sleigh-sdk 0.1.2

Python SDK for Sleigh runtime server.

Sleigh Python SDK

Python SDK for the Sleigh runtime server.

Two client variants are included:

• LangChain Tool variant: one-call as_langchain_tool() (returns StructuredTool)
• MCP variant: expose runtime APIs as MCP tools over stdio

---

1. Install

pip install .

Optional extras:

pip install ".[langchain]"
pip install ".[mcp]"

---

2. Base Python Client

from sdk import SleighClient

client = SleighClient(base_url="http://127.0.0.1:8080")
session_token = "sess_demo"
created = client.create_sandbox(session_token=session_token, image="alpine:3.20")
sandbox_id = created["id"]

---

3. Ordered Workflow (AI Coding)

Run multiple steps in one request and stop early on failure/timeout:

result = client.run_workflow(
    session_token=session_token,
    steps=[
        {"action": "create_sandbox", "image": "alpine:3.20", "memory_limit_mb": 512},
        {"action": "exec_command", "command": "echo hello", "wait": True, "wait_timeout_seconds": 10},
        {"action": "create_snapshot"},
        {"action": "exec_command", "command": "uname -a", "wait": True},
    ],
)
print(result["stopped_early"], result["steps"])

---

4. Sandbox Read API (AI Coding)

read_result = client.read_sandbox(
    session_token=session_token,
    sandbox_id=sandbox_id,
    command="rg",
    args=["TODO", "/workspace"],
    timeout_seconds=10,
    max_output_bytes=65536,
    max_lines=200,
)
print(read_result)

---

5. Patch API (Sandbox Semantic)

patch_workspace targets:

• POST /sandboxes/{id}/ops/patch
• validates sandbox auth + mounted workspace path ownership
• cwd is required and must point to a mounted host directory for that sandbox

result = client.patch_workspace(
    session_token=session_token,
    sandbox_id=sandbox_id,
    cwd="/home/hxluo/sleigh_env",
    patch="*** Begin Patch\n*** End Patch\n",
)

---

6. Low-Memory Create Guard

When host available memory ratio:

• < 5%: create is blocked
• >= 5% and < 8%: create requires confirm_low_memory=True

created = client.create_sandbox(
    session_token=session_token,
    image="alpine:3.20",
    confirm_low_memory=True,
)

---

7. More Examples

• LangChain integration: ../README_langchain.md
• MCP integration: ../README_mcp.md

8. Session Exec History

list_session_exec_tasks accepts optional session_id. If omitted, SDK uses session_token as session_id automatically:

history = client.list_session_exec_tasks(
    session_token=session_token,
    limit=20,
)
print(history)