[isaacslee]

Why

The first real-host acceptance runs on a disposable Ubuntu 24.04 aarch64 VM (Lima, Apple Virtualization, on the maintainer's laptop). It is turning the fixtures' assumptions into observations. Each assumption the host contradicted gets a commit here. This PR stays open while the acceptance continues, and its final head is the commit the acceptance record names.

Stacked on #16.

What changed

• fix: expect what systemd prints for restricted namespaces. systemd 255 prints RestrictNamespaces=yes for a restricted unit, no for an unrestricted or unset one, and a list of types otherwise. The check expected an empty value for baseline, so a correct host failed its own check. baseline is now checked as yes, and userns (newly checked) as no.

How it was verified

On the VM. These were observed while the acceptance commit was being prepared:

• systemctl show -p RestrictNamespaces gave yes, no, net user, and no when unset;
• a stand-in process carrying the rendered unit's sandbox (same User, Slice, masks, ProtectHome, BindPaths, RestrictNamespaces) reported RestrictNamespaces=yes, and unshare -U inside it failed with "Operation not permitted".

Locally: check drill, 45 passed. It adds cases for a baseline runner reporting no, or net user, and for a userns runner reporting yes.

GitHub-hosted CI: BLOCKED (see #3).

Evidence level

RUNTIME VERIFIED for what systemd 255 prints. The fix is SYNTHETICALLY TESTED, until runner-slice check passes against the registered runner on the VM.

Limitations

The behaviour is observed on systemd 255 only.

Risks

None.

Compatibility impact

None.

Remaining work

Continue the acceptance after registration. Further findings will be added here as commits.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.

[isaacslee]

Added: f28b6a5, fix: register a runner without replacing one of the same name.

The registration command RunnerSlice printed used --replace. That silently takes over any runner already registered under the same name to the repository, which could be another host's runner. The first registration is now printed without it. If the name is already registered, config.sh refuses, and the printed note explains what --replace would do.

Verified locally:

• The new apply-drill case fails against the previous command and passes now.
• tests/run.sh passes: shellcheck reports 20 scripts clean, and the drills run 434 cases with 0 failed.

The acceptance subject is now f28b6a553749588a438510c1104928fce613b4fd (tree bc54b3c).

• It is pushed to the dogfood repository as subject/<sha>.
• The VM's installed tree equals git archive of this commit, file by file (48 files).

Also checked before the first registration, by reading the code and the host:

• Token through the environment. Runner v2.338.0 reads ACTIONS_RUNNER_INPUT_TOKEN as the token argument and masks it as a secret. It then removes it from its own environment (src/Runner.Listener/CommandSettings.cs).
• sudo forwarding. sudo --preserve-env=VAR passes that one variable to the runner account, and only that one. This was tried with a harmless test variable.
• The registration does not exist yet. GitHub lists 0 runners for the repository. The runner directory has no .runner, .credentials or .credentials_rsaparams.
• The runner files are in place. config.sh is present and executable. The directories have the profile's owners and modes. --work equals the profile's work-dir.

[isaacslee]

Added: 9208151, fix: run config.sh from its own directory.

What happened. During the first real registration (runner rs-arm64-dogfood-01, registered by the owner), config.sh printed this three times:

ldd: ./bin/libcoreclr.so: No such file or directory

The other two lines named libSystem.Security.Cryptography.Native.OpenSsl.so and libSystem.IO.Compression.Native.so. Authentication and registration still succeeded.

What was established on the VM, and in the runner's source:

• The libraries are present. All three exist in runner/bin/. Run as the runner account from the runner directory, ldd reports 0 unresolved dependencies for each of them.
• The check uses relative paths. In config.sh (v2.338.0, and unchanged on main), the ldd ./bin/... checks on lines 23–37 run before cd "$DIR" on line 72. The registration ran with the working directory / (limactl shell --workdir /), so ./bin/... did not exist.
• The check silently passed. ldd writes "No such file" to stderr, and grep 'not found' reads only stdout. The dependency check therefore passed without checking anything.
• Reproduced:
  • config.sh --help run from / prints 3 ldd: lines;
  • run from the runner directory, it prints 0;
  • run through env -C <runner-dir>, it prints 0.
• Known upstream: actions/runner #945 (this exact message), #2373 ("Change directories first", closed) and #4460 (open).

Conclusion. This is a working-directory problem, not a missing library. The registration is valid and was not repeated.

Fix. The registration and deregistration commands RunnerSlice prints now run env -C <runner-dir> HOME=… ./config.sh …. The note and the operations guide say why. The apply drill fails without the change, and passes now with 141 cases.

Verification:

• tests/run.sh passes: shellcheck reports 20 scripts clean, and the drills run 436 cases with 0 failed.
• The VM's installed tree equals git archive of the new commit, file by file (48 files).

The acceptance subject is now 92081511e1a59e982b8ee5ba017d58174a7fb9af (tree 57413b0). Evidence gathered before this commit is recorded against the commit it ran on:

• the registration ran from the command f28b6a5 printed;
• the earlier probes ran on 364a9c3.

[isaacslee]

The ARM64 acceptance is complete. All 16 steps were run on a real host. The record is in docs/acceptance/2026-10-arm64.md (commit 01ca427).

Added since the last comment:

• 2bc81c0 fix: deregister without a flag config.sh remove rejects. During the real deregistration of both runners, config.sh remove --unattended printed "Unrecognized command-line input arguments for command remove: 'unattended'". Runner v2.338.0 accepts only --token, --pat and --local for remove. Both removals still succeeded, because the token came from the environment.
  • A new apply-drill case fails without the fix.
  • The VM's installed tree was compared file by file with this commit before the final retirement ran on it.
• 01ca427 docs: record the arm64 acceptance and what it verified. It changes documentation only (README.md, docs/*).

Real-host results, recorded against the commit that ran them:

• On 9208151: two GitHub runners (ids 2 and 22) took real dogfood jobs.
  • drills: 436 cases on each runner;
  • boundary: 14/14 on each runner;
  • pool assignment observed on both runners;
  • shared ceiling under concurrent load;
  • busy guard with real hold jobs: deferred, uninterrupted, exactly one restart afterwards;
  • guest reboot, after which services, limits and filters held and boundary jobs passed.
• On 2bc81c0: final retirement.
  • GitHub lists 0 runners.
  • The host matches its pre-test baseline, except the shared slice unit, which is the documented residue.
  • The foreign unit, the other account and the secret file are intact.

Still not verified:

• unknown job state on a real host (synthetic only);
• a misregistration on a real host;
• a runner self-update;
• runner-slice status on the host;
• x86_64.

GitHub-hosted CI remains BLOCKED (see #3). The dogfood runs are separate evidence and do not count as that gate.