name	write-numba-friendly-python-code
description	Skill to write Numba friendly Python code. Use it whenever you write or edit an `@njit` function in `nucs/`.

Write Numba friendly Python code

These apply to every @njit function in nucs/ — most of the codebase.

Use @njit(cache=True). cache=True is what makes warm starts fast; never remove it.
No Python objects in jitted code: no dict, no exceptions, no isinstance, no strings other than literals. Pass typed NDArrays and ints.
Mutate arrays in place (domains[i][MIN] = ...). Never rebind a slot (domains[i] = ...) — Numba can't always type that.
Functions passed as values go through _get_wrapper_address (see nucs/numba_helper.py); the address is recovered to a typed callable at runtime. If you're adding a callable-typed parameter, this is the mechanism.
When a Numba compile error is cryptic, re-run with NUMBA_DISABLE_JIT=1 — the real Python traceback points to the line.

name	write-numba-friendly-python-code
description	Skill to write Numba friendly Python code. Use it whenever you write or edit an `@njit` function in `nucs/`.

These apply to every @njit function in nucs/ — most of the codebase.

Use @njit(cache=True). cache=True is what makes warm starts fast; never remove it.
No Python objects in jitted code: no dict, no exceptions, no isinstance, no strings other than literals. Pass typed NDArrays and ints.
Mutate arrays in place (domains[i][MIN] = ...). Never rebind a slot (domains[i] = ...) — Numba can't always type that.
Functions passed as values go through _get_wrapper_address (see nucs/numba_helper.py); the address is recovered to a typed callable at runtime. If you're adding a callable-typed parameter, this is the mechanism.
When a Numba compile error is cryptic, re-run with NUMBA_DISABLE_JIT=1 — the real Python traceback points to the line.

write-numba-friendly-python-code