submilli

The Python SDK for submilli-server. It runs TypeScript programs under a Blueprint's policy, on behalf of a user, from any Python harness: an agent framework, a job queue, or a script.

pip install submilli

It needs Python 3.11 or later and submilli-server 0.4.1 or later. For LangChain and deepagents, langchain-submilli builds on it.

Quick start

import asyncio

from submilli import AsyncSubmilli

PROGRAM = """
export function main(): string {
    return `${1 + 1}`;
}
"""


async def main() -> None:
    async with AsyncSubmilli() as client:
        session = await client.sessions.create("research", variables={"userId": "u_ada"})
        result = await client.sessions.run(session.id, PROGRAM, idempotency_key="call-1")
        print(result.result)  # "2"


asyncio.run(main())

Submilli has the same API without await:

from submilli import Submilli

with Submilli() as client:
    session = client.sessions.create("research", variables={"userId": "u_ada"})
    print(client.sessions.run(session.id, PROGRAM).result)

The client reads the server's URL from $SUBMILLI_SERVER (default http://127.0.0.1:8128) and its API token from $SUBMILLI_SERVER_TOKEN, or from base_url= and token=.

Sessions

A session binds a Blueprint to one user's variables and harness secrets. Programs in it run one after another and share its files.

session = await client.sessions.create(
    "research",
    variables={"userId": "u_ada"},
    secrets={"GITHUB_TOKEN": token},  # the Blueprint's `harness` secrets
)
state = await client.sessions.execute(session.id, code, idempotency_key=call_id)
await client.sessions.rebind(session.id, {"GITHUB_TOKEN": new_token})
last = await client.sessions.last_run(session.id)  # the full console, or None
await client.sessions.delete(session.id)

Results, not exceptions. A program that fails or is denied is a result:

result = await client.sessions.run(session.id, code)
if result.denied:
    print(result.error.capability, result.error.source)  # e.g. "http.get", "policy"
elif result.error:
    print(result.error.kind, result.error.message)  # e.g. "compile_error"
else:
    print(result.result, result.console)

Idempotency. Pass an idempotency_key with every execute that may be retried. When the same request reaches the server twice, the program runs once and the retry gets the recorded result. idempotency_key(call_id) turns any id, such as a model's tool call id, into a key the server accepts.

Your tools in programs

A session can declare the application's own tools. Programs import them as @tools, typed from their schemas, and the server checks each call against the Blueprint's tools.call capability:

from submilli import PendingCall, ToolResult

QUERY_ORDERS = {
    "name": "query_orders",
    "description": "Orders of one customer.",
    "input_schema": {"type": "object", "properties": {"customer": {"type": "string"}}, "required": ["customer"]},
    "output_schema": {"type": "object", "properties": {"count": {"type": "number"}}, "required": ["count"]},
}


async def answer(call: PendingCall) -> ToolResult:
    orders = await db.orders(call.args["customer"])
    return ToolResult.success(call.call_id, {"count": len(orders)})


session = await client.sessions.create("shop", variables={"userId": "u_ada"}, tools=[QUERY_ORDERS])
result = await client.sessions.run(session.id, PROGRAM, idempotency_key="call-1", on_tool_call=answer)

import tools from "@tools";

function main(): string {
    return `${tools.query_orders({ customer: "c_42" }).count} orders`;
}

• The loop. Each call stops the run on the server. sessions.run hands the call to on_tool_call, sync or async, and resumes the run with the answer, until the program finishes. The time a run waits doesn't count against its time limit.
• Answers. ToolResult.success(call_id, value) returns value to the program. ToolResult.failed(call_id, message) throws Error there. ToolResult.denied(call_id, message) throws PermissionDeniedError, reported with source: "harness".
• Pausing. If on_tool_call raises, the exception propagates and the run keeps waiting on the server, up to the Blueprint's tool_call_timeout. That is how a harness pauses for a person's approval. Executing again with the same idempotency_key answers the run's current state, and sessions.run carries on from the call it waits on.
• Declarations. session.warnings lists what programs won't see of the declared tools, such as a Blueprint that grants no tools.call, or a tool whose result is typed unknown. The server doesn't resolve $ref, so pass Pydantic schemas through submilli.schema.inline_refs.
• Lower level. runs.resume, runs.retrieve, and runs.complete drive a run yourself.
• Restarts. A suspended run is lost if the server restarts, and resuming it raises UnknownRunError.

Discovery

What a model needs to write programs for a Blueprint:

prompt = await client.blueprints.prompt("research")  # tool descriptions for a model
await client.blueprints.search_packages("research", "github")
docs = await client.blueprints.package_docs("research", "@submilli/jina")  # docs.found, docs.text
await client.blueprints.builtins("research")
await client.blueprints.builtin_docs("research", ["Map", "Temporal"])

sessions.search_packages and sessions.package_docs answer the same for one session, including its @tools.

Errors

Every error is a SubmilliError:

Each APIStatusError carries status_code, body, the body's code, and the execution_id it concerns, when there is one.

Harness secrets are kept encrypted with the server's secret-store key, so sessions keep them across restarts. A server whose key changed refuses such a session with a ConflictError ("cannot decrypt session secrets").

Retries and timeouts

A connection failure, or a 502, 503, or 504, is retried up to max_retries times (default 2), with backoff, but only for requests that are safe to repeat: reads, deletes, rebinds, resumes, and executes that carry an idempotency key. An execute without a key and sessions.create are never retried. The default timeout is 330 seconds, because an execute may wait up to 300 seconds for an earlier attempt with the same key.

client = AsyncSubmilli(timeout=60.0, max_retries=0, http_client=httpx.AsyncClient(...))

Development

This package lives in the submilli-python workspace, next to langchain-submilli. See the repository's README for setup, checks, and releases. In packages/submilli:

uv run pytest                                   # unit tests, against a fake of the HTTP API
SUBMILLI_BIN=/path/to/bin uv run pytest tests/integration   # starts its own submilli-server
uv run python scripts/generate_sync.py          # after changing _async_client.py

License

Apache-2.0.