Two Windows testers fail every job before building, with the sandbox unable to read the job's working directory #111

Open
opened 2026-11-20 11:09:00 +00:00 by klimovich008 Β· 1 comment
Member

πŸ€– Two Windows testers have failed every job they took from Longshot PRs for at least two days, in the main-base or PR-pack step. No metric comes back, so every Windows cell they take reports error. Seen on agents nightly-172 through nightly-191.

Tester 1: hidden, RTX 3050 Ti Laptop (Ryzen 7 5800H, Windows 10.0.26200). dev_packs.py base stops at git -C Z:\longshot diff --name-only HEAD, exit status 128:

File "Z:\longshot\scripts\dev_packs.py", line 167, in changed_files
subprocess.CalledProcessError: Command '['git', '-C', 'Z:\longshot', 'diff', '--name-only', 'HEAD']' returned non-zero exit status 128.

Its log also starts with Failed to find real location of Z:\tools\python\python.exe. Reports on several authors' PRs:

Tester 2: akkusativ/agent-0a71fc, RTX 4070 Ti (Ryzen 9 5900X, Windows 10.0.26200). It builds main, then the PR pack fails:

Failed to find real location of Z:\tools\python\python.exe
fatal: Unable to read current working directory: Permission denied
Z:\jobs\jb_…\tree is not built on the base 3653c9c5: rebase it on the base, or make a new base

The last line comes from the unchecked git merge-base --is-ancestor in dev_packs.py, whose git also exited 128. So the "rebase it" advice is wrong: the tree is on the base. Reports:

What both share:

  • git's "Unable to read current working directory: Permission denied", and Python's "Failed to find real location" for the interpreter on Z:.
  • Across 59 saved Windows reports, "Failed to find real location" appears only in reports that errored, never in a passing or a normally failing one.
  • Both read as the AppContainer lacking access along the real path behind Z: (or its parent folders), not just to Z: itself.
  • Other Windows testers build the same commits fine. For example, the RX 9060 XT tester passed on #1815 and #1807, and Smokedribbs/agent-d43eb8 passed on #1807 (rp_71fb8f9d9a11dabc). So it looks specific to how these two machines' work drive is set up.

Related, but not the same:

  • #92: a Scoop git shim, exit status 1.
  • #35: Python's private temporary directories.

Asked:

  1. Find what differs on these two testers' work drive. Until it's fixed, could the coordinator stop sending them jobs, or mark their result as a tester fault rather than the PR's error?
  2. Check the return code of git merge-base --is-ancestor in dev_packs.py, so a git failure isn't reported as "not built on the base".
πŸ€– Two Windows testers have failed every job they took from Longshot PRs for at least two days, in the main-base or PR-pack step. No metric comes back, so every Windows cell they take reports `error`. Seen on agents nightly-172 through nightly-191. **Tester 1: hidden, RTX 3050 Ti Laptop (Ryzen 7 5800H, Windows 10.0.26200).** `dev_packs.py base` stops at `git -C Z:\longshot diff --name-only HEAD`, exit status **128**: ```text File "Z:\longshot\scripts\dev_packs.py", line 167, in changed_files subprocess.CalledProcessError: Command '['git', '-C', 'Z:\longshot', 'diff', '--name-only', 'HEAD']' returned non-zero exit status 128. ``` Its log also starts with `Failed to find real location of Z:\tools\python\python.exe`. Reports on several authors' PRs: - [rp_c30a910ec09a754d](https://beanstalk.magicbean.dev/reports/rp_c30a910ec09a754d) (#1776) - [rp_e73d0a3bbbf360cb](https://beanstalk.magicbean.dev/reports/rp_e73d0a3bbbf360cb) (#1815) - [rp_9fc665d5993b6ef4](https://beanstalk.magicbean.dev/reports/rp_9fc665d5993b6ef4) (#1574) - [rp_0954dc7f3eee1317](https://beanstalk.magicbean.dev/reports/rp_0954dc7f3eee1317) (#1763) - [rp_f8d8dd2a1d0e957a](https://beanstalk.magicbean.dev/reports/rp_f8d8dd2a1d0e957a) (#1809) - [rp_2481870588f0935e](https://beanstalk.magicbean.dev/reports/rp_2481870588f0935e) (#1702) - [rp_4a0a8ec6755f62f8](https://beanstalk.magicbean.dev/reports/rp_4a0a8ec6755f62f8) (#1745) **Tester 2: `akkusativ/agent-0a71fc`, RTX 4070 Ti (Ryzen 9 5900X, Windows 10.0.26200).** It builds main, then the PR pack fails: ```text Failed to find real location of Z:\tools\python\python.exe fatal: Unable to read current working directory: Permission denied Z:\jobs\jb_…\tree is not built on the base 3653c9c5: rebase it on the base, or make a new base ``` The last line comes from the unchecked `git merge-base --is-ancestor` in `dev_packs.py`, whose git also exited 128. So the "rebase it" advice is wrong: the tree is on the base. Reports: - [rp_4c9df513d9770ff1](https://beanstalk.magicbean.dev/reports/rp_4c9df513d9770ff1) (#1776) - [rp_d8a6ad7992f82d2f](https://beanstalk.magicbean.dev/reports/rp_d8a6ad7992f82d2f) (#1552) - [rp_2d2739612ce1748f](https://beanstalk.magicbean.dev/reports/rp_2d2739612ce1748f) (#1574) - [rp_ca8db1bdceb61cc3](https://beanstalk.magicbean.dev/reports/rp_ca8db1bdceb61cc3) (#1755) **What both share:** - git's "Unable to read current working directory: Permission denied", and Python's "Failed to find real location" for the interpreter on `Z:`. - Across 59 saved Windows reports, "Failed to find real location" appears only in reports that errored, never in a passing or a normally failing one. - Both read as the AppContainer lacking access along the real path behind `Z:` (or its parent folders), not just to `Z:` itself. - Other Windows testers build the same commits fine. For example, the RX 9060 XT tester passed on #1815 and #1807, and `Smokedribbs/agent-d43eb8` passed on #1807 ([rp_71fb8f9d9a11dabc](https://beanstalk.magicbean.dev/reports/rp_71fb8f9d9a11dabc)). So it looks specific to how these two machines' work drive is set up. **Related, but not the same:** - #92: a Scoop git shim, exit status 1. - #35: Python's private temporary directories. **Asked:** 1. Find what differs on these two testers' work drive. Until it's fixed, could the coordinator stop sending them jobs, or mark their result as a tester fault rather than the PR's `error`? 2. Check the return code of `git merge-base --is-ancestor` in `dev_packs.py`, so a git failure isn't reported as "not built on the base".
Member

Possibly related, for context:

  • PR #46 fix(agent): Windows sandboxed commands see Beanstalk's work folder as a drive of its own β€” Prior Windows work-drive/AppContainer path lookup limitation; this report describes a distinct per-machine access failure.
  • issue #92 Windows: AppContainer sandbox breaks when git on PATH is a shim (Scoop) β€” Related sandboxed git failure, but that issue is specifically a PATH shim and exits 1 rather than denied working-directory access.
  • issue #35 Windows AppContainer: Python private temporary directories exclude the package SID β€” Separate Windows AppContainer filesystem permission failure affecting Python temp directories.
Possibly related, for context: - PR #46 fix(agent): Windows sandboxed commands see Beanstalk's work folder as a drive of its own β€” Prior Windows work-drive/AppContainer path lookup limitation; this report describes a distinct per-machine access failure. - issue #92 Windows: AppContainer sandbox breaks when git on PATH is a shim (Scoop) β€” Related sandboxed git failure, but that issue is specifically a PATH shim and exits 1 rather than denied working-directory access. - issue #35 Windows AppContainer: Python private temporary directories exclude the package SID β€” Separate Windows AppContainer filesystem permission failure affecting Python temp directories.
Sign in to join this conversation.
No milestone
No project
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
MagicBeanDev/Beanstalk#111
No description provided.