apply --check drift report tells you to run socket-patch apply without the -g / --global-prefix / --cwd it was given, so following it patches nothing and exits 0 #1219

[agent] Found by the scheduled Pipenv bug-hunt routine (ledger #313).

Summary

When apply --check finds drift, the human report always ends with the fixed line:

Error: Patches are OUT OF SYNC:
  pkg:pypi/six@1.16.0: patch not applied (an installed copy is still unpatched)
Run `socket-patch apply` to regenerate them.

The suggested command drops every scope flag the check ran with. The check itself is correct (exit 1 on the stale copy), but running the remedy as printed targets a different tree:

• apply --check -g --global-prefix <site-packages> (the Docker pipenv install --system shape): plain socket-patch apply crawls the cwd project, prints 0 of 1 targeted patch applied, … 1 not found on disk, and exits 0. The global copy stays unpatched. apply -g without the prefix patches the default system interpreter instead. In the sandbox that was the distro's /usr/lib/python3/dist-packages/six.py, which isn't the tree that was checked.
• apply --check --cwd app (CI running from a parent directory): plain socket-patch apply prints No patch manifest found; nothing to apply. and exits 0. The Pipenv venv stays unpatched.

This is the same class as #464 (the report-only scan -g hint dropped -g), which was fixed in #777. The apply --check report wasn't covered by that fix.

Impact

apply --check is the CI / GitHub-App audit gate (CLI_CONTRACT apply --check row). A user or CI job that follows its remedy gets exit 0 and believes the drift is fixed. The next apply --check fails again, and the installed copy keeps running unpatched bytes in the meantime. There's no false VEX: this is a misleading remedy, not a wrong verdict.

Repro (Linux, main f3c6313, Pipenv 2026.8.0, local mock of the patch API serving a patched six-1.16.0 wheel)

# global / Docker `pipenv install --system` shape
G=$(mktemp -d)/sys; python3.11 -m venv "$G"; SPK=$G/lib/python3.11/site-packages
"$G/bin/pip" install six==1.16.0
cd my-pipenv-project                       # Pipfile + Pipfile.lock pinning six==1.16.0
socket-patch scan -g --global-prefix "$SPK" --mode agent --yes      # patches $SPK/six.py
"$G/bin/pip" install --force-reinstall --no-deps six==1.16.0       # image rebuilt / reinstall
socket-patch apply --check -g --global-prefix "$SPK"               # exit 1, "Run `socket-patch apply` to regenerate them."
socket-patch apply                                                 # the remedy verbatim: exit 0, "1 not found on disk"
head -1 "$SPK/six.py"                                              # still the upstream bytes

# --cwd shape (agent mode, Pipenv WORKON_HOME venv)
socket-patch scan --cwd app --mode agent --yes
(cd app && pipenv run pip install --force-reinstall --no-deps six==1.16.0)
socket-patch apply --check --cwd app       # exit 1, same remedy line
socket-patch apply                         # exit 0, "No patch manifest found; nothing to apply."

I reproduced the -g --global-prefix lane 3 times and the --cwd lane once. In both, the remedy with the original flags (apply -g --global-prefix "$SPK", apply --cwd app) heals the tree, and the next --check exits 0.

Expected vs actual

• Expected: the remedy names the command that fixes what the check just reported. That means the same scope as the check: -g, --global-prefix <dir>, --cwd <dir>, and --manifest-path / --ecosystems when given. That's how A report-only scan -g tells you to run socket-patch scan --mode agent [PATHS] without -g, so following the hint scans the cwd project instead of the global install #464 / Fix report-only scan -g hint dropping -g (#464) #777 fixed the scan -g hint. CLI_CONTRACT's apply --check row specifies the Error: Patches are OUT OF SYNC: report, and its purpose is to be acted on.
• Actual: a fixed string with no scope flags. Following it exits 0 and patches nothing, or patches a different interpreter (-g without the prefix).

OS × version

This isn't Pipenv- or PyPI-specific: every ecosystem's apply --check goes through the same branch.

Suspect code

crates/socket-patch-cli/src/commands/apply.rs:674, in run_check: eprintln!("Run \socket-patch apply` to regenerate them.");doesn't consultargs.common (global, global_prefix, cwd, manifest_path, ecosystems). Compare report_only_hint` after #777.