[field-failure] no_distribution on Windows (py 3.11)

[github-actions]

Aggregated desktop bootstrap-failure telemetry crossed the reporting threshold for this signature.

signature	value
stage	bootstrap_failed
failure class	no_distribution
OS	Windows
Python	3.11
distinct installs (30d)	1
events (30d)	1
last seen	2026-09-07T11:23:32.215661+00:00
latest shell version seen	0.12.826

These are aggregates over a closed enum — no machine ever sends paths, usernames, or log text. The class is the same one _explain_pip_failure shows on the machine's splash; desktop/app.py maps each class to its likely fix surface.

Live view: https://app.clawmetry.com/api/desktop/_failures

Auto-filed by field-failure-issues.yml (Requirement: Field Failure Reporting and Auto-Triage, AC-FFR-002).

[vivekchand]

Investigated, could not build a failing reproduction, standing down without a fix.

no_distribution on Windows py3.11 is already the exact class _retry_on_known_good_python in desktop/app.py exists to handle (added for #5628/#5639/#5711): it auto-retries against the pinned python.org 3.12 interpreter via winget, and the fallback hint already avoids blaming the interpreter version, naming the two real causes instead (a private mirror that doesn't proxy pypi.org, or an interpreter PyPI ships no wheel for). tests/test_desktop_bootstrap_resilience.py has extensive regression coverage of this exact classify+retry path, including the invariant that no hint ever blames the Python version.

Checked setup.py's environment markers for cryptography/cffi/truststore against Python 3.11: all resolve correctly there (cffi>=2, cryptography>=50.0.0), so there's no marker bug denying py3.11 a wheel of our own dependencies.

With only 1 install and 1 event in 30 days, and the telemetry a closed enum carrying no pip log text by design (so this report cannot tell us which of the two named causes it was), I can't distinguish an actual code defect from the environmental cases the code already handles and reports honestly. What's needed to go further: the pip output tail from bootstrap.log on the affected machine, which never leaves it.

---

Generated by Claude Code

[vivekchand]

Hi! Took a careful look at the telemetry and the code path. Here's the plan:

The problem
When Python 3.12 is installed system-wide (not per-user) on a Windows machine that also has Python 3.11, _known_good_python() in desktop/app.py only checks the per-user %LOCALAPPDATA%\Programs\Python\Python312\python.exe path — it misses any Python 3.12 that lives elsewhere (system-wide install, a different user path, etc.). The retry mechanism then silently aborts (if not good: return rc, out), bootstrap_python_version stays "3.11", and the field-failure report keeps firing on every launch with bootstrap_python = 3.11.

Approach

1. desktop/app.py:830–855 — In _known_good_python(), after the per-user LOCALAPPDATA path check comes up empty or fails the probe, add a fallback that asks the py launcher with an explicit version pin: py -3.12 -c "import sys, venv, ensurepip; print(sys.executable)". Unlike bare py (which resolves to the newest interpreter — exactly the one whose wheels are missing — per the existing comment), py -3.12 resolves only to Python 3.12 regardless of what else is installed, covering system-wide installs that winget did not make.
2. tests/test_desktop_bootstrap_resilience.py:523+ — Add two tests: one that verifies the py-launcher fallback is used when the per-user path is absent, and one that verifies subprocess is never spawned when the py launcher itself is not on PATH.

Test plan

• Both new tests use monkeypatch to control shutil.which("py") and subprocess.run, run on Linux CI without touching real interpreters.
• Existing tests (test_original_failure_survives_when_no_interpreter_can_be_provided, test_no_wheel_failure_installs_a_supported_python_and_retries, etc.) mock _known_good_python entirely and are unaffected.
• python3 -c 'import ast; ast.parse(open("desktop/app.py").read())' to confirm no syntax errors.

Estimated change
~35 lines across 2 files (desktop/app.py, tests/test_desktop_bootstrap_resilience.py).

Drafting a PR now — will link here when ready. If you'd rather drive this yourself, drop a comment and I'll back off.

---

Generated by Claude Code

[vivekchand]

Draft PR opened: #5864 — fix(desktop): probe py -3.12 launcher as fallback in _known_good_python()

What changed (2 files, ~73 net lines):

• desktop/app.py — after the per-user %LOCALAPPDATA%\Programs\Python\Python312\python.exe check, the function now falls through to shutil.which("py") + py -3.12 when that path is absent or fails the probe. The explicit -3.12 flag avoids the "newest interpreter" trap that bare py poses.
• tests/test_desktop_bootstrap_resilience.py — two new tests covering the system-wide-install path and the no-py-launcher path.

65 tests passed locally before the PR was opened.

---

Generated by Claude Code