Kunda

Install

pip install kunda

Select an interpreter

python_for searches parent folders up to stop, then uses the supplied default or a virtual environment under roots. venv_env prepares the child’s environment.

selected = python_for(root/'src', stop=root)
choices = find_pythons(roots=[root], current=selected)
environment = venv_env(selected, env={'PATH': '/usr/bin'})
Path(selected).relative_to(root), [row['label'] for row in choices], environment['VIRTUAL_ENV']

(Path('.venv/bin/python'),
 ['this one', 'tmps35syx1f/.venv', 'python3 on PATH', 'python on PATH'],
 '/private/var/folders/kg/9vdw4mdd1fs58svgh4k1qhr09x7dqh/T/tmps35syx1f/.venv')

Run one kernel

Execution errors are returned in ExecOutcome and leave the kernel available.

kernel = Kernel(cwd=root, inspect=True)
await kernel.start()
outcome = await kernel.execute('value = 21 * 2; value')
completion = await kernel.complete('val', 3)
outcome.ok, outcome.text, completion['matches'][:3]

(True, '42', ['value'])

Keep several kernels alive

Each key has at most one kernel. RuntimeBroker enforces a process-wide limit, and idle=0 disables reaping.

class RustRunner: pass

broker = RuntimeBroker(max_kernels=12)
pool = KernelPool(broker=broker, idle=0,
                  runner_for=lambda lang: RustRunner if lang == 'rust' else None,
                  known_kernels={'julia': 'julia-1.10'})
broker.status(), pool._class_for({'lang': 'rust'}).__name__

({'live': 0, 'limit': 12, 'runtimes': []}, 'RustRunner')

Kernel support

Support probes do not install packages. Installation uses the host’s installed package versions.

support = kernel_support(python)
{k: v['available'] for k, v in support.items()}, installable(python)

({'ipykernel': False, 'ipymini': False}, True)

Inspect live variables

With inspect=True, Kunda starts Dhrishti inside the kernel and reads its registry.

await kernel.execute('visible_value = 42')
'name' in (await kernel.names()), 'visible_value' in (await kernel.names())

(False, True)

Build a host on Kunda

A host that starts its own helper processes asks Kunda the same questions a kernel does. runtime_python names the interpreter to spawn, which inside a frozen app is the bundled helper, and frozen_pythonpath gives the module paths that helper needs. kernel_spec builds the kernelspec from the host’s own table of languages. venv_roots lists the folders searched for a project’s environment. A host’s runner subclasses InspectorClient for the inspector calls and idle clock, and give_up_after bounds any close that might not answer. Each of these had a leading underscore before 0.0.4, and the old names still work.