everyinc__compound-engineering-plugin/docs/solutions/conventions/resolve-python-interpreter-not-python3.md

title, date, category, module, problem_type, component, severity, applies_when, tags

title	date	category	module	problem_type	component	severity	applies_when	tags
Resolve the Python interpreter by probing execution — never hardcode `python3` in agent-facing prose	2026-07-24	conventions	skills (agent-facing reference prose and SKILL.md shell blocks across every skill with bundled scripts)	convention	tooling	high	Writing a SKILL.md or reference doc that instructs an agent to run a bundled script A bundled Python script works when run by hand but the documented invocation fails Supporting native Windows contributors (not WSL, not Git Bash with a POSIX Python) A feature degrades to a fallback path on one platform and never reports why	windows python interpreter-resolution portability skill-authoring silent-failure store-stub

Resolve the Python interpreter by probing execution — never hardcode python3 in agent-facing prose

Context

While adding native Windows support to peer-job-runner.py (issue #1243), the runner itself was made to work — verified end to end on Windows 11. The feature still did not work, because every skill that invokes it tells the agent to run:

python3 "$SKILL_DIR/scripts/peer-job-runner.py" start ...

On native Windows, python3 resolves to the Microsoft Store App Execution Alias stub. It prints an install advertisement and exits non-zero without running Python. The standard python.org Windows installer creates python.exe and py.exe — it never creates python3.exe. So this is the default state of an extremely common setup, not an edge case.

Observed on Windows 11 / Python 3.11:

$ python3 --version
Python was not found; run without arguments to install from the Microsoft Store, ...
$ echo $?
49
$ python --version
Python 3.11.0

The failure surfaced while attempting a sanctioned cross-model review pass: it died at the interpreter, before the runner's own preflight ever ran.

Guidance

Do not hardcode an interpreter name in prose an agent will execute. Resolve it, and resolve it by probing execution, not presence.

PY="$(for c in python3 python py; do command -v "$c" >/dev/null 2>&1 && "$c" -c '' >/dev/null 2>&1 && { echo "$c"; break; }; done)"; [ -n "$PY" ] || { echo "no working Python 3 interpreter on PATH" >&2; exit 1; };

Then invoke "$PY" "$SKILL_DIR/scripts/…".

Three rules make this work:

1. Probe execution, not existence. This is the whole trap. command -v python3 succeeds — the Store stub is a real file on PATH. Every existence check (command -v, which, test -x) passes and the call still fails. Only actually running the interpreter ("$c" -c '') distinguishes them.
2. Repeat the resolution in every self-contained shell block. Agent harnesses run each tool call in a fresh shell, so a $PY exported in one block does not exist in the next. The repetition is deliberate, not redundancy to factor out — the same reason SKILL_DIR is already set inline in every block in these docs.
3. Fail loudly when nothing resolves. A silent fallback to a broken interpreter is how this bug survived; an explicit non-zero exit with a message is the point.

Order matters: try python3 first so POSIX hosts keep their canonical name, then python, then py (the Windows launcher, the most reliable there). Note python can also be a Store stub if python.org Python is not installed — which is why the probe, not the order, is what guarantees correctness.

Why This Matters

The failure is silent in the worst way: the calling workflow degrades gracefully. In ce-code-review, a failed cross-model dispatch falls back to the in-process adversarial reviewer and the run completes normally — so a Windows user gets a quietly weaker review forever, with no error and no indication that an entire independent-model pass never ran.

It also invalidates smoke evidence. This project's own Phase C smoke commands invoke python $runner start ..., so the runner passed its verification while the shipped invocation path — the one an agent actually follows — stayed broken. Verify the literal documented invocation, not a hand-adjusted variant.

When to Apply

Any time agent-facing prose or a bundled script names an interpreter or external tool that the host may resolve differently. Python is the instance encountered here; the class is larger. The general form: the name on PATH is not proof the thing runs.

Related repo scope: cross-model review/elevation paths and the remaining bundled-script invocation sites (pr-snapshot, sweep-state.py, validate-*.py, session-history extractors, ce-optimize shell helpers) reuse this exact snippet rather than inventing a variant. Issue #1247 tracks the Store-stub failure mode.

Examples

Before — works on macOS/Linux, silently unreachable on native Windows:

SKILL_DIR="<absolute path…>";
python3 "$SKILL_DIR/scripts/peer-job-runner.py" status "<job-id>" --json

After — self-contained, correct on every host:

SKILL_DIR="<absolute path…>";
PY="$(for c in python3 python py; do command -v "$c" >/dev/null 2>&1 && "$c" -c '' >/dev/null 2>&1 && { echo "$c"; break; }; done)"; [ -n "$PY" ] || { echo "no working Python 3 interpreter on PATH" >&2; exit 1; };
"$PY" "$SKILL_DIR/scripts/peer-job-runner.py" status "<job-id>" --json

The wrong fix, for the record — this passes and still breaks, because the stub satisfies it:

PY=$(command -v python3 || command -v python)   # WRONG: existence, not execution