@agenticmail/enterprise 0.5.327 → 0.5.328

package/src/runtime/agent-loop.ts

@@ -1,779 +0,0 @@
-/**
- * Core Agent Loop
- *
- * The heart of the runtime. Implements the think-act cycle:
- *   1. Build messages (system prompt + history + user message)
- *   2. Call hooks.beforeLLMCall → inject KB context, policy, memory
- *   3. Stream LLM response via llm-client
- *   4. If tool_use → check permissions via hooks, execute, loop back
- *   5. If end_turn → done
- *   6. If max_tokens → compact context, loop back
- *   7. If maxTurns exceeded → force stop
- */
-
-import type { AgentConfig, AgentMessage, RuntimeHooks, SessionState, StreamEvent, ToolCall } from './types.js';
-import { callLLM, toolsToDefinitions, estimateMessageTokens, type LLMResponse } from './llm-client.js';
-import { ToolRegistry, executeTool } from './tool-executor.js';
-import { compactContext, COMPACTION_THRESHOLD } from './compaction.js';
-import { buildModelChain, withModelFallback, type ModelFallbackConfig } from '../engine/model-fallback.js';
-
-// ─── Constants ───────────────────────────────────────────
-
-const DEFAULT_MAX_TURNS = 0; // 0 = unlimited
-const DEFAULT_MAX_TOKENS = 8192;
-const DEFAULT_TEMPERATURE = 0.7;
-const DEFAULT_CONTEXT_WINDOW = 2_000_000; // 1M — most frontier models support this (Feb 2026)
-// COMPACTION_THRESHOLD imported from ./compaction.js
-
-/**
- * Fix ALL tool_use / tool_result pairing issues in message history.
- * 
- * Anthropic requires:
- * 1. Every tool_result must reference a tool_use in the preceding assistant message
- * 2. Every tool_use in an assistant message must have a tool_result in the immediately following user message
- * 
- * After compaction or message injection (e.g. meeting monitor), either side can be orphaned.
- */
-function fixOrphanedToolBlocks(messages: AgentMessage[]): AgentMessage[] {
-  let fixed = false;
-
-  // ─── Pass 1: Collect all tool_use ids and all tool_result ids ───
-  const allToolUseIds = new Set<string>();
-  const allToolResultIds = new Set<string>();
-  for (const m of messages) {
-    if (!Array.isArray(m.content)) continue;
-    for (const block of m.content as any[]) {
-      if (block.type === 'tool_use' && block.id) allToolUseIds.add(block.id);
-      if (block.type === 'tool_result' && block.tool_use_id) allToolResultIds.add(block.tool_use_id);
-    }
-  }
-
-  // ─── Pass 2: Remove orphaned blocks ───
-  let result = messages.map(m => {
-    if (!Array.isArray(m.content)) return m;
-
-    if (m.role === 'user') {
-      const filtered = (m.content as any[]).filter(block => {
-        if (block.type === 'tool_result' && block.tool_use_id && !allToolUseIds.has(block.tool_use_id)) {
-          fixed = true;
-          return false;
-        }
-        return true;
-      });
-      if (filtered.length === 0) return null;
-      if (filtered.length !== (m.content as any[]).length) return { ...m, content: filtered };
-      return m;
-    }
-
-    if (m.role === 'assistant') {
-      const filtered = (m.content as any[]).filter(block => {
-        if (block.type === 'tool_use' && block.id && !allToolResultIds.has(block.id)) {
-          fixed = true;
-          return false;
-        }
-        return true;
-      });
-      if (filtered.length === 0) return null;
-      if (filtered.length !== (m.content as any[]).length) return { ...m, content: filtered };
-      return m;
-    }
-
-    return m;
-  }).filter(Boolean) as AgentMessage[];
-
-  // ─── Pass 3: Fix adjacency — tool_result MUST immediately follow its tool_use ───
-  // If user messages got injected between assistant tool_use and user tool_result, reorder.
-  for (let i = 0; i < result.length; i++) {
-    const m = result[i];
-    if (m.role !== 'assistant' || !Array.isArray(m.content)) continue;
-    const toolUseIds = (m.content as any[])
-      .filter((b: any) => b.type === 'tool_use' && b.id)
-      .map((b: any) => b.id);
-    if (toolUseIds.length === 0) continue;
-
-    // The very next message must be a user message with tool_results for these IDs
-    const next = result[i + 1];
-    if (next && next.role === 'user' && Array.isArray(next.content)) {
-      const hasResults = (next.content as any[]).some((b: any) => b.type === 'tool_result' && toolUseIds.includes(b.tool_use_id));
-      if (hasResults) continue; // adjacency is fine
-    }
-
-    // Search further for the tool_result message and move it
-    for (let j = i + 2; j < result.length; j++) {
-      const candidate = result[j];
-      if (candidate.role !== 'user' || !Array.isArray(candidate.content)) continue;
-      const matchingResults = (candidate.content as any[]).filter((b: any) =>
-        b.type === 'tool_result' && toolUseIds.includes(b.tool_use_id));
-      if (matchingResults.length === 0) continue;
-
-      // Extract matching tool_results from their current position
-      const remaining = (candidate.content as any[]).filter((b: any) =>
-        !(b.type === 'tool_result' && toolUseIds.includes(b.tool_use_id)));
-
-      // Build the tool_result message to insert right after the tool_use
-      const toolResultMsg: AgentMessage = { role: 'user', content: matchingResults };
-
-      // Update or remove the source message
-      if (remaining.length === 0) {
-        result.splice(j, 1); // remove empty message
-      } else {
-        result[j] = { ...candidate, content: remaining };
-      }
-
-      // Insert right after the tool_use message
-      result.splice(i + 1, 0, toolResultMsg);
-      fixed = true;
-      console.log(`[agent-loop] Reordered tool_result to fix adjacency (moved from index ${j} to ${i + 1})`);
-      break;
-    }
-
-    // If no tool_result found anywhere, remove the tool_use blocks (truly orphaned)
-    // Check if adjacency is STILL broken for this specific message
-    const nextAfterFix = result[i + 1];
-    const adjacencyOk = nextAfterFix && nextAfterFix.role === 'user' && Array.isArray(nextAfterFix.content) &&
-      (nextAfterFix.content as any[]).some((b: any) => b.type === 'tool_result' && toolUseIds.includes(b.tool_use_id));
-    if (!adjacencyOk) {
-      const filtered = (m.content as any[]).filter((b: any) =>
-        !(b.type === 'tool_use' && toolUseIds.includes(b.id)));
-      if (filtered.length === 0) {
-        result.splice(i, 1);
-        i--;
-      } else {
-        result[i] = { ...m, content: filtered };
-      }
-      fixed = true;
-    }
-  }
-
-  if (fixed) console.log(`[agent-loop] Fixed orphaned/non-adjacent tool_use/tool_result blocks`);
-  return result;
-}
-
-// ─── Agent Loop ──────────────────────────────────────────
-
-export interface AgentLoopOptions {
-  apiKey: string;
-  signal?: AbortSignal;
-  onEvent?: (event: StreamEvent) => void;
-  /** Tool execution timeout per call in ms (default: 30000) */
-  toolTimeoutMs?: number;
-  /** Called after each turn completes — saves messages incrementally to DB */
-  onCheckpoint?: (data: { messages: AgentMessage[]; turnCount: number; tokenCount: number }) => Promise<void>;
-  /** Called periodically to keep the session alive */
-  onHeartbeat?: (data: { turnCount: number; tokenCount: number; timestamp: number }) => Promise<void>;
-  /** Session ID for hook calls */
-  sessionId?: string;
-  /** Retry config for LLM calls */
-  retryConfig?: { maxRetryDurationMs?: number; maxRetries?: number; baseDelayMs?: number; maxDelayMs?: number };
-  /** Runtime reference for tool access */
-  runtime?: any;
-  /** Fallback models to try if primary model fails */
-  fallbackModels?: string[];
-  /** Resolve API key for a provider dynamically */
-  resolveApiKey?: (provider: string) => Promise<string | undefined>;
-}
-
-export interface AgentLoopResult {
-  messages: AgentMessage[];
-  status: SessionState['status'];
-  tokenCount: number;
-  turnCount: number;
-  textContent: string;
-  stopReason: string;
-  usage?: { inputTokens: number; outputTokens: number; costUsd: number };
-}
-
-export async function runAgentLoop(
-  config: AgentConfig,
-  initialMessages: AgentMessage[],
-  hooks: RuntimeHooks,
-  options: AgentLoopOptions,
-): Promise<AgentLoopResult> {
-  var maxTurns = config.maxTurns ?? DEFAULT_MAX_TURNS;
-  var maxTokens = config.maxTokens ?? DEFAULT_MAX_TOKENS;
-  var temperature = config.temperature ?? DEFAULT_TEMPERATURE;
-  var contextWindowSize = config.contextWindowSize ?? DEFAULT_CONTEXT_WINDOW;
-  var toolTimeout = options.toolTimeoutMs ?? 120_000; // 2 minutes — meeting_join, browser actions need time
-
-  // Build tool registry
-  var registry = new ToolRegistry();
-  registry.register(config.tools);
-  var toolDefs = toolsToDefinitions(config.tools);
-  // Expose tools globally for cross-tool calls (e.g., browser → meeting_join redirect)
-  (globalThis as any).__currentSessionTools = config.tools;
-
-  // Initialize messages
-  var messages: AgentMessage[] = [];
-
-  // Add system prompt
-  if (config.systemPrompt) {
-    messages.push({ role: 'system', content: config.systemPrompt });
-  }
-
-  // Add history
-  messages.push(...initialMessages);
-
-  var turnCount = 0;
-  var totalTextContent = '';
-  var lastStopReason = 'end_turn';
-  var sessionId = options.sessionId || ''; // Set by caller
-  var cumulativeUsage = { inputTokens: 0, outputTokens: 0, costUsd: 0 };
-
-  // Emit session start
-  var sessionStartEvent: StreamEvent = { type: 'session_start', sessionId: config.agentId };
-  options.onEvent?.(sessionStartEvent);
-
-  while (maxTurns === 0 || turnCount < maxTurns) {
-    // Check abort
-    if (options.signal?.aborted) {
-      return buildResult(messages, 'paused', turnCount, totalTextContent, 'aborted', cumulativeUsage);
-    }
-
-    turnCount++;
-
-    // Heartbeat — keep session alive
-    if (options.onHeartbeat) {
-      try {
-        await options.onHeartbeat({ turnCount, tokenCount: estimateMessageTokens(messages), timestamp: Date.now() });
-      } catch {}
-    }
-
-    // Emit turn start
-    var turnStartEvent: StreamEvent = { type: 'turn_start', turnNumber: turnCount };
-    options.onEvent?.(turnStartEvent);
-
-    // Hook: inject context before LLM call
-    try {
-      messages = await hooks.beforeLLMCall(messages, config.agentId, sessionId);
-    } catch (err: any) {
-      console.warn(`[runtime] beforeLLMCall hook error: ${err.message}`);
-    }
-
-    // Check context window — compact if needed
-    var estimatedTokens = estimateMessageTokens(messages);
-    if (estimatedTokens > contextWindowSize * COMPACTION_THRESHOLD) {
-      messages = await compactContext(messages, config, hooks, { apiKey: options.apiKey, sessionId: options.sessionId });
-    }
-
-    // Always fix orphaned tool blocks before LLM call (can happen from compaction,
-    // meeting monitor injections, or session message injection)
-    messages = fixOrphanedToolBlocks(messages);
-
-    // Budget check — stop if budget exceeded
-    if (hooks.checkBudget) {
-      try {
-        var budgetResult = await hooks.checkBudget(config.agentId, config.orgId, estimateMessageTokens(messages));
-        if (!budgetResult.allowed) {
-          console.log(`[agent-loop] BUDGET EXCEEDED for agent ${config.agentId}: ${budgetResult.reason}`);
-          var budgetEvent: StreamEvent = { type: 'budget_exceeded', reason: budgetResult.reason || 'Budget limit reached' };
-          options.onEvent?.(budgetEvent);
-          return buildResult(messages, 'completed', turnCount, totalTextContent, 'budget_exceeded', cumulativeUsage);
-        }
-        if (budgetResult.remainingUsd !== undefined && budgetResult.remainingUsd < 1.0) {
-          var warnEvent: StreamEvent = { type: 'budget_warning', remainingUsd: budgetResult.remainingUsd, usedUsd: 0 };
-          options.onEvent?.(warnEvent);
-        }
-      } catch {}
-    }
-
-    // Call LLM with retry for transient errors
-    var llmCallStart = Date.now();
-    var llmResponse: LLMResponse = undefined as any;
-    var llmRetryMax = 3;
-    var llmRetryMaxRateLimit = 5; // More retries for rate limits (with longer backoff)
-    var llmRetryDelay = 2000; // start at 2s, doubles each retry
-    for (var llmAttempt = 0; llmAttempt <= llmRetryMax; llmAttempt++) {
-      try {
-        llmResponse = await callLLM(
-          {
-            provider: config.model.provider,
-            modelId: config.model.modelId,
-            apiKey: options.apiKey,
-            thinkingLevel: config.model.thinkingLevel,
-            baseUrl: config.model.baseUrl,
-            headers: config.model.headers,
-            authMode: config.model.authMode,
-          },
-          messages,
-          toolDefs,
-          { maxTokens, temperature, signal: options.signal },
-          options.onEvent,
-          options.retryConfig,
-        );
-        console.log(`[agent-loop] LLM responded in ${Date.now() - llmCallStart}ms (${estimateMessageTokens(messages)} input tokens)`);
-        break; // success
-      } catch (err: any) {
-        // Recover from orphaned tool_result errors (e.g. after compaction)
-        var isOrphanError = /tool_use_id.*tool_result|tool_use.*without.*tool_result|tool_result.*without.*tool_use/i.test(err.message);
-        if (isOrphanError && llmAttempt < llmRetryMax) {
-          // Extract specific tool IDs from the error message for targeted removal
-          var idMatch = err.message.match(/toolu_[A-Za-z0-9_]+/g);
-          var orphanIds = idMatch ? new Set(idMatch) : null;
-          console.warn(`[agent-loop] Orphaned tool blocks detected (ids: ${idMatch?.join(', ') || 'unknown'}) — force-removing and retrying`);
-          messages = fixOrphanedToolBlocks(messages);
-          // If fixOrphanedToolBlocks didn't catch it (e.g. user message injected between
-          // tool_use and tool_result), do a targeted force-removal of the specific IDs
-          if (orphanIds) {
-            messages = messages.map(m => {
-              if (!Array.isArray(m.content)) return m;
-              var filtered = (m.content as any[]).filter(block => {
-                if (block.type === 'tool_use' && block.id && orphanIds!.has(block.id)) return false;
-                if (block.type === 'tool_result' && block.tool_use_id && orphanIds!.has(block.tool_use_id)) return false;
-                return true;
-              });
-              if (filtered.length === 0) return null as any;
-              if (filtered.length !== (m.content as any[]).length) return { ...m, content: filtered };
-              return m;
-            }).filter(Boolean) as AgentMessage[];
-          }
-          continue;
-        }
-        var isRateLimit = /429|rate.limit|rate_limit/i.test(err.message);
-        var isTransient = isRateLimit || /premature close|ECONNRESET|ETIMEDOUT|socket hang up|fetch failed|network|502|503|529|overloaded/i.test(err.message);
-        var effectiveMax = isRateLimit ? llmRetryMaxRateLimit : llmRetryMax;
-        if (isTransient && llmAttempt < effectiveMax) {
-          // Rate limits get longer backoff (30s base) vs transient errors (normal backoff)
-          var delay = isRateLimit ? Math.min(30000 * Math.pow(2, llmAttempt), 120000) : llmRetryDelay * Math.pow(2, llmAttempt);
-          console.warn(`[agent-loop] LLM call failed (attempt ${llmAttempt + 1}/${llmRetryMax + 1}): ${err.message} — retrying in ${delay}ms`);
-          await new Promise(r => setTimeout(r, delay));
-          continue;
-        }
-        // ─── Model Fallback: Try backup models before giving up ───
-        var fallbackModels = (options as any).fallbackModels as string[] | undefined;
-        if (fallbackModels && fallbackModels.length > 0) {
-          var primaryModel = config.model.provider + '/' + config.model.modelId;
-          var remainingFallbacks = fallbackModels.filter(function(m) { return m !== primaryModel; });
-          for (var fi = 0; fi < remainingFallbacks.length; fi++) {
-            var fbModel = remainingFallbacks[fi];
-            var fbParts = fbModel.split('/');
-            var fbProvider = fbParts[0];
-            var fbModelId = fbParts.slice(1).join('/');
-            console.log(`[agent-loop] Primary model failed — trying fallback ${fi + 1}/${remainingFallbacks.length}: ${fbModel}`);
-            try {
-              var fbApiKey = options.apiKey;
-              // Resolve API key for fallback provider if different
-              if (fbProvider !== config.model.provider && (options as any).resolveApiKey) {
-                var resolved = await (options as any).resolveApiKey(fbProvider);
-                if (resolved) fbApiKey = resolved;
-              }
-              llmResponse = await callLLM(
-                { provider: fbProvider, modelId: fbModelId, apiKey: fbApiKey, thinkingLevel: config.model.thinkingLevel },
-                messages, toolDefs,
-                { maxTokens, temperature, signal: options.signal },
-                options.onEvent,
-                options.retryConfig,
-              );
-              console.log(`[agent-loop] Fallback model ${fbModel} succeeded in ${Date.now() - llmCallStart}ms`);
-              break;
-            } catch (fbErr: any) {
-              console.warn(`[agent-loop] Fallback model ${fbModel} also failed: ${fbErr.message}`);
-              if (fi === remainingFallbacks.length - 1) {
-                // All fallbacks exhausted
-                console.error(`[agent-loop] All models failed (primary + ${remainingFallbacks.length} fallbacks): ${err.message}`);
-                var errorEvent: StreamEvent = { type: 'error', message: `All models failed. Primary: ${err.message}`, retryable: false };
-                options.onEvent?.(errorEvent);
-                return buildResult(messages, 'failed', turnCount, totalTextContent, 'error', cumulativeUsage);
-              }
-            }
-          }
-          if (llmResponse) break; // fallback succeeded, exit retry loop
-        }
-
-        if (!llmResponse) {
-          console.error(`[agent-loop] LLM call failed after ${llmAttempt + 1} attempts: ${err.message}`);
-          var errorEvent2: StreamEvent = { type: 'error', message: err.message, retryable: isTransient };
-          options.onEvent?.(errorEvent2);
-          return buildResult(messages, 'failed', turnCount, totalTextContent, 'error', cumulativeUsage);
-        }
-      }
-    }
-
-    // Record LLM usage for budget tracking
-    if (hooks.recordLLMUsage) {
-      try {
-        var usageInput = llmResponse.usage?.inputTokens || 0;
-        var usageOutput = llmResponse.usage?.outputTokens || 0;
-        // Fallback: estimate tokens if provider didn't return usage (common with streaming)
-        if (usageInput === 0 && usageOutput === 0) {
-          usageInput = Math.ceil(estimateMessageTokens(messages) * 0.9);
-          usageOutput = Math.ceil((llmResponse.textContent?.length || 100) / 4);
-        }
-        var costUsd = await estimateCostAsync(hooks, config.model, usageInput, usageOutput);
-        await hooks.recordLLMUsage(config.agentId, config.orgId, {
-          inputTokens: usageInput,
-          outputTokens: usageOutput,
-          costUsd,
-        });
-        cumulativeUsage.inputTokens += usageInput;
-        cumulativeUsage.outputTokens += usageOutput;
-        cumulativeUsage.costUsd += costUsd;
-      } catch {}
-    }
-
-    // Build assistant message from response
-    var assistantMessage = buildAssistantMessage(llmResponse);
-    messages.push(assistantMessage);
-    totalTextContent += llmResponse.textContent;
-
-    // Handle stop reasons
-    if (llmResponse.stopReason === 'end_turn') {
-      lastStopReason = 'end_turn';
-      var turnEndEvent: StreamEvent = { type: 'turn_end', stopReason: 'end_turn' };
-      options.onEvent?.(turnEndEvent);
-
-      // Incremental checkpoint — persist messages to DB after each turn
-      if (options.onCheckpoint) {
-        try {
-          var currentTokens = estimateMessageTokens(messages);
-          await options.onCheckpoint({ messages, turnCount, tokenCount: currentTokens });
-        } catch (err: any) {
-          console.warn('[runtime] Checkpoint save error:', err.message);
-        }
-      }
-
-      break;
-    }
-
-    if (llmResponse.stopReason === 'tool_use' && llmResponse.toolCalls.length > 0) {
-      lastStopReason = 'tool_use';
-
-      // Execute each tool call
-      var toolResults: { tool_use_id: string; content: string; is_error: boolean }[] = [];
-
-      for (var toolCall of llmResponse.toolCalls) {
-        // Hook: check permissions before tool execution
-        var hookResult = await hooks.beforeToolCall({
-          toolCallId: toolCall.id,
-          toolName: toolCall.name,
-          parameters: toolCall.input,
-          agentId: config.agentId,
-          orgId: config.orgId,
-          sessionId,
-        });
-
-        if (!hookResult.allowed) {
-          // Tool blocked by hook
-          console.log(`[agent-loop] 🚫 Tool ${toolCall.name} BLOCKED: ${hookResult.reason} (session: ${sessionId})`);
-          var blockedEvent: StreamEvent = {
-            type: 'tool_call_end',
-            toolName: toolCall.name,
-            result: hookResult.reason,
-            blocked: true,
-          };
-          options.onEvent?.(blockedEvent);
-
-          toolResults.push({
-            tool_use_id: toolCall.id,
-            content: `Tool call blocked: ${hookResult.reason}`,
-            is_error: true,
-          });
-
-          await hooks.afterToolCall(
-            { toolCallId: toolCall.id, toolName: toolCall.name, parameters: toolCall.input, agentId: config.agentId, orgId: config.orgId, sessionId },
-            { success: false, error: hookResult.reason },
-          );
-          continue;
-        }
-
-        // Apply modified parameters if hook changed them
-        var effectiveInput = hookResult.modifiedParameters || toolCall.input;
-        var effectiveToolCall: ToolCall = { ...toolCall, input: effectiveInput };
-
-        // Look up and execute the tool
-        var tool = registry.get(toolCall.name);
-        if (!tool) {
-          toolResults.push({
-            tool_use_id: toolCall.id,
-            content: `Unknown tool: "${toolCall.name}"`,
-            is_error: true,
-          });
-
-          var unknownEvent: StreamEvent = {
-            type: 'tool_call_end',
-            toolName: toolCall.name,
-            result: 'unknown tool',
-            blocked: true,
-          };
-          options.onEvent?.(unknownEvent);
-          continue;
-        }
-
-        console.log(`[agent-loop] 🔧 Executing tool: ${toolCall.name} (session: ${sessionId})`);
-
-        // Track tool start in real-time status
-        try {
-          const { describeToolActivity } = await import('../engine/agent-status.js');
-          if ((options as any).runtime?._reportStatus) {
-            (options as any).runtime._reportStatus({
-              status: 'online',
-              currentActivity: { type: 'tool_call', detail: describeToolActivity(toolCall.name), tool: toolCall.name, startedAt: new Date().toISOString() },
-            });
-          }
-        } catch { /* non-blocking */ }
-
-        var { result, content } = await executeTool(tool, effectiveToolCall, {
-          timeoutMs: toolTimeout,
-          signal: options.signal,
-        });
-
-        if (!result.success) {
-          console.log(`[agent-loop] ❌ Tool ${toolCall.name} failed: ${result.error?.slice(0, 200)}`);
-        } else {
-          console.log(`[agent-loop] ✅ Tool ${toolCall.name} succeeded (${content.length} chars): ${content.slice(0, 300)}`);
-        }
-
-        // Record tool call to activity tracker (single lightweight event, no extra DB writes)
-        try {
-          const { activity } = await import('../engine/routes.js');
-          activity.recordToolCallCompact({
-            agentId: config.agentId, orgId: config.orgId, sessionId,
-            tool: toolCall.name, success: !!result.success,
-            durationMs: result.durationMs, error: result.error?.slice(0, 200),
-          });
-        } catch { /* non-blocking */ }
-
-        // Track tool end in real-time status
-        try {
-          if ((options as any).runtime?._reportStatus) {
-            (options as any).runtime._reportStatus({ currentActivity: null });
-          }
-        } catch { /* non-blocking */ }
-
-        // Dynamic tool injection — request_tools returns new tools to add to the session
-        if ((result as any)._dynamicTools?.length) {
-          var newTools = (result as any)._dynamicTools as any[];
-          var existingNames = new Set(config.tools.map((t: any) => t.name));
-          var added = 0;
-          for (var nt of newTools) {
-            if (!existingNames.has(nt.name)) {
-              config.tools.push(nt);
-              toolDefs.push({ name: nt.name, description: nt.description, input_schema: nt.parameters });
-              existingNames.add(nt.name);
-              added++;
-            }
-          }
-          if (added) {
-            console.log(`[agent-loop] Dynamically loaded ${added} new tools`);
-            (globalThis as any).__currentSessionTools = config.tools;
-          }
-        }
-
-        toolResults.push({
-          tool_use_id: toolCall.id,
-          content,
-          is_error: !result.success,
-        });
-
-        var tcEndEvent: StreamEvent = {
-          type: 'tool_call_end',
-          toolName: toolCall.name,
-          result: result.success ? 'success' : result.error,
-        };
-        options.onEvent?.(tcEndEvent);
-
-        // Hook: record tool call
-        await hooks.afterToolCall(
-          { toolCallId: toolCall.id, toolName: toolCall.name, parameters: effectiveInput, agentId: config.agentId, orgId: config.orgId, sessionId },
-          result,
-        );
-      }
-
-      // Add tool results as user message (Anthropic format)
-      messages.push({
-        role: 'user',
-        content: toolResults.map(function(tr) {
-          return { type: 'tool_result' as const, tool_use_id: tr.tool_use_id, content: tr.content, is_error: tr.is_error };
-        }),
-        tool_results: toolResults.map(function(tr) {
-          return { tool_use_id: tr.tool_use_id, content: tr.content, is_error: tr.is_error };
-        }),
-      });
-
-      // Incremental checkpoint — persist messages to DB after each turn
-      if (options.onCheckpoint) {
-        try {
-          var currentTokens = estimateMessageTokens(messages);
-          await options.onCheckpoint({ messages, turnCount, tokenCount: currentTokens });
-        } catch (err: any) {
-          console.warn('[runtime] Checkpoint save error:', err.message);
-        }
-      }
-
-      // ── Auto-end optimization: skip second LLM call for terminal tools ──
-      // If every tool call in this turn was a "terminal" tool (sends a message, no further
-      // reasoning needed) AND all succeeded, end the session without another LLM round-trip.
-      // Saves ~$0.03 and ~3s per chat message.
-      var TERMINAL_TOOLS = new Set([
-        'google_chat_send_message', 'google_chat_send_card', 'google_chat_send_dm',
-        'google_chat_reply_message',
-      ]);
-      var allTerminal = llmResponse.toolCalls.length > 0
-        && llmResponse.toolCalls.every(function(tc: any) { return TERMINAL_TOOLS.has(tc.name); })
-        && toolResults.every(function(tr: any) { return !tr.is_error; });
-      if (allTerminal) {
-        console.log(`[agent-loop] ⚡ Auto-end: all tools were terminal send tools, skipping confirmation LLM call`);
-        lastStopReason = 'end_turn';
-        var autoEndEvent: StreamEvent = { type: 'turn_end', stopReason: 'end_turn' };
-        options.onEvent?.(autoEndEvent);
-        break;
-      }
-
-      // Continue loop — LLM will process tool results
-      continue;
-    }
-
-    if (llmResponse.stopReason === 'max_tokens') {
-      lastStopReason = 'max_tokens';
-      // Compact and continue
-      messages = await compactContext(messages, config, hooks, { apiKey: options.apiKey, sessionId: options.sessionId });
-      messages = fixOrphanedToolBlocks(messages);
-
-      // Incremental checkpoint — persist messages to DB after each turn
-      if (options.onCheckpoint) {
-        try {
-          var currentTokens = estimateMessageTokens(messages);
-          await options.onCheckpoint({ messages, turnCount, tokenCount: currentTokens });
-        } catch (err: any) {
-          console.warn('[runtime] Checkpoint save error:', err.message);
-        }
-      }
-
-      continue;
-    }
-
-    // Unknown stop reason — break
-    lastStopReason = llmResponse.stopReason;
-    break;
-  }
-
-  if (maxTurns > 0 && turnCount >= maxTurns) {
-    lastStopReason = 'max_turns';
-  }
-
-  // Emit session end
-  var sessionEndEvent: StreamEvent = {
-    type: 'session_end',
-    summary: totalTextContent.slice(0, 200),
-  };
-  options.onEvent?.(sessionEndEvent);
-
-  return buildResult(
-    messages,
-    lastStopReason === 'end_turn' ? 'completed' : lastStopReason === 'max_turns' ? 'completed' : 'completed',
-    turnCount,
-    totalTextContent,
-    lastStopReason,
-    cumulativeUsage,
-  );
-}
-
-// ─── Helpers ─────────────────────────────────────────────
-
-// Built-in fallback pricing (used when no custom pricing configured in dashboard)
-var FALLBACK_PRICES: Record<string, { input: number; output: number }> = {
-  // Anthropic (Feb 2026 — 1M context window)
-  'claude-opus-4-6': { input: 5, output: 25 },
-  'claude-sonnet-4-6': { input: 3, output: 15 },
-  'claude-sonnet-4-5-20250929': { input: 3, output: 15 },
-  'claude-haiku-4-5-20251001': { input: 0.8, output: 4 },
-  // OpenAI
-  'gpt-4o': { input: 2.5, output: 10 },
-  'gpt-4o-mini': { input: 0.15, output: 0.6 },
-  'gpt-4.1': { input: 2, output: 8 },
-  'gpt-4.1-mini': { input: 0.4, output: 1.6 },
-  'gpt-4.1-nano': { input: 0.1, output: 0.4 },
-  'o3': { input: 10, output: 40 },
-  'o4-mini': { input: 1.1, output: 4.4 },
-  // Google Gemini (up to 2M context)
-  'gemini-2.5-pro': { input: 2.5, output: 15 },
-  'gemini-2.5-flash': { input: 0.15, output: 0.6 },
-  'gemini-2.0-flash': { input: 0.1, output: 0.4 },
-  'gemini-3-pro': { input: 2.5, output: 15 },
-  // DeepSeek (128K context)
-  'deepseek-chat': { input: 0.14, output: 0.28 },
-  'deepseek-chat-v3': { input: 0.14, output: 0.28 },
-  'deepseek-reasoner': { input: 0.55, output: 2.19 },
-  // xAI Grok (2M context window)
-  'grok-4': { input: 3, output: 15 },
-  'grok-4-fast': { input: 0.2, output: 0.5 },
-  'grok-3': { input: 3, output: 15 },
-  'grok-3-mini': { input: 0.3, output: 0.5 },
-  // Mistral
-  'mistral-large-latest': { input: 2, output: 6 },
-  'mistral-small-latest': { input: 0.1, output: 0.3 },
-  'codestral-latest': { input: 0.3, output: 0.9 },
-  // Meta Llama (via Groq, Together, etc.)
-  'llama-3.3-70b-versatile': { input: 0.59, output: 0.79 },
-  'llama-3.1-8b-instant': { input: 0.05, output: 0.08 },
-  'meta-llama/Llama-3.3-70B-Instruct-Turbo': { input: 0.88, output: 0.88 },
-};
-
-// In-memory cache for DB pricing (refreshed every 5 min)
-var _pricingCache: { data: Record<string, { input: number; output: number }>; expiresAt: number } | null = null;
-var _PRICING_CACHE_TTL = 5 * 60_000;
-
-async function estimateCostAsync(
-  hooks: import('./types.js').RuntimeHooks,
-  model: import('./types.js').ModelConfig,
-  inputTokens: number,
-  outputTokens: number,
-): Promise<number> {
-  // Try DB pricing via hook
-  if (hooks.getModelPricing) {
-    try {
-      var dbPricing = await hooks.getModelPricing(model.provider, model.modelId);
-      if (dbPricing) {
-        return (inputTokens * dbPricing.inputCostPerMillion + outputTokens * dbPricing.outputCostPerMillion) / 1_000_000;
-      }
-    } catch {}
-  }
-  // Fall back to built-in pricing
-  var p = FALLBACK_PRICES[model.modelId] || { input: 3, output: 15 };
-  return (inputTokens * p.input + outputTokens * p.output) / 1_000_000;
-}
-
-function buildAssistantMessage(response: LLMResponse): AgentMessage {
-  var content: import('./types.js').ContentBlock[] = [];
-
-  if (response.thinkingContent) {
-    content.push({ type: 'thinking', thinking: response.thinkingContent });
-  }
-
-  if (response.textContent) {
-    content.push({ type: 'text', text: response.textContent });
-  }
-
-  for (var tc of response.toolCalls) {
-    content.push({ type: 'tool_use', id: tc.id, name: tc.name, input: tc.input });
-  }
-
-  return {
-    role: 'assistant',
-    content: content.length === 1 && content[0].type === 'text'
-      ? (content[0] as { type: 'text'; text: string }).text
-      : content,
-    tool_calls: response.toolCalls.length > 0 ? response.toolCalls : undefined,
-  };
-}
-
-function buildResult(
-  messages: AgentMessage[],
-  status: SessionState['status'],
-  turnCount: number,
-  textContent: string,
-  stopReason: string,
-  usage?: { inputTokens: number; outputTokens: number; costUsd: number },
-): AgentLoopResult {
-  return {
-    messages,
-    status,
-    tokenCount: estimateMessageTokens(messages),
-    turnCount,
-    textContent,
-    stopReason,
-    usage,
-  };
-}
-
-// Context compaction is now in ./compaction.ts
-// Exported: compactContext, needsCompaction, COMPACTION_THRESHOLD