@geminixiang/mama 0.1.10 → 0.2.0-beta.0

package/dist/agent.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"agent.d.ts","sourceRoot":"","sources":["../src/agent.ts"],"names":[],"mappings":"AAkBA,OAAO,KAAK,EAAE,WAAW,EAAE,mBAAmB,EAAE,YAAY,EAAE,MAAM,cAAc,CAAC;AAInF,OAAO,EAAkB,KAAK,aAAa,EAAE,MAAM,cAAc,CAAC;AAGlE,MAAM,WAAW,cAAc;IAC7B,QAAQ,EAAE,MAAM,CAAC;IACjB,IAAI,EAAE,MAAM,CAAC;IACb,WAAW,EAAE;QAAE,KAAK,EAAE,MAAM,CAAA;KAAE,EAAE,CAAC;IACjC,SAAS,EAAE,MAAM,CAAC;CACnB;AAED,MAAM,WAAW,WAAW;IAC1B,GAAG,CACD,OAAO,EAAE,WAAW,EACpB,WAAW,EAAE,mBAAmB,EAChC,QAAQ,EAAE,YAAY,GACrB,OAAO,CAAC;QAAE,UAAU,EAAE,MAAM,CAAC;QAAC,YAAY,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC;IAC1D,KAAK,IAAI,IAAI,CAAC;IACd,6DAA6D;IAC7D,cAAc,IAAI;QAAE,QAAQ,CAAC,EAAE,MAAM,CAAC;QAAC,KAAK,CAAC,EAAE,MAAM,CAAA;KAAE,GAAG,SAAS,CAAC;CACrE;AAgWD;;;;;;GAMG;AACH,wBAAsB,YAAY,CAChC,aAAa,EAAE,aAAa,EAC5B,UAAU,EAAE,MAAM,EAClB,SAAS,EAAE,MAAM,EACjB,UAAU,EAAE,MAAM,EAClB,YAAY,EAAE,MAAM,GACnB,OAAO,CAAC,WAAW,CAAC,CA8jBtB","sourcesContent":["import { Agent, type AgentEvent } from \"@mariozechner/pi-agent-core\";\nimport { getModel, type ImageContent } from \"@mariozechner/pi-ai\";\nimport {\n  AgentSession,\n  AuthStorage,\n  convertToLlm,\n  createExtensionRuntime,\n  formatSkillsForPrompt,\n  loadSkillsFromDir,\n  ModelRegistry,\n  type ResourceLoader,\n  SessionManager,\n  type Skill,\n} from \"@mariozechner/pi-coding-agent\";\nimport { existsSync, mkdirSync, readFileSync } from \"fs\";\nimport { mkdir, readFile, writeFile } from \"fs/promises\";\nimport { homedir } from \"os\";\nimport { join } from \"path\";\nimport type { ChatMessage, ChatResponseContext, PlatformInfo } from \"./adapter.js\";\nimport { loadAgentConfig } from \"./config.js\";\nimport { createMamaSettingsManager, syncLogToSessionManager } from \"./context.js\";\nimport * as log from \"./log.js\";\nimport { createExecutor, type SandboxConfig } from \"./sandbox.js\";\nimport { createMamaTools } from \"./tools/index.js\";\n\nexport interface PendingMessage {\n  userName: string;\n  text: string;\n  attachments: { local: string }[];\n  timestamp: number;\n}\n\nexport interface AgentRunner {\n  run(\n    message: ChatMessage,\n    responseCtx: ChatResponseContext,\n    platform: PlatformInfo,\n  ): Promise<{ stopReason: string; errorMessage?: string }>;\n  abort(): void;\n  /** Get current step info (tool name, label) for debugging */\n  getCurrentStep(): { toolName?: string; label?: string } | undefined;\n}\n\nconst IMAGE_MIME_TYPES: Record<string, string> = {\n  jpg: \"image/jpeg\",\n  jpeg: \"image/jpeg\",\n  png: \"image/png\",\n  gif: \"image/gif\",\n  webp: \"image/webp\",\n};\n\nfunction getImageMimeType(filename: string): string | undefined {\n  return IMAGE_MIME_TYPES[filename.toLowerCase().split(\".\").pop() || \"\"];\n}\n\nasync function getMemory(channelDir: string): Promise<string> {\n  const parts: string[] = [];\n\n  // Read workspace-level memory (shared across all channels)\n  const workspaceMemoryPath = join(channelDir, \"..\", \"MEMORY.md\");\n  if (existsSync(workspaceMemoryPath)) {\n    try {\n      const content = (await readFile(workspaceMemoryPath, \"utf-8\")).trim();\n      if (content) {\n        parts.push(`### Global Workspace Memory\\n${content}`);\n      }\n    } catch (error) {\n      log.logWarning(\"Failed to read workspace memory\", `${workspaceMemoryPath}: ${error}`);\n    }\n  }\n\n  // Read channel-specific memory\n  const channelMemoryPath = join(channelDir, \"MEMORY.md\");\n  if (existsSync(channelMemoryPath)) {\n    try {\n      const content = (await readFile(channelMemoryPath, \"utf-8\")).trim();\n      if (content) {\n        parts.push(`### Channel-Specific Memory\\n${content}`);\n      }\n    } catch (error) {\n      log.logWarning(\"Failed to read channel memory\", `${channelMemoryPath}: ${error}`);\n    }\n  }\n\n  if (parts.length === 0) {\n    return \"(no working memory yet)\";\n  }\n\n  return parts.join(\"\\n\\n\");\n}\n\nfunction loadMamaSkills(channelDir: string, workspacePath: string): Skill[] {\n  const skillMap = new Map<string, Skill>();\n\n  // channelDir is the host path (e.g., /Users/.../data/C0A34FL8PMH)\n  // hostWorkspacePath is the parent directory on host\n  // workspacePath is the container path (e.g., /workspace)\n  const hostWorkspacePath = join(channelDir, \"..\");\n\n  // Helper to translate host paths to container paths\n  const translatePath = (hostPath: string): string => {\n    if (hostPath.startsWith(hostWorkspacePath)) {\n      return workspacePath + hostPath.slice(hostWorkspacePath.length);\n    }\n    return hostPath;\n  };\n\n  // Load workspace-level skills (global)\n  const workspaceSkillsDir = join(hostWorkspacePath, \"skills\");\n  for (const skill of loadSkillsFromDir({ dir: workspaceSkillsDir, source: \"workspace\" }).skills) {\n    // Translate paths to container paths for system prompt\n    skill.filePath = translatePath(skill.filePath);\n    skill.baseDir = translatePath(skill.baseDir);\n    skillMap.set(skill.name, skill);\n  }\n\n  // Load channel-specific skills (override workspace skills on collision)\n  const channelSkillsDir = join(channelDir, \"skills\");\n  for (const skill of loadSkillsFromDir({ dir: channelSkillsDir, source: \"channel\" }).skills) {\n    skill.filePath = translatePath(skill.filePath);\n    skill.baseDir = translatePath(skill.baseDir);\n    skillMap.set(skill.name, skill);\n  }\n\n  return Array.from(skillMap.values());\n}\n\nfunction buildSystemPrompt(\n  workspacePath: string,\n  channelId: string,\n  memory: string,\n  sandboxConfig: SandboxConfig,\n  platform: PlatformInfo,\n  skills: Skill[],\n): string {\n  const channelPath = `${workspacePath}/${channelId}`;\n  const isDocker = sandboxConfig.type === \"docker\";\n  const isFirecracker = sandboxConfig.type === \"firecracker\";\n\n  // Format channel mappings\n  const channelMappings =\n    platform.channels.length > 0\n      ? platform.channels.map((c) => `${c.id}\\t#${c.name}`).join(\"\\n\")\n      : \"(no channels loaded)\";\n\n  // Format user mappings\n  const userMappings =\n    platform.users.length > 0\n      ? platform.users.map((u) => `${u.id}\\t@${u.userName}\\t${u.displayName}`).join(\"\\n\")\n      : \"(no users loaded)\";\n\n  const envDescription = isDocker\n    ? `You are running inside a Docker container (Alpine Linux).\n- Bash working directory: / (use cd or absolute paths)\n- Install tools with: apk add <package>\n- Your changes persist across sessions`\n    : isFirecracker\n      ? `You are running inside a Firecracker microVM.\n- Bash working directory: / (use cd or absolute paths)\n- Install tools with: apt-get install <package> (Debian-based)\n- Your changes persist across sessions`\n      : `You are running directly on the host machine.\n- Bash working directory: ${process.cwd()}\n- Be careful with system modifications`;\n\n  return `You are mama, a ${platform.name} bot assistant. Be concise. No emojis.\n\n## Context\n- For current date/time, use: date\n- You have access to previous conversation context including tool results from prior turns.\n- For older history beyond your context, search log.jsonl (contains user messages and your final responses, but not tool results).\n- User messages include a \\`[in-thread:TS]\\` marker when sent from within a Slack thread (TS is the root message timestamp). Without this marker, the message is a top-level channel message.\n\n${platform.formattingGuide}\n\n## Platform IDs\nChannels: ${channelMappings}\n\nUsers: ${userMappings}\n\nWhen mentioning users, use <@username> format (e.g., <@mario>).\n\n## Environment\n${envDescription}\n\n## Workspace Layout\n${workspacePath}/\n├── MEMORY.md                    # Global memory (all channels)\n├── skills/                      # Global CLI tools you create\n└── ${channelId}/                # This channel\n    ├── MEMORY.md                # Channel-specific memory\n    ├── log.jsonl                # Message history (no tool results)\n    ├── attachments/             # User-shared files\n    ├── scratch/                 # Your working directory\n    └── skills/                  # Channel-specific tools\n\n## Skills (Custom CLI Tools)\nYou can create reusable CLI tools for recurring tasks (email, APIs, data processing, etc.).\n\n### Creating Skills\nStore in \\`${workspacePath}/skills/<name>/\\` (global) or \\`${channelPath}/skills/<name>/\\` (channel-specific).\nEach skill directory needs a \\`SKILL.md\\` with YAML frontmatter:\n\n\\`\\`\\`markdown\n---\nname: skill-name\ndescription: Short description of what this skill does\n---\n\n# Skill Name\n\nUsage instructions, examples, etc.\nScripts are in: {baseDir}/\n\\`\\`\\`\n\n\\`name\\` and \\`description\\` are required. Use \\`{baseDir}\\` as placeholder for the skill's directory path.\n\n### Available Skills\n${skills.length > 0 ? formatSkillsForPrompt(skills) : \"(no skills installed yet)\"}\n\n## Events\nYou can schedule events that wake you up at specific times or when external things happen. Events are JSON files in \\`${workspacePath}/events/\\`.\n\n### Event Types\n\n**Immediate** - Triggers as soon as harness sees the file. Use in scripts/webhooks to signal external events.\n\\`\\`\\`json\n{\"type\": \"immediate\", \"platform\": \"${platform.name}\", \"channelId\": \"${channelId}\", \"text\": \"New GitHub issue opened\"}\n\\`\\`\\`\n\n**One-shot** - Triggers once at a specific time. Use for reminders.\n\\`\\`\\`json\n{\"type\": \"one-shot\", \"platform\": \"${platform.name}\", \"channelId\": \"${channelId}\", \"text\": \"Remind Mario about dentist\", \"at\": \"2025-12-15T09:00:00+01:00\"}\n\\`\\`\\`\n\n**Periodic** - Triggers on a cron schedule. Use for recurring tasks.\n\\`\\`\\`json\n{\"type\": \"periodic\", \"platform\": \"${platform.name}\", \"channelId\": \"${channelId}\", \"text\": \"Check inbox and summarize\", \"schedule\": \"0 9 * * 1-5\", \"timezone\": \"${Intl.DateTimeFormat().resolvedOptions().timeZone}\"}\n\\`\\`\\`\n\n### Cron Format\n\\`minute hour day-of-month month day-of-week\\`\n- \\`0 9 * * *\\` = daily at 9:00\n- \\`0 9 * * 1-5\\` = weekdays at 9:00\n- \\`30 14 * * 1\\` = Mondays at 14:30\n- \\`0 0 1 * *\\` = first of each month at midnight\n\n### Timezones\nAll \\`at\\` timestamps must include offset (e.g., \\`+01:00\\`). Periodic events use IANA timezone names. The harness runs in ${Intl.DateTimeFormat().resolvedOptions().timeZone}. When users mention times without timezone, assume ${Intl.DateTimeFormat().resolvedOptions().timeZone}.\n\n### Platform Routing\nSet \\`platform\\` to the target bot platform (\\`${platform.name}\\` for this conversation). When only one platform is running, omitting \\`platform\\` is allowed for backward compatibility, but include it by default to avoid ambiguity.\n\n### Creating Events\nUse unique filenames to avoid overwriting existing events. Include a timestamp or random suffix:\n\\`\\`\\`bash\ncat > ${workspacePath}/events/dentist-reminder-$(date +%s).json << 'EOF'\n{\"type\": \"one-shot\", \"platform\": \"${platform.name}\", \"channelId\": \"${channelId}\", \"text\": \"Dentist tomorrow\", \"at\": \"2025-12-14T09:00:00+01:00\"}\nEOF\n\\`\\`\\`\nOr check if file exists first before creating.\n\n### Managing Events\n- List: \\`ls ${workspacePath}/events/\\`\n- View: \\`cat ${workspacePath}/events/foo.json\\`\n- Delete/cancel: \\`rm ${workspacePath}/events/foo.json\\`\n\n### When Events Trigger\nYou receive a message like:\n\\`\\`\\`\n[EVENT:dentist-reminder.json:one-shot:2025-12-14T09:00:00+01:00] Dentist tomorrow\n\\`\\`\\`\nImmediate and one-shot events auto-delete after triggering. Periodic events persist until you delete them.\n\n### Silent Completion\nFor periodic events where there's nothing to report, respond with just \\`[SILENT]\\` (no other text). This deletes the status message and posts nothing to the platform. Use this to avoid spamming the channel when periodic checks find nothing actionable.\n\n### Debouncing\nWhen writing programs that create immediate events (email watchers, webhook handlers, etc.), always debounce. If 50 emails arrive in a minute, don't create 50 immediate events. Instead collect events over a window and create ONE immediate event summarizing what happened, or just signal \"new activity, check inbox\" rather than per-item events. Or simpler: use a periodic event to check for new items every N minutes instead of immediate events.\n\n### Limits\nMaximum 5 events can be queued. Don't create excessive immediate or periodic events.\n\n## Memory\nWrite to MEMORY.md files to persist context across conversations.\n- Global (${workspacePath}/MEMORY.md): skills, preferences, project info\n- Channel (${channelPath}/MEMORY.md): channel-specific decisions, ongoing work\nUpdate when you learn something important or when asked to remember something.\n\n### Current Memory\n${memory}\n\n## System Configuration Log\nMaintain ${workspacePath}/SYSTEM.md to log all environment modifications:\n- Installed packages (apk add, npm install, pip install)\n- Environment variables set\n- Config files modified (~/.gitconfig, cron jobs, etc.)\n- Skill dependencies installed\n\nUpdate this file whenever you modify the environment. On fresh container, read it first to restore your setup.\n\n## Log Queries (for older history)\nFormat: \\`{\"date\":\"...\",\"ts\":\"...\",\"user\":\"...\",\"userName\":\"...\",\"text\":\"...\",\"isBot\":false}\\`\nThe log contains user messages and your final responses (not tool calls/results).\n${isDocker ? \"Install jq: apk add jq\" : \"\"}\n${isFirecracker ? \"Install jq: apt-get install jq\" : \"\"}\n\n\\`\\`\\`bash\n# Recent messages\ntail -30 log.jsonl | jq -c '{date: .date[0:19], user: (.userName // .user), text}'\n\n# Search for specific topic\ngrep -i \"topic\" log.jsonl | jq -c '{date: .date[0:19], user: (.userName // .user), text}'\n\n# Messages from specific user\ngrep '\"userName\":\"mario\"' log.jsonl | tail -20 | jq -c '{date: .date[0:19], text}'\n\\`\\`\\`\n\n## Tools\n- bash: Run shell commands (primary tool). Install packages as needed.\n- read: Read files\n- write: Create/overwrite files\n- edit: Surgical file edits\n- attach: Share files to the platform\n\nEach tool requires a \"label\" parameter (shown to user).\n`;\n}\n\nfunction truncate(text: string, maxLen: number): string {\n  if (text.length <= maxLen) return text;\n  return `${text.substring(0, maxLen - 3)}...`;\n}\n\nfunction extractToolResultText(result: unknown): string {\n  if (typeof result === \"string\") {\n    return result;\n  }\n\n  if (\n    result &&\n    typeof result === \"object\" &&\n    \"content\" in result &&\n    Array.isArray((result as { content: unknown }).content)\n  ) {\n    const content = (result as { content: Array<{ type: string; text?: string }> }).content;\n    const textParts: string[] = [];\n    for (const part of content) {\n      if (part.type === \"text\" && part.text) {\n        textParts.push(part.text);\n      }\n    }\n    if (textParts.length > 0) {\n      return textParts.join(\"\\n\");\n    }\n  }\n\n  return JSON.stringify(result);\n}\n\nfunction formatToolArgsForSlack(_toolName: string, args: Record<string, unknown>): string {\n  const lines: string[] = [];\n\n  for (const [key, value] of Object.entries(args)) {\n    if (key === \"label\") continue;\n\n    if (key === \"path\" && typeof value === \"string\") {\n      const offset = args.offset as number | undefined;\n      const limit = args.limit as number | undefined;\n      if (offset !== undefined && limit !== undefined) {\n        lines.push(`${value}:${offset}-${offset + limit}`);\n      } else {\n        lines.push(value);\n      }\n      continue;\n    }\n\n    if (key === \"offset\" || key === \"limit\") continue;\n\n    if (typeof value === \"string\") {\n      lines.push(value);\n    } else {\n      lines.push(JSON.stringify(value));\n    }\n  }\n\n  return lines.join(\"\\n\");\n}\n\n// ============================================================================\n// Agent runner\n// ============================================================================\n\n/**\n * Create a new AgentRunner for a channel.\n * Sets up the session and subscribes to events once.\n *\n * Runner caching is handled by the caller (channelStates in main.ts).\n * This is a stateless factory function.\n */\nexport async function createRunner(\n  sandboxConfig: SandboxConfig,\n  sessionKey: string,\n  channelId: string,\n  channelDir: string,\n  workspaceDir: string,\n): Promise<AgentRunner> {\n  const agentConfig = loadAgentConfig(workspaceDir);\n\n  // Initialize logger with settings from config\n  log.initLogger({\n    logFormat: agentConfig.logFormat,\n    logLevel: agentConfig.logLevel,\n  });\n\n  const executor = createExecutor(sandboxConfig);\n  const workspacePath = executor.getWorkspacePath(channelDir.replace(`/${channelId}`, \"\"));\n\n  // Create tools (per-runner, with per-runner upload function setter)\n  const { tools, setUploadFunction } = createMamaTools(executor);\n\n  // Resolve model from config\n  // Use 'as any' cast because agentConfig.provider/model are plain strings,\n  // while getModel() has constrained generic types for known providers.\n  // eslint-disable-next-line @typescript-eslint/no-explicit-any\n  const model = (getModel as any)(agentConfig.provider, agentConfig.model);\n\n  // Initial system prompt (will be updated each run with fresh memory/channels/users/skills)\n  const memory = await getMemory(channelDir);\n  const skills = loadMamaSkills(channelDir, workspacePath);\n  const emptyPlatform: PlatformInfo = {\n    name: \"slack\",\n    formattingGuide: \"\",\n    channels: [],\n    users: [],\n  };\n  const systemPrompt = buildSystemPrompt(\n    workspacePath,\n    channelId,\n    memory,\n    sandboxConfig,\n    emptyPlatform,\n    skills,\n  );\n\n  // Create session manager and settings manager\n  // Per-session context file: {channelDir}/sessions/{rootTs}/context.jsonl\n  const rootTs = sessionKey.includes(\":\") ? sessionKey.split(\":\").pop()! : sessionKey;\n  const sessionDir = join(channelDir, \"sessions\", rootTs);\n  mkdirSync(sessionDir, { recursive: true });\n  const contextFile = join(sessionDir, \"context.jsonl\");\n  const sessionManager = SessionManager.open(contextFile, channelDir);\n  const settingsManager = createMamaSettingsManager(join(channelDir, \"..\"));\n\n  // Create AuthStorage and ModelRegistry\n  // Auth stored outside workspace so agent can't access it\n  const authStorage = AuthStorage.create(join(homedir(), \".pi\", \"mama\", \"auth.json\"));\n  const modelRegistry = new ModelRegistry(authStorage);\n\n  // Create agent\n  const agent = new Agent({\n    initialState: {\n      systemPrompt,\n      model,\n      thinkingLevel:\n        (agentConfig.thinkingLevel as \"off\" | \"low\" | \"medium\" | \"high\" | undefined) ?? \"off\",\n      tools,\n    },\n    convertToLlm,\n    getApiKey: async () => {\n      const key = await modelRegistry.getApiKeyForProvider(model.provider);\n      if (!key)\n        throw new Error(\n          `No API key for provider \"${model.provider}\". Set the appropriate environment variable or configure via auth.json`,\n        );\n      return key;\n    },\n  });\n\n  // Load existing messages\n  const loadedSession = sessionManager.buildSessionContext();\n  if (loadedSession.messages.length > 0) {\n    agent.replaceMessages(loadedSession.messages);\n    log.logInfo(\n      `[${channelId}] Loaded ${loadedSession.messages.length} messages from context.jsonl`,\n    );\n  }\n\n  const resourceLoader: ResourceLoader = {\n    getExtensions: () => ({ extensions: [], errors: [], runtime: createExtensionRuntime() }),\n    getSkills: () => ({ skills: [], diagnostics: [] }),\n    getPrompts: () => ({ prompts: [], diagnostics: [] }),\n    getThemes: () => ({ themes: [], diagnostics: [] }),\n    getAgentsFiles: () => ({ agentsFiles: [] }),\n    getSystemPrompt: () => systemPrompt,\n    getAppendSystemPrompt: () => [],\n    extendResources: () => {},\n    reload: async () => {},\n  };\n\n  const baseToolsOverride = Object.fromEntries(tools.map((tool) => [tool.name, tool]));\n\n  // Create AgentSession wrapper\n  const session = new AgentSession({\n    agent,\n    sessionManager,\n    settingsManager,\n    cwd: process.cwd(),\n    modelRegistry,\n    resourceLoader,\n    baseToolsOverride,\n  });\n\n  // Mutable per-run state - event handler references this\n  const runState = {\n    responseCtx: null as ChatResponseContext | null,\n    logCtx: null as { channelId: string; userName?: string; channelName?: string } | null,\n    queue: null as {\n      enqueue(fn: () => Promise<void>, errorContext: string): void;\n      enqueueMessage(\n        text: string,\n        target: \"main\" | \"thread\",\n        errorContext: string,\n        doLog?: boolean,\n      ): void;\n    } | null,\n    pendingTools: new Map<string, { toolName: string; args: unknown; startTime: number }>(),\n    totalUsage: {\n      input: 0,\n      output: 0,\n      cacheRead: 0,\n      cacheWrite: 0,\n      cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n    },\n    stopReason: \"stop\",\n    errorMessage: undefined as string | undefined,\n  };\n\n  // Subscribe to events ONCE\n  session.subscribe(async (event) => {\n    // Skip if no active run\n    if (!runState.responseCtx || !runState.logCtx || !runState.queue) return;\n\n    const { responseCtx, logCtx, queue, pendingTools } = runState;\n\n    if (event.type === \"tool_execution_start\") {\n      const agentEvent = event as AgentEvent & { type: \"tool_execution_start\" };\n      const args = agentEvent.args as { label?: string };\n      const label = args.label || agentEvent.toolName;\n\n      pendingTools.set(agentEvent.toolCallId, {\n        toolName: agentEvent.toolName,\n        args: agentEvent.args,\n        startTime: Date.now(),\n      });\n\n      log.logToolStart(\n        logCtx,\n        agentEvent.toolName,\n        label,\n        agentEvent.args as Record<string, unknown>,\n      );\n      // Tool labels are omitted from the main message to reduce Slack noise.\n      // Tool execution details are still posted to the thread (see tool_execution_end).\n    } else if (event.type === \"tool_execution_end\") {\n      const agentEvent = event as AgentEvent & { type: \"tool_execution_end\" };\n      const resultStr = extractToolResultText(agentEvent.result);\n      const pending = pendingTools.get(agentEvent.toolCallId);\n      pendingTools.delete(agentEvent.toolCallId);\n\n      const durationMs = pending ? Date.now() - pending.startTime : 0;\n\n      if (agentEvent.isError) {\n        log.logToolError(logCtx, agentEvent.toolName, durationMs, resultStr);\n      } else {\n        log.logToolSuccess(logCtx, agentEvent.toolName, durationMs, resultStr);\n      }\n\n      // Post args + result to thread\n      const label = pending?.args ? (pending.args as { label?: string }).label : undefined;\n      const argsFormatted = pending\n        ? formatToolArgsForSlack(agentEvent.toolName, pending.args as Record<string, unknown>)\n        : \"(args not found)\";\n      const duration = (durationMs / 1000).toFixed(1);\n      let threadMessage = `*${agentEvent.isError ? \"✗\" : \"✓\"} ${agentEvent.toolName}*`;\n      if (label) threadMessage += `: ${label}`;\n      threadMessage += ` (${duration}s)\\n`;\n      if (argsFormatted) threadMessage += `\\`\\`\\`\\n${argsFormatted}\\n\\`\\`\\`\\n`;\n      threadMessage += `*Result:*\\n\\`\\`\\`\\n${resultStr}\\n\\`\\`\\``;\n\n      // Only post thread details for tools with meaningful output (bash, attach).\n      // Skip read/write/edit to reduce Slack noise — their results are in the log.\n      const quietTools = new Set([\"read\", \"write\", \"edit\"]);\n      if (!quietTools.has(agentEvent.toolName)) {\n        queue.enqueueMessage(threadMessage, \"thread\", \"tool result thread\", false);\n      }\n\n      if (agentEvent.isError) {\n        queue.enqueue(\n          () => responseCtx.respond(`_Error: ${truncate(resultStr, 200)}_`),\n          \"tool error\",\n        );\n      }\n    } else if (event.type === \"message_start\") {\n      const agentEvent = event as AgentEvent & { type: \"message_start\" };\n      if (agentEvent.message.role === \"assistant\") {\n        log.logResponseStart(logCtx);\n      }\n    } else if (event.type === \"message_end\") {\n      const agentEvent = event as AgentEvent & { type: \"message_end\" };\n      if (agentEvent.message.role === \"assistant\") {\n        const assistantMsg = agentEvent.message as any;\n\n        if (assistantMsg.stopReason) {\n          runState.stopReason = assistantMsg.stopReason;\n        }\n        if (assistantMsg.errorMessage) {\n          runState.errorMessage = assistantMsg.errorMessage;\n        }\n\n        if (assistantMsg.usage) {\n          runState.totalUsage.input += assistantMsg.usage.input;\n          runState.totalUsage.output += assistantMsg.usage.output;\n          runState.totalUsage.cacheRead += assistantMsg.usage.cacheRead;\n          runState.totalUsage.cacheWrite += assistantMsg.usage.cacheWrite;\n          runState.totalUsage.cost.input += assistantMsg.usage.cost.input;\n          runState.totalUsage.cost.output += assistantMsg.usage.cost.output;\n          runState.totalUsage.cost.cacheRead += assistantMsg.usage.cost.cacheRead;\n          runState.totalUsage.cost.cacheWrite += assistantMsg.usage.cost.cacheWrite;\n          runState.totalUsage.cost.total += assistantMsg.usage.cost.total;\n        }\n\n        const content = agentEvent.message.content;\n        const thinkingParts: string[] = [];\n        const textParts: string[] = [];\n        for (const part of content) {\n          if (part.type === \"thinking\") {\n            thinkingParts.push((part as any).thinking);\n          } else if (part.type === \"text\") {\n            textParts.push((part as any).text);\n          }\n        }\n\n        const text = textParts.join(\"\\n\");\n\n        for (const thinking of thinkingParts) {\n          log.logThinking(logCtx, thinking);\n          queue.enqueueMessage(`_${thinking}_`, \"main\", \"thinking main\");\n          queue.enqueueMessage(`_${thinking}_`, \"thread\", \"thinking thread\", false);\n        }\n\n        if (text.trim()) {\n          log.logResponse(logCtx, text);\n          queue.enqueueMessage(text, \"main\", \"response main\");\n          // Only overflow to thread for texts that will be truncated in main\n          if (text.length > SLACK_MAX_LENGTH) {\n            queue.enqueueMessage(text, \"thread\", \"response thread\", false);\n          }\n        }\n      }\n    } else if (event.type === \"compaction_start\") {\n      log.logInfo(`Auto-compaction started (reason: ${(event as any).reason})`);\n      queue.enqueue(() => responseCtx.respond(\"_Compacting context..._\"), \"compaction start\");\n    } else if (event.type === \"compaction_end\") {\n      const compEvent = event as any;\n      if (compEvent.result) {\n        log.logInfo(`Auto-compaction complete: ${compEvent.result.tokensBefore} tokens compacted`);\n      } else if (compEvent.aborted) {\n        log.logInfo(\"Auto-compaction aborted\");\n      }\n    } else if (event.type === \"auto_retry_start\") {\n      const retryEvent = event as any;\n      log.logWarning(\n        `Retrying (${retryEvent.attempt}/${retryEvent.maxAttempts})`,\n        retryEvent.errorMessage,\n      );\n      queue.enqueue(\n        () =>\n          responseCtx.respond(`_Retrying (${retryEvent.attempt}/${retryEvent.maxAttempts})..._`),\n        \"retry\",\n      );\n    }\n  });\n\n  // Message limit constant\n  const SLACK_MAX_LENGTH = 40000;\n  const splitForSlack = (text: string): string[] => {\n    if (text.length <= SLACK_MAX_LENGTH) return [text];\n    const parts: string[] = [];\n    let remaining = text;\n    let partNum = 1;\n    while (remaining.length > 0) {\n      const chunk = remaining.substring(0, SLACK_MAX_LENGTH - 50);\n      remaining = remaining.substring(SLACK_MAX_LENGTH - 50);\n      const suffix = remaining.length > 0 ? `\\n_(continued ${partNum}...)_` : \"\";\n      parts.push(chunk + suffix);\n      partNum++;\n    }\n    return parts;\n  };\n\n  return {\n    async run(\n      message: ChatMessage,\n      responseCtx: ChatResponseContext,\n      platform: PlatformInfo,\n    ): Promise<{ stopReason: string; errorMessage?: string }> {\n      // Extract channelId from sessionKey (format: \"channelId:rootTs\" or just \"channelId\")\n      const sessionChannel = message.sessionKey.split(\":\")[0];\n\n      // Ensure channel directory exists\n      await mkdir(channelDir, { recursive: true });\n\n      // Sync messages from log.jsonl that arrived while we were offline or busy\n      // Exclude the current message (it will be added via prompt())\n      // Default sync range is 10 days (handled by syncLogToSessionManager)\n      // Thread filter ensures only messages from this session's thread are synced\n      const syncedCount = await syncLogToSessionManager(\n        sessionManager,\n        channelDir,\n        message.id,\n        undefined,\n        { rootTs, threadTs: message.threadTs },\n      );\n      if (syncedCount > 0) {\n        log.logInfo(`[${channelId}] Synced ${syncedCount} messages from log.jsonl`);\n      }\n\n      // Reload messages from context.jsonl\n      // This picks up any messages synced above\n      const reloadedSession = sessionManager.buildSessionContext();\n      if (reloadedSession.messages.length > 0) {\n        agent.replaceMessages(reloadedSession.messages);\n        log.logInfo(\n          `[${channelId}] Reloaded ${reloadedSession.messages.length} messages from context`,\n        );\n      }\n\n      // Update system prompt with fresh memory, channel/user info, and skills\n      const memory = await getMemory(channelDir);\n      const skills = loadMamaSkills(channelDir, workspacePath);\n      const systemPrompt = buildSystemPrompt(\n        workspacePath,\n        channelId,\n        memory,\n        sandboxConfig,\n        platform,\n        skills,\n      );\n      session.agent.setSystemPrompt(systemPrompt);\n\n      // Set up file upload function\n      setUploadFunction(async (filePath: string, title?: string) => {\n        const hostPath = translateToHostPath(filePath, channelDir, workspacePath, channelId);\n        await responseCtx.uploadFile(hostPath, title);\n      });\n\n      // Reset per-run state\n      runState.responseCtx = responseCtx;\n      runState.logCtx = {\n        channelId: sessionChannel,\n        userName: message.userName,\n        channelName: undefined,\n      };\n      runState.pendingTools.clear();\n      runState.totalUsage = {\n        input: 0,\n        output: 0,\n        cacheRead: 0,\n        cacheWrite: 0,\n        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n      };\n      runState.stopReason = \"stop\";\n      runState.errorMessage = undefined;\n\n      // Create queue for this run\n      let queueChain = Promise.resolve();\n      runState.queue = {\n        enqueue(fn: () => Promise<void>, errorContext: string): void {\n          queueChain = queueChain.then(async () => {\n            try {\n              await fn();\n            } catch (err) {\n              const errMsg = err instanceof Error ? err.message : String(err);\n              log.logWarning(`API error (${errorContext})`, errMsg);\n              try {\n                // Split long error messages to avoid msg_too_long\n                const errParts = splitForSlack(`_Error: ${errMsg}_`);\n                for (const part of errParts) {\n                  await responseCtx.respondInThread(part);\n                }\n              } catch {\n                // Ignore\n              }\n            }\n          });\n        },\n        enqueueMessage(\n          text: string,\n          target: \"main\" | \"thread\",\n          errorContext: string,\n          _doLog = true,\n        ): void {\n          const parts = splitForSlack(text);\n          for (const part of parts) {\n            this.enqueue(\n              () =>\n                target === \"main\" ? responseCtx.respond(part) : responseCtx.respondInThread(part),\n              errorContext,\n            );\n          }\n        },\n      };\n\n      // Log context info\n      log.logInfo(\n        `Context sizes - system: ${systemPrompt.length} chars, memory: ${memory.length} chars`,\n      );\n      log.logInfo(`Channels: ${platform.channels.length}, Users: ${platform.users.length}`);\n\n      // Build user message with timestamp and username prefix\n      // Format: \"[YYYY-MM-DD HH:MM:SS+HH:MM] [username]: message\" so LLM knows when and who\n      const now = new Date();\n      const pad = (n: number) => n.toString().padStart(2, \"0\");\n      const offset = -now.getTimezoneOffset();\n      const offsetSign = offset >= 0 ? \"+\" : \"-\";\n      const offsetHours = pad(Math.floor(Math.abs(offset) / 60));\n      const offsetMins = pad(Math.abs(offset) % 60);\n      const timestamp = `${now.getFullYear()}-${pad(now.getMonth() + 1)}-${pad(now.getDate())} ${pad(now.getHours())}:${pad(now.getMinutes())}:${pad(now.getSeconds())}${offsetSign}${offsetHours}:${offsetMins}`;\n      const threadContext = message.threadTs ? ` [in-thread:${message.threadTs}]` : \"\";\n      let userMessage = `[${timestamp}] [${message.userName || \"unknown\"}]${threadContext}: ${message.text}`;\n\n      const imageAttachments: ImageContent[] = [];\n      const nonImagePaths: string[] = [];\n\n      for (const a of message.attachments || []) {\n        // a.localPath is the path relative to the workspace (same as old a.local)\n        const fullPath = `${workspacePath}/${a.localPath}`;\n        const mimeType = getImageMimeType(a.localPath);\n\n        if (mimeType && existsSync(fullPath)) {\n          try {\n            imageAttachments.push({\n              type: \"image\",\n              mimeType,\n              data: readFileSync(fullPath).toString(\"base64\"),\n            });\n          } catch {\n            nonImagePaths.push(fullPath);\n          }\n        } else {\n          nonImagePaths.push(fullPath);\n        }\n      }\n\n      if (nonImagePaths.length > 0) {\n        userMessage += `\\n\\n<slack_attachments>\\n${nonImagePaths.join(\"\\n\")}\\n</slack_attachments>`;\n      }\n\n      // Debug: write context to last_prompt.jsonl\n      const debugContext = {\n        systemPrompt,\n        messages: session.messages,\n        newUserMessage: userMessage,\n        imageAttachmentCount: imageAttachments.length,\n      };\n      await writeFile(join(channelDir, \"last_prompt.jsonl\"), JSON.stringify(debugContext, null, 2));\n\n      await session.prompt(\n        userMessage,\n        imageAttachments.length > 0 ? { images: imageAttachments } : undefined,\n      );\n\n      // Wait for queued messages\n      await queueChain;\n\n      // Handle error case - update main message and post error to thread\n      if (runState.stopReason === \"error\" && runState.errorMessage) {\n        try {\n          await responseCtx.replaceResponse(\"_Sorry, something went wrong_\");\n          // Split long error messages to avoid msg_too_long\n          const errorParts = splitForSlack(`_Error: ${runState.errorMessage}_`);\n          for (const part of errorParts) {\n            await responseCtx.respondInThread(part);\n          }\n        } catch (err) {\n          const errMsg = err instanceof Error ? err.message : String(err);\n          log.logWarning(\"Failed to post error message\", errMsg);\n        }\n      } else {\n        // Final message update\n        const messages = session.messages;\n        const lastAssistant = messages.filter((m) => m.role === \"assistant\").pop();\n        const finalText =\n          lastAssistant?.content\n            .filter((c): c is { type: \"text\"; text: string } => c.type === \"text\")\n            .map((c) => c.text)\n            .join(\"\\n\") || \"\";\n\n        // Check for [SILENT] marker - delete message and thread instead of posting\n        if (finalText.trim() === \"[SILENT]\" || finalText.trim().startsWith(\"[SILENT]\")) {\n          try {\n            await responseCtx.deleteResponse();\n            log.logInfo(\"Silent response - deleted message and thread\");\n          } catch (err) {\n            const errMsg = err instanceof Error ? err.message : String(err);\n            log.logWarning(\"Failed to delete message for silent response\", errMsg);\n          }\n        } else if (finalText.trim()) {\n          try {\n            const mainText =\n              finalText.length > SLACK_MAX_LENGTH\n                ? `${finalText.substring(0, SLACK_MAX_LENGTH - 50)}\\n\\n_(see thread for full response)_`\n                : finalText;\n            await responseCtx.replaceResponse(mainText);\n          } catch (err) {\n            const errMsg = err instanceof Error ? err.message : String(err);\n            log.logWarning(\"Failed to replace message with final text\", errMsg);\n          }\n        }\n      }\n\n      // Log usage summary with context info\n      if (runState.totalUsage.cost.total > 0) {\n        // Get last non-aborted assistant message for context calculation\n        const messages = session.messages;\n        const lastAssistantMessage = messages\n          .slice()\n          .reverse()\n          .find((m) => m.role === \"assistant\" && (m as any).stopReason !== \"aborted\") as any;\n\n        const contextTokens = lastAssistantMessage\n          ? lastAssistantMessage.usage.input +\n            lastAssistantMessage.usage.output +\n            lastAssistantMessage.usage.cacheRead +\n            lastAssistantMessage.usage.cacheWrite\n          : 0;\n        const contextWindow = model.contextWindow || 200000;\n\n        const summary = log.logUsageSummary(\n          runState.logCtx!,\n          runState.totalUsage,\n          contextTokens,\n          contextWindow,\n        );\n        // Split long summaries to avoid msg_too_long\n        const summaryParts = splitForSlack(summary);\n        for (const part of summaryParts) {\n          runState.queue!.enqueue(\n            () => responseCtx.respondInThread(part, { style: \"muted\" }),\n            \"usage summary\",\n          );\n        }\n        await queueChain;\n      }\n\n      // Clear run state\n      runState.responseCtx = null;\n      runState.logCtx = null;\n      runState.queue = null;\n\n      return { stopReason: runState.stopReason, errorMessage: runState.errorMessage };\n    },\n\n    abort(): void {\n      session.abort();\n    },\n\n    getCurrentStep(): { toolName?: string; label?: string } | undefined {\n      const pending = runState.pendingTools;\n      if (pending.size === 0) return undefined;\n      // Get the first pending tool\n      const first = pending.values().next().value;\n      if (!first) return undefined;\n      return {\n        toolName: first.toolName,\n        label: (first.args as { label?: string })?.label,\n      };\n    },\n  };\n}\n\n/**\n * Translate container path back to host path for file operations\n */\nfunction translateToHostPath(\n  containerPath: string,\n  channelDir: string,\n  workspacePath: string,\n  channelId: string,\n): string {\n  if (workspacePath === \"/workspace\") {\n    const prefix = `/workspace/${channelId}/`;\n    if (containerPath.startsWith(prefix)) {\n      return join(channelDir, containerPath.slice(prefix.length));\n    }\n    if (containerPath.startsWith(\"/workspace/\")) {\n      return join(channelDir, \"..\", containerPath.slice(\"/workspace/\".length));\n    }\n  }\n  return containerPath;\n}\n"]}
+{"version":3,"file":"agent.d.ts","sourceRoot":"","sources":["../src/agent.ts"],"names":[],"mappings":"AAiBA,OAAO,KAAK,EAAE,WAAW,EAAE,mBAAmB,EAAE,YAAY,EAAE,MAAM,cAAc,CAAC;AAInF,OAAO,EAAkB,KAAK,aAAa,EAAE,MAAM,cAAc,CAAC;AAiBlE,MAAM,WAAW,cAAc;IAC7B,QAAQ,EAAE,MAAM,CAAC;IACjB,IAAI,EAAE,MAAM,CAAC;IACb,WAAW,EAAE;QAAE,KAAK,EAAE,MAAM,CAAA;KAAE,EAAE,CAAC;IACjC,SAAS,EAAE,MAAM,CAAC;CACnB;AAED,MAAM,WAAW,WAAW;IAC1B,GAAG,CACD,OAAO,EAAE,WAAW,EACpB,WAAW,EAAE,mBAAmB,EAChC,QAAQ,EAAE,YAAY,GACrB,OAAO,CAAC;QAAE,UAAU,EAAE,MAAM,CAAC;QAAC,YAAY,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC;IAC1D,KAAK,IAAI,IAAI,CAAC;IACd,6DAA6D;IAC7D,cAAc,IAAI;QAAE,QAAQ,CAAC,EAAE,MAAM,CAAC;QAAC,KAAK,CAAC,EAAE,MAAM,CAAA;KAAE,GAAG,SAAS,CAAC;CACrE;AAgWD;;;;;;GAMG;AACH,wBAAsB,YAAY,CAChC,aAAa,EAAE,aAAa,EAC5B,UAAU,EAAE,MAAM,EAClB,SAAS,EAAE,MAAM,EACjB,UAAU,EAAE,MAAM,EAClB,YAAY,EAAE,MAAM,GACnB,OAAO,CAAC,WAAW,CAAC,CAguBtB","sourcesContent":["import { Agent, type AgentEvent } from \"@mariozechner/pi-agent-core\";\nimport { getModel, type ImageContent } from \"@mariozechner/pi-ai\";\nimport {\n  AgentSession,\n  AuthStorage,\n  convertToLlm,\n  DefaultResourceLoader,\n  formatSkillsForPrompt,\n  loadSkillsFromDir,\n  ModelRegistry,\n  SessionManager,\n  type Skill,\n} from \"@mariozechner/pi-coding-agent\";\nimport { existsSync, readFileSync } from \"fs\";\nimport { mkdir, readFile, writeFile } from \"fs/promises\";\nimport { homedir } from \"os\";\nimport { join } from \"path\";\nimport type { ChatMessage, ChatResponseContext, PlatformInfo } from \"./adapter.js\";\nimport { loadAgentConfig } from \"./config.js\";\nimport { createMamaSettingsManager, syncLogToSessionManager } from \"./context.js\";\nimport * as log from \"./log.js\";\nimport { createExecutor, type SandboxConfig } from \"./sandbox.js\";\nimport { addLifecycleBreadcrumb, metricAttributes } from \"./sentry.js\";\nimport {\n  createManagedSessionFileAtPath,\n  extractSessionSuffix,\n  extractSessionUuid,\n  forkThreadSessionFile,\n  getSessionDir,\n  getThreadSessionFile,\n  openManagedSession,\n  resolveChannelSessionFile,\n  resolveManagedSessionFile,\n  tryResolveThreadSession,\n} from \"./session-store.js\";\nimport { createMamaTools } from \"./tools/index.js\";\nimport * as Sentry from \"@sentry/node\";\n\nexport interface PendingMessage {\n  userName: string;\n  text: string;\n  attachments: { local: string }[];\n  timestamp: number;\n}\n\nexport interface AgentRunner {\n  run(\n    message: ChatMessage,\n    responseCtx: ChatResponseContext,\n    platform: PlatformInfo,\n  ): Promise<{ stopReason: string; errorMessage?: string }>;\n  abort(): void;\n  /** Get current step info (tool name, label) for debugging */\n  getCurrentStep(): { toolName?: string; label?: string } | undefined;\n}\n\nconst IMAGE_MIME_TYPES: Record<string, string> = {\n  jpg: \"image/jpeg\",\n  jpeg: \"image/jpeg\",\n  png: \"image/png\",\n  gif: \"image/gif\",\n  webp: \"image/webp\",\n};\n\nfunction getImageMimeType(filename: string): string | undefined {\n  return IMAGE_MIME_TYPES[filename.toLowerCase().split(\".\").pop() || \"\"];\n}\n\nasync function getMemory(channelDir: string): Promise<string> {\n  const parts: string[] = [];\n\n  // Read workspace-level memory (shared across all channels)\n  const workspaceMemoryPath = join(channelDir, \"..\", \"MEMORY.md\");\n  if (existsSync(workspaceMemoryPath)) {\n    try {\n      const content = (await readFile(workspaceMemoryPath, \"utf-8\")).trim();\n      if (content) {\n        parts.push(`### Global Workspace Memory\\n${content}`);\n      }\n    } catch (error) {\n      log.logWarning(\"Failed to read workspace memory\", `${workspaceMemoryPath}: ${error}`);\n    }\n  }\n\n  // Read channel-specific memory\n  const channelMemoryPath = join(channelDir, \"MEMORY.md\");\n  if (existsSync(channelMemoryPath)) {\n    try {\n      const content = (await readFile(channelMemoryPath, \"utf-8\")).trim();\n      if (content) {\n        parts.push(`### Channel-Specific Memory\\n${content}`);\n      }\n    } catch (error) {\n      log.logWarning(\"Failed to read channel memory\", `${channelMemoryPath}: ${error}`);\n    }\n  }\n\n  if (parts.length === 0) {\n    return \"(no working memory yet)\";\n  }\n\n  return parts.join(\"\\n\\n\");\n}\n\nfunction loadMamaSkills(channelDir: string, workspacePath: string): Skill[] {\n  const skillMap = new Map<string, Skill>();\n\n  // channelDir is the host path (e.g., /Users/.../data/C0A34FL8PMH)\n  // hostWorkspacePath is the parent directory on host\n  // workspacePath is the container path (e.g., /workspace)\n  const hostWorkspacePath = join(channelDir, \"..\");\n\n  // Helper to translate host paths to container paths\n  const translatePath = (hostPath: string): string => {\n    if (hostPath.startsWith(hostWorkspacePath)) {\n      return workspacePath + hostPath.slice(hostWorkspacePath.length);\n    }\n    return hostPath;\n  };\n\n  // Load workspace-level skills (global)\n  const workspaceSkillsDir = join(hostWorkspacePath, \"skills\");\n  for (const skill of loadSkillsFromDir({ dir: workspaceSkillsDir, source: \"workspace\" }).skills) {\n    // Translate paths to container paths for system prompt\n    skill.filePath = translatePath(skill.filePath);\n    skill.baseDir = translatePath(skill.baseDir);\n    skillMap.set(skill.name, skill);\n  }\n\n  // Load channel-specific skills (override workspace skills on collision)\n  const channelSkillsDir = join(channelDir, \"skills\");\n  for (const skill of loadSkillsFromDir({ dir: channelSkillsDir, source: \"channel\" }).skills) {\n    skill.filePath = translatePath(skill.filePath);\n    skill.baseDir = translatePath(skill.baseDir);\n    skillMap.set(skill.name, skill);\n  }\n\n  return Array.from(skillMap.values());\n}\n\nfunction buildSystemPrompt(\n  workspacePath: string,\n  channelId: string,\n  memory: string,\n  sandboxConfig: SandboxConfig,\n  platform: PlatformInfo,\n  skills: Skill[],\n): string {\n  const channelPath = `${workspacePath}/${channelId}`;\n  const isDocker = sandboxConfig.type === \"docker\";\n  const isFirecracker = sandboxConfig.type === \"firecracker\";\n\n  // Format channel mappings\n  const channelMappings =\n    platform.channels.length > 0\n      ? platform.channels.map((c) => `${c.id}\\t#${c.name}`).join(\"\\n\")\n      : \"(no channels loaded)\";\n\n  // Format user mappings\n  const userMappings =\n    platform.users.length > 0\n      ? platform.users.map((u) => `${u.id}\\t@${u.userName}\\t${u.displayName}`).join(\"\\n\")\n      : \"(no users loaded)\";\n\n  const envDescription = isDocker\n    ? `You are running inside a Docker container (Alpine Linux).\n- Bash working directory: / (use cd or absolute paths)\n- Install tools with: apk add <package>\n- Your changes persist across sessions`\n    : isFirecracker\n      ? `You are running inside a Firecracker microVM.\n- Bash working directory: / (use cd or absolute paths)\n- Install tools with: apt-get install <package> (Debian-based)\n- Your changes persist across sessions`\n      : `You are running directly on the host machine.\n- Bash working directory: ${process.cwd()}\n- Be careful with system modifications`;\n\n  return `You are mama, a ${platform.name} bot assistant. Be concise. No emojis.\n\n## Context\n- For current date/time, use: date\n- You have access to previous conversation context including tool results from prior turns.\n- For older history beyond your context, search log.jsonl (contains user messages and your final responses, but not tool results).\n- User messages include a \\`[in-thread:TS]\\` marker when sent from within a Slack thread (TS is the root message timestamp). Without this marker, the message is a top-level channel message.\n\n${platform.formattingGuide}\n\n## Platform IDs\nChannels: ${channelMappings}\n\nUsers: ${userMappings}\n\nWhen mentioning users, use <@username> format (e.g., <@mario>).\n\n## Environment\n${envDescription}\n\n## Workspace Layout\n${workspacePath}/\n├── MEMORY.md                    # Global memory (all channels)\n├── skills/                      # Global CLI tools you create\n└── ${channelId}/                # This channel\n    ├── MEMORY.md                # Channel-specific memory\n    ├── log.jsonl                # Message history (no tool results)\n    ├── attachments/             # User-shared files\n    ├── scratch/                 # Your working directory\n    └── skills/                  # Channel-specific tools\n\n## Skills (Custom CLI Tools)\nYou can create reusable CLI tools for recurring tasks (email, APIs, data processing, etc.).\n\n### Creating Skills\nStore in \\`${workspacePath}/skills/<name>/\\` (global) or \\`${channelPath}/skills/<name>/\\` (channel-specific).\nEach skill directory needs a \\`SKILL.md\\` with YAML frontmatter:\n\n\\`\\`\\`markdown\n---\nname: skill-name\ndescription: Short description of what this skill does\n---\n\n# Skill Name\n\nUsage instructions, examples, etc.\nScripts are in: {baseDir}/\n\\`\\`\\`\n\n\\`name\\` and \\`description\\` are required. Use \\`{baseDir}\\` as placeholder for the skill's directory path.\n\n### Available Skills\n${skills.length > 0 ? formatSkillsForPrompt(skills) : \"(no skills installed yet)\"}\n\n## Events\nYou can schedule events that wake you up at specific times or when external things happen. Events are JSON files in \\`${workspacePath}/events/\\`.\n\n### Event Types\n\n**Immediate** - Triggers as soon as harness sees the file. Use in scripts/webhooks to signal external events.\n\\`\\`\\`json\n{\"type\": \"immediate\", \"platform\": \"${platform.name}\", \"channelId\": \"${channelId}\", \"text\": \"New GitHub issue opened\"}\n\\`\\`\\`\n\n**One-shot** - Triggers once at a specific time. Use for reminders.\n\\`\\`\\`json\n{\"type\": \"one-shot\", \"platform\": \"${platform.name}\", \"channelId\": \"${channelId}\", \"text\": \"Remind Mario about dentist\", \"at\": \"2025-12-15T09:00:00+01:00\"}\n\\`\\`\\`\n\n**Periodic** - Triggers on a cron schedule. Use for recurring tasks.\n\\`\\`\\`json\n{\"type\": \"periodic\", \"platform\": \"${platform.name}\", \"channelId\": \"${channelId}\", \"text\": \"Check inbox and summarize\", \"schedule\": \"0 9 * * 1-5\", \"timezone\": \"${Intl.DateTimeFormat().resolvedOptions().timeZone}\"}\n\\`\\`\\`\n\n### Cron Format\n\\`minute hour day-of-month month day-of-week\\`\n- \\`0 9 * * *\\` = daily at 9:00\n- \\`0 9 * * 1-5\\` = weekdays at 9:00\n- \\`30 14 * * 1\\` = Mondays at 14:30\n- \\`0 0 1 * *\\` = first of each month at midnight\n\n### Timezones\nAll \\`at\\` timestamps must include offset (e.g., \\`+01:00\\`). Periodic events use IANA timezone names. The harness runs in ${Intl.DateTimeFormat().resolvedOptions().timeZone}. When users mention times without timezone, assume ${Intl.DateTimeFormat().resolvedOptions().timeZone}.\n\n### Platform Routing\nSet \\`platform\\` to the target bot platform (\\`${platform.name}\\` for this conversation). When only one platform is running, omitting \\`platform\\` is allowed for backward compatibility, but include it by default to avoid ambiguity.\n\n### Creating Events\nUse unique filenames to avoid overwriting existing events. Include a timestamp or random suffix:\n\\`\\`\\`bash\ncat > ${workspacePath}/events/dentist-reminder-$(date +%s).json << 'EOF'\n{\"type\": \"one-shot\", \"platform\": \"${platform.name}\", \"channelId\": \"${channelId}\", \"text\": \"Dentist tomorrow\", \"at\": \"2025-12-14T09:00:00+01:00\"}\nEOF\n\\`\\`\\`\nOr check if file exists first before creating.\n\n### Managing Events\n- List: \\`ls ${workspacePath}/events/\\`\n- View: \\`cat ${workspacePath}/events/foo.json\\`\n- Delete/cancel: \\`rm ${workspacePath}/events/foo.json\\`\n\n### When Events Trigger\nYou receive a message like:\n\\`\\`\\`\n[EVENT:dentist-reminder.json:one-shot:2025-12-14T09:00:00+01:00] Dentist tomorrow\n\\`\\`\\`\nImmediate and one-shot events auto-delete after triggering. Periodic events persist until you delete them.\n\n### Silent Completion\nFor periodic events where there's nothing to report, respond with just \\`[SILENT]\\` (no other text). This deletes the status message and posts nothing to the platform. Use this to avoid spamming the channel when periodic checks find nothing actionable.\n\n### Debouncing\nWhen writing programs that create immediate events (email watchers, webhook handlers, etc.), always debounce. If 50 emails arrive in a minute, don't create 50 immediate events. Instead collect events over a window and create ONE immediate event summarizing what happened, or just signal \"new activity, check inbox\" rather than per-item events. Or simpler: use a periodic event to check for new items every N minutes instead of immediate events.\n\n### Limits\nMaximum 5 events can be queued. Don't create excessive immediate or periodic events.\n\n## Memory\nWrite to MEMORY.md files to persist context across conversations.\n- Global (${workspacePath}/MEMORY.md): skills, preferences, project info\n- Channel (${channelPath}/MEMORY.md): channel-specific decisions, ongoing work\nUpdate when you learn something important or when asked to remember something.\n\n### Current Memory\n${memory}\n\n## System Configuration Log\nMaintain ${workspacePath}/SYSTEM.md to log all environment modifications:\n- Installed packages (apk add, npm install, pip install)\n- Environment variables set\n- Config files modified (~/.gitconfig, cron jobs, etc.)\n- Skill dependencies installed\n\nUpdate this file whenever you modify the environment. On fresh container, read it first to restore your setup.\n\n## Log Queries (for older history)\nFormat: \\`{\"date\":\"...\",\"ts\":\"...\",\"user\":\"...\",\"userName\":\"...\",\"text\":\"...\",\"isBot\":false}\\`\nThe log contains user messages and your final responses (not tool calls/results).\n${isDocker ? \"Install jq: apk add jq\" : \"\"}\n${isFirecracker ? \"Install jq: apt-get install jq\" : \"\"}\n\n\\`\\`\\`bash\n# Recent messages\ntail -30 log.jsonl | jq -c '{date: .date[0:19], user: (.userName // .user), text}'\n\n# Search for specific topic\ngrep -i \"topic\" log.jsonl | jq -c '{date: .date[0:19], user: (.userName // .user), text}'\n\n# Messages from specific user\ngrep '\"userName\":\"mario\"' log.jsonl | tail -20 | jq -c '{date: .date[0:19], text}'\n\\`\\`\\`\n\n## Tools\n- bash: Run shell commands (primary tool). Install packages as needed.\n- read: Read files\n- write: Create/overwrite files\n- edit: Surgical file edits\n- attach: Share files to the platform\n\nEach tool requires a \"label\" parameter (shown to user).\n`;\n}\n\nfunction truncate(text: string, maxLen: number): string {\n  if (text.length <= maxLen) return text;\n  return `${text.substring(0, maxLen - 3)}...`;\n}\n\nfunction extractToolResultText(result: unknown): string {\n  if (typeof result === \"string\") {\n    return result;\n  }\n\n  if (\n    result &&\n    typeof result === \"object\" &&\n    \"content\" in result &&\n    Array.isArray((result as { content: unknown }).content)\n  ) {\n    const content = (result as { content: Array<{ type: string; text?: string }> }).content;\n    const textParts: string[] = [];\n    for (const part of content) {\n      if (part.type === \"text\" && part.text) {\n        textParts.push(part.text);\n      }\n    }\n    if (textParts.length > 0) {\n      return textParts.join(\"\\n\");\n    }\n  }\n\n  return JSON.stringify(result);\n}\n\nfunction formatToolArgsForSlack(_toolName: string, args: Record<string, unknown>): string {\n  const lines: string[] = [];\n\n  for (const [key, value] of Object.entries(args)) {\n    if (key === \"label\") continue;\n\n    if (key === \"path\" && typeof value === \"string\") {\n      const offset = args.offset as number | undefined;\n      const limit = args.limit as number | undefined;\n      if (offset !== undefined && limit !== undefined) {\n        lines.push(`${value}:${offset}-${offset + limit}`);\n      } else {\n        lines.push(value);\n      }\n      continue;\n    }\n\n    if (key === \"offset\" || key === \"limit\") continue;\n\n    if (typeof value === \"string\") {\n      lines.push(value);\n    } else {\n      lines.push(JSON.stringify(value));\n    }\n  }\n\n  return lines.join(\"\\n\");\n}\n\n// ============================================================================\n// Agent runner\n// ============================================================================\n\n/**\n * Create a new AgentRunner for a channel.\n * Sets up the session and subscribes to events once.\n *\n * Runner caching is handled by the caller (channelStates in main.ts).\n * This is a stateless factory function.\n */\nexport async function createRunner(\n  sandboxConfig: SandboxConfig,\n  sessionKey: string,\n  channelId: string,\n  channelDir: string,\n  workspaceDir: string,\n): Promise<AgentRunner> {\n  const agentConfig = loadAgentConfig(workspaceDir);\n\n  // Initialize logger with settings from config\n  log.initLogger({\n    logFormat: agentConfig.logFormat,\n    logLevel: agentConfig.logLevel,\n  });\n\n  const executor = createExecutor(sandboxConfig);\n  const workspacePath = executor.getWorkspacePath(channelDir.replace(`/${channelId}`, \"\"));\n\n  // Create tools (per-runner, with per-runner upload function setter)\n  const { tools, setUploadFunction } = createMamaTools(executor);\n\n  // Resolve model from config\n  // Use 'as any' cast because agentConfig.provider/model are plain strings,\n  // while getModel() has constrained generic types for known providers.\n  // eslint-disable-next-line @typescript-eslint/no-explicit-any\n  const model = (getModel as any)(agentConfig.provider, agentConfig.model);\n\n  // Initial system prompt (will be updated each run with fresh memory/channels/users/skills)\n  const memory = await getMemory(channelDir);\n  const skills = loadMamaSkills(channelDir, workspacePath);\n  const emptyPlatform: PlatformInfo = {\n    name: \"slack\",\n    formattingGuide: \"\",\n    channels: [],\n    users: [],\n  };\n  const systemPrompt = buildSystemPrompt(\n    workspacePath,\n    channelId,\n    memory,\n    sandboxConfig,\n    emptyPlatform,\n    skills,\n  );\n\n  // Create session manager and settings manager\n  // Channel sessions use {channelDir}/sessions/current.\n  // Thread sessions use fixed files: {channelDir}/sessions/{threadTs}.jsonl\n  const sessionDir = getSessionDir(channelDir, sessionKey);\n  const isThread = sessionKey.includes(\":\");\n\n  let sessionManager!: SessionManager;\n  let contextFile!: string;\n\n  if (isThread) {\n    const threadFile = getThreadSessionFile(channelDir, sessionKey);\n    const existing = tryResolveThreadSession(threadFile);\n    if (existing) {\n      contextFile = existing;\n      sessionManager = openManagedSession(contextFile, sessionDir, channelDir);\n    } else {\n      const channelSource = resolveChannelSessionFile(channelDir);\n      if (channelSource) {\n        try {\n          contextFile = forkThreadSessionFile(channelSource, threadFile, channelDir);\n          sessionManager = openManagedSession(contextFile, sessionDir, channelDir);\n        } catch {\n          contextFile = createManagedSessionFileAtPath(threadFile, channelDir);\n          sessionManager = openManagedSession(contextFile, sessionDir, channelDir);\n        }\n      } else {\n        contextFile = createManagedSessionFileAtPath(threadFile, channelDir);\n        sessionManager = openManagedSession(contextFile, sessionDir, channelDir);\n      }\n    }\n  } else {\n    // Channel/DM session: normal resolve\n    contextFile = resolveManagedSessionFile(sessionDir, channelDir);\n    sessionManager = openManagedSession(contextFile, sessionDir, channelDir);\n  }\n  const sessionUuid = extractSessionUuid(contextFile);\n  // Used for Slack thread filtering — for non-Slack platforms this is effectively a no-op\n  const rootTs = extractSessionSuffix(sessionKey);\n  const settingsManager = createMamaSettingsManager(join(channelDir, \"..\"));\n\n  // Create AuthStorage and ModelRegistry\n  // Auth stored outside workspace so agent can't access it\n  const authStorage = AuthStorage.create(join(homedir(), \".pi\", \"mama\", \"auth.json\"));\n  const modelRegistry = new ModelRegistry(authStorage);\n\n  // Create agent\n  const agent = new Agent({\n    initialState: {\n      systemPrompt,\n      model,\n      thinkingLevel:\n        (agentConfig.thinkingLevel as \"off\" | \"low\" | \"medium\" | \"high\" | undefined) ?? \"off\",\n      tools,\n    },\n    convertToLlm,\n    getApiKey: async () => {\n      const key = await modelRegistry.getApiKeyForProvider(model.provider);\n      if (!key)\n        throw new Error(\n          `No API key for provider \"${model.provider}\". Set the appropriate environment variable or configure via auth.json`,\n        );\n      return key;\n    },\n  });\n\n  // Load existing messages\n  const loadedSession = sessionManager.buildSessionContext();\n  if (loadedSession.messages.length > 0) {\n    agent.replaceMessages(loadedSession.messages);\n    log.logInfo(\n      `[${channelId}] Loaded ${loadedSession.messages.length} messages from context.jsonl`,\n    );\n  }\n\n  // Load extensions, skills, prompts, themes via DefaultResourceLoader\n  // This reads ~/.pi/agent/settings.json (packages, extensions enable/disable)\n  // and discovers resources from standard locations + npm/git packages.\n  const resourceLoader = new DefaultResourceLoader({\n    cwd: workspaceDir,\n    systemPrompt,\n  });\n  try {\n    await resourceLoader.reload();\n    const extResult = resourceLoader.getExtensions();\n    if (extResult.errors.length > 0) {\n      for (const err of extResult.errors) {\n        log.logWarning(`[${channelId}] Extension load error: ${err.path}`, err.error);\n      }\n    }\n    log.logInfo(\n      `[${channelId}] Loaded ${extResult.extensions.length} extension(s): ${extResult.extensions.map((e) => e.path).join(\", \")}`,\n    );\n  } catch (error) {\n    log.logWarning(`[${channelId}] Failed to load resources`, String(error));\n  }\n\n  const baseToolsOverride = Object.fromEntries(tools.map((tool) => [tool.name, tool]));\n\n  // Create AgentSession wrapper\n  const session = new AgentSession({\n    agent,\n    sessionManager,\n    settingsManager,\n    cwd: workspaceDir,\n    modelRegistry,\n    resourceLoader,\n    baseToolsOverride,\n  });\n\n  // Mutable per-run state - event handler references this\n  const runState = {\n    responseCtx: null as ChatResponseContext | null,\n    logCtx: null as {\n      channelId: string;\n      userName?: string;\n      channelName?: string;\n      sessionId?: string;\n    } | null,\n    queue: null as {\n      enqueue(fn: () => Promise<void>, errorContext: string): void;\n      enqueueMessage(\n        text: string,\n        target: \"main\" | \"thread\",\n        errorContext: string,\n        doLog?: boolean,\n      ): void;\n    } | null,\n    pendingTools: new Map<string, { toolName: string; args: unknown; startTime: number }>(),\n    totalUsage: {\n      input: 0,\n      output: 0,\n      cacheRead: 0,\n      cacheWrite: 0,\n      cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n    },\n    llmCallCount: 0,\n    stopReason: \"stop\",\n    errorMessage: undefined as string | undefined,\n  };\n\n  // Subscribe to events ONCE\n  session.subscribe(async (event) => {\n    // Skip if no active run\n    if (!runState.responseCtx || !runState.logCtx || !runState.queue) return;\n\n    const { responseCtx, logCtx, queue, pendingTools } = runState;\n    const baseAttrs = { channel_id: logCtx.channelId, session_id: logCtx.sessionId };\n\n    if (event.type === \"tool_execution_start\") {\n      const agentEvent = event as AgentEvent & { type: \"tool_execution_start\" };\n      const args = agentEvent.args as { label?: string };\n      const label = args.label || agentEvent.toolName;\n\n      pendingTools.set(agentEvent.toolCallId, {\n        toolName: agentEvent.toolName,\n        args: agentEvent.args,\n        startTime: Date.now(),\n      });\n      addLifecycleBreadcrumb(\"agent.tool.started\", {\n        tool: agentEvent.toolName,\n        ...baseAttrs,\n      });\n\n      log.logToolStart(\n        logCtx,\n        agentEvent.toolName,\n        label,\n        agentEvent.args as Record<string, unknown>,\n      );\n      // Tool labels are omitted from the main message to reduce Slack noise.\n      // Tool execution details are still posted to the thread (see tool_execution_end).\n    } else if (event.type === \"tool_execution_end\") {\n      const agentEvent = event as AgentEvent & { type: \"tool_execution_end\" };\n      const resultStr = extractToolResultText(agentEvent.result);\n      const pending = pendingTools.get(agentEvent.toolCallId);\n      pendingTools.delete(agentEvent.toolCallId);\n\n      const durationMs = pending ? Date.now() - pending.startTime : 0;\n\n      Sentry.metrics.count(\"agent.tool.calls\", 1, {\n        attributes: metricAttributes({\n          tool: agentEvent.toolName,\n          error: String(agentEvent.isError),\n          ...baseAttrs,\n        }),\n      });\n      Sentry.metrics.distribution(\"agent.tool.duration\", durationMs, {\n        unit: \"millisecond\",\n        attributes: metricAttributes({\n          tool: agentEvent.toolName,\n          ...baseAttrs,\n        }),\n      });\n      addLifecycleBreadcrumb(\"agent.tool.completed\", {\n        tool: agentEvent.toolName,\n        error: agentEvent.isError,\n        duration_ms: durationMs,\n        ...baseAttrs,\n      });\n\n      if (agentEvent.isError) {\n        log.logToolError(logCtx, agentEvent.toolName, durationMs, resultStr);\n      } else {\n        log.logToolSuccess(logCtx, agentEvent.toolName, durationMs, resultStr);\n      }\n\n      // Post args + result to thread\n      const label = pending?.args ? (pending.args as { label?: string }).label : undefined;\n      const argsFormatted = pending\n        ? formatToolArgsForSlack(agentEvent.toolName, pending.args as Record<string, unknown>)\n        : \"(args not found)\";\n      const duration = (durationMs / 1000).toFixed(1);\n      let threadMessage = `*${agentEvent.isError ? \"✗\" : \"✓\"} ${agentEvent.toolName}*`;\n      if (label) threadMessage += `: ${label}`;\n      threadMessage += ` (${duration}s)\\n`;\n      if (argsFormatted) threadMessage += `\\`\\`\\`\\n${argsFormatted}\\n\\`\\`\\`\\n`;\n      threadMessage += `*Result:*\\n\\`\\`\\`\\n${resultStr}\\n\\`\\`\\``;\n\n      // Only post thread details for tools with meaningful output (bash, attach).\n      // Skip read/write/edit to reduce Slack noise — their results are in the log.\n      const quietTools = new Set([\"read\", \"write\", \"edit\"]);\n      if (!quietTools.has(agentEvent.toolName)) {\n        queue.enqueueMessage(threadMessage, \"thread\", \"tool result thread\", false);\n      }\n\n      if (agentEvent.isError) {\n        queue.enqueue(\n          () => responseCtx.respond(`_Error: ${truncate(resultStr, 200)}_`),\n          \"tool error\",\n        );\n      }\n    } else if (event.type === \"message_start\") {\n      const agentEvent = event as AgentEvent & { type: \"message_start\" };\n      if (agentEvent.message.role === \"assistant\") {\n        runState.llmCallCount += 1;\n        addLifecycleBreadcrumb(\"agent.llm.call.started\", {\n          call_index: runState.llmCallCount,\n          provider: model.provider,\n          model: agentConfig.model,\n          ...baseAttrs,\n        });\n        log.logResponseStart(logCtx);\n      }\n    } else if (event.type === \"message_end\") {\n      const agentEvent = event as AgentEvent & { type: \"message_end\" };\n      if (agentEvent.message.role === \"assistant\") {\n        const assistantMsg = agentEvent.message as any;\n\n        if (assistantMsg.stopReason) {\n          runState.stopReason = assistantMsg.stopReason;\n        }\n        if (assistantMsg.errorMessage) {\n          runState.errorMessage = assistantMsg.errorMessage;\n        }\n\n        if (assistantMsg.usage) {\n          runState.totalUsage.input += assistantMsg.usage.input;\n          runState.totalUsage.output += assistantMsg.usage.output;\n          runState.totalUsage.cacheRead += assistantMsg.usage.cacheRead;\n          runState.totalUsage.cacheWrite += assistantMsg.usage.cacheWrite;\n          runState.totalUsage.cost.input += assistantMsg.usage.cost.input;\n          runState.totalUsage.cost.output += assistantMsg.usage.cost.output;\n          runState.totalUsage.cost.cacheRead += assistantMsg.usage.cost.cacheRead;\n          runState.totalUsage.cost.cacheWrite += assistantMsg.usage.cost.cacheWrite;\n          runState.totalUsage.cost.total += assistantMsg.usage.cost.total;\n\n          // Per-turn LLM metrics\n          const llmAttributes = metricAttributes({\n            provider: model.provider,\n            model: agentConfig.model,\n            ...baseAttrs,\n            stop_reason: assistantMsg.stopReason,\n            error: Boolean(assistantMsg.errorMessage),\n          });\n          Sentry.metrics.count(\"agent.llm.calls\", 1, { attributes: llmAttributes });\n          Sentry.metrics.distribution(\"agent.llm.tokens_in\", assistantMsg.usage.input, {\n            attributes: llmAttributes,\n          });\n          Sentry.metrics.distribution(\"agent.llm.tokens_out\", assistantMsg.usage.output, {\n            attributes: llmAttributes,\n          });\n          if (assistantMsg.usage.cacheRead > 0) {\n            Sentry.metrics.distribution(\"agent.llm.cache_read\", assistantMsg.usage.cacheRead, {\n              attributes: llmAttributes,\n            });\n          }\n          if (assistantMsg.usage.cacheWrite > 0) {\n            Sentry.metrics.distribution(\"agent.llm.cache_write\", assistantMsg.usage.cacheWrite, {\n              attributes: llmAttributes,\n            });\n          }\n          Sentry.metrics.distribution(\"agent.llm.cost_per_turn\", assistantMsg.usage.cost.total, {\n            attributes: llmAttributes,\n          });\n          addLifecycleBreadcrumb(\"agent.llm.call.completed\", {\n            call_index: runState.llmCallCount,\n            provider: model.provider,\n            model: agentConfig.model,\n            stop_reason: assistantMsg.stopReason,\n            error: Boolean(assistantMsg.errorMessage),\n            input_tokens: assistantMsg.usage.input,\n            output_tokens: assistantMsg.usage.output,\n            cost_total_usd: assistantMsg.usage.cost.total,\n          });\n        }\n\n        const content = agentEvent.message.content;\n        const thinkingParts: string[] = [];\n        const textParts: string[] = [];\n        for (const part of content) {\n          if (part.type === \"thinking\") {\n            thinkingParts.push((part as any).thinking);\n          } else if (part.type === \"text\") {\n            textParts.push((part as any).text);\n          }\n        }\n\n        const text = textParts.join(\"\\n\");\n\n        for (const thinking of thinkingParts) {\n          log.logThinking(logCtx, thinking);\n          queue.enqueueMessage(`_${thinking}_`, \"main\", \"thinking main\");\n          queue.enqueueMessage(`_${thinking}_`, \"thread\", \"thinking thread\", false);\n        }\n\n        if (text.trim()) {\n          log.logResponse(logCtx, text);\n          queue.enqueueMessage(text, \"main\", \"response main\");\n          // Only overflow to thread for texts that will be truncated in main\n          if (text.length > SLACK_MAX_LENGTH) {\n            queue.enqueueMessage(text, \"thread\", \"response thread\", false);\n          }\n        }\n      }\n    } else if (event.type === \"compaction_start\") {\n      log.logInfo(`Auto-compaction started (reason: ${(event as any).reason})`);\n      queue.enqueue(() => responseCtx.respond(\"_Compacting context..._\"), \"compaction start\");\n    } else if (event.type === \"compaction_end\") {\n      const compEvent = event as any;\n      if (compEvent.result) {\n        log.logInfo(`Auto-compaction complete: ${compEvent.result.tokensBefore} tokens compacted`);\n      } else if (compEvent.aborted) {\n        log.logInfo(\"Auto-compaction aborted\");\n      }\n    } else if (event.type === \"auto_retry_start\") {\n      const retryEvent = event as any;\n      log.logWarning(\n        `Retrying (${retryEvent.attempt}/${retryEvent.maxAttempts})`,\n        retryEvent.errorMessage,\n      );\n      queue.enqueue(\n        () =>\n          responseCtx.respond(`_Retrying (${retryEvent.attempt}/${retryEvent.maxAttempts})..._`),\n        \"retry\",\n      );\n    }\n  });\n\n  // Message limit constant\n  const SLACK_MAX_LENGTH = 40000;\n  const splitForSlack = (text: string): string[] => {\n    if (text.length <= SLACK_MAX_LENGTH) return [text];\n    const parts: string[] = [];\n    let remaining = text;\n    let partNum = 1;\n    while (remaining.length > 0) {\n      const chunk = remaining.substring(0, SLACK_MAX_LENGTH - 50);\n      remaining = remaining.substring(SLACK_MAX_LENGTH - 50);\n      const suffix = remaining.length > 0 ? `\\n_(continued ${partNum}...)_` : \"\";\n      parts.push(chunk + suffix);\n      partNum++;\n    }\n    return parts;\n  };\n\n  return {\n    async run(\n      message: ChatMessage,\n      responseCtx: ChatResponseContext,\n      platform: PlatformInfo,\n    ): Promise<{ stopReason: string; errorMessage?: string }> {\n      // Extract channelId from sessionKey (format: \"channelId:rootTs\" or just \"channelId\")\n      const sessionChannel = message.sessionKey.split(\":\")[0];\n\n      // Ensure channel directory exists\n      await mkdir(channelDir, { recursive: true });\n\n      // Sync messages from log.jsonl that arrived while we were offline or busy\n      // Exclude the current message (it will be added via prompt())\n      // Default sync range is 10 days (handled by syncLogToSessionManager)\n      // Thread filter ensures only messages from this session's thread are synced\n      const threadFilter = message.sessionKey.includes(\":\")\n        ? { scope: \"thread\" as const, rootTs, threadTs: message.threadTs }\n        : { scope: \"top-level\" as const, rootTs };\n      const syncedCount = await syncLogToSessionManager(\n        sessionManager,\n        channelDir,\n        message.id,\n        undefined,\n        threadFilter,\n      );\n      if (syncedCount > 0) {\n        log.logInfo(`[${channelId}] Synced ${syncedCount} messages from log.jsonl`);\n      }\n\n      // Reload messages from context.jsonl\n      // This picks up any messages synced above\n      const reloadedSession = sessionManager.buildSessionContext();\n      if (reloadedSession.messages.length > 0) {\n        agent.replaceMessages(reloadedSession.messages);\n        log.logInfo(\n          `[${channelId}] Reloaded ${reloadedSession.messages.length} messages from context`,\n        );\n      }\n\n      // Update system prompt with fresh memory, channel/user info, and skills\n      const memory = await getMemory(channelDir);\n      const skills = loadMamaSkills(channelDir, workspacePath);\n      const systemPrompt = buildSystemPrompt(\n        workspacePath,\n        channelId,\n        memory,\n        sandboxConfig,\n        platform,\n        skills,\n      );\n      session.agent.setSystemPrompt(systemPrompt);\n\n      // Set up file upload function\n      setUploadFunction(async (filePath: string, title?: string) => {\n        const hostPath = translateToHostPath(filePath, channelDir, workspacePath, channelId);\n        await responseCtx.uploadFile(hostPath, title);\n      });\n\n      // Reset per-run state\n      runState.responseCtx = responseCtx;\n      runState.logCtx = {\n        channelId: sessionChannel,\n        userName: message.userName,\n        channelName: undefined,\n        sessionId: sessionUuid,\n      };\n      runState.pendingTools.clear();\n      runState.totalUsage = {\n        input: 0,\n        output: 0,\n        cacheRead: 0,\n        cacheWrite: 0,\n        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n      };\n      runState.llmCallCount = 0;\n      runState.stopReason = \"stop\";\n      runState.errorMessage = undefined;\n\n      // Create queue for this run\n      let queueChain = Promise.resolve();\n      runState.queue = {\n        enqueue(fn: () => Promise<void>, errorContext: string): void {\n          queueChain = queueChain.then(async () => {\n            try {\n              await fn();\n            } catch (err) {\n              const errMsg = err instanceof Error ? err.message : String(err);\n              log.logWarning(`API error (${errorContext})`, errMsg);\n              try {\n                // Split long error messages to avoid msg_too_long\n                const errParts = splitForSlack(`_Error: ${errMsg}_`);\n                for (const part of errParts) {\n                  await responseCtx.respondInThread(part);\n                }\n              } catch {\n                // Ignore\n              }\n            }\n          });\n        },\n        enqueueMessage(\n          text: string,\n          target: \"main\" | \"thread\",\n          errorContext: string,\n          _doLog = true,\n        ): void {\n          const parts = splitForSlack(text);\n          for (const part of parts) {\n            this.enqueue(\n              () =>\n                target === \"main\" ? responseCtx.respond(part) : responseCtx.respondInThread(part),\n              errorContext,\n            );\n          }\n        },\n      };\n\n      // Log context info\n      log.logInfo(\n        `Context sizes - system: ${systemPrompt.length} chars, memory: ${memory.length} chars`,\n      );\n      log.logInfo(`Channels: ${platform.channels.length}, Users: ${platform.users.length}`);\n\n      // Build user message with timestamp and username prefix\n      // Format: \"[YYYY-MM-DD HH:MM:SS+HH:MM] [username]: message\" so LLM knows when and who\n      const now = new Date();\n      const pad = (n: number) => n.toString().padStart(2, \"0\");\n      const offset = -now.getTimezoneOffset();\n      const offsetSign = offset >= 0 ? \"+\" : \"-\";\n      const offsetHours = pad(Math.floor(Math.abs(offset) / 60));\n      const offsetMins = pad(Math.abs(offset) % 60);\n      const timestamp = `${now.getFullYear()}-${pad(now.getMonth() + 1)}-${pad(now.getDate())} ${pad(now.getHours())}:${pad(now.getMinutes())}:${pad(now.getSeconds())}${offsetSign}${offsetHours}:${offsetMins}`;\n      const threadContext = message.threadTs ? ` [in-thread:${message.threadTs}]` : \"\";\n      let userMessage = `[${timestamp}] [${message.userName || \"unknown\"}]${threadContext}: ${message.text}`;\n\n      const imageAttachments: ImageContent[] = [];\n      const nonImagePaths: string[] = [];\n\n      for (const a of message.attachments || []) {\n        // a.localPath is the path relative to the workspace (same as old a.local)\n        const fullPath = `${workspacePath}/${a.localPath}`;\n        const mimeType = getImageMimeType(a.localPath);\n\n        if (mimeType && existsSync(fullPath)) {\n          try {\n            imageAttachments.push({\n              type: \"image\",\n              mimeType,\n              data: readFileSync(fullPath).toString(\"base64\"),\n            });\n          } catch {\n            nonImagePaths.push(fullPath);\n          }\n        } else {\n          nonImagePaths.push(fullPath);\n        }\n      }\n\n      if (nonImagePaths.length > 0) {\n        userMessage += `\\n\\n<slack_attachments>\\n${nonImagePaths.join(\"\\n\")}\\n</slack_attachments>`;\n      }\n\n      // Debug: write context to last_prompt.jsonl\n      const debugContext = {\n        systemPrompt,\n        messages: session.messages,\n        newUserMessage: userMessage,\n        imageAttachmentCount: imageAttachments.length,\n      };\n      await writeFile(join(channelDir, \"last_prompt.jsonl\"), JSON.stringify(debugContext, null, 2));\n      addLifecycleBreadcrumb(\"agent.prompt.sent\", {\n        provider: model.provider,\n        model: agentConfig.model,\n        channel_id: sessionChannel,\n        session_id: sessionUuid,\n        attachment_count: message.attachments?.length ?? 0,\n        image_attachment_count: imageAttachments.length,\n      });\n\n      await session.prompt(\n        userMessage,\n        imageAttachments.length > 0 ? { images: imageAttachments } : undefined,\n      );\n\n      // Wait for queued messages\n      await queueChain;\n\n      // Handle error case - update main message and post error to thread\n      if (runState.stopReason === \"error\" && runState.errorMessage) {\n        try {\n          await responseCtx.replaceResponse(\"_Sorry, something went wrong_\");\n          // Split long error messages to avoid msg_too_long\n          const errorParts = splitForSlack(`_Error: ${runState.errorMessage}_`);\n          for (const part of errorParts) {\n            await responseCtx.respondInThread(part);\n          }\n        } catch (err) {\n          const errMsg = err instanceof Error ? err.message : String(err);\n          log.logWarning(\"Failed to post error message\", errMsg);\n        }\n      } else {\n        // Final message update\n        const messages = session.messages;\n        const lastAssistant = messages.filter((m) => m.role === \"assistant\").pop();\n        const finalText =\n          lastAssistant?.content\n            .filter((c): c is { type: \"text\"; text: string } => c.type === \"text\")\n            .map((c) => c.text)\n            .join(\"\\n\") || \"\";\n\n        // Check for [SILENT] marker - delete message and thread instead of posting\n        if (finalText.trim() === \"[SILENT]\" || finalText.trim().startsWith(\"[SILENT]\")) {\n          try {\n            await responseCtx.deleteResponse();\n            log.logInfo(\"Silent response - deleted message and thread\");\n          } catch (err) {\n            const errMsg = err instanceof Error ? err.message : String(err);\n            log.logWarning(\"Failed to delete message for silent response\", errMsg);\n          }\n        } else if (finalText.trim()) {\n          try {\n            const mainText =\n              finalText.length > SLACK_MAX_LENGTH\n                ? `${finalText.substring(0, SLACK_MAX_LENGTH - 50)}\\n\\n_(see thread for full response)_`\n                : finalText;\n            await responseCtx.replaceResponse(mainText);\n          } catch (err) {\n            const errMsg = err instanceof Error ? err.message : String(err);\n            log.logWarning(\"Failed to replace message with final text\", errMsg);\n          }\n        }\n      }\n\n      // Log usage summary with context info\n      if (runState.totalUsage.cost.total > 0) {\n        // Get last non-aborted assistant message for context calculation\n        const messages = session.messages;\n        const lastAssistantMessage = messages\n          .slice()\n          .reverse()\n          .find((m) => m.role === \"assistant\" && (m as any).stopReason !== \"aborted\") as any;\n\n        const contextTokens = lastAssistantMessage\n          ? lastAssistantMessage.usage.input +\n            lastAssistantMessage.usage.output +\n            lastAssistantMessage.usage.cacheRead +\n            lastAssistantMessage.usage.cacheWrite\n          : 0;\n        const contextWindow = model.contextWindow || 200000;\n\n        // Run-level Sentry metrics\n        const { totalUsage } = runState;\n        const runMetricAttributes = metricAttributes({\n          provider: model.provider,\n          model: agentConfig.model,\n          channel_id: sessionChannel,\n          session_id: sessionUuid,\n          stop_reason: runState.stopReason,\n          llm_calls: runState.llmCallCount,\n        });\n        Sentry.metrics.distribution(\"agent.run.tokens_in\", totalUsage.input, {\n          attributes: runMetricAttributes,\n        });\n        Sentry.metrics.distribution(\"agent.run.tokens_out\", totalUsage.output, {\n          attributes: runMetricAttributes,\n        });\n        Sentry.metrics.distribution(\"agent.run.cache_read\", totalUsage.cacheRead, {\n          attributes: runMetricAttributes,\n        });\n        Sentry.metrics.distribution(\"agent.run.cache_write\", totalUsage.cacheWrite, {\n          attributes: runMetricAttributes,\n        });\n        Sentry.metrics.distribution(\"agent.run.cost\", totalUsage.cost.total, {\n          attributes: runMetricAttributes,\n        });\n        Sentry.metrics.gauge(\"agent.context.utilization\", contextTokens / contextWindow, {\n          unit: \"ratio\",\n          attributes: runMetricAttributes,\n        });\n\n        const summary = log.logUsageSummary(\n          runState.logCtx!,\n          runState.totalUsage,\n          contextTokens,\n          contextWindow,\n        );\n        // Split long summaries to avoid msg_too_long\n        const summaryParts = splitForSlack(summary);\n        for (const part of summaryParts) {\n          runState.queue!.enqueue(\n            () => responseCtx.respondInThread(part, { style: \"muted\" }),\n            \"usage summary\",\n          );\n        }\n        await queueChain;\n      }\n\n      // Clear run state\n      runState.responseCtx = null;\n      runState.logCtx = null;\n      runState.queue = null;\n\n      return { stopReason: runState.stopReason, errorMessage: runState.errorMessage };\n    },\n\n    abort(): void {\n      session.abort();\n    },\n\n    getCurrentStep(): { toolName?: string; label?: string } | undefined {\n      const pending = runState.pendingTools;\n      if (pending.size === 0) return undefined;\n      // Get the first pending tool\n      const first = pending.values().next().value;\n      if (!first) return undefined;\n      return {\n        toolName: first.toolName,\n        label: (first.args as { label?: string })?.label,\n      };\n    },\n  };\n}\n\n/**\n * Translate container path back to host path for file operations\n */\nfunction translateToHostPath(\n  containerPath: string,\n  channelDir: string,\n  workspacePath: string,\n  channelId: string,\n): string {\n  if (workspacePath === \"/workspace\") {\n    const prefix = `/workspace/${channelId}/`;\n    if (containerPath.startsWith(prefix)) {\n      return join(channelDir, containerPath.slice(prefix.length));\n    }\n    if (containerPath.startsWith(\"/workspace/\")) {\n      return join(channelDir, \"..\", containerPath.slice(\"/workspace/\".length));\n    }\n  }\n  return containerPath;\n}\n"]}

package/dist/agent.js

@@ -1,7 +1,7 @@
 import { Agent } from "@mariozechner/pi-agent-core";
 import { getModel } from "@mariozechner/pi-ai";
-import { AgentSession, AuthStorage, convertToLlm, createExtensionRuntime, formatSkillsForPrompt, loadSkillsFromDir, ModelRegistry, SessionManager, } from "@mariozechner/pi-coding-agent";
-import { existsSync, mkdirSync, readFileSync } from "fs";
+import { AgentSession, AuthStorage, convertToLlm, DefaultResourceLoader, formatSkillsForPrompt, loadSkillsFromDir, ModelRegistry, } from "@mariozechner/pi-coding-agent";
+import { existsSync, readFileSync } from "fs";
 import { mkdir, readFile, writeFile } from "fs/promises";
 import { homedir } from "os";
 import { join } from "path";
@@ -9,7 +9,10 @@ import { loadAgentConfig } from "./config.js";
 import { createMamaSettingsManager, syncLogToSessionManager } from "./context.js";
 import * as log from "./log.js";
 import { createExecutor } from "./sandbox.js";
+import { addLifecycleBreadcrumb, metricAttributes } from "./sentry.js";
+import { createManagedSessionFileAtPath, extractSessionSuffix, extractSessionUuid, forkThreadSessionFile, getSessionDir, getThreadSessionFile, openManagedSession, resolveChannelSessionFile, resolveManagedSessionFile, tryResolveThreadSession, } from "./session-store.js";
 import { createMamaTools } from "./tools/index.js";
+import * as Sentry from "@sentry/node";
 const IMAGE_MIME_TYPES = {
     jpg: "image/jpeg",
     jpeg: "image/jpeg",
@@ -361,12 +364,45 @@ export async function createRunner(sandboxConfig, sessionKey, channelId, channel
     };
     const systemPrompt = buildSystemPrompt(workspacePath, channelId, memory, sandboxConfig, emptyPlatform, skills);
     // Create session manager and settings manager
-    // Per-session context file: {channelDir}/sessions/{rootTs}/context.jsonl
-    const rootTs = sessionKey.includes(":") ? sessionKey.split(":").pop() : sessionKey;
-    const sessionDir = join(channelDir, "sessions", rootTs);
-    mkdirSync(sessionDir, { recursive: true });
-    const contextFile = join(sessionDir, "context.jsonl");
-    const sessionManager = SessionManager.open(contextFile, channelDir);
+    // Channel sessions use {channelDir}/sessions/current.
+    // Thread sessions use fixed files: {channelDir}/sessions/{threadTs}.jsonl
+    const sessionDir = getSessionDir(channelDir, sessionKey);
+    const isThread = sessionKey.includes(":");
+    let sessionManager;
+    let contextFile;
+    if (isThread) {
+        const threadFile = getThreadSessionFile(channelDir, sessionKey);
+        const existing = tryResolveThreadSession(threadFile);
+        if (existing) {
+            contextFile = existing;
+            sessionManager = openManagedSession(contextFile, sessionDir, channelDir);
+        }
+        else {
+            const channelSource = resolveChannelSessionFile(channelDir);
+            if (channelSource) {
+                try {
+                    contextFile = forkThreadSessionFile(channelSource, threadFile, channelDir);
+                    sessionManager = openManagedSession(contextFile, sessionDir, channelDir);
+                }
+                catch {
+                    contextFile = createManagedSessionFileAtPath(threadFile, channelDir);
+                    sessionManager = openManagedSession(contextFile, sessionDir, channelDir);
+                }
+            }
+            else {
+                contextFile = createManagedSessionFileAtPath(threadFile, channelDir);
+                sessionManager = openManagedSession(contextFile, sessionDir, channelDir);
+            }
+        }
+    }
+    else {
+        // Channel/DM session: normal resolve
+        contextFile = resolveManagedSessionFile(sessionDir, channelDir);
+        sessionManager = openManagedSession(contextFile, sessionDir, channelDir);
+    }
+    const sessionUuid = extractSessionUuid(contextFile);
+    // Used for Slack thread filtering — for non-Slack platforms this is effectively a no-op
+    const rootTs = extractSessionSuffix(sessionKey);
     const settingsManager = createMamaSettingsManager(join(channelDir, ".."));
     // Create AuthStorage and ModelRegistry
     // Auth stored outside workspace so agent can't access it
@@ -394,24 +430,33 @@ export async function createRunner(sandboxConfig, sessionKey, channelId, channel
         agent.replaceMessages(loadedSession.messages);
         log.logInfo(`[${channelId}] Loaded ${loadedSession.messages.length} messages from context.jsonl`);
     }
-    const resourceLoader = {
-        getExtensions: () => ({ extensions: [], errors: [], runtime: createExtensionRuntime() }),
-        getSkills: () => ({ skills: [], diagnostics: [] }),
-        getPrompts: () => ({ prompts: [], diagnostics: [] }),
-        getThemes: () => ({ themes: [], diagnostics: [] }),
-        getAgentsFiles: () => ({ agentsFiles: [] }),
-        getSystemPrompt: () => systemPrompt,
-        getAppendSystemPrompt: () => [],
-        extendResources: () => { },
-        reload: async () => { },
-    };
+    // Load extensions, skills, prompts, themes via DefaultResourceLoader
+    // This reads ~/.pi/agent/settings.json (packages, extensions enable/disable)
+    // and discovers resources from standard locations + npm/git packages.
+    const resourceLoader = new DefaultResourceLoader({
+        cwd: workspaceDir,
+        systemPrompt,
+    });
+    try {
+        await resourceLoader.reload();
+        const extResult = resourceLoader.getExtensions();
+        if (extResult.errors.length > 0) {
+            for (const err of extResult.errors) {
+                log.logWarning(`[${channelId}] Extension load error: ${err.path}`, err.error);
+            }
+        }
+        log.logInfo(`[${channelId}] Loaded ${extResult.extensions.length} extension(s): ${extResult.extensions.map((e) => e.path).join(", ")}`);
+    }
+    catch (error) {
+        log.logWarning(`[${channelId}] Failed to load resources`, String(error));
+    }
     const baseToolsOverride = Object.fromEntries(tools.map((tool) => [tool.name, tool]));
     // Create AgentSession wrapper
     const session = new AgentSession({
         agent,
         sessionManager,
         settingsManager,
-        cwd: process.cwd(),
+        cwd: workspaceDir,
         modelRegistry,
         resourceLoader,
         baseToolsOverride,
@@ -429,6 +474,7 @@ export async function createRunner(sandboxConfig, sessionKey, channelId, channel
             cacheWrite: 0,
             cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
         },
+        llmCallCount: 0,
         stopReason: "stop",
         errorMessage: undefined,
     };
@@ -438,6 +484,7 @@ export async function createRunner(sandboxConfig, sessionKey, channelId, channel
         if (!runState.responseCtx || !runState.logCtx || !runState.queue)
             return;
         const { responseCtx, logCtx, queue, pendingTools } = runState;
+        const baseAttrs = { channel_id: logCtx.channelId, session_id: logCtx.sessionId };
         if (event.type === "tool_execution_start") {
             const agentEvent = event;
             const args = agentEvent.args;
@@ -447,6 +494,10 @@ export async function createRunner(sandboxConfig, sessionKey, channelId, channel
                 args: agentEvent.args,
                 startTime: Date.now(),
             });
+            addLifecycleBreadcrumb("agent.tool.started", {
+                tool: agentEvent.toolName,
+                ...baseAttrs,
+            });
             log.logToolStart(logCtx, agentEvent.toolName, label, agentEvent.args);
             // Tool labels are omitted from the main message to reduce Slack noise.
             // Tool execution details are still posted to the thread (see tool_execution_end).
@@ -457,6 +508,26 @@ export async function createRunner(sandboxConfig, sessionKey, channelId, channel
             const pending = pendingTools.get(agentEvent.toolCallId);
             pendingTools.delete(agentEvent.toolCallId);
             const durationMs = pending ? Date.now() - pending.startTime : 0;
+            Sentry.metrics.count("agent.tool.calls", 1, {
+                attributes: metricAttributes({
+                    tool: agentEvent.toolName,
+                    error: String(agentEvent.isError),
+                    ...baseAttrs,
+                }),
+            });
+            Sentry.metrics.distribution("agent.tool.duration", durationMs, {
+                unit: "millisecond",
+                attributes: metricAttributes({
+                    tool: agentEvent.toolName,
+                    ...baseAttrs,
+                }),
+            });
+            addLifecycleBreadcrumb("agent.tool.completed", {
+                tool: agentEvent.toolName,
+                error: agentEvent.isError,
+                duration_ms: durationMs,
+                ...baseAttrs,
+            });
             if (agentEvent.isError) {
                 log.logToolError(logCtx, agentEvent.toolName, durationMs, resultStr);
             }
@@ -489,6 +560,13 @@ export async function createRunner(sandboxConfig, sessionKey, channelId, channel
         else if (event.type === "message_start") {
             const agentEvent = event;
             if (agentEvent.message.role === "assistant") {
+                runState.llmCallCount += 1;
+                addLifecycleBreadcrumb("agent.llm.call.started", {
+                    call_index: runState.llmCallCount,
+                    provider: model.provider,
+                    model: agentConfig.model,
+                    ...baseAttrs,
+                });
                 log.logResponseStart(logCtx);
             }
         }
@@ -512,6 +590,44 @@ export async function createRunner(sandboxConfig, sessionKey, channelId, channel
                     runState.totalUsage.cost.cacheRead += assistantMsg.usage.cost.cacheRead;
                     runState.totalUsage.cost.cacheWrite += assistantMsg.usage.cost.cacheWrite;
                     runState.totalUsage.cost.total += assistantMsg.usage.cost.total;
+                    // Per-turn LLM metrics
+                    const llmAttributes = metricAttributes({
+                        provider: model.provider,
+                        model: agentConfig.model,
+                        ...baseAttrs,
+                        stop_reason: assistantMsg.stopReason,
+                        error: Boolean(assistantMsg.errorMessage),
+                    });
+                    Sentry.metrics.count("agent.llm.calls", 1, { attributes: llmAttributes });
+                    Sentry.metrics.distribution("agent.llm.tokens_in", assistantMsg.usage.input, {
+                        attributes: llmAttributes,
+                    });
+                    Sentry.metrics.distribution("agent.llm.tokens_out", assistantMsg.usage.output, {
+                        attributes: llmAttributes,
+                    });
+                    if (assistantMsg.usage.cacheRead > 0) {
+                        Sentry.metrics.distribution("agent.llm.cache_read", assistantMsg.usage.cacheRead, {
+                            attributes: llmAttributes,
+                        });
+                    }
+                    if (assistantMsg.usage.cacheWrite > 0) {
+                        Sentry.metrics.distribution("agent.llm.cache_write", assistantMsg.usage.cacheWrite, {
+                            attributes: llmAttributes,
+                        });
+                    }
+                    Sentry.metrics.distribution("agent.llm.cost_per_turn", assistantMsg.usage.cost.total, {
+                        attributes: llmAttributes,
+                    });
+                    addLifecycleBreadcrumb("agent.llm.call.completed", {
+                        call_index: runState.llmCallCount,
+                        provider: model.provider,
+                        model: agentConfig.model,
+                        stop_reason: assistantMsg.stopReason,
+                        error: Boolean(assistantMsg.errorMessage),
+                        input_tokens: assistantMsg.usage.input,
+                        output_tokens: assistantMsg.usage.output,
+                        cost_total_usd: assistantMsg.usage.cost.total,
+                    });
                 }
                 const content = agentEvent.message.content;
                 const thinkingParts = [];
@@ -586,7 +702,10 @@ export async function createRunner(sandboxConfig, sessionKey, channelId, channel
             // Exclude the current message (it will be added via prompt())
             // Default sync range is 10 days (handled by syncLogToSessionManager)
             // Thread filter ensures only messages from this session's thread are synced
-            const syncedCount = await syncLogToSessionManager(sessionManager, channelDir, message.id, undefined, { rootTs, threadTs: message.threadTs });
+            const threadFilter = message.sessionKey.includes(":")
+                ? { scope: "thread", rootTs, threadTs: message.threadTs }
+                : { scope: "top-level", rootTs };
+            const syncedCount = await syncLogToSessionManager(sessionManager, channelDir, message.id, undefined, threadFilter);
             if (syncedCount > 0) {
                 log.logInfo(`[${channelId}] Synced ${syncedCount} messages from log.jsonl`);
             }
@@ -613,6 +732,7 @@ export async function createRunner(sandboxConfig, sessionKey, channelId, channel
                 channelId: sessionChannel,
                 userName: message.userName,
                 channelName: undefined,
+                sessionId: sessionUuid,
             };
             runState.pendingTools.clear();
             runState.totalUsage = {
@@ -622,6 +742,7 @@ export async function createRunner(sandboxConfig, sessionKey, channelId, channel
                 cacheWrite: 0,
                 cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
             };
+            runState.llmCallCount = 0;
             runState.stopReason = "stop";
             runState.errorMessage = undefined;
             // Create queue for this run
@@ -702,6 +823,14 @@ export async function createRunner(sandboxConfig, sessionKey, channelId, channel
                 imageAttachmentCount: imageAttachments.length,
             };
             await writeFile(join(channelDir, "last_prompt.jsonl"), JSON.stringify(debugContext, null, 2));
+            addLifecycleBreadcrumb("agent.prompt.sent", {
+                provider: model.provider,
+                model: agentConfig.model,
+                channel_id: sessionChannel,
+                session_id: sessionUuid,
+                attachment_count: message.attachments?.length ?? 0,
+                image_attachment_count: imageAttachments.length,
+            });
             await session.prompt(userMessage, imageAttachments.length > 0 ? { images: imageAttachments } : undefined);
             // Wait for queued messages
             await queueChain;
@@ -767,6 +896,35 @@ export async function createRunner(sandboxConfig, sessionKey, channelId, channel
                         lastAssistantMessage.usage.cacheWrite
                     : 0;
                 const contextWindow = model.contextWindow || 200000;
+                // Run-level Sentry metrics
+                const { totalUsage } = runState;
+                const runMetricAttributes = metricAttributes({
+                    provider: model.provider,
+                    model: agentConfig.model,
+                    channel_id: sessionChannel,
+                    session_id: sessionUuid,
+                    stop_reason: runState.stopReason,
+                    llm_calls: runState.llmCallCount,
+                });
+                Sentry.metrics.distribution("agent.run.tokens_in", totalUsage.input, {
+                    attributes: runMetricAttributes,
+                });
+                Sentry.metrics.distribution("agent.run.tokens_out", totalUsage.output, {
+                    attributes: runMetricAttributes,
+                });
+                Sentry.metrics.distribution("agent.run.cache_read", totalUsage.cacheRead, {
+                    attributes: runMetricAttributes,
+                });
+                Sentry.metrics.distribution("agent.run.cache_write", totalUsage.cacheWrite, {
+                    attributes: runMetricAttributes,
+                });
+                Sentry.metrics.distribution("agent.run.cost", totalUsage.cost.total, {
+                    attributes: runMetricAttributes,
+                });
+                Sentry.metrics.gauge("agent.context.utilization", contextTokens / contextWindow, {
+                    unit: "ratio",
+                    attributes: runMetricAttributes,
+                });
                 const summary = log.logUsageSummary(runState.logCtx, runState.totalUsage, contextTokens, contextWindow);
                 // Split long summaries to avoid msg_too_long
                 const summaryParts = splitForSlack(summary);