# multillm-claude

Claude Agent SDK provider for [multillm](../multillm).

This provider enables autonomous agents that can use tools and perform multi-step tasks.

## Installation

```bash
pip install multillm-claude
```

## Configuration

Choose one of the following methods (priority: direct config > env var > config file):

**Option 1: Claude CLI OAuth (recommended)**
```bash
claude login
```

**Option 2: Config file**

Create `~/.config/multillm/providers/claude.json`:

```json
{
  "api_key": "sk-ant-..."
}
```

Or for OAuth token:

```json
{
  "oauth_token": "sk-ant-oat01-..."
}
```

Set permissions:
```bash
chmod 600 ~/.config/multillm/providers/claude.json
```

**Option 3: Environment variables**

```bash
# API Key
export ANTHROPIC_API_KEY=sk-ant-...

# Or OAuth Token
export CLAUDE_CODE_OAUTH_TOKEN=your-token
```

**Option 4: Direct config**

```python
client = multillm.Client({
    "claude": {"api_key": "sk-ant-..."}
})
```

### Config File Fields

| Field | Required | Description |
|-------|----------|-------------|
| `api_key` | No* | Your Anthropic API key |
| `oauth_token` | No* | Your Claude OAuth token |

*One of `api_key` or `oauth_token` must be provided, unless using `claude login`.

## Usage

```python
import asyncio
import multillm

async def main():
    client = multillm.Client()

    # Simple query with tools
    answer = await client.single(
        "claude/default",
        "What Python version is installed?",
        allowed_tools=["Bash"],
        permission_mode="acceptEdits",
    )
    print(answer)

    # Streaming agent events
    options = multillm.AgentOptions(
        system_prompt="You are a helpful assistant.",
        max_turns=10,
        allowed_tools=["Read", "Write", "Bash"],
        permission_mode="acceptEdits",
    )

    async for msg in client.run("claude/default", "Create hello.py", options):
        if msg.type == "text":
            print(msg.content, end="")
        elif msg.type == "tool_use":
            print(f"\n[{msg.tool_name}]", end="")

asyncio.run(main())
```

## Model Selection

Specify the model after the provider prefix:

- `claude/default` - Uses the default model
- `claude/claude-sonnet-4-20250514` - Claude Sonnet
- `claude/claude-opus-4-20250514` - Claude Opus

## Tool Support

### Supported Tools

These tools work correctly with the Claude Agent SDK provider:

| Tool | Description | Permission Required |
|------|-------------|---------------------|
| `Bash` | Execute bash commands | Yes |
| `Read` | Read files from filesystem | No |
| `Write` | Create or overwrite files | Yes |
| `Edit` | Edit existing files | Yes |
| `Glob` | Find files by pattern | No |
| `Grep` | Search file contents | No |
| `Task` | Launch sub-agents | Varies |
| `WebFetch` | Fetch web content | No |
| `WebSearch` | Search the web | No |
| `NotebookEdit` | Edit Jupyter notebooks | Yes |
| `KillShell` | Kill background shells | Yes |
| `EnterPlanMode` | Enter planning mode | No |
| `ExitPlanMode` | Exit planning mode | No |

### Interactive Tools

| Tool | CLI Support | Notes |
|------|-------------|-------|
| `AskUserQuestion` | ✅ **Auto-converts** | CLI automatically uses custom `ask_user` tool |
| Custom tools | ✅ Full support | Provide Tool objects with interactive handlers |

**About Interactive Tools with Claude:**

The Claude Agent SDK's built-in `AskUserQuestion` runs in a subprocess and can't access our terminal's stdin/stdout for interactive prompting. To solve this, **multillm-cli automatically provides a custom interactive tool** when you request `AskUserQuestion`.

**What happens:**

When you use `--allowed-tools AskUserQuestion`, the CLI:
1. Removes the built-in AskUserQuestion (which doesn't work interactively)
2. Adds a custom `ask_user` tool with an interactive handler
3. The agent uses this tool instead, with full interactive support

**Usage:**

```bash
# Request AskUserQuestion - CLI auto-provides working alternative
multillm -m claude/default \
  -p "Ask me about my preferences and create a summary" \
  --allowed-tools AskUserQuestion \
  --permission-mode acceptEdits
```

**What you'll see:**
```
ℹ️  Using custom 'ask_user' tool instead of AskUserQuestion for interactive prompting

======================================================================
❓ QUESTION FROM ASSISTANT
======================================================================

What is your favorite programming language?

Suggested options:
  1. Python
  2. JavaScript
  3. Rust

Your answer: 1
======================================================================
```

**Programmatic usage:**

For programmatic use, provide your own `ask_user` tool with an interactive handler:

```python
import multillm

# Define interactive ask_user tool
ask_user_tool = multillm.Tool(
    name="ask_user",
    description="Ask the user a question",
    parameters={
        "type": "object",
        "properties": {
            "question": {"type": "string"},
            "options": {"type": "array", "items": {"type": "string"}}
        },
        "required": ["question"]
    },
    handler=lambda args: {
        "answer": input(f"\n{args['question']}\nYour answer: ")
    }
)

# Use with Claude
async for msg in client.run(
    "claude/default",
    "Ask me questions",
    options=multillm.AgentOptions(max_turns=10),
    tools=[ask_user_tool]  # Provide custom tool
):
    if msg.type == "text":
        print(msg.content)
```

**Why not use the built-in AskUserQuestion?**

The SDK's built-in `AskUserQuestion` is designed for Claude Code CLI's interactive mode where the subprocess has special stdin/stdout handling. In multillm, this doesn't work because:
- The SDK subprocess can't access our terminal
- We can't intercept and respond to built-in tool calls
- The tool returns an error instead of prompting

**Solution:** Use custom tools (which the CLI provides automatically!)

**Comparison:**

| Approach | Works? | How |
|----------|--------|-----|
| `--allowed-tools AskUserQuestion` | ✅ Yes | CLI auto-converts to custom tool |
| Custom `ask_user` Tool | ✅ Yes | Provide Tool object with handler |
| SDK built-in (direct) | ❌ No | Subprocess can't access stdin/stdout |
| Chat provider `ask_user` | ✅ Yes | Via agentwrap |

## Agent Options

```python
options = multillm.AgentOptions(
    system_prompt="...",           # System prompt
    max_turns=10,                  # Maximum agent turns
    allowed_tools=["Read", "Bash"], # Tools the agent can use
    permission_mode="acceptEdits", # Permission handling
    working_directory="/path",     # Working directory
)
```

## Message Types

When streaming with `client.run()`:

| Type | Description |
|------|-------------|
| `text` | Text content from the agent |
| `tool_use` | Agent is using a tool |
| `tool_result` | Result from a tool |
| `result` | Final result |
| `system` | System messages |

## Debugging

### Enable Debug Mode

Set the `MULTILLM_DEBUG` environment variable to see detailed error information and SDK options:

```bash
export MULTILLM_DEBUG=1
python your_script.py
```

This will show:
- Detailed error messages with full stderr/stdout
- SDK configuration options
- Full Python tracebacks

### Common Issues

**Error: "Command failed with exit code 1"**

The provider now captures and displays all available error information to stderr. Look for:

```
======================================================================
CLAUDE AGENT SDK ERROR
======================================================================
Error: <detailed error message>

Error Details:
  stderr: <actual error from subprocess>
  stdout: <subprocess output>
  exit_code: 1
======================================================================
```

**Authentication Errors:**
```bash
# Check authentication status
claude login --check

# Re-authenticate if needed
claude login
```

**Permission Errors:**

Always specify `permission_mode` when using tools:

```python
result = await client.single(
    "claude/default",
    "List files",
    allowed_tools=["Bash"],
    permission_mode="acceptEdits"  # Required!
)
```

### Full Debugging Guide

See [DEBUGGING.md](./DEBUGGING.md) for comprehensive debugging information, including:
- How to read error messages
- Common issues and solutions
- Debug mode usage
- Testing error handling
- Reporting bugs