SDK Integration¶

Current Implementation¶

The current implementation uses the Claude CLI in --print mode:

cmd = [
    "claude",
    "--print",
    "--dangerously-skip-permissions",
    prompt,
]

process = await asyncio.create_subprocess_exec(
    *cmd,
    stdout=asyncio.subprocess.PIPE,
    stderr=asyncio.subprocess.PIPE,
    cwd=session.cwd,
)

This approach:

Uses --print to get plain text output
Uses --dangerously-skip-permissions for non-interactive use
Streams output line by line back to Telegram
Uses Claude CLI's own authentication (via claude login)

Authentication¶

Claude CLI handles its own authentication. Before running voice-agent:

claude login

This stores credentials that the CLI uses automatically.

Future: Claude Agent SDK¶

The plan is to migrate to the official Claude Agent SDK for:

Structured responses: Native message types instead of text parsing
Permission callbacks: canUseTool callback for approval flow
Session persistence: Resume sessions via session_id

Planned SDK Usage¶

from claude_code_sdk import ClaudeCodeClient

async def can_use_tool(tool_name: str, input_data: dict) -> bool:
    if is_safe_tool_call(tool_name, input_data):
        return True

    # Queue for user approval
    await notify_user(tool_name, input_data)
    return await wait_for_approval()

client = ClaudeCodeClient(
    cwd="/code/project",
    can_use_tool=can_use_tool,
)

async for message in client.query("list files"):
    process_message(message)

Benefits of SDK¶

CLI Approach	SDK Approach
Parse text output	Structured message objects
`--dangerously-skip-permissions`	`canUseTool` callback
Text streaming	Typed message objects
One process per prompt	Persistent client instance

Migration Path¶

Keep CLI implementation as fallback
Add SDK client alongside CLI
Switch based on configuration/availability
Remove CLI fallback once SDK is stable