sbo3l-langchain

LangChain Python tool wrapping SBO3L. Mirror of @sbo3l/langchain (TypeScript).

⚠ DRAFT (T-1-2): depends on F-10 (sbo3l-sdk) merging + publishing to PyPI.

Install

pip install sbo3l-langchain sbo3l-sdk langchain-core

Quick start

from langchain_core.tools import StructuredTool
from sbo3l_sdk import SBO3LClientSync, bearer
from sbo3l_langchain import sbo3l_tool

client = SBO3LClientSync(
    "http://localhost:8730",
    auth=bearer("my-bearer-token"),
)
descriptor = sbo3l_tool(client=client)
tool = StructuredTool.from_function(
    func=descriptor.func,
    name=descriptor.name,
    description=descriptor.description,
)
# Pass `tool` into your LangChain agent's tool list.

What it does

The agent emits a tool call with a JSON-stringified APRP. The tool forwards to client.submit() and returns a JSON envelope:

{
  "decision": "allow",
  "deny_code": null,
  "matched_rule_id": "allow-low-risk-x402",
  "execution_ref": "kh-...",
  "audit_event_id": "evt-...",
  "request_hash": "...",
  "policy_hash": "..."
}

On deny, the LLM sees deny_code (policy.budget_exceeded, policy.token_unsupported, etc.) and can self-correct or escalate.

Idempotency

descriptor = sbo3l_tool(
    client=client,
    idempotency_key=lambda body: f"{body['task_id']}-{body['nonce']}",
)

Async clients

This factory returns a sync callback. Use SBO3LClientSync (httpx.Client under the hood — safe inside any event loop). If you must use the async client, create your own async tool that awaits client.submit(...).

Errors

Transport / auth failures surface as a JSON envelope with error (RFC 7807 domain code, e.g. auth.required) and status.

License

MIT