taylorai/lm-deluge/skill

from lm_deluge import LLMClient, Conversation, Tool, APIResponse

llm = LLMClient(model_names="claude-4.5-haiku", max_new_tokens=1024)
response = await llm.start(Conversation().user("Hello!"))
print(response.completion)  # Text output - NOT .text

llm = LLMClient(model_names="claude-4.5-haiku", max_new_tokens=1024)
conv = Conversation().user("Search for X and summarize")
final_conv, response = await llm.run_agent_loop(conv, tools=my_tools, max_rounds=5)

conv = Conversation()
conv.system("You are helpful")
conv.user("Question here")
conv.ai("Previous response")  # For multi-turn
conv.user("Follow-up")

conv.user("Analyze this:", image="path/to/img.png")
conv.user("Summarize:", file="path/to/doc.pdf")

async def search(query: str, limit: int = 10) -> list[dict]:
    """Search the database."""
    return results
tool = Tool.from_function(search)

tool = Tool(
    name="search",
    description="Search database",
    parameters={
        "query": {"type": "string", "description": "Search query"},
        "limit": {"type": "integer", "description": "Max results"},
    },
    required=["query"],
    run=search_function,
)

from pydantic import BaseModel
class SearchParams(BaseModel):
    query: str
    limit: int = 10
tool = Tool(name="search", parameters=SearchParams, run=search_fn)

llm = LLMClient(
    model_names="claude-4.5-sonnet",  # or list for fallback
    max_new_tokens=1024,
    temperature=0.7,
    json_mode=False,           # Force JSON output
    reasoning_effort="high",   # For reasoning models: low/medium/high
    thinking_budget=10000,     # Token budget for extended thinking
    max_requests_per_minute=1000,
    max_concurrent_requests=225,
)

prompts = ["Q1", "Q2", Conversation().user("Q3")]
responses = await llm.process_prompts_async(prompts)
# Or sync: llm.process_prompts_sync(prompts)

llm.open(total=100, show_progress=True)
responses = await llm.process_prompts_async(prompts)
llm.close()

task_id = llm.start_nowait(prompt)
# ... do other work ...
response = await llm.wait_for(task_id)

from pydantic import BaseModel
class Person(BaseModel):
    name: str
    age: int
response = await llm.start(prompt, output_schema=Person)

response = await llm.start(conv, cache="system_and_tools")
# Options: "tools_only", "system_and_tools", "last_user_message"

from lm_deluge.tool.prefab.sandbox import DockerSandbox, JustBashSandbox, SeatbeltSandbox
# Docker (cross-platform)
async with DockerSandbox() as sandbox:
    tools = sandbox.get_tools()
    conv, resp = await llm.run_agent_loop(conv, tools=tools)
# JustBash (cross-platform, optional dependency: npm install -g just-bash)
async with JustBashSandbox(root_dir=".", working_dir=".") as sandbox:
    tools = sandbox.get_tools()
# Seatbelt (macOS only, lighter)
async with SeatbeltSandbox(network_access=False) as sandbox:
    tools = sandbox.get_tools()

from lm_deluge import MCPServer
server = MCPServer(name="search", url="https://example.com/mcp")
tools = await server.to_tools()  # Convert to Tool objects
# Or pass directly
conv, resp = await llm.run_agent_loop(conv, mcp_servers=[server])