camelagi 0.5.0 → 0.5.17

package/LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2026 CamelAGI Contributors
+Copyright (c) 2026 Nawaf Almutairi & Brickell Lab, Kuwait
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

package/README.md

@@ -16,27 +16,186 @@
 
 <br>
 
-CamelAGI is a self-hosted AI assistant that runs on your server and puts you in full control from your phone. One command to set up, then manage everything from Telegram — no terminal needed.
+CamelAGI is a self-hosted AI assistant that runs on your server and puts you in full control from your phone. One command to install, then manage everything — create agents, switch models, approve tools, monitor usage — all from Telegram, Discord, or your terminal.
 
 <p align="center">
-  <a href="https://camelagi.net"><strong>🌐 CamelAGI.net</strong></a>
+  <img src="assets/Claude_Logo.png" alt="Claude Logo" width="200" />
+</p>
+<p align="center">
+Powered by <a href="https://platform.claude.com/docs/en/agent-sdk/overview">Claude Agent SDK</a> — the same runtime behind Claude Code.
+</p>
+
+<p align="center">
+  <a href="https://camelagi.net"><strong>camelagi.net</strong></a>
 </p>
 
 ## Contents
 
-- [Built on Claude Agent SDK](#built-on-claude-agent-sdk)
-- [Admin Bot — BotFather for Your AI Server](#admin-bot--botfather-for-your-ai-server)
-- [Claude Agent SDK vs pi-agent](#claude-agent-sdk-vs-pi-agent)
 - [Quick Start](#quick-start)
 - [Features](#features)
-- [Terminal UI — `camel chat`](#terminal-ui--camel-chat)
-- [Developer Experience](#developer-experience)
-- [Roadmap](#roadmap)
+- [Channels](#channels)
+- [Admin Bot Commands](#admin-bot-commands)
+- [Agent Bot Commands](#agent-bot-commands)
+- [CLI Commands](#cli-commands)
+- [Built on Claude Agent SDK](#built-on-claude-agent-sdk)
+- [Claude Agent SDK vs pi-agent](#claude-agent-sdk-vs-pi-agent)
+- [Configuration](#configuration)
+- [Architecture](#architecture)
 - [Documentation](#documentation)
+- [Roadmap](#roadmap)
 - [License](#license)
 
 <br>
 
+## Quick Start
+
+**Install** (no Node.js required)
+```bash
+curl -fsSL https://raw.githubusercontent.com/inawafalm/CamelAGI/main/install.sh | bash
+```
+
+**Setup & Run**
+```bash
+camel bootstrap     # Full setup: admin bot + pairing + API config
+camel serve         # Start the gateway server
+camel chat          # Terminal UI
+```
+
+`bootstrap` walks you through everything: create a Telegram admin bot, pair your account with OTP, pick your AI provider and model. After that, use `/newagent` in Telegram to create your first AI agent.
+
+<br>
+
+## Features
+
+> 3 channels, 18 CLI commands, 10 built-in tools — terminal, Telegram, or Discord.
+
+| | Feature | Description |
+|---|---|---|
+| 🤖 | **Telegram — Admin Bot** | BotFather for AI agents. Create, configure, clone, and manage agents entirely from Telegram — instant commands, zero tokens |
+| 💬 | **Telegram — Agent Bots** | Each agent gets its own Telegram bot. Message it like any chat — it runs tools, reads files, remembers context |
+| 🎮 | **Discord Bots** | Per-agent Discord bots with mention-only mode, role filtering, and channel restrictions |
+| ⌨️ | **Terminal UI** | Full TUI with streaming, slash commands, model switching, session management, and markdown rendering |
+| 🧠 | **Agent Memory** | Isolated two-tier memory per agent — curated MEMORY.md + daily auto-journaling with recency-boosted search |
+| 🎙️ | **Voice Transcription** | Send voice messages to agent bots — transcribed via Groq, OpenAI, or Deepgram and processed as text |
+| 🔌 | **MCP Servers** | Connect external tool servers (stdio, HTTP, SSE). Global or per-agent. Add/remove from Telegram with `/mcp` |
+| 💭 | **Extended Thinking** | Claude reasons step by step before answering. Configure depth: off, low, medium, high |
+| ⏰ | **Cron Jobs** | Schedule AI tasks — daily summaries, monitoring, automations. Intervals, cron expressions, or one-shot timers |
+| 📋 | **Brief Mode** | Toggle short text-message-style replies per chat or per agent — ideal for Telegram |
+| 📊 | **Usage Tracking** | Per-agent token usage and cost breakdown — input, output, cache reads, API calls |
+| 🧬 | **Agent Cloning** | Clone an existing agent with all its config, personality, memory, and MCP servers |
+| 🔐 | **Secure Pairing** | OTP-based user verification. No hardcoded IDs — pairing code + 5-digit OTP from Telegram |
+| 🔁 | **Multi-Provider** | Anthropic, OpenAI, OpenRouter, Ollama — any OpenAI-compatible endpoint. Zero vendor lock-in |
+| 🛡️ | **Tool Approvals** | Human-in-the-loop safety. Approve dangerous operations from Telegram with inline buttons |
+| 🪝 | **Skills & Hooks** | Teach agents skills via Markdown. Run shell/JS hooks before and after tool calls |
+| 🔄 | **Auto Compaction** | Summarizes old turns at 80% capacity. Flushes facts to memory first — nothing is lost |
+| ⚙️ | **Same Engine** | All channels run the same agent loop, same tools, same memory. Switch freely between them |
+
+<br>
+
+## Channels
+
+CamelAGI runs across three channels — all sharing the same agent runtime, tools, and memory.
+
+### Telegram
+
+Two bot types work together:
+
+- **Admin Bot** — A non-AI command bot (zero tokens burned). Create agents, manage config, approve users, monitor sessions, configure MCP servers, set up voice — all from Telegram. Think [@BotFather](https://t.me/BotFather), but for your entire AI infrastructure.
+- **Agent Bots** — Each agent gets its own Telegram bot. Message it like any contact — it runs tools, reads files, remembers context, and supports voice messages.
+
+### Discord
+
+Per-agent Discord bots with:
+- **Mention-only mode** — responds only to @mentions in guild channels, all messages in DMs
+- **Role filtering** — optional allowlist of Discord roles
+- **Channel restrictions** — optional allowlist of channels
+
+### Terminal
+
+`camel chat` gives you a full TUI with streaming, slash commands, keyboard shortcuts, model switching, session management, and one-shot mode (`camel "your question"`).
+
+<br>
+
+## Admin Bot Commands
+
+The Admin Bot is a non-AI Telegram bot for managing your entire CamelAGI server.
+
+| Category | Command | Description |
+|----------|---------|-------------|
+| **Setup** | `/setup` | Configure API provider, key, model |
+| | `/config` | View configuration |
+| | `/config <key> <value>` | Update config |
+| **Agents** | `/newagent` | Create agent wizard |
+| | `/agents` | List all agents |
+| | `/agent` | View/edit agent config |
+| | `/deleteagent` | Delete an agent |
+| | `/soul` | View/edit agent personality |
+| **Tools** | `/mcp` | Manage MCP servers (add/list/remove) |
+| | `/voice` | Configure voice transcription provider |
+| **Monitor** | `/status` | System health & stats |
+| | `/sessions` | List & manage sessions |
+| | `/usage` | Per-agent usage & cost summary |
+| | `/restart` | Restart agent bots |
+| **Security** | `/pairing` | Manage access requests |
+| **Utility** | `/help` | List all commands |
+| | `/cancel` | Cancel active wizard |
+
+<br>
+
+## Agent Bot Commands
+
+Each agent bot supports these commands in Telegram:
+
+| Category | Command | Description |
+|----------|---------|-------------|
+| **Chat** | `/clear` | Clear this chat's history |
+| | `/compact` | Force compaction of chat history |
+| | `/brief` | Toggle brief response mode |
+| | `/export` | Export session as markdown file |
+| **Model** | `/model` | Switch model for this chat |
+| | `/think` | Set thinking level (off/low/medium/high) |
+| | `/effort` | Set effort level (low/medium/high/max) |
+| **Session** | `/session` | Show or switch session |
+| | `/status` | Show model, message count, token usage |
+| | `/usage` | Token usage for this session |
+| **Tools** | `/skills` | List active skills |
+| | `/mcp` | Manage MCP tool servers |
+| | `/voice` | Voice transcription info |
+| **Utility** | `/help` | List commands and current config |
+
+<br>
+
+## CLI Commands
+
+```
+camel <command> [options]
+```
+
+| Category | Command | Description |
+|----------|---------|-------------|
+| **Getting Started** | `bootstrap` | First-time setup wizard |
+| | `setup` | Interactive setup (re-run anytime) |
+| | `chat` | Interactive terminal UI |
+| **Server** | `serve` | Start the gateway server |
+| | `daemon` | Manage launchd daemon (install/uninstall/status) |
+| | `logs` | Tail server request log |
+| | `status` | System health overview |
+| **Agents & Sessions** | `agents` | List configured agents |
+| | `soul` | View/edit agent SOUL.md in $EDITOR |
+| | `sessions` | List saved sessions |
+| | `pairing` | List and approve/deny pending requests |
+| **Configuration** | `config` | View/edit config (get/set/list) |
+| | `cron` | Manage cron jobs (list/add/rm/run) |
+| **Maintenance** | `doctor` | Run health checks |
+| | `reset` | Delete all config, sessions, agents |
+| | `install` | Install to ~/.camelagi/versions/ and add to PATH |
+| | `uninstall` | Remove CamelAGI completely |
+| | `update` | Update to the latest version |
+
+One-shot mode: `camel "your question"` — spins up an ephemeral gateway and answers inline.
+
+<br>
+
 ## Built on Claude Agent SDK
 
 <p align="center">
@@ -99,123 +258,101 @@ CamelAGI uses **Claude Agent SDK**. OpenClaw uses **pi-agent-core**. Here's why
 
 <br>
 
-## Quick Start
-
-**One-Liner**
-```bash
-curl -fsSL https://camelAGI.ai/install.sh | bash
+## Configuration
+
+CamelAGI uses a single YAML config file at `~/.camelagi/config.yaml`, validated with Zod.
+
+```yaml
+# Provider & Model
+provider: anthropic          # anthropic | openai | openrouter | ollama
+model: claude-sonnet-4-20250514
+anthropicApiKey: sk-ant-...
+
+# Telegram
+telegramBotToken: "123456:ABC..."
+allowedTelegramUsers: [123456789]
+
+# Agents
+agents:
+  coder:
+    model: claude-sonnet-4-20250514
+    thinkingLevel: medium
+    briefMode: true
+    telegramBotToken: "654321:XYZ..."
+    mcp:
+      servers:
+        github:
+          type: stdio
+          command: npx
+          args: ["-y", "@modelcontextprotocol/server-github"]
+
+# MCP Servers (global)
+mcp:
+  servers:
+    supabase:
+      type: http
+      url: https://mcp.supabase.com/...
 ```
 
-**npm**
-```bash
-npm install -g camelagi
-```
+Environment variables override file values: `ANTHROPIC_API_KEY`, `OPENAI_API_KEY`, `CAMELAGI_MODEL`, `CAMELAGI_PROVIDER`, `CAMELAGI_TOKEN`, `TELEGRAM_BOT_TOKEN`.
 
-**Usage**
-```bash
-camel bootstrap     # Full setup: admin bot + pairing + API config
-camel serve         # Start the gateway (after bootstrap)
-camel chat          # Terminal UI
-```
+See [featuresDocs/configuration.md](featuresDocs/configuration.md) for the full configuration reference.
 
 <br>
 
-## Admin Bot — BotFather for Your AI Server
-
-Create agents, manage config, approve users, monitor sessions, restart bots — all from Telegram. The Admin Bot is a **non-AI Telegram bot** — no LLM calls, no tokens burned, just instant commands. Think [@BotFather](https://t.me/BotFather), but for your entire AI infrastructure.
-
-### Getting Started
-
-1. Create a bot with [@BotFather](https://t.me/BotFather) on Telegram
-2. Run `camel bootstrap` and paste the bot token
-3. Send a message to your bot on Telegram — the CLI detects you
-4. Approve yourself in the CLI, then enter the OTP in Telegram
-5. Pick your AI provider and model (or skip and do it later via `/setup` in Telegram)
-6. Done — use `/newagent` in Telegram to create your first AI agent
-
-### Commands
-
-| Category | Command | Description |
-|----------|---------|-------------|
-| **Agents** | `/newagent` | Create agent wizard |
-| | `/agents` | List all agents |
-| | `/deleteagent` | Delete an agent |
-| | `/soul` | View/edit agent personality |
-| **Config** | `/config` | View configuration |
-| | `/config <key> <value>` | Update config |
-| | `/setup` | API provider wizard |
-| **Monitor** | `/status` | System health & stats |
-| | `/sessions` | List & manage sessions |
-| | `/restart` | Restart agent bots |
-| **Security** | `/pairing` | Manage access requests |
+## Architecture
 
-<br>
-
-## Features
+```
+~/.camelagi/
+├── config.yaml              ← Single config file
+├── workspace/               ← Default agent workspace
+│   ├── SOUL.md              ← Default personality
+│   ├── MEMORY.md            ← Default curated memory
+│   └── memory/              ← Daily notes
+├── agents/                  ← Per-agent isolated directories
+│   └── <agent-id>/
+│       ├── SOUL.md
+│       ├── TOOLS.md
+│       ├── MEMORY.md
+│       └── memory/
+├── sessions/                ← JSONL conversation history
+├── skills/                  ← Markdown skill files
+└── usage/                   ← Token usage data
+```
 
-> Terminal or Telegram — same agent, same tools, same memory.
+**Request flow:**
 
-| | Feature | Description |
-|---|---|---|
-| ⌨️ | **camel chat — Terminal UI** | Full TUI with streaming, slash commands, model switching, session management, tool output, and markdown rendering |
-| 🤖 | **Telegram — Admin Bot** | @BotFather for AI agents. Create, configure, and manage agents entirely from Telegram — instant commands, zero tokens burned |
-| 💬 | **Telegram — Agent Bots** | Each agent gets its own Telegram bot. Message it like any chat — it runs tools, reads files, remembers context |
-| ⚙️ | **Same Engine** | Both interfaces run the same agent loop, same 10 tools, same two-tier memory. Switch between terminal and Telegram anytime |
-| 🧠 | **Agent Memory** | Each agent gets isolated two-tier memory — curated MEMORY.md + daily auto-journaling with recency-boosted search |
-| ⏰ | **Cron Jobs** | Schedule AI tasks — daily summaries, monitoring, automations. Manage from Telegram, CLI, or the agent itself |
-| 🛡️ | **Tool Approvals** | Human-in-the-loop safety. Approve dangerous operations from Telegram with inline buttons — even headless |
-| 💭 | **Extended Thinking** | Claude reasons step by step before answering. Configure depth: off, low, medium, high |
-| 🔌 | **Multi-Provider** | Anthropic, OpenAI, OpenRouter, Ollama — any OpenAI-compatible endpoint. Zero vendor lock-in |
-| 🔐 | **Secure Pairing** | OTP-based user verification. No hardcoded IDs — pairing code + 5-digit OTP from Telegram |
-| 🪝 | **Skills & Hooks** | Teach agents skills via Markdown. Run shell/JS hooks before and after tool calls |
-| 🔄 | **Auto Compaction** | Summarizes old turns at 80% capacity. Flushes facts to memory first — nothing is lost |
+```
+Inbound message (TUI / REST / WS / Telegram / Discord)
+  → Queue check (per-session, prevents concurrent runs)
+  → Lane acquisition (concurrency limits: main/cron/subagent)
+  → History loading + compaction (summarize at 80% of maxTokens)
+  → Agent execution with retry (classify error → backoff or compact → retry)
+  → Message persistence (JSONL sessions)
+  → Cleanup (release lane, drain queued messages)
+```
 
 <br>
 
-## Terminal UI — `camel chat`
-
-Don't want to use Telegram? `camel chat` gives you a full terminal interface with the same agent capabilities.
-
-- Streaming responses with markdown rendering
-- Slash commands (`/model`, `/sessions`, `/tools`, `/compact`, `/status`, `/context`)
-- Model selector overlay (`Ctrl+L`), session switcher (`Ctrl+P`)
-- Tool output toggle (`Ctrl+O`), abort with `Escape`
-- Shell execution with `!command`
-- Agent creation wizard, SOUL.md editing
-- Thinking indicators, subagent progress, approval overlay
-- One-shot mode: `camel "your question"` for quick answers
-
-<br>
+## Documentation
 
-## Developer Experience
+| Document | Description |
+|----------|-------------|
+| [DOCS.md](DOCS.md) | Full reference documentation |
+| [GUIDE.md](GUIDE.md) | User guide with examples |
+| [featuresDocs/](featuresDocs/) | Deep-dive feature docs |
 
-| | **CamelAGI** | **OpenClaw** |
-|---|---|---|
-| **Codebase size** | ~10K LOC | ~700K+ LOC |
-| **AI Agent runtime** | [Claude Agent SDK](https://platform.claude.com/docs/en/agent-sdk/overview) — fully documented by Anthropic, easy to extend | pi-agent & custom abstractions |
+Feature docs cover: [agent system](featuresDocs/agent-system.md), [memory](featuresDocs/memory-system.md), [Telegram bots](featuresDocs/telegram-bots.md), [gateway server](featuresDocs/gateway-server.md), [runtime](featuresDocs/runtime.md), [tools](featuresDocs/tools.md), [extensions](featuresDocs/extensions.md), [configuration](featuresDocs/configuration.md), [CLI](featuresDocs/cli-commands.md), [TUI](featuresDocs/tui.md), [pairing](featuresDocs/pairing-otp.md).
 
 <br>
 
 ## Roadmap
 
-We're building CamelAGI to be the most capable open-source AI platform — for individuals and businesses alike.
-
 | | Feature | Description |
 |---|---|---|
-| 📡 | **More Channels** | WhatsApp, Discord, Slack — connect your AI agents to every platform your team already uses |
-| 🧩 | **ClawHub Skills** | Browse and install community skills from [clawhub.io](https://clawhub.io) — one command to add new capabilities to any agent |
-| 🖥️ | **Native Desktop Apps** | Standalone macOS and Windows apps to run CamelAGI natively — no terminal, no companion app required |
-| 🏢 | **Business Ready** | Deploy CamelAGI for your business — finance, accounting, operations, customer support. AI agents that understand your workflows |
-
-<br>
-
-## Documentation
-
-| Document | Description |
-|----------|-------------|
-| [DOCS.md](DOCS.md) | Full reference documentation |
-| [GUIDE.md](GUIDE.md) | User guide with examples |
-| [featuresDocs/](featuresDocs/) | Deep-dive feature docs |
+| 📡 | **More Channels** | WhatsApp, Slack — connect your AI agents to every platform your team already uses |
+| 🧩 | **ClawHub Skills** | Browse and install community skills from [clawhub.io](https://clawhub.io) — one command to add new capabilities |
+| 🏢 | **Business Ready** | Deploy for your business — finance, operations, customer support. AI agents that understand your workflows |
 
 <br>
 

package/dist/agent/agent-sdk.js

@@ -32,6 +32,16 @@ const BUILTIN_TOOLS = [
     "WebSearch", "WebFetch",
     "Agent",
 ];
+/** Extract text from an SDK "assistant" message's content blocks */
+function extractAssistantText(msg) {
+    if (msg.type !== "assistant" || !msg.message?.content)
+        return "";
+    const parts = Array.isArray(msg.message.content) ? msg.message.content : [];
+    return parts
+        .filter((b) => b.type === "text" && b.text)
+        .map((b) => b.text)
+        .join("");
+}
 export async function runAgentSdk(apiKey, model, systemPrompt, history, userMessage, opts) {
     // Build effective prompt: if resuming, just use user message.
     // Otherwise, prepend history as structured context in the prompt itself
@@ -130,80 +140,141 @@ export async function runAgentSdk(apiKey, model, systemPrompt, history, userMess
     }
     let result = "";
     let sdkSessionId;
-    const q = query({
-        prompt: effectivePrompt,
-        options: {
-            model,
-            systemPrompt,
-            allowedTools: [...BUILTIN_TOOLS],
-            ...(disallowedTools && { disallowedTools }),
-            mcpServers: { camelagi: mcpServer, ...(opts?.mcpServers ?? {}) },
-            maxTurns: opts?.maxTurns ?? DEFAULT_MAX_TURNS,
-            permissionMode: "bypassPermissions",
-            allowDangerouslySkipPermissions: true,
-            cwd: process.cwd(),
-            env: { ANTHROPIC_API_KEY: apiKey },
-            thinking,
-            ...(opts?.effort && { effort: opts.effort }),
-            ...(opts?.maxBudgetUsd && { maxBudgetUsd: opts.maxBudgetUsd }),
-            ...(opts?.resumeSessionId && { resume: opts.resumeSessionId }),
-            ...(abortController && { abortController }),
-            includePartialMessages: !!emit,
-            settingSources: ["project"],
-            hooks: {
-                PreToolUse: [{ matcher: ".*", hooks: [preToolHook] }],
-                PostToolUse: [{ matcher: ".*", hooks: [postToolHook] }],
-            },
+    // Build options object step-by-step to isolate any spread errors
+    const mcpServers = { camelagi: mcpServer, ...(opts?.mcpServers ?? {}) };
+    const allowedTools = [
+        ...BUILTIN_TOOLS,
+        "mcp__camelagi__*",
+        ...Object.keys(opts?.mcpServers ?? {}).map(name => `mcp__${name}__*`),
+    ];
+    const envVars = { ...(process.env ?? {}), ...buildSdkEnv(apiKey, opts?.provider, opts?.baseUrl) };
+    const queryOptions = {
+        model,
+        systemPrompt,
+        allowedTools,
+        mcpServers,
+        maxTurns: opts?.maxTurns ?? DEFAULT_MAX_TURNS,
+        permissionMode: "bypassPermissions",
+        allowDangerouslySkipPermissions: true,
+        cwd: process.cwd(),
+        env: envVars,
+        thinking,
+        includePartialMessages: !!emit,
+        settingSources: ["project"],
+        hooks: {
+            PreToolUse: [{ matcher: ".*", hooks: [preToolHook] }],
+            PostToolUse: [{ matcher: ".*", hooks: [postToolHook] }],
         },
-    });
-    for await (const message of q) {
-        if (message.type === "result") {
-            const resultMsg = message;
-            result = resultMsg.result;
-            sdkSessionId = resultMsg.session_id;
-            if (resultMsg.usage && emit) {
-                emit({ type: "usage", inputTokens: resultMsg.usage.input_tokens ?? 0, outputTokens: resultMsg.usage.output_tokens ?? 0, cacheReadTokens: resultMsg.usage.cache_read_input_tokens, cacheWriteTokens: resultMsg.usage.cache_creation_input_tokens });
-            }
-            if (resultMsg.usage && opts?.sessionId) {
-                recordUsage(opts.sessionId, { inputTokens: resultMsg.usage.input_tokens ?? 0, outputTokens: resultMsg.usage.output_tokens ?? 0, cacheReadTokens: resultMsg.usage.cache_read_input_tokens ?? 0, cacheWriteTokens: resultMsg.usage.cache_creation_input_tokens ?? 0 });
-            }
-            emit?.({ type: "chunk", text: result });
-        }
-        else if (message.type === "system" && emit) {
-            const sysMsg = message;
-            if (sysMsg.subtype === "init" && sysMsg.session_id) {
-                sdkSessionId = sysMsg.session_id;
-                emit({ type: "init", sessionId: sysMsg.session_id });
-            }
-            else if (sysMsg.subtype === "task_started") {
-                emit({ type: "subagent_start", agentId: sysMsg.agent_id ?? "subagent", taskId: sysMsg.task_id });
-            }
-            else if (sysMsg.subtype === "task_progress") {
-                emit({ type: "subagent_progress", agentId: sysMsg.agent_id ?? "subagent", toolCount: sysMsg.tool_count, duration: sysMsg.duration_ms ? Math.round(sysMsg.duration_ms / 1000) : undefined });
-            }
-            else if (sysMsg.subtype === "task_notification") {
-                emit({ type: "subagent_done", agentId: sysMsg.agent_id ?? "subagent", toolUseId: sysMsg.tool_use_id });
+    };
+    if (disallowedTools)
+        queryOptions.disallowedTools = disallowedTools;
+    if (opts?.effort)
+        queryOptions.effort = opts.effort;
+    if (opts?.maxBudgetUsd)
+        queryOptions.maxBudgetUsd = opts.maxBudgetUsd;
+    if (opts?.resumeSessionId)
+        queryOptions.resume = opts.resumeSessionId;
+    if (abortController)
+        queryOptions.abortController = abortController;
+    const q = query({ prompt: effectivePrompt, options: queryOptions });
+    // The for-await loop may throw if the SDK subprocess exits unexpectedly
+    // (e.g. non-Claude models via OpenRouter). If we already captured a result
+    // from "assistant" messages, we still want to return it.
+    try {
+        for await (const message of q) {
+            const msg = message;
+            // "assistant" messages: non-Claude models (via OpenRouter) emit these
+            // with the response text instead of a final "result" message.
+            const assistantText = extractAssistantText(msg);
+            if (assistantText) {
+                result = assistantText;
+                if (msg.session_id)
+                    sdkSessionId = msg.session_id;
             }
-        }
-        else if (message.type === "stream_event" && emit) {
-            const streamMsg = message;
-            const event = streamMsg.event;
-            if (event.type === "content_block_start" && event.content_block?.type === "thinking") {
-                emit({ type: "thinking", state: "start" });
+            if (message.type === "result") {
+                const resultMsg = message;
+                // Only overwrite if non-empty (non-Claude models return text via
+                // "assistant" messages; the trailing "result" may be empty).
+                if (resultMsg.result) {
+                    result = resultMsg.result;
+                }
+                sdkSessionId = resultMsg.session_id ?? sdkSessionId;
+                if (resultMsg.usage && emit) {
+                    emit({ type: "usage", inputTokens: resultMsg.usage.input_tokens ?? 0, outputTokens: resultMsg.usage.output_tokens ?? 0, cacheReadTokens: resultMsg.usage.cache_read_input_tokens, cacheWriteTokens: resultMsg.usage.cache_creation_input_tokens });
+                }
+                if (resultMsg.usage && opts?.sessionId) {
+                    recordUsage(opts.sessionId, { inputTokens: resultMsg.usage.input_tokens ?? 0, outputTokens: resultMsg.usage.output_tokens ?? 0, cacheReadTokens: resultMsg.usage.cache_read_input_tokens ?? 0, cacheWriteTokens: resultMsg.usage.cache_creation_input_tokens ?? 0 });
+                }
+                emit?.({ type: "chunk", text: result });
             }
-            else if (event.type === "content_block_delta") {
-                if (event.delta?.type === "text_delta" && event.delta.text)
-                    emit({ type: "stream_text", text: event.delta.text });
-                else if (event.delta?.type === "thinking_delta" && event.delta.thinking)
-                    emit({ type: "thinking_delta", text: event.delta.thinking });
+            else if (message.type === "system" && emit) {
+                const sysMsg = message;
+                if (sysMsg.subtype === "init" && sysMsg.session_id) {
+                    sdkSessionId = sysMsg.session_id;
+                    emit({ type: "init", sessionId: sysMsg.session_id });
+                }
+                else if (sysMsg.subtype === "task_started") {
+                    emit({ type: "subagent_start", agentId: sysMsg.agent_id ?? "subagent", taskId: sysMsg.task_id });
+                }
+                else if (sysMsg.subtype === "task_progress") {
+                    emit({ type: "subagent_progress", agentId: sysMsg.agent_id ?? "subagent", toolCount: sysMsg.tool_count, duration: sysMsg.duration_ms ? Math.round(sysMsg.duration_ms / 1000) : undefined });
+                }
+                else if (sysMsg.subtype === "task_notification") {
+                    emit({ type: "subagent_done", agentId: sysMsg.agent_id ?? "subagent", toolUseId: sysMsg.tool_use_id });
+                }
             }
-            else if (event.type === "content_block_stop") {
-                emit({ type: "thinking", state: "end" });
+            else if (message.type === "stream_event" && emit) {
+                const streamMsg = message;
+                const event = streamMsg.event;
+                if (event.type === "content_block_start" && event.content_block?.type === "thinking") {
+                    emit({ type: "thinking", state: "start" });
+                }
+                else if (event.type === "content_block_delta") {
+                    if (event.delta?.type === "text_delta" && event.delta.text)
+                        emit({ type: "stream_text", text: event.delta.text });
+                    else if (event.delta?.type === "thinking_delta" && event.delta.thinking)
+                        emit({ type: "thinking_delta", text: event.delta.thinking });
+                }
+                else if (event.type === "content_block_stop") {
+                    emit({ type: "thinking", state: "end" });
+                }
             }
         }
     }
+    catch (err) {
+        // If the subprocess exited but we already have a response, return it.
+        // Otherwise, re-throw so the retry/error handling picks it up.
+        if (!result)
+            throw err;
+    }
     const userMsg = { role: "user", content: userMessage };
     const aiMsg = { role: "assistant", content: result };
     return { response: result, newMessages: [userMsg, aiMsg], usage: null, sessionId: sdkSessionId };
 }
+/** Build environment variables for the SDK subprocess based on provider */
+function buildSdkEnv(apiKey, provider, baseUrl) {
+    // OpenRouter: use ANTHROPIC_BASE_URL + ANTHROPIC_AUTH_TOKEN
+    if (baseUrl?.includes("openrouter")) {
+        // SDK expects /api not /api/v1
+        const sdkBaseUrl = baseUrl.replace(/\/v1\/?$/, "");
+        return {
+            ANTHROPIC_BASE_URL: sdkBaseUrl,
+            ANTHROPIC_AUTH_TOKEN: apiKey,
+            ANTHROPIC_API_KEY: "",
+        };
+    }
+    // Direct Anthropic
+    if (provider === "anthropic" || !provider) {
+        return { ANTHROPIC_API_KEY: apiKey };
+    }
+    // Custom base URL (non-OpenRouter)
+    if (baseUrl) {
+        return {
+            ANTHROPIC_API_KEY: apiKey,
+            ANTHROPIC_BASE_URL: baseUrl,
+        };
+    }
+    // Default
+    return { ANTHROPIC_API_KEY: apiKey };
+}
 //# sourceMappingURL=agent-sdk.js.map