetSystemPrompt string
	if req.PresetID != "" {
		preset := ResolvePreset(req.PresetID, userID)
		if preset == nil {
			c.JSON(http.StatusBadRequest, gin.H{"error": "preset not found or not accessible"})
			return
		}
		// Preset provides defaults; explicit request fields take priority
		if req.Model == "" {
			req.Model = preset.BaseModelID
		}
		if req.APIConfigID == "" && preset.ProviderConfigID != nil {
			req.APIConfigID = *preset.ProviderConfigID
		}
		if req.Temperature == nil && preset.Temperature != nil {
			req.Temperature = preset.Temperature
		}
		if req.MaxTokens == 0 && preset.MaxTokens != nil {
			req.MaxTokens = *preset.MaxTokens
		}
		if preset.SystemPrompt != "" {
			presetSystemPrompt = preset.SystemPrompt
		}
	}

	// Resolve provider config
	providerCfg, providerID, model, configID, providerScope, err := h.resolveConfig(userID, channelID, req)
	if err != nil {
		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
		return
	}

	// ── Team policy: require_private_providers ──
	if err := enforcePrivateProviderPolicy(userID, configID); err != nil {
		c.JSON(http.StatusForbidden, gin.H{"error": err.Error()})
		return
	}

	provider, err := providers.Get(providerID)
	if err != nil {
		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
		return
	}

	// Load conversation history
	messages, err := h.loadConversation(channelID, userID, presetSystemPrompt)
	if err != nil {
		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to load conversation"})
		return
	}

	// Resolve capabilities early — needed for vision gating below
	caps := h.getModelCapabilities(c, model, configID)

	// Build user message — multimodal if attachments are present
	userMsg := providers.Message{
		Role:    "user",
		Content: req.Content,
	}

	if len(req.AttachmentIDs) > 0 && h.objStore != nil {
		parts, augContent, _, err := h.buildMultimodalParts(c, channelID, req.Content, req.AttachmentIDs, caps)
		if err != nil {
			c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
			return
		}
		if len(parts) > 0 {
			// Multimodal (images present): use ContentParts
			userMsg.ContentParts = parts
		} else if augContent != req.Content {
			// Doc-only: use augmented text content
			userMsg.Content = augContent
		}
	}

	messages = append(messages, userMsg)

	// Persist user message (text-only for storage — multimodal parts are ephemeral)
	msgID, err := h.persistMessage(channelID, userID, "user", req.Content, "", 0, 0, nil, nil)
	if err != nil {
		log.Printf("Failed to persist user message: %v", err)
	}

	// Link attachments to the persisted message
	if msgID != "" && len(req.AttachmentIDs) > 0 {
		for _, attID := range req.AttachmentIDs {
			if err := h.stores.Attachments.SetMessageID(c.Request.Context(), attID, msgID); err != nil {
				log.Printf("Failed to link attachment %s to message %s: %v", attID, msgID, err)
			}
		}
	}

	// Build provider request
	provReq := providers.CompletionRequest{
		Model:    model,
		Messages: messages,
	}

	if req.MaxTokens > 0 {
		provReq.MaxTokens = req.MaxTokens
	} else {
		// ResolveMaxOutput checks: caps → known models → context/8 → 4096
		provReq.MaxTokens = capspkg.ResolveMaxOutput(model, caps)
	}

	if req.Temperature != nil {
		provReq.Temperature = req.Temperature
	}
	if req.TopP != nil {
		provReq.TopP = req.TopP
	}

	// Attach tool definitions if model supports tool calling and tools are available
	hasBrowserTools := h.hub != nil && h.hub.IsConnected(userID)
	if caps.ToolCalling && (tools.HasTools() || hasBrowserTools) {
		provReq.Tools = h.buildToolDefs(c.Request.Context(), userID, hasBrowserTools, req.DisabledTools)
	}

	// Determine streaming
	stream := true // default
	if req.Stream != nil {
		stream = *req.Stream
	}

	if stream {
		h.streamCompletion(c, provider, providerCfg, provReq, channelID, userID, model, configID, providerScope)
	} else {
		h.syncCompletion(c, provider, providerCfg, provReq, channelID, userID, model, configID, providerScope)
	}
}

// buildToolDefs converts registered tools to provider-format tool definitions.
// If includeBrowser is true, also fetches tool schemas from enabled browser extensions.
// Tools whose names appear in disabledTools are excluded.
func (h *CompletionHandler) buildToolDefs(ctx context.Context, userID string, includeBrowser bool, disabledTools []string) []providers.ToolDef {
	// Build disabled set for O(1) lookup
	disabled := make(map[string]bool, len(disabledTools))
	for _, name := range disabledTools {
		disabled[name] = true
	}

	allTools := tools.AllDefinitionsFiltered(disabled)
	defs := make([]providers.ToolDef, 0, len(allTools))
	for _, t := range allTools {
		defs = append(defs, providers.ToolDef{
			Type: "function",
			Function: providers.FunctionDef{
				Name:        t.Name,
				Description: t.Description,
				Parameters:  t.Parameters,
			},
		})
	}

	// Append browser-defined tool schemas from extensions
	if includeBrowser {
		exts, err := h.stores.Extensions.ListForUser(ctx, userID)
		if err != nil {
			log.Printf("⚠️ Failed to load extensions for tools: %v", err)
			return defs
		}
		for _, ext := range exts {
			if ext.Tier != "browser" {
				continue
			}
			var manifest struct {
				Tools []struct {
					Name        string          `json:"name"`
					Description string          `json:"description"`
					Parameters  json.RawMessage `json:"parameters"`
					Tier        string          `json:"tier"`
				} `json:"tools"`
			}
			if err := json.Unmarshal(ext.Manifest, &manifest); err != nil {
				continue
			}
			for _, t := range manifest.Tools {
				if disabled[t.Name] {
					continue
				}
				defs = append(defs, providers.ToolDef{
					Type: "function",
					Function: providers.FunctionDef{
						Name:        t.Name,
						Description: t.Description,
						Parameters:  t.Parameters,
					},
				})
			}
		}
	}

	return defs
}

// ── List Available Tools ────────────────────
// GET /api/v1/tools
//
// Returns all registered server-side tools plus browser extension tools
// for the current user. Used by the frontend to build the tools toggle menu.

type toolInfo struct {
	Name        string `json:"name"`
	DisplayName string `json:"display_name"`
	Description string `json:"description"`
	Category    string `json:"category"`
}

func (h *CompletionHandler) ListTools(c *gin.Context) {
	userID := c.GetString("user_id")

	// Server-side tools
	allDefs := tools.AllDefinitions()
	result := make([]toolInfo, 0, len(allDefs))
	for _, d := range allDefs {
		result = append(result, toolInfo{
			Name:        d.Name,
			DisplayName: d.DisplayName,
			Description: d.Description,
			Category:    d.Category,
		})
	}

	// Browser extension tools
	if h.hub != nil && h.hub.IsConnected(userID) {
		exts, err := h.stores.Extensions.ListForUser(c.Request.Context(), userID)
		if err == nil {
			for _, ext := range exts {
				if ext.Tier != "browser" {
					continue
				}
				var manifest struct {
					Tools []struct {
						Name        string `json:"name"`
						Description string `json:"description"`
					} `json:"tools"`
				}
				if err := json.Unmarshal(ext.Manifest, &manifest); err != nil {
					continue
				}
				for _, t := range manifest.Tools {
					result = append(result, toolInfo{
						Name:        t.Name,
						Description: t.Description,
						Category:    "browser",
					})
				}
			}
		}
	}

	c.JSON(http.StatusOK, gin.H{"data": result})
}

// ── Streaming Completion (SSE) with Tool Loop ──

const maxToolIterations = 10 // safety limit on tool call rounds

func (h *CompletionHandler) streamCompletion(
	c *gin.Context,
	provider providers.Provider,
	cfg providers.ProviderConfig,
	req providers.CompletionRequest,
	channelID, userID, model, configID, providerScope string,
) {
	result := streamWithToolLoop(c, provider, cfg, &req, model, userID, channelID, h.hub)

	// Persist assistant response
	if result.Content != "" {
		if _, err := h.persistMessage(channelID, userID, "assistant", result.Content, model, result.InputTokens, result.OutputTokens, nil, toolActivityJSON(result.ToolActivity)); err != nil {
			log.Printf("Failed to persist assistant message: %v", err)
		}
	}

	// Log usage
	h.logUsage(c, channelID, userID, configID, providerScope, model,
		result.InputTokens, result.OutputTokens,
		result.CacheCreationTokens, result.CacheReadTokens)
}

// ── Non-Streaming Completion with Tool Loop ──

func (h *CompletionHandler) syncCompletion(
	c *gin.Context,
	provider providers.Provider,
	cfg providers.ProviderConfig,
	req providers.CompletionRequest,
	channelID, userID, model, configID, providerScope string,
) {
	var totalInput, totalOutput int
	var totalCacheCreation, totalCacheRead int
	var finalContent string
	var allToolActivity []map[string]interface{}

	for iteration := 0; iteration < maxToolIterations; iteration++ {
		resp, err := provider.ChatCompletion(c.Request.Context(), cfg, req)
		if err != nil {
			c.JSON(http.StatusBadGateway, gin.H{"error": "provider error: " + err.Error()})
			return
		}

		totalInput += resp.InputTokens
		totalOutput += resp.OutputTokens
		totalCacheCreation += resp.CacheCreationTokens
		totalCacheRead += resp.CacheReadTokens

		// No tool calls — normal response
		if len(resp.ToolCalls) == 0 || resp.FinishReason != "tool_calls" {
			finalContent = resp.Content

			// Persist assistant response
			if _, err := h.persistMessage(channelID, userID, "assistant", finalContent, model, totalInput, totalOutput, nil, toolActivityJSON(allToolActivity)); err != nil {
				log.Printf("Failed to persist assistant message: %v", err)
			}

			// Log usage
			h.logUsage(c, channelID, userID, configID, providerScope, model,
				totalInput, totalOutput, totalCacheCreation, totalCacheRead)

			// Return OpenAI-compatible response
			c.JSON(http.StatusOK, gin.H{
				"model": resp.Model,
				"choices": []gin.H{
					{
						"message": gin.H{
							"role":    "assistant",
							"content": finalContent,
						},
						"finish_reason": resp.FinishReason,
					},
				},
				"usage": gin.H{
					"prompt_tokens":     totalInput,
					"completion_tokens": totalOutput,
					"total_tokens":      totalInput + totalOutput,
				},
			})
			return
		}

		// ── Tool execution round ────────────────
		assistantMsg := providers.Message{
			Role:    "assistant",
			Content: resp.Content,
		}
		for _, tc := range resp.ToolCalls {
			assistantMsg.ToolCalls = append(assistantMsg.ToolCalls, tc)
		}
		req.Messages = append(req.Messages, assistantMsg)

		execCtx := tools.ExecutionContext{
			UserID:    userID,
			ChannelID: channelID,
		}
		for _, tc := range resp.ToolCalls {
			call := tools.ToolCall{
				ID:        tc.ID,
				Name:      tc.Function.Name,
				Arguments: tc.Function.Arguments,
			}
			log.Printf("🔧 Executing tool: %s (call %s)", call.Name, call.ID)
			result := tools.ExecuteCall(c.Request.Context(), execCtx, call)

			// Collect for persistence
			allToolActivity = append(allToolActivity, map[string]interface{}{
				"id":        tc.ID,
				"name":      tc.Function.Name,
				"arguments": tc.Function.Arguments,
				"result":    result.Content,
				"is_error":  result.IsError,
			})

			req.Messages = append(req.Messages, providers.Message{
				Role:       "tool",
				Content:    result.Content,
				ToolCallID: result.ToolCallID,
				Name:       result.Name,
			})
		}
	}

	// Max iterations reached
	c.JSON(http.StatusOK, gin.H{
		"model": model,
		"choices": []gin.H{
			{
				"message": gin.H{
					"role":    "assistant",
					"content": "[Tool execution limit reached]",
				},
				"finish_reason": "stop",
			},
		},
	})
}

// ── Model Capabilities ──────────────────────

// escapeJSON escapes a string for safe embedding in a JSON string value.
func escapeJSON(s string) string {
	s = strings.ReplaceAll(s, `\`, `\\`)
	s = strings.ReplaceAll(s, `"`, `\"`)
	s = strings.ReplaceAll(s, "\n", `\n`)
	s = strings.ReplaceAll(s, "\r", `\r`)
	s = strings.ReplaceAll(s, "\t", `\t`)
	return s
}

// getModelCapabilities looks up capabilities from model_catalog DB,
// then overlays with known model defaults and heuristic detection.
func (h *CompletionHandler) getModelCapabilities(c *gin.Context, model, apiConfigID string) models.ModelCapabilities {
	return ResolveModelCaps(c, model, apiConfigID)
}

// ── Multimodal Assembly ─────────────────────
// Builds content parts from attachments for the user message.
//
// Returns:
//   - parts: ContentParts array (non-nil only when images are present)
//   - augContent: enriched text content with document context (for doc-only case)
//   - validIDs: attachment IDs that were successfully processed
//   - error: if any attachment is invalid or vision is needed but missing
//
// Rules:
//   - Images → base64 data URI (requires vision capability)
//   - Documents with extracted_text → text injection
//   - Documents without extraction → filename placeholder
//   - Text is always the first part

func (h *CompletionHandler) buildMultimodalParts(
	c *gin.Context,
	channelID, textContent string,
	attachmentIDs []string,
	caps models.ModelCapabilities,
) ([]providers.ContentPart, string, []string, error) {

	parts := []providers.ContentPart{
		{Type: "text", Text: textContent},
	}
	var docTexts []string
	var validIDs []string
	hasImage := false

	for _, attID := range attachmentIDs {
		att, err := h.stores.Attachments.GetByID(c.Request.Context(), attID)
		if err != nil {
			return nil, "", nil, fmt.Errorf("attachment %s not found", attID)
		}

		// Security: verify attachment belongs to this channel
		if att.ChannelID != channelID {
			return nil, "", nil, fmt.Errorf("attachment %s does not belong to this channel", attID)
		}

		if isImageContentType(att.ContentType) {
			// Vision gating: reject images if model lacks vision
			if !caps.Vision {
				return nil, "", nil, fmt.Errorf("model does not support image input; remove image attachments or choose a vision-capable model")
			}

			// Read image from storage, base64 encode, build data URI
			reader, _, _, err := h.objStore.Get(c.Request.Context(), att.StorageKey)
			if err != nil {
				log.Printf("Failed to read attachment %s from storage: %v", attID, err)
				parts = append(parts, providers.ContentPart{
					Type: "text",
					Text: fmt.Sprintf("[Image: %s — failed to read from storage]", att.Filename),
				})
				validIDs = append(validIDs, attID)
				continue
			}
			data, err := io.ReadAll(reader)
			reader.Close()
			if err != nil {
				log.Printf("Failed to read attachment %s bytes: %v", attID, err)
				validIDs = append(validIDs, attID)
				continue
			}

			b64 := base64.StdEncoding.EncodeToString(data)
			dataURI := fmt.Sprintf("data:%s;base64,%s", att.ContentType, b64)

			parts = append(parts, providers.ContentPart{
				Type: "image_url",
				ImageURL: &providers.ImageURL{
					URL:    dataURI,
					Detail: "auto",
				},
			})
			hasImage = true

		} else if att.ExtractedText != nil && *att.ExtractedText != "" {
			// Document with extracted text → inject as context
			docText := fmt.Sprintf("[Document: %s]\n%s", att.Filename, *att.ExtractedText)
			parts = append(parts, providers.ContentPart{
				Type: "text",
				Text: docText,
			})
			docTexts = append(docTexts, docText)

		} else {
			// Document without extraction (pending, failed, or not extractable)
			status := "pending"
			if s, ok := att.Metadata["extraction_status"].(string); ok {
				status = s
			}
			placeholder := fmt.Sprintf("[Attached file: %s (extraction %s)]", att.Filename, status)
			parts = append(parts, providers.ContentPart{
				Type: "text",
				Text: placeholder,
			})
			docTexts = append(docTexts, placeholder)
		}

		validIDs = append(validIDs, attID)
	}

	// If images present → use ContentParts (multimodal array)
	if hasImage {
		return parts, textContent, validIDs, nil
	}

	// Doc-only: merge document context into a single text string (more efficient)
	augmented := textContent
	for _, dt := range docTexts {
		augmented += "\n\n" + dt
	}
	return nil, augmented, validIDs, nil
}

// isImageContentType returns true for MIME types that should be sent as
// base64 image content parts rather than text extraction.
func isImageContentType(ct string) bool {
	return strings.HasPrefix(ct, "image/")
}

// ── Config Resolution ───────────────────────
// Priority: request.provider_config_id → chat.provider_config_id → user's first active config

func (h *CompletionHandler) resolveConfig(userID string, channelID string, req completionRequest) (providers.ProviderConfig, string, string, string, string, error) {
	var configID string

	// 1. Explicit config from request
	if req.APIConfigID != "" {
		configID = req.APIConfigID
	}

	// 2. Config from channel
	if configID == "" {
		var channelConfigID *string
		err := database.DB.QueryRow(
			`SELECT provider_config_id FROM channels WHERE id = $1`, channelID,
		).Scan(&channelConfigID)
		if err == nil && channelConfigID != nil {
			configID = *channelConfigID
		}
	}

	// 3. User's first active config (personal first, then global — excludes team providers)
	if configID == "" {
		err := database.DB.QueryRow(`
			SELECT id FROM provider_configs
			WHERE is_active = true AND (
				(scope = 'personal' AND owner_id = $1)
				OR scope = 'global'
			)
			ORDER BY scope ASC, created_at ASC
			LIMIT 1
		`, userID).Scan(&configID)
		if err != nil {
			return providers.ProviderConfig{}, "", "", "", "", fmt.Errorf("no API config found — add one at /api-configs")
		}
	}

	// Load the config — allow personal, global, OR team configs the user belongs to
	var providerID, endpoint string
	var providerScope string
	var modelDefault *string
	var apiKeyEnc, keyNonce []byte
	var keyScope string
	var customHeadersJSON, providerSettingsJSON []byte
	err := database.DB.QueryRow(`
		SELECT provider, endpoint, scope, api_key_enc, key_nonce, key_scope,
		       model_default, headers, settings
		FROM provider_configs
		WHERE id = $1 AND is_active = true
		  AND (scope = 'global'
		       OR (scope = 'personal' AND owner_id = $2)
		       OR (scope = 'team' AND owner_id IN (SELECT team_id FROM team_members WHERE user_id = $2)))
	`, configID, userID).Scan(&providerID, &endpoint, &providerScope, &apiKeyEnc, &keyNonce, &keyScope,
		&modelDefault, &customHeadersJSON, &providerSettingsJSON)

	if err == sql.ErrNoRows {
		return providers.ProviderConfig{}, "", "", "", "", fmt.Errorf("API config not found or not accessible")
	}
	if err != nil {
		return providers.ProviderConfig{}, "", "", "", "", fmt.Errorf("failed to load API config: %w", err)
	}

	// Resolve model: request > config default
	model := req.Model
	if model == "" && modelDefault != nil {
		model = *modelDefault
	}
	if model == "" {
		return providers.ProviderConfig{}, "", "", "", "", fmt.Errorf("no model specified and no default model in config")
	}

	// Decrypt API key using the appropriate tier
	key := ""
	if len(apiKeyEnc) > 0 {
		if h.vault != nil {
			var err error
			key, err = h.vault.Decrypt(apiKeyEnc, keyNonce, keyScope, userID)
			if err != nil {
				if err == crypto.ErrVaultLocked {
					return providers.ProviderConfig{}, "", "", "", "", fmt.Errorf("personal vault is locked — please log in again")
				}
				return providers.ProviderConfig{}, "", "", "", "", fmt.Errorf("failed to decrypt API key: %w", err)
			}
		} else {
			// No vault — key stored as raw bytes (unencrypted fallback)
			key = string(apiKeyEnc)
		}
	}

	// Parse custom headers
	customHeaders := make(map[string]string)
	if customHeadersJSON != nil {
		_ = json.Unmarshal(customHeadersJSON, &customHeaders)
	}

	// Parse provider-specific settings
	providerSettings := make(map[string]interface{})
	if providerSettingsJSON != nil {
		_ = json.Unmarshal(providerSettingsJSON, &providerSettings)
	}

	return providers.ProviderConfig{
		Endpoint:      endpoint,
		APIKey:        key,
		CustomHeaders: customHeaders,
		Settings:      providerSettings,
	}, providerID, model, configID, providerScope, nil
}

// ── Conversation Loader ─────────────────────

// loadConversation builds the message history for the LLM by walking the
// active path through the message tree. Only the current branch is sent —
// sibling branches are never included in context.
//
// Summary-aware: if the path contains a summary node (metadata.type = "summary"),
// messages before it are replaced by the summary content as a system message.
func (h *CompletionHandler) loadConversation(channelID, userID, presetSystemPrompt string) ([]providers.Message, error) {
	messages := make([]providers.Message, 0)

	// ── Admin system prompt (always injected first, no opt out) ──
	adminPrompt, err := h.stores.GlobalConfig.Get(context.Background(), "system_prompt")
	if err == nil {
		if content, ok := adminPrompt["content"].(string); ok && content != "" {
			messages = append(messages, providers.Message{
				Role:    "system",
				Content: content,
			})
		}
	}

	// ── User/preset system prompt (appended after admin prompt) ──
	var systemPrompt *string
	_ = database.DB.QueryRow(
		`SELECT system_prompt FROM channels WHERE id = $1`, channelID,
	).Scan(&systemPrompt)

	// Preset system prompt takes priority; channel system prompt is fallback
	if presetSystemPrompt != "" {
		messages = append(messages, providers.Message{
			Role:    "system",
			Content: presetSystemPrompt,
		})
	} else if systemPrompt != nil && *systemPrompt != "" {
		messages = append(messages, providers.Message{
			Role:    "system",
			Content: *systemPrompt,
		})
	}

	// ── KB hint (nudge LLM to use kb_search when KBs are active) ──
	if kbHint := BuildKBHint(context.Background(), h.stores, channelID, userID); kbHint != "" {
		messages = append(messages, providers.Message{
			Role:    "system",
			Content: kbHint,
		})
	}

	// Walk the active path (root → leaf) instead of loading all messages
	path, err := getActivePath(channelID, userID)
	if err != nil {
		return nil, err
	}

	// Find the most recent summary node in the path
	summaryIdx := -1
	for i, m := range path {
		if isSummaryMessage(&m) {
			summaryIdx = i
		}
	}

	// If we found a summary, inject it as a system message and skip prior messages
	startIdx := 0
	if summaryIdx >= 0 {
		messages = append(messages, providers.Message{
			Role:    "system",
			Content: "Previous conversation summary:\n" + path[summaryIdx].Content,
		})
		startIdx = summaryIdx + 1
	}

	for _, m := range path[startIdx:] {
		if m.Role == "system" {
			continue // system prompts handled above
		}
		// Skip summary nodes that aren't the boundary (shouldn't happen, but defensive)
		if isSummaryMessage(&m) {
			continue
		}
		messages = append(messages, providers.Message{
			Role:    m.Role,
			Content: m.Content,
		})
	}

	return messages, nil
}

// isSummaryMessage checks if a PathMessage has summary metadata.
func isSummaryMessage(m *PathMessage) bool {
	if m.Metadata == nil {
		return false
	}
	var meta map[string]interface{}
	if err := json.Unmarshal(*m.Metadata, &meta); err != nil {
		return false
	}
	return meta["type"] == "summary"
}

// ── Message Persistence ─────────────────────

// persistMessage inserts a message into the tree, updates the cursor, and
// returns the new message's ID.
//
// Parent resolution:
//   - parentOverride (explicit parent for edit/regen operations)
//   - cursor's active_leaf_id (normal conversation flow)
//   - latest message by created_at (fallback for legacy/missing cursor)
// toolActivityJSON marshals tool activity for persistence. Returns nil if empty.
func toolActivityJSON(activity []map[string]interface{}) json.RawMessage {
	if len(activity) == 0 {
		return nil
	}
	b, _ := json.Marshal(activity)
	return b
}

func (h *CompletionHandler) persistMessage(channelID, userID, role, content, model string, inputTokens, outputTokens int, parentOverride *string, toolCallsJSON json.RawMessage) (string, error) {
	var tokensUsed *int
	if inputTokens > 0 || outputTokens > 0 {
		total := inputTokens + outputTokens
		tokensUsed = &total
	}

	// Determine parent
	var parentID *string
	if parentOverride != nil {
		parentID = parentOverride
	} else {
		parentID, _ = getActiveLeaf(channelID, userID)
	}

	// Compute sibling_index
	siblingIdx := nextSiblingIndex(channelID, parentID)

	// Determine participant
	participantType := "user"
	participantID := userID
	if role == "assistant" {
		participantType = "model"
		participantID = model
	}

	// Prepare tool_calls JSONB (nil → NULL)
	var tcVal interface{}
	if len(toolCallsJSON) > 0 {
		tcVal = string(toolCallsJSON)
	}

	// Insert with RETURNING id
	var newID string
	err := database.DB.QueryRow(`
		INSERT INTO messages (channel_id, role, content, model, tokens_used,
		                      tool_calls, parent_id, participant_type, participant_id, sibling_index)
		VALUES ($1, $2, $3, NULLIF($4, ''), $5, $6, $7, $8, $9, $10)
		RETURNING id
	`, channelID, role, content, model, tokensUsed,
		tcVal, parentID, participantType, participantID, siblingIdx,
	).Scan(&newID)
	if err != nil {
		return "", err
	}

	// Update cursor to point at new message
	if userID != "" {
		_ = updateCursor(channelID, userID, newID)
	}

	// Touch channel updated_at
	_, _ = database.DB.Exec(`UPDATE channels SET updated_at = NOW() WHERE id = $1`, channelID)
	return newID, nil
}

// ── Usage Logging ─────────────────────────

// logUsage records token usage and cost in the usage_log table.
// Always logs even if tokens are zero — the request happened.
func (h *CompletionHandler) logUsage(
	c *gin.Context,
	channelID, userID, configID, providerScope, modelID string,
	inputTokens, outputTokens, cacheCreation, cacheRead int,
) {
	if h.stores.Usage == nil {
		return
	}

	entry := &models.UsageEntry{
		ChannelID:           &channelID,
		UserID:              userID,
		ProviderConfigID:    &configID,
		ProviderScope:       providerScope,
		ModelID:             modelID,
		PromptTokens:        inputTokens,
		CompletionTokens:    outputTokens,
		CacheCreationTokens: cacheCreation,
		CacheReadTokens:     cacheRead,
	}

	// Look up pricing for cost calculation
	if h.stores.Pricing != nil {
		pricing, err := h.stores.Pricing.GetForModel(c.Request.Context(), configID, modelID)
		if err == nil && pricing != nil {
			entry.CostInput = calcCost(inputTokens, pricing.InputPerM)
			entry.CostOutput = calcCost(outputTokens, pricing.OutputPerM)
		}
	}

	if err := h.stores.Usage.Log(c.Request.Context(), entry); err != nil {
		log.Printf("⚠ Failed to log usage: %v", err)
	}
}

// calcCost computes the cost for a given token count and price-per-million.
func calcCost(tokens int, pricePerM *float64) *float64 {
	if pricePerM == nil || tokens == 0 {
		return nil
	}
	cost := float64(tokens) * *pricePerM / 1_000_000.0
	return &cost
}