[cgfixit]

Proposed changes

tests/test_installer_python_contract.py::test_powershell_probes_exact_py312_launcher_and_rejects_313 hit its 60s subprocess.TimeoutExpired on the windows-latest leg of #1722 (run 39416581582). #1722 only touched utils/health.py. Main's Windows leg was green on the same tree, and the one re-run passed. The failure showed nothing useful, because TimeoutExpired discards the installer's partial output, so there was no way to see where it stalled.

Both installer runners, _run_macos_installer and _run_powershell_installer, now go through one helper, _run_installer:

Change	Effect
_INSTALLER_TIMEOUT_SEC = 180 (was 60, in two places)	headroom for Windows PowerShell 5.1 start-up and module autoload on a loaded runner
stdin=subprocess.DEVNULL	anything that reads stdin gets EOF at once instead of waiting
-NonInteractive on powershell.exe	a prompt (Read-Host etc.) errors immediately instead of hanging until the timeout
On TimeoutExpired, raise AssertionError(...) from exc carrying the last 4,000 chars of stdout and stderr	the next timeout shows the last thing the installer printed

No assertion changed, and no test was skipped or loosened. I grepped Install-CyClaw.ps1 and CyClaw-SecretStore.ps1 and found no prompts in the install path, so -NonInteractive changes no behaviour these tests check.

The second commit (b685469) answers CodeQL py/mixed-returns on this PR. The except branch first ended in pytest.fail(), which never returns but which CodeQL reads as an implicit None. It now raises explicitly, with the same message.

Root cause, stated plainly: I cannot run Windows PowerShell here, so "slow PowerShell start-up on a loaded hosted runner" is an inference, not a measurement. Every fake command in these tests answers at once, which points at host start-up rather than the installer logic. If it times out again at 180s, the new failure message will show where it stopped.

Invariant / Governance Impact: none. Test-only change.

Types of changes

• Refactor or tests (behavior-preserving, no new capability)

Scope: Infrastructure / CI only.

Benefits / why

• One slow Windows runner no longer turns an unrelated PR red. [fix] - Keep health probes available for malformed model catalogs #1722 cost a re-run cycle.
• A real hang still fails, and now the installer's own output comes with the failure.

Risks to monitor

• A real regression that makes the installer slow but not hung would now pass for up to 180s instead of failing at 60s. The four PowerShell tests run one after another, so the worst case adds about 8 minutes to a failing Windows leg.
• Signal: the Windows-leg duration for this file. An AssertionError: installer exceeded 180s ... means a real stall, and the stdout tail shows where it stopped.

Checklist

• I have read docs/THREAT_MODEL.md, INVARIANTS.md and SECURITY.md (and any relevant Phase docs)
• This change preserves all 6 security invariants and I6 module isolation (test-only)
• Full sandbox validation has been run (per CLAUDE.md §5, left to CI)
• No new external network dependencies or mandatory online LLM assumptions were introduced
• Commit messages follow the title prefix convention above

Further comments

Run directly (Linux, Python 3.12):

• ruff check and ruff format --check on the file: clean.
• The file under pytest: the 4 bash-installer tests pass and the 4 PowerShell tests skip (no powershell.exe here).
• _run_installer called directly:
  • with the timeout forced to 1s around echo started; sleep 5, it raised AssertionError with __cause__ set to TimeoutExpired and the message installer exceeded 1s | --- stdout --- | started;
  • read x || echo stdin-closed printed stdin-closed.

Only CI can run: the four PowerShell tests on windows-latest. That leg is the real test of this PR.

Suggested merge order of open PRs

Trial merges verified (git merge --no-commit --no-ff + abort, each on its own) against main @ cdcef76 on 2026-10-20.

#1726 · Blocked (draft)
Impact: .github/dependabot.yml only.
Action: Independent; either order. Clean.

#1727 (this PR) · Blocked (draft)
Impact: installer contract test timeout and diagnostics.
Action: Merge once windows-latest is green. Clean.

ELI5

A Windows test runs CyClaw's PowerShell installer against fake Python programs that all answer instantly, and it gave the whole thing 60 seconds. On a busy Windows CI machine PowerShell itself can be slow to start, and once it ran out of time on a PR that had nothing to do with installers. Worse, when it timed out it threw away everything the installer had printed, so nobody could see where it got stuck.

Now tests/test_installer_python_contract.py gives it 180 seconds. It also tells PowerShell never to wait for a person to type (-NonInteractive, plus closed input), so a prompt fails at once instead of hanging. A timeout now shows the installer's last output.

• What it enhances: fewer false reds, and a readable message when one is real.
• What it makes obsolete: two copy-pasted subprocess.run blocks.
• What it retrieves: nothing.
• What it flips into when pushed too far: a slow-but-working regression stays hidden longer before it fails.

I ran the Linux half and the helper itself. Only Windows CI can run the PowerShell half.

🤖 Generated with Claude Code

https://claude.ai/code/session_014FZ8JTF2v1eQd1L5CKQR8j

Last updated: 2026-10-20 09:45 ET