Writing Python Code

Prerequisites

This skill extends myai's engineering-principles. Load that first. using-my-skills and engineering-principles are assumed already loaded via myai bootstrap.

For the general engineering philosophy (pit of success, fail fast, error handling as control flow, testing philosophy, architecture separation), see myai's engineering-principles. This skill covers only Python-specific tooling, type system configuration, and patterns.

Run project-local tools through uv, not system-installed binaries: uv run python, uv run pytest, uv run ruff, uv run basedpyright, uv run poe, uv run pre-commit. The normal verification pair is uv run poe lint_full and uv run poe test.

Type System

basedpyright Configuration

[tool.pyright]
pythonVersion = "3.14"
typeCheckingMode = "strict"
reportAny = "error"
reportImportCycles = "error"
reportImplicitStringConcatenation = "none"
reportUnusedCallResult = "none"
reportUnnecessaryIsInstance = "none"

Additional strict options (for medium-big projects):

reportExplicitAny = "error"
reportUnnecessaryTypeIgnoreComment = "error"
reportMissingModuleSource = "error"
reportPrivateUsage = "error"
reportOptionalMemberAccess = "error"
reportOptionalCall = "error"
reportAttributeAccessIssue = "error"

Banned Patterns

object is allowed only in: TypeIs/TypeGuard params, *args: object, Coroutine[object, None, T], PySide6 Signal(object). Everywhere else, use Protocol, TypeVar, union types, or typed structures.

Type Patterns

External data → msgspec.Struct:

import msgspec

class UserConfig(msgspec.Struct):
    name: str
    port: int
    debug: bool = False

# JSON → typed object (validates at decode time)
config = msgspec.json.decode(raw_bytes, type=UserConfig)

# Dict/YAML → typed object
config = msgspec.convert(raw_dict, type=UserConfig)

TypedDict is still valid when dict compatibility is needed (e.g., **unpacking, APIs expecting dicts).

FastAPI HTTP edge → pydantic DTOs are used at the transport boundary. Convert them immediately into framework-free typed structures and do not let framework models become domain models.

Domain objects → dataclass:

@dataclass(frozen=True, slots=True)
class Profile:
    name: str
    version: str
    active: bool = True

Duck typing → Protocol:

class Renderable(Protocol):
    def render(self) -> str: ...

Constants → Final:

MAX_RETRIES: Final = 3
CONFIG_PATH: Final[Path] = Path("~/.config/app").expanduser()

Handling Any at Library Boundaries

0. Typed deserialization (for external data):

msgspec.json.decode(data, type=MyStruct) eliminates Any for JSON/API responses — the return type is MyStruct, not Any. Use this before reaching for wrappers when the boundary is data deserialization.

1. Typed wrappers (preferred for library APIs):

class WhisperModelWrapper:
    def __init__(self, model_size: str, device: str = "auto") -> None:
        from faster_whisper import WhisperModel as _WhisperModel
        self._model = _WhisperModel(model_size, device=device)

    def transcribe(self, audio: np.ndarray, language: str | None = None) -> TranscriptionResult:
        segments_gen, info = self._model.transcribe(audio, language=language)
        return TranscriptionResult(
            text="".join(s.text for s in segments_gen),
            language=str(info.language),
        )

Enforce wrapper usage via ruff:

[tool.ruff.lint.flake8-tidy-imports.banned-api]
"faster_whisper" = { msg = "Use src/wrappers/whisper_wrapper instead" }

2. Type stubs in src/stubs/:

# src/stubs/some_library.pyi
def some_function(arg: str) -> list[int]: ...

Configure: stubPath = "src/stubs" in basedpyright config.

3. Inline type narrowing for one-off cases:

raw_value = untyped_lib.get_value()  # returns Any
assert isinstance(raw_value, str)    # narrows to str

TypeIs Guards

Note: TypeIs guards are unnecessary for msgspec-decoded data (already fully typed). Use them for narrowing in-memory objects of unknown type.

from typing import TypeIs, TypedDict, Required

class ValidResponse(TypedDict):
    status: Required[str]
    data: Required[dict[str, str | int | bool | list[str]]]

# Simplified for brevity — production code should use msgspec.convert() for full validation
def is_valid_response(obj: object) -> TypeIs[ValidResponse]:
    return (
        isinstance(obj, dict)
        and isinstance(obj.get("status"), str)
        and isinstance(obj.get("data"), dict)
    )

def process(response: object) -> Result[str, str]:
    if not is_valid_response(response):
        return Err("Invalid response format")
    return Ok(response["data"]["key"])  # type-safe access

Pattern Matching for Type Safety

@dataclass
class Success:
    value: str

@dataclass
class Failure:
    error: str
    code: int

type Outcome = Success | Failure

def handle(outcome: Outcome) -> str:
    match outcome:
        case Success(value=v):
            return f"OK: {v}"
        case Failure(error=e, code=c):
            return f"Error {c}: {e}"

# type: ignore Policy

Every # type: ignore must have a specific error code and rationale:

# BAD
result = some_call()  # type: ignore

# GOOD
result = some_call()  # type: ignore[no-any-return]  # rationale: lib returns Any, validated below
assert isinstance(result, ExpectedType)

TYPE_CHECKING Guard

For imports that cause circular dependencies or are only needed for annotations:

from __future__ import annotations
from typing import TYPE_CHECKING

if TYPE_CHECKING:
    from src.managers.audio_manager import AudioManager

class TranscriptionService:
    def __init__(self, audio: "AudioManager") -> None: ...

Note: TYPECHECKING is for forward references within the same layer. If two modules import each other, that's a circular dependency — fix the architecture (extract a Protocol, move types to a common module), don't hide the cycle with TYPECHECKING.

Error Handling

For the general principle of error handling as control flow, see myai's engineering-principles. Python-specific: use Result[T, E] from rusty-results for expected failures. Exceptions are reserved for programming errors (bugs).

Decision Table

Pattern

import msgspec
from rusty_results import Result, Ok, Err

class Config(msgspec.Struct):
    name: str

def load_config(path: Path) -> Result[Config, str]:
    if not path.exists():
        return Err(f"Config not found: {path}")
    try:
        return Ok(msgspec.json.decode(path.read_bytes(), type=Config))
    except (msgspec.DecodeError, OSError) as e:
        return Err(f"Failed to load: {e}")

Rules

1. Early returns — handle error first, keep success path linear:

``python result = loaddata(path) if result.iserr: return Err(f"Cannot proceed: {result.unwrap_err()}") data = result.unwrap() ``

1. Three error boundaries:

- Library: catch third-party exceptions → Result - Component: each subsystem returns Result to caller - Global: UI/CLI top-level catches everything, shows user message

1. Never swallow errors — no except: pass, no ignored Results. Every Result[T, E] must be checked — at minimum log the error + show toast/alert in GUI or print to CLI.

1. Cleanup on failure — if multi-step operation fails midway, clean up partial state

1. Custom error types for complex domains:

``python @dataclass class ConfigError: path: Path reason: str line: int | None = None ``

1. Handle received error values gracefully: show UI warning, or log, or do early return or propagate with context

Async Task Boundaries

Any coroutine launched via asyncio.ensurefuture() or createtask() is a fire-and-forget boundary. If an exception escapes, nobody retrieves it and the UI gets stuck in an intermediate state (e.g. "processing" forever).

Mandatory pattern: wrap the entire coroutine body in try/except Exception as a safety net:

async def _do_work(self, path: Path) -> None:
    try:
        await self._do_work_inner(path)
    except Exception as exc:
        logger.exception("Unexpected error for %s", path)
        self._set_error_state(path, f"Unexpected error: {exc}")

The inner method handles expected errors (Result checks, specific exceptions). The outer method guarantees the UI always transitions to a terminal state.

CLI Error Boundary

def main() -> int:
    result = run_app()
    if result.is_err:
        typer.echo(f"Error: {result.unwrap_err()}", err=True)
        return 1
    return 0

Async Patterns

When to Use

HTTP Requests

async def fetch_data(url: str) -> Result[bytes, str]:
    try:
        async with httpx.AsyncClient() as client:
            resp = await client.get(url, timeout=30.0)
            resp.raise_for_status()
            return Ok(resp.content)
    except httpx.HTTPError as e:
        return Err(f"HTTP error: {e}")

Subprocess Execution

async def run_command(args: list[str]) -> Result[str, str]:
    try:
        process = await asyncio.create_subprocess_exec(
            *args,
            stdout=asyncio.subprocess.PIPE,
            stderr=asyncio.subprocess.PIPE,
        )
        stdout, stderr = await process.communicate()
        if process.returncode != 0:
            return Err(f"Command failed: {stderr.decode()}")
        return Ok(stdout.decode())
    except FileNotFoundError:
        return Err(f"Command not found: {args[0]}")

Concurrent Operations

async def fetch_all(urls: list[str]) -> list[Result[bytes, str]]:
    return await asyncio.gather(*(fetch_data(url) for url in urls))

Timeouts

async def fetch_with_timeout(url: str) -> Result[bytes, str]:
    try:
        async with asyncio.timeout(10):
            return await fetch_data(url)
    except TimeoutError:
        return Err(f"Timeout fetching {url}")

Async Rules

1. Never call subprocess.run(), time.sleep(), or synchronous HTTP in async context
2. Never use shell=True in subprocess calls
3. Always use asyncio.createsubprocessexec() for subprocesses
4. Always use async with for resource management (HTTP clients, file handles)
5. Never mix sync and async in the same call chain
6. Always handle TimeoutError for operations that may hang

Code Style

Naming

Documentation

Google-style docstrings on public APIs. Comments explain why, not what.

Preconditions & Validation

For the general fail-fast principle, see myai's engineering-principles. Python-specific: validate at subsystem entry points using msgspec.Struct for data shape, isinstance/TypeIs for type narrowing, and Result[T,E] for expected validation failures. Check permissions, external deps, config validity, and input ranges before proceeding with business logic.

Architecture

For the general architecture separation principle, see myai's engineering-principles and architecting-changes. Python-specific layout:

Presentation (Qt GUI / CLI / API)
        |
        v
Domain (Managers, Models, Business Rules)
        |
        v
Utilities (Helpers, Wrappers, Common)

• Dependencies flow downward only
• UI is a plugin — adding CLI, GUI, or API should not change business logic
• Domain never imports from presentation
• For architecture decisions, load architecting-python-changes (which extends myai's architecting-changes)

Security

• No shell=True in subprocess calls
• Validate paths (symlink-aware):

``python def issafepath(path: Path, base: Path) -> bool: try: resolved = path.resolve(strict=True) return resolved.isrelativeto(base.resolve(strict=True)) except (OSError, ValueError): return False ``

• No hardcoded secrets — use environment variables or dotenv
• Input validation at system boundaries
• Never interpolate user input into subprocess commands
• Cleanup on failure — if multi-step operation fails midway, clean up partial state

Performance

• Clear code first, optimize only with profiling data
• slots on frequently instantiated dataclasses
• Batch I/O operations (don't read files one by one in a loop)
• Lazy loading for expensive resources (load on first use, not import time)
• Concurrent I/O with asyncio.gather()

Logging

import colorlog

def get_logger(name: str) -> logging.Logger:
    handler = colorlog.StreamHandler()
    handler.setFormatter(colorlog.ColoredFormatter(
        "%(log_color)s%(levelname)-8s%(reset)s %(name)s: %(message)s"
    ))
    logger = logging.getLogger(name)
    logger.addHandler(handler)
    logger.setLevel(logging.INFO)
    return logger

Usage: logger = get_logger(name)

Jinja2 Templating

Use when generating text output (HTML, configs, reports, markdown):

Tooling

Always enable -h for help — typer only supports --help by default:

app = typer.Typer(context_settings={"help_option_names": ["-h", "--help"]})

Run uv run poe lint_full continuously, not just at the end. Treat it as the main sweep for basedpyright, Ruff check/format, and custom linters; pair it with uv run poe test for verification.

Related myai Skills

• engineering-principles — Parent skill. Language-agnostic philosophy: pit of success, fail fast, error handling as control flow, testing philosophy, architecture separation.
• architecting-changes — For architecture decisions before writing code. Load architecting-python-changes for the Python-specific extension.
• code-simplification — For refactoring Python code for clarity.