sleigh-sdk 0.1.10

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 = client.create_session_token()["session_token"]
created = client.create_sandbox(session_token=session_token, image="alpine:3.20")
sandbox_id = created["sandbox_id"]

---

3. Ordered Workflow (AI Coding)

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

sandbox_id = created["sandbox_id"]
result = client.run_workflow(
    session_token=session_token,
    steps=[
        {"action": "exec_command", "sandbox_id": sandbox_id, "command": "echo hello", "wait": True, "wait_timeout_seconds": 10},
        {"action": "create_snapshot", "sandbox_id": sandbox_id},
        {"action": "exec_command", "sandbox_id": sandbox_id, "command": "uname -a", "wait": True},
    ],
)
print(result["stopped_early"], result["steps"])

Note: in SDK validation, every workflow step must include sandbox_id.

---

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 and targets directory inside sandbox filesystem
• sandbox_path is required and must be an absolute directory path in sandbox
• service exports sandbox dir to host temp workspace, applies edit, and syncs back
• quality checks: run pre-commit when config exists; otherwise auto-detect language for fallback checks
• write_mode=context_edit is default for partial edits; pass raw snippets with target_file_path, old_text, new_text, and optional before_context/after_context/occurrence
• write_mode=replace_file is supported for full overwrite by raw source content

result = client.patch_workspace(
    session_token=session_token,
    sandbox_id=sandbox_id,
    sandbox_path="/app",
    write_mode="context_edit",
    target_file_path="calculator.py",
    before_context="    def multiply(self, a, b):\n        return a * b\n\n",
    old_text="    def multiply(self, a, b):\n        return a * b\n",
    new_text="    def multiply(self, a, b):\n        return a * b\n\n    def sqrt(self, a):\n        if a < 0:\n            raise ValueError('Cannot sqrt negative number!')\n        return a ** 0.5\n",
)

# Full overwrite mode (raw source content)
rewrite_result = client.patch_workspace(
    session_token=session_token,
    sandbox_id=sandbox_id,
    sandbox_path="/app",
    write_mode="replace_file",
    target_file_path="calculator.py",
    content="print('hello from overwrite mode')\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,
    request_timeout_seconds=180,
)

---

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)