Imported from dylan-sutton-chavez/edge-python (skill/SKILL.md). Install upstream with npx skills add dylan-sutton-chavez/edge-python --skill skill. Copyright stays with the author.

Edge Python

This document is self-verifying and its examples follow the cells v1 grammar. A python or yml block followed immediately by a text block is a runnable cell, and the skill crate in this directory executes every cell through the edge CLI and compares it against the text block. The tag on the text block picks the engine, Output runs on both, Native on the native engine only, Web on the web runtime only, and Error expects a failing run whose stderr contains the given text. A python block tagged skip never runs on any engine and never pairs with a text block, and it always says why with one comment at the exact construct that is nondeterministic. A yml block tagged actor runs a trusted actor pool through edge actor, while one tagged untrusted runs eval groups. Any python block without a text pair is illustrative only. Verify the whole file from the repository root with cargo run -p skill -- skill/SKILL.md --engine both.

Edge Python is a sandboxed Python subset compiled in a single pass to bytecode and executed by a stack VM. It runs in the browser as WebAssembly and in the edge CLI as an in-process native engine. There is no bundled stdlib, every module is an external package resolved at compile time. Programs are deterministic, there is no file, network or environment access unless a system module grants it.

Use this skill to write correct Edge Python on the first try. The language looks like Python 3 but is a strict subset, and the differences matter more than the similarities. Read the delta section before writing non-trivial code.

The working loop

A project is any folder with .py files and an optional packages.json. The loop is always the same.

1. Write or edit the .py files.
2. Run the entry point with edge run main.py.
3. Add tests in *_test.py files and run edge test.
4. Pack a release with edge build when the program must run elsewhere.

edge init myapp        # scaffold main.py, packages.json and index.html
cd myapp
edge run main.py       # execute in the native engine
edge test              # discover and run every *_test.py
edge build             # pack a standalone ./app.edge binary

Piping a script works too, which is how the cells of this document run.

echo 'print(6 * 7)' | edge run

When a file path is given, piped stdin instead feeds input(), one line per call.

CLI reference

Bare edge prints help and exits 0. edge -v prints the version. Ctrl+C exits 130. Errors print to stderr and exit 1.

Global flags

edge run

edge run [file] executes a .py script, a packed .edge binary or a .package bundle, auto-detected by content. With no file it reads the script from stdin. A bare edge run in a terminal with no pipe errors. edge run -c 'print(1)' runs inline code instead of a file or stdin, and piped stdin then feeds input().

Native-only flags, combining them with --web is an error.

raise SystemExit(code) with no argument or an integer exits cleanly with that code. Any other uncaught error prints a traceback and exits 1.

edge repl

A persistent interpreter across prompts. Imports, definitions and mutations survive between lines, and an input that raises keeps the effects made before the error. One line is one eval, so compound statements go on a single line. Expression results are not auto-printed, use print(). Dot commands are .reset to wipe state and .exit to quit. History lives for the session only.

edge test

edge test [path] discovers *_test.py recursively, skipping hidden dirs, node_modules, target and dist. A file argument runs exactly that file. Each file executes in a fresh interpreter and state never leaks between files. Exit code is 0 when everything passes, 1 when a file fails or no tests are found, 2 when the engine cannot start. See the test package section for the API.

edge init, edge add, edge remove

edge init [name] scaffolds main.py, packages.json and index.html, with --bare skipping the HTML. edge add json network writes manifest entries for known packages, and edge add foo=<url> registers a custom URL, a .wasm or .py URL is treated as a std package and anything else as a system module. edge remove deletes entries. Unknown names abort the whole command before any write.

edge serve

A static dev server with live reload for the current directory. --host defaults to 127.0.0.1, --port to 5173, --open opens a browser.

edge build

Three mutually exclusive modes.

--out <path> overrides the default. The bundle contains every .py under the project plus packages.json, and the entry is main.py, app.py or index.py when present. An .edge binary accepts only the snapshot flags --save-state, --restore-state, --preempt and --events.

edge actor

edge actor <file> runs a actor pool from a actor.yml manifest. See the actors section for the schema and the two execution models.

edge uninstall

Interactive removal of the binary, PATH entries and caches.

Environment variables

The Python delta

Edge Python parses like Python 3 but deliberately drops parts of the language. This section is the one to internalize, because everything here is valid CPython that fails or behaves differently in Edge Python.

Not supported at all

• No stdlib. Every module is an external package, so import os, import sys and import asyncio fail at compile time.
• No dynamic code. exec, eval, compile and __import__ do not exist.
• No open. input() reads from a host fed buffer with no prompt argument.

open("data.txt")

NameError

• No complex numbers. 1j lexes as 1 followed by the name j.
• No metaclasses, descriptors, __slots__, __new__, __init_subclass__ or __set_name__. Some parse but are never dispatched.
• No augmented assignment dunders. a += b desugars to a = a + b for user classes, except list += and set |=, &=, ^=, -= which mutate in place.
• No bytearray and no memoryview.
• No exception chaining. raise X from Y evaluates Y but the cause is discarded.
• No gen.send, gen.throw or gen.close. Generators are one-way producers.

Eager where Python is lazy

Generator expressions lower eagerly to lists. Write def plus yield when real laziness matters.

g = (i * 2 for i in range(3))
print(g)

[0, 2, 4]

reversed, map, filter and enumerate return eager lists, not iterators, and iter(x) materializes a snapshot.

print(map(str, [1, 2]))
print(reversed([1, 2, 3]))

['1', '2']
[3, 2, 1]

Dict views are concrete list snapshots taken at call time, not live views.

d = {"a": 1}
keys = d.keys()
d["b"] = 2
print(keys)

['a']

Numbers are bounded

Integers are 48-bit inline with automatic promotion to 128-bit. Past ±2^127 the run raises OverflowError.

print(2**126)

85070591730234615865843651857942052864

print(2**127)

OverflowError

pow(a, b, m) requires a modulus below 2^63, and the int_to_bytes and int_from_bytes builtins cap at 8 bytes while the int.to_bytes and int.from_bytes methods do not.

Reduced pattern matching

match supports literal patterns, captures, the _ wildcard, OR patterns with |, guards with if, and flat sequence patterns like [x, y] or [first, *rest]. Sequence patterns match only list and tuple subjects. There are no nested sequence patterns, no mapping patterns, no class patterns and no as captures.

def describe(value):
    match value:
        case 0 | 1:
            return "small"
        case [first, *rest]:
            return f"list of {len(rest) + 1}"
        case n if n < 0:
            return "negative"
        case _:
            return "other"

print(describe([10, 20, 30]))

list of 3

match is a soft keyword. A parenthesized subject like match (a, b): works as a statement, and match(a, b) in expression position still parses as a call.

Missing pieces by type

• tuple and frozenset have no methods at all. (1, 2).count(1) raises AttributeError, use operators or convert to list or set first.
• str lacks translate, maketrans, format_map, isascii, isidentifier, isnumeric, isdecimal and isprintable. str.format accepts positional fields only, no {name} keyword fields.
• bytes.split requires an explicit separator, bytes.replace has no count, and codecs are limited to utf-8 and ascii with strict, ignore and replace error handling.
• zip has no strict flag. round uses ties-to-even and always returns int for one argument, float for two.

Async without asyncio

There is no asyncio and no event loop object. The async primitives are top-level builtins, run, gather, sleep, with_timeout, cancel, frame and receive. There are no async comprehensions, no async dunders and no background tasks, create_task does not exist. See the async section.

Compile time versus run time

Import failures and syntax errors are compile-time diagnostics and can never be caught with try. Everything else raises normal catchable exceptions at run time.

Imports

Every import resolves at compile time through a host resolver. The compiler flattens each module into the bytecode and the VM fetches nothing at run time.

import math
from json import dumps, loads
from math import sqrt as root
from re import *

print(root(16.0), loads(dumps({"ok": True}))["ok"])

4.0 True

Dotted specs import files. A leading dot anchors at the importing file, one extra dot per directory up. Without it the spec anchors at the nearest packages.json dir. The .py suffix is implicit.

from .lib.helpers import slugify
from ..shared.util import chunks
from lib.helpers import slugify as sl

Not supported. from . import x and any form of dynamic import.

Bare names resolve through packages.json, walking up from the importing file with the nearest manifest winning. The manifest maps names to paths or URLs under imports, JS system modules under system, and may extend a parent manifest.

{
  "imports": {
    "utils": "./lib/utils.py",
    "mypkg": "https://example.com/mypkg.wasm"
  },
  "system": {
    "charts": "https://example.com/charts.js"
  }
}

The names json, re, math, struct, test, dom, network, storage and time resolve with no manifest at all, as official defaults. Modules are singletons with shared mutable state, an import cycle raises RuntimeError at startup, and inside an imported module __name__ is its canonical spec so if __name__ == "__main__": blocks are skipped on import. import_module(name) looks up a module already bound by a plain import in scope.

Builtins

The global namespace holds exactly 68 builtin functions, the type objects, the exception classes, NotImplemented, __name__ and the async primitives. Nothing else exists, and names like dir, help or exit are simply undefined.

Output and input

print(*args, sep=' ', end='\n') accepts file and flush and ignores them. input() reads one host fed line with no prompt.

print("a", "b", sep="-", end="!\n")

a-b!

Numeric

abs, round, min, max, sum, pow, divmod, bin, oct, hex. round breaks ties to even. min and max accept variadic args or one iterable plus key and default.

print(round(2.5), round(3.5), round(1.55, 1))
print(divmod(7, 2), max("xy", "abcde", key=len))
print(pow(2, 10, 100))

2 4 1.6
(3, 1) abcde
24

Conversion

int, float, str, bool, list, tuple, set, frozenset, dict, bytes, chr, ord. int truncates toward zero and parses bases 2 to 36 or 0 for auto-detect. int("nan") style failures raise ValueError and int(float("inf")) raises OverflowError.

print(int("ff", 16), int("0b101", 0), int(-3.7))
print(float("inf") > 1e308, ord("A"), chr(97))

255 5 -3
True 65 a

Iteration

len, range, sorted, reversed, enumerate, zip, iter, next, map, filter, all, any, slice. range is genuinely lazy and everything else eager, see the delta section.

print(list(enumerate("ab", start=1)))
print(zip([1, 2, 3], "ab"))
print(sorted([3, 1, 2], reverse=True), any([0, "", 3]))

[(1, 'a'), (2, 'b')]
[(1, 'a'), (2, 'b')]
[3, 2, 1] True

Types and attributes

type, object, isinstance, issubclass, callable, id, hash, repr, format, getattr, hasattr, setattr, delattr, vars, globals, locals, import_module, super, property, staticmethod, classmethod. isinstance accepts a tuple of types and bool is a subclass of int. vars(x) returns a snapshot of instance attributes, and globals() and locals() return copies whose mutation binds nothing.

print(isinstance(True, int), callable(len))
print(format(255, "08x"), repr("it's"))

True True
000000ff "it's"

Bytes helpers

bytes_fromhex, int_from_bytes(b, order) and int_to_bytes(n, length, order) with a limit of 8 bytes and unsigned values.

print(bytes_fromhex("ff00"), int_from_bytes(b"\x01\x00", "little"))

b'\xff\x00' 1

Exceptions

The catchable tree under Exception is ArithmeticError with OverflowError and ZeroDivisionError, LookupError with IndexError and KeyError, RuntimeError with RecursionError and NotImplementedError, plus ValueError, TypeError, AttributeError, NameError, OSError, StopIteration, StopAsyncIteration, AssertionError, MemoryError and TimeoutError. Under BaseException sit SystemExit and CancelledError, which except Exception does not catch.

Handlers name one class, a tuple or nothing, and a bare except must come last. except X as e binds the exception and e.args is its argument tuple. finally runs on every exit path including return, break and continue.

try:
    {}["missing"]
except (KeyError, IndexError) as e:
    print(type(e).__name__, e.args)
finally:
    print("always")

KeyError ('missing',)
always

User exception classes support inheritance among themselves for except matching but do not join the builtin tree.

Type methods

Methods live on the builtin types. tuple, frozenset, bool and NoneType have none.

str

encode, upper, lower, strip, lstrip, rstrip, capitalize, title, casefold, swapcase, isdigit, isalpha, isalnum, isspace, isupper, islower, istitle, startswith, endswith, find, rfind, index, rindex, count, split, rsplit, join, replace, removeprefix, removesuffix, splitlines, partition, rpartition, center, ljust, rjust, zfill, expandtabs, format. Indices count code points. startswith and endswith accept a tuple of prefixes. format takes positional fields with specs, never keyword fields.

print(" hello ".strip(), "a,b,c".split(",", 1))
print("-".join(["x", "y"]), "Hello".casefold(), "{0}{1}{0}".format("a", "b"))
print("file.py".removesuffix(".py"), "5".zfill(3))

hello ['a', 'b,c']
x-y hello aba
file 005

list

append, extend, insert, remove, pop, clear, copy, reverse, index, count, sort with key and reverse. Slice assignment resizes, and += extends in place.

xs = ["bb", "a", "ccc"]
xs.sort(key=len)
xs[1:1] = ["z"]
print(xs, xs.pop())

['a', 'z', 'bb'] ccc

dict

Insertion ordered. keys, values, items return list snapshots, plus get, update, pop, popitem, setdefault, fromkeys, copy, clear. popitem removes the most recently inserted pair. Numerically equal keys collapse, so 1, 1.0 and True are one key.

d = dict.fromkeys(["a", "b"], 0)
d.update({"c": 1})
print(d.pop("a"), d.setdefault("d", 4), list(d))
print({1: "x", True: "y"})

0 4 ['b', 'c', 'd']
{1: 'y'}

set

add, remove, discard, pop, clear, update, copy, union, intersection, difference, symmetric_difference, intersection_update, difference_update, symmetric_difference_update, issubset, issuperset, isdisjoint. Named methods accept any iterable while the operators require sets on both sides. Iteration order is hash based, never rely on it and print through sorted.

print(sorted({1, 2, 3} & {2, 3, 4}))
print({1, 2}.issubset({1, 2, 3}), {1}.isdisjoint({2}))

[2, 3]
True True

int and float

int has bit_length, bit_count, to_bytes and the classmethod from_bytes. float has is_integer.

print((255).bit_length(), (5).bit_count(), (258).to_bytes(2).hex())
print((4.0).is_integer(), int.from_bytes(b"\x01\x02", "big"))

8 2 0102
True 258

bytes

decode, encode via str, hex, fromhex, startswith, endswith, find, index, count, replace, split, lower, upper, strip, lstrip, rstrip, join. Case methods are ASCII only.

print(b"\x00\x01".hex(), b"a,b".split(b","), b"ABC".lower())

0001 [b'a', b'b'] b'abc'

Functions and classes

Functions support defaults, keyword arguments, *args, **kwargs and bare-* keyword-only parameters, plus call-site unpacking. The positional-only marker / parses but is not enforced, so never rely on it. Lambdas are single expressions. Decorators work on functions and classes, stacked bottom-up. Closures capture variables by reference, see the gotchas section.

def greet(name, *, punct="!"):
    return f"hi {name}{punct}"

print(greet("edge", punct="?"))

hi edge?

Classes support single and multiple inheritance with C3 linearization, zero-argument super(), property with setters, staticmethod and classmethod, and class decorators. There is no two-argument super() form. Dunders are looked up on the class, assigning one on an instance has no effect.

The supported dunders are __init__, __call__, __repr__, __str__, __format__, __bool__, __len__, __hash__, __iter__, __next__, __getitem__, __setitem__, __delitem__, __contains__, __getattr__, __enter__, __exit__, __index__, __int__, __float__, __abs__, the arithmetic and bitwise operators with their reflected forms, and the six comparisons. Returning NotImplemented from an arithmetic dunder triggers the reflected fallback.

class Vector:
    def __init__(self, x, y):
        self.x = x
        self.y = y

    def __add__(self, other):
        return Vector(self.x + other.x, self.y + other.y)

    def __repr__(self):
        return f"Vector({self.x}, {self.y})"

class Scaled(Vector):
    pass

print(Scaled(1, 2) + Vector(10, 20))

Vector(11, 22)

Context managers implement __enter__ and __exit__(exc_type, exc_value, traceback) where the traceback argument is always None. A truthy __exit__ suppresses the exception.

Pure functions are memoized automatically after two identical calls. The VM detects purity by the absence of I/O, mutation, raising and free-name reads, so naive recursive code is fast and side-effecting calls skip the cache safely.

Async

The module body runs as an implicit coroutine, so top-level code can call suspending functions directly. A plain def called from a coroutine can also call them.

async def dbl(n):
    await sleep(0)
    return n * 2

print(gather(dbl(1), dbl(2), dbl(3)))

[2, 4, 6]

The primitives are builtins, no import needed.

async def slow():
    await sleep(10)
    return "done"

try:
    with_timeout(0.01, slow())
except TimeoutError:
    print("timed out")

timed out

Scheduling is cooperative. A tight loop without a suspending call cannot be cancelled or preempted unless the engine runs with --preempt. There is no create_task and no preemption between coroutines.

Snapshots

The native engine can serialize the full interpreter state, heap, globals, suspended coroutines and scheduler, and restore it later. This is how long-running or event-driven programs survive process restarts.

edge run app.py --save-state state.bin     # writes the blob when the script parks
edge run --restore-state state.bin         # resumes from the blob
edge run app.py --preempt 500              # makes even while-True snapshottable

A snapshot is taken when the script parks on a wait the engine cannot serve, for example receive() with no events left. Without --save-state such a park is an error. The blob embeds a bytecode fingerprint and only restores into the same program. Feed a resumed run with --events file, one receive() line per call.

Std packages

Five official packages import by bare name with no manifest, on both engines. edge add <name> writes the manifest entry explicitly when a project should pin it.

json

loads(s) with optional object_hook, object_pairs_hook, parse_float, parse_int and parse_constant. dumps(obj) with indent, sort_keys, ensure_ascii, check_circular, allow_nan, skipkeys, default, separators and cls. Parse failures raise ValueError, non-serializable values raise TypeError unless default handles them. Integers round-trip at 128-bit and non-finite floats map to NaN and Infinity.

import json

data = json.loads('{"n": 21, "xs": [1, 2]}')
print(json.dumps(data, sort_keys=True))
print(json.dumps({"bad": object()}, default=str))

{"n":21,"xs":[1,2]}
{"bad":"<object instance>"}

math

Constants pi, e, tau, inf, nan. Functions sqrt, cbrt, exp, exp2, expm1, pow, log, log2, log10, log1p, the trig and hyperbolic families, atan2, hypot, dist, degrees, radians, erf, erfc, gamma, lgamma, fabs, fmod, remainder, copysign, ldexp, modf, frexp, floor, ceil, trunc, isnan, isinf, isfinite, fsum, prod, and the integer functions factorial, gcd, lcm, isqrt, comb, perm at 128-bit. Domain errors raise ValueError. A batch family (sqrt_all, add_all, dot_all, matvec and friends) operates on bytes buffers of little-endian f64, pair it with struct.pack.

import math

print(math.gcd(12, 18), math.factorial(10), math.isqrt(17))
print(math.floor(math.pi), math.isfinite(math.inf))

6 3628800 4
3 False

re

Backtracking engine with a step budget that raises RuntimeError against catastrophic backtracking. Module functions take (pattern, string) and compile(pattern) returns a pattern object with the same operations as methods. There are no Match objects, matchers return the matched string or None.

Supported syntax covers classes, anchors, quantifiers with lazy forms, capturing, non-capturing and named groups, backreferences, alternation, lookahead and fixed-width lookbehind, plus inline flags (?i), (?s) and (?m). Not supported, \p{...}, atomic groups, possessive quantifiers, conditionals and scoped flags.

import re

print(re.findall(r"\d+", "a1b22"))
print(re.sub(r"(\w+)@(\w+)", r"\2@\1", "user@host"))
print(re.groups(r"(\d+)-(\d+)", "12-34"))

['1', '22']
host@user
['12', '34']

struct

pack(fmt, *values) returns bytes, unpack(fmt, data) returns a list, calcsize(fmt) returns an int. Codes are x b B ? h H i I q Q f d with repeat counts, prefixes < (default), =, > and !. String codes s and p, half-float e, native-size n and N, and the pack_into family are not implemented.

import struct

buf = struct.pack("<hh", 258, -1)
print(buf.hex(), struct.unpack("<hh", buf), struct.calcsize("<hh"))

0201ffff [258, -1] 4

test

The test framework, imported by bare name and driven by edge test discovery. Test files do not need to call run() themselves, the runner evaluates the file and then invokes the driver, so a file that registers no tests fails.

• @fixture registers a factory under its function name, built fresh per test.
• @test("description", *uses) registers a test and injects named fixtures by keyword.
• with raises(ExcType): asserts the block raises, accepts a class or a tuple.
• Assertions are plain assert.
• run() executes everything registered, prints verdicts and raises SystemExit(0) or SystemExit(1).

from test import fixture, test, raises, run

@fixture
def numbers():
    return [1, 2, 3]

@test("sum adds up", "numbers")
def total(numbers):
    assert sum(numbers) == 6

@test("division by zero raises")
def div():
    with raises(ZeroDivisionError):
        1 / 0

run()

PASS - sum adds up
PASS - division by zero raises
2 passed, 0 failed

System modules

Four system libraries plus the actor module. Availability differs by engine, and importing a web-only module natively is a compile-time error telling you to rerun with --web.

time

time(), time_ns(), monotonic(), monotonic_ns(), perf_counter(), perf_counter_ns(), and a suspending sleep(secs). gmtime and localtime return a JSON string of the nine struct_time fields, decode it with json.loads. mktime, strftime, strptime, asctime and ctime convert between forms. timezone(), altzone(), daylight() and tzname() are calls. The native engine has no timezone database, tzname() is always "UTC" and localtime equals gmtime.

from time import tzname, time

print(tzname())
print(type(time()).__name__)

UTC
float

network

fetch(url[, options_json]) returns a JSON string with id, ok, status, headers and body. fetch_text and fetch_json return the body directly and raise on non-2xx responses. All three suspend until the response arrives. abort_request(id) cancels an in-flight request and exists only on the web runtime. WebSockets use ws_open, ws_send, ws_close and ws_state, SSE uses sse_open, sse_close and sse_state, and both stream events through receive() as JSON payloads with a type field.

from network import fetch_json

data = fetch_json("https://api.example.com/items")
items = json.loads(data)

storage

Web only. Synchronous key-value access through local_get, local_set, local_remove, local_clear, local_keys and the session_* twins, values are strings so encode structured data with json.dumps. IndexedDB through suspending calls, idb_open, idb_put, idb_get, idb_delete, idb_keys and idb_close.

import storage

storage.local_set("k", "v")
print(storage.local_get("k"))

v

dom

Web only. Handles are opaque ints, multi-result queries return CSV strings of handles, structured results return JSON strings, and async results arrive through receive(). The surface covers selection and traversal (query, query_all, closest, parent, children, siblings), creation and mutation (create_element, append_child, insert_before, remove, replace_children, clone_node), content and attributes (get_text, set_text, get_html, set_html, get_attribute, set_attribute, class and data helpers), style and layout (set_style, rect, scroll_top, focus), events (bind_event, unbind_event, dispatch_event, click), forms and files, observers, animations, media and platform dialogs. A batch() context manager buffers the mutating calls and applies them with one host call on exit.

import dom

print(dom.tag_name(dom.body()))

body

Actors

edge actor runs many isolated programs as cooperative actors over a few threads, share-nothing with message passing. There are two execution models and the manifest chooses per group.

runtime:
  listen: tcp://127.0.0.1:7777   # optional, its presence makes the actor a live server
  durable: tmp/actor/log         # WAL path, replays unprocessed messages on restart
  schedulers: auto               # or a fixed number
  max_actors: 1000000

groups:
  actor:
    run: app                     # script path or project directory
    replicas: 100                # ceiling, actors spawn on demand
    retry: 2                     # re-deliveries before a message is dropped dead
    seed: ["first message"]      # delivered before the pool starts
    out: stdout                  # stdout, null or file://path
    limits:                      # per-actor sandbox overrides
      heap: 65536
      preempt: 500

Each group picks exactly one of run, code or eval: true. Without listen: the pool runs until every message is processed and exits.

The trusted model

Actors keep state between messages, pick work up with the receive() builtin and forward with send from the actor module, strings only, never blocking.

groups:
  actor:
    code: |
      msg = receive()
      print(f"got {msg}")
    replicas: 4
    seed: ["hello"]

got hello

The untrusted model

eval: true groups compile each message as its own program in a fresh interpreter, on a thread locked to a seccomp allowlist. No state survives between messages, the actor module is unavailable, and neither the code nor a native plugin it loads can open a socket, run a process, or make any syscall outside pure computation. The allowlist needs Linux, so an eval group is refused elsewhere. A code or run group is trusted instead, it runs in the host process with no syscall isolation, so reach for eval when the code is not yours.

groups:
  guest:
    eval: true
    seed: ["print(6 * 7)"]

42

With listen: the actor accepts one <group> <body> line per TCP message and exposes an HTTP control endpoint, GET /stats, POST /pub/<group> and POST /eval/<group> for eval groups. Untrusted bundles arrive as base64 .package payloads behind an EDGEPKG: marker and run materialized in an isolated temp dir.

Semantics that surprise Python programmers

Closures capture loop variables by reference. Bind the value with a default argument when building functions in a loop.

fns = [(lambda i=i: i) for i in range(3)]
print([f() for f in fns])

[0, 1, 2]

Inline integers compare by value under is, so is on numbers does not mean identity. Reserve is for None and sentinels.

a = 1000
b = 1000
print(a is b)

True

List += and set |=, &=, ^=, -= mutate in place and aliases see the change. Every other augmented assignment rebinds.

a = [1]
b = a
a += [2]
print(b)

[1, 2]

Set iteration and repr order is hash based. Always present sets through sorted.

s = {"a", "b", "c"}  # skip: set iteration order is hash based
print(s)

id() reuses heap slots and varies between runs. Never print it in examples or tests.

print(id(object()))  # skip: heap slots are reused, the value changes between runs

Truthiness follows Python, the falsy set is None, False, 0, 0.0, "", b"", [], (), {}, set(), frozenset() and range(0). bool subclasses int so True + True == 2. User objects in builtin dicts and sets compare by identity even when they define __hash__ and __eq__, so look them up by the same reference. len on strings counts code points. Same source and input give the same output on every run, id() aside.

Sandbox limits

Programs run under a fixed budget. Exceeding one raises the matching exception, catchable except the op-limit RuntimeError whose handler re-raises on its first operation because the budget is still exhausted.