Learn lclang: start here¶

lclang is for configuration that has relationships, not just stored values. A service URL depends on an environment. A price depends on quantity, customer policy, and perhaps an asynchronous rate provider. Those relationships should stay visible and testable without becoming scattered Python glue.

LCL stands for Lazy Context Language:

Lazy: a definition runs only when a requested result needs it.
Context: the same definitions can receive different inputs for each request, command, tenant, or test.
Language: relationships are parsed into lclang's own immutable syntax and interpreted by its async runtime. They are not passed to Python eval or exec.

This introduction gives you a working mental model and four complete examples. They progress from a one-line calculation to a reusable, file-backed pricing policy. By the end, you should know whether lclang fits your problem and which tutorial to read next.

The lclang philosophy: lclang defines a strong way of working and provides strong flexibility within that way. It does not invite every application to invent its own parsing, lookup, caching, and lifecycle rules. It fixes those rules, then lets you be creative with definitions, dependency shapes, parameters, host capabilities, and outputs. The result is expressive configuration that remains recognizable, inspectable, and testable.

Every chapter is published, and the documentation tests execute each marked example directly from its Markdown source.

The tutorial series¶

Start here: mental model and first application (this page) -- decide when lclang is useful, then follow the path from one expression to a configuration file.
Expressions and values -- name calculations in a Module and evaluate them with host inputs and canonical builtins in a Frame.
Modules and Frames -- define policy once, evaluate it in many independent contexts, and manage each context's lifecycle.
The LCL language -- build expressions with collections, conditionals, comprehensions, formatted strings, arrow functions, and control forms.
Configuration files -- write .lclcfg, compose files with using, keep source provenance, and control loading.
Async Python integration -- expose narrow host functions, await external work automatically, and clean up resources deterministically.
Caching and recalculation -- reason about lazy success and failure snapshots, concurrency sharing, mixins, and explicit refresh.
Errors and inspection -- preserve source-aware diagnostics and explain a value without accidentally evaluating it.
Dependency analysis -- compare possible static dependencies with lookups observed during a real evaluation.
Command-line applications -- combine typed commands, configuration, overrides, dry runs, logging, and stable exit behavior.
Workflow status -- report nested task and step outcomes while preserving failure and cleanup semantics.
Business-day calendars -- compose three-state calendars, map dates, and load named policies from strict JSON.
Production patterns -- organize modules, factories, tests, limits, trust boundaries, and application ownership.
Scoped values and Frame evaluation -- organize qualified bindings through lazy proxies, materialize override-friendly dataclass lists, and run unnamed expressions in an existing Frame.
Tree workflows -- connect typed async actions, scoped resources, explicit data flow, status, cleanup, and CLI inference in one inspectable task tree.
Case Study: Energy Settlement Workflow -- build and exercise a complete configured workflow command with result files, logs, masking, overrides, dry-run, verbose, and errors.
Python utilities for downstream applications -- use the live environment, standalone logging, reviewed data helpers, and calendars without adopting LCL evaluation or the CLI.
Configuration composition -- choose shared expansion or isolated imports, supply required and runtime values, compose Modules, and trace loading and evaluation failures.
Application logging: own process logging, borrow existing loggers, trace selected operations, and report final failures with compact or full Python stacks.

The sequence is deliberate: lessons 1-4 establish everyday use, lessons 5-8 make behavior predictable in real applications, and lessons 9-15 apply the same model to larger systems. Lesson 16 shows which supporting Python APIs can stand alone, lesson 17 brings configuration composition and diagnostics together, and lesson 18 explains downstream application logging. Read the lessons in order or jump to one topic after finishing this introduction.

Is lclang a good fit?¶

Use lclang when all or most of these statements are true:

Your configuration contains derived values that must stay consistent.
Stable policy should be reusable with different per-run Python inputs.
Some calculations are optional, expensive, or asynchronous.
You need to inspect definitions, dependencies, cached values, or failures.
Configuration authors and host callables are trusted by the application.

Prefer JSON, TOML, YAML, or ordinary Python data when the values are already static. Do not use LCL as a sandbox for hostile input. The interpreter avoids Python dynamic-code execution, but trusted host values and callables still have the authority your application gives them.

The two ideas to remember¶

A Module is an immutable collection of named, unevaluated definitions. It answers what can be calculated?

A Frame combines a Module with inputs, name lookup, a lazy result cache, and owned asynchronous work. It answers what do those definitions mean for this run?

                        define once
Python/LCL source  ---------------------->  Module
                                                  \
                                                   + inputs for one run
                                                  /
                        evaluate lazily           v
requested name  --------------------------->  Frame  ----> value snapshot

Modules are safe to reuse because they have no evaluation state. Frames are intentionally stateful and belong to one event loop. Create a fresh Frame for each independent run and use async with so its work is closed cleanly.

Before you begin¶

lclang requires Python 3.14 or newer and has no third-party runtime dependencies.

python -m pip install lclang

The examples use assertions so that success is unambiguous. Save any complete Python block as a .py file and run it; no output means every documented result was correct.

Example 1: calculate one expression¶

Name the result in a Module, then create a Frame with inputs for one run.

import asyncio

from lclang.lang import define_frame, define_module


TOTAL = define_module("total", {"result": "unit_price * quantity"})


async def main() -> None:
    async with define_frame(
        TOTAL,
        preset={"unit_price": 6, "quantity": 4},
    ) as frame:
        assert await frame.get("result") == 24


asyncio.run(main())

unit_price resolves to 6 and quantity resolves to 4 from the Frame's preset. Requesting result follows those lookups and multiplies them to produce 24. Leaving the context closes the Frame and its owned work.

lclang parses the source into its own immutable abstract syntax tree and then interprets it in the Frame with the supplied values and canonical builtins. Names are explicit: if quantity is missing, evaluation raises LclNameError instead of silently inventing a default.

asyncio.run(main()) starts the event loop at a synchronous application's outer boundary. Servers, workers, and other async applications await their coroutine in the existing loop. The Module and Frame APIs stay the same.

Example 2: define once, evaluate in context¶

Suppose a deployment tool needs a URL and timeout for several environments. The relationships are stable, while service and environment change for each run.

import asyncio

from lclang.lang import define_frame, define_module


SERVICE_POLICY = define_module(
    "service-policy",
    {
        "summary": 'f"{base_url} (timeout={timeout}s)"',
        "base_url": 'f"https://{host}/v1"',
        "host": 'f"{service}.{environment}.example.com"',
        "timeout": "30 if environment == 'production' else 5",
    },
)


async def describe(service: str, environment: str) -> str:
    async with define_frame(
        SERVICE_POLICY,
        preset={"service": service, "environment": environment},
    ) as frame:
        result = await frame.get("summary")
        assert isinstance(result, str)
        return result


async def main() -> None:
    assert await describe("billing", "test") == (
        "https://billing.test.example.com/v1 (timeout=5s)"
    )
    assert await describe("billing", "production") == (
        "https://billing.production.example.com/v1 (timeout=30s)"
    )


asyncio.run(main())

For the test run, host joins billing, test, and the fixed domain; base_url adds the scheme and path; and the false conditional branch selects a five-second timeout. The production Frame repeats the same dependency path with different inputs, and its true branch selects 30. The two assertions show the complete strings produced at the end of those paths.

Three details matter here:

define_module parses every definition up front. A malformed expression fails during application setup, not halfway through a request.
Definition order is irrelevant. Asking for summary makes the Frame follow summary -> base_url -> host and summary -> timeout as needed.
The Module is reused, but each call creates a fresh Frame with independent inputs, cached results, dependency observations, and lifecycle.

The preset name means host-provided bindings; it does not mutate the Module. Keep this boundary narrow. Pass the primitive values and purpose-built functions that policy needs rather than exposing an entire application object.

Example 3: skip work that is not needed¶

Lazy evaluation becomes valuable when a host function performs asynchronous or expensive work. This pricing policy converts a subtotal only for non-USD orders. lclang automatically awaits an async host function when the selected branch calls it.

import asyncio

from lclang.lang import define_frame, define_module


PRICING = define_module(
    "pricing",
    {
        "subtotal": "unit_price * quantity",
        "converted_total": (
            "subtotal if currency == 'USD' "
            "else subtotal * exchange_rate(currency)"
        ),
        "label": 'f"{currency} {converted_total:.2f}"',
    },
)


async def main() -> None:
    calls: list[str] = []

    async def exchange_rate(currency: str) -> float:
        calls.append(currency)
        return {"EUR": 0.92}[currency]

    async with define_frame(
        PRICING,
        preset={
            "unit_price": 10,
            "quantity": 3,
            "currency": "EUR",
            "exchange_rate": exchange_rate,
        },
    ) as euro_frame:
        assert await euro_frame.get("subtotal") == 30
        assert calls == []

        assert await euro_frame.get("label") == "EUR 27.60"
        assert calls == ["EUR"]

        assert await euro_frame.get("label") == "EUR 27.60"
        assert calls == ["EUR"]  # The value is a cached snapshot.

    async with define_frame(
        PRICING,
        preset={
            "unit_price": 10,
            "quantity": 3,
            "currency": "USD",
            "exchange_rate": exchange_rate,
        },
    ) as usd_frame:
        assert await usd_frame.get("label") == "USD 30.00"
        assert calls == ["EUR"]  # The USD branch never calls the provider.


asyncio.run(main())

The EUR Frame first multiplies 10 * 3, so requesting only subtotal leaves the provider untouched. Requesting label reaches converted_total, selects the non-USD branch, awaits the 0.92 rate, and formats 27.6 as 27.60. The second label read uses that snapshot. In the USD Frame the conditional returns the subtotal directly, so the provider call list still contains only EUR.

The first request for a definition stores either its value or its ordinary failure. Later reads in that Frame receive the same snapshot, and concurrent callers in the same event loop share in-flight work. A new Frame starts a new run.

This is deliberately not spreadsheet-style reactivity. Changing an input does not invalidate downstream snapshots behind your back. Later tutorials will show when to create a new Frame and when explicit mixin plus recalculate is the clearer choice.

Example 4: move policy into a configuration file¶

Once non-Python users should edit the policy, or source locations matter for diagnostics, put definitions in a UTF-8 .lclcfg file. Create pricing.lclcfg beside the Python program:

__LCL_VERSION__: 1

subtotal: unit_price * quantity
discount: subtotal * discount_rate if vip else 0
total: subtotal - discount
message: f"{customer}: {currency} {total:.2f}"

Each definition has a name, a colon, and one complete LCL expression. The file stores relationships, while the application still owns request-specific values.

import asyncio
from pathlib import Path

from lclang.config import load_config


async def main(config_path: Path) -> None:
    config = await load_config(config_path)
    async with config.to_frame(
        preset={
            "unit_price": 25,
            "quantity": 4,
            "discount_rate": 0.10,
            "vip": True,
            "customer": "Ada",
            "currency": "USD",
        },
    ) as frame:
        assert await frame.get("message") == "Ada: USD 90.00"


if __name__ == "__main__":
    asyncio.run(main(Path("pricing.lclcfg")))

The supplied values make subtotal equal to 25 * 4, or 100. Because vip is true, discount becomes ten percent of that subtotal. total subtracts 10 from 100, and the f-string combines the customer, currency, and two-decimal value into the asserted Ada: USD 90.00 message.

This example uses no dynamic using target, so load_config does not evaluate its definitions. Dynamic targets can evaluate earlier definitions in temporary Frames to choose an included file. The immutable result keeps source provenance for diagnostics and can be reused. config.to_frame(preset=...) preserves that source information and supplies the same canonical lookup, caching, and lifecycle as define_frame(config.to_module(), preset=...). Request further names from the same Frame when they should share snapshots in one run.

Larger configurations can compose files with source-ordered using declarations. That belongs in the focused configuration tutorial; the important boundary for now is simple: the file defines trusted policy, and Python supplies only the context for the current run.

What you have built¶

The four examples use the same design at increasing scale:

one expression
    -> related named definitions
        -> lazy asynchronous host integration
            -> file-backed reusable policy with source provenance

The design stays understandable because parsing, reusable definitions, per-run inputs, cached evaluation, and cleanup have separate owners. That separation is the core of lclang's code structure too: lexer/parser and immutable AST, interpreter, runtime Module/Frame layer, configuration loader, then optional application frameworks such as CLI, workflow, and calendars.

Keep these rules nearby:

Parse or load stable definitions once.
Create one Frame for each independent run.
Supply only the values and functions the policy needs.
Ask only for the outputs you need; evaluation follows dependencies lazily.
Treat cached results as snapshots, not reactive cells.
Use owned Frames as async context managers.
Catch lclang.error.LclError at an application boundary when you need one family for expected language, configuration, and runtime failures.
Treat LCL source and host capabilities as trusted configuration.

For exact syntax and API contracts, use the reference index. For an even smaller first run, see the Quick Start Guide.