For some reasons, after util-linux is bumped from 2.41.4-r0 to 2.42-r0,
the 'su' command from util-linux-login seems to not correctly run commands in
https://github.com/jirutka/setup-alpine/blob/v1.4.1/alpine.sh
and causes the following spurious failure:
```
2026-05-15T21:19:15.6539432Z ##[group]Set up user runner
2026-05-15T21:19:15.6981963Z /bin/sh: line 0: ��: not found
2026-05-15T21:19:15.6982503Z /bin/sh: line 1: ␡ELF␂␁␁␃: not found
2026-05-15T21:19:15.6985788Z /bin/sh: line 10: ␒␐␆␒B␈␒�␄␒y␄␒�␁␒␞␇␒:␁␒�␃␒�␄␒@␁␒9␈␒?␆␒␚␈␒x: not found
2026-05-15T21:19:15.7010731Z /bin/sh: line 33: can't open ␂␒-␂␒�: no such file
2026-05-15T21:19:15.7016026Z /bin/sh: line 33: syntax error: unexpected word (expecting ")")
2026-05-15T21:19:15.7049583Z
2026-05-15T21:19:15.7050199Z ␛[1;31mError occurred at line 338:␛[0m
2026-05-15T21:19:15.7050830Z 335 | echo 'permit nopass keepenv $SUDO_USER' | tee /etc/doas.d/root.conf
2026-05-15T21:19:15.7051287Z 336 | fi
2026-05-15T21:19:15.7051549Z 337 | SHELL
2026-05-15T21:19:15.7052039Z ␛[1;31m> 338 | abin/"$INPUT_SHELL_NAME" --root /.setup.sh␛[0m
2026-05-15T21:19:15.7052506Z 339 |
2026-05-15T21:19:15.7052796Z 340 | rm .setup.sh
2026-05-15T21:19:15.7053172Z 341 | endgroup
2026-05-15T21:19:15.7096322Z ##[error]Error occurred at line 338: abin/"$INPUT_SHELL_NAME" --root /.setup.sh (see the job log for more information)
2026-05-15T21:19:15.7101400Z ##[error]Process completed with exit code 1.
```
Let's not install the package. It seems no command provided by the
package is used.
The cdn mirror is preferred by SUSE for clouds/CIs. There have been issues with some
mirrors, which fail to download from GHA quite often lately, so hopefully this will
make it reliable again.
Checking for pefile required that module to be made available for the
Python used to build systemd, even though it's only used at runtime,
potentially via a different Python installation.
Furthermore, Meson's Python module doesn't do the right thing when cross
compiling and looking up a Python for the host system, so this would end
up uselessly checking whether the build Python had the pefile module,
which is not needed. Even if it were made to check the host Python using
find_program, it still relies on being able to run its Python, which in
a cross scenario it probably wouldn't be able to do.
All in all, this check does more harm than good, and prevents building
ukify in valid configurations, so remove it.
Jammy's kernel is too old at this point, and doesn't even provide a
vmlinux.h, so disable the feature in the build smoketests to let us
add new features
Co-developed-by: Luca Boccassi <luca.boccassi@gmail.com>
Signed-off-by: Christian Brauner <brauner@kernel.org>
Commit a65ebc3ff9 ("claude-review: improve review quality for large
PRs") dropped the `Claude: **<severity>**: ` prefix from posted inline
comments on the theory that Claude was also adding the severity into
`body`, producing duplicates. But nothing in the prompt or schema
actually asks the subagent to include severity in `body` — severity
is a separate structured field. The result is that inline comments
no longer show must-fix/suggestion/nit classification.
Restore the prefix in the posting step, and add an explicit instruction
to the subagent prompt telling it not to repeat severity inside `body`
so the two don't collide.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Opus 4.7 is in research preview on Bedrock and the Invoke API rejects
the beta headers Claude Code sends ("invalid beta flag"). Enable the
Mantle endpoint, which serves Claude via the native Anthropic API shape
and accepts those headers, and switch the model ID to the Mantle form
(no region prefix or version suffix).
- Use persist-credentials: false for actions/checkout, so we don't
leak the github token credentials to subsequent jobs.
- Remove one / from the Edit/Write permissions. Currently, with the
absolute path from github.workspace, we expand to three slashes while
we only need two.
Several issues were identified from analyzing logs of a large (52-commit) PR
review:
- Claude was batching multiple commits into a single review agent instead of
one per worktree. Strengthen the prompt to explicitly prohibit grouping.
- Claude was reading pr-context.json and commit messages before spawning
agents despite instructions not to, wasting time. Tighten the pre-spawn
rules to only allow listing worktrees/ and reading review-schema.json.
- Subagents were spawned with model "sonnet" instead of "opus". Add explicit
instruction to use opus.
- After agents returned, Claude spent 9 minutes re-verifying findings with
bash/grep/sed commands, duplicating the agents' work. Add instruction to
trust subagent findings and only read pr-context.json in phase 2.
- Subagents returned markdown-wrapped JSON instead of raw JSON arrays. Add
instruction requiring raw JSON output only.
- Each subagent was independently reading review-schema.json. Instead have
the main agent read it once and paste it into each subagent prompt.
- The "drop low-confidence findings" instruction was being used to justify
dropping findings that Claude itself acknowledged as valid ("solid cleanup
suggestions", "reasonable consistency improvement"). Remove the instruction.
- Simplify the deduplication instructions
- Stop adding the severity to the body in the post processing job as claude is
also adding it so they end up duplicated.
The main agent doesn't need to read pr-context.json until all
reviews have finished. This should prevent it from passing unnecessary
data from pr-context.json in the prompt to its subagents, which can just
read that file themselves when needed.
Replace claude-code-action with a direct claude CLI invocation. This
gives us explicit control over settings, permissions, and output
handling.
Other changes:
- Prepare per-commit git worktrees with pre-generated commit.patch and
commit-message.txt files, replacing the pr-review branch approach.
- Use structured JSON output (--output-format stream-json --json-schema)
instead of having Claude write review-result.json directly.
- Use jq instead of python3 for JSON prettification.
- Add timeout-minutes: 60 to the review job.
- List tool permissions explicitly instead of using a wildcard.
- Fix sandbox filesystem paths to use regular paths instead of the "//"
prefix.
The fix in 8f1751a111 made me wonder if we could automatically detect
when pointers are accessed but when this might not be safe. Systemd
is already using a lot of `assert(dst)` and this change now forces
us to use them.
So this commit (ab)uses coccinelle to flag any pointer parameter
dereference not preceded by assert(param), ASSERT_PTR(param), or an
explicit NULL check. It adds integration into meson as a new "coccinelle"
test suite (just like clang-tidy) and is run in CI. The check is not
perfect but seems a reasonable heuristic.
For this RFC commit it is scoped to a subset, it excludes 25 dirs right
now and includes around 100. About 300 warnings left. Busywork that I am
happy to do if there is agreement that it is worth it.
With this in place we would have caught the bug from 8f1751a111 in CI:
```
FAIL: check-pointer-deref.cocci found issues in systemd/src/boot:
diff -u -p systemd/src/boot/measure.c /tmp/nothing/measure.c
--- systemd/src/boot/measure.c
+++ /tmp/nothing/measure.c
@@ -312,7 +312,6 @@ EFI_STATUS tpm_log_tagged_event(
if (err != EFI_SUCCESS)
return err;
- *ret_measured = true;
return EFI_SUCCESS;
}
```
This also adds a new POINTER_MAY_BE_NULL() for the cases when the
called function will do the NULL check (like `iovec_is_set()`).
Pass side, start_line, and start_side through to createReviewComment()
when present, enabling multi-line review comments. Update the prompt to
document all positioning fields using JSON Schema and make line required.
In https://github.com/systemd/systemd/pull/40980 claude hallucinated
and used "path" instead of "file" as the JSON key. Since "path" is
arguably more correct than "file" anyway, let's switch to that.
After analyzing all 218 CodeQL alerts across the project's history, the
workflow has not justified its CI cost:
- The most impactful query (PotentiallyDangerousFunction) was a custom
systemd-specific query that has already been replaced by clang-tidy's
bugprone-unsafe-functions check (6fb5ec3dd1).
- Of the remaining C++ queries, 6 never triggered at all
(bad-strncpy-size, unsafe-strcat, unsafe-strncat,
suspicious-pointer-scaling, suspicious-pointer-scaling-void,
inconsistent-null-check).
- Several high-value-sounding queries had extreme false positive rates:
toctou-race-condition (95% FP), use-after-free (88% FP),
cleartext-transmission (100% FP).
- Many queries that did trigger are already covered by compiler warnings
(-Wshadow, -Wformat, -Wunused-variable, -Wreturn-type,
-Wtautological-compare) or existing clang-tidy checks
(bugprone-sizeof-expression).
- Across all alerts, only 3 genuinely useful C++ fixes can be
attributed to CodeQL: 1 tainted-format-string, 2
incorrectly-checked-scanf. The rest were either false positives or
incidental fixes during refactoring that weren't prompted by CodeQL.
- The Python queries are largely superseded by ruff (already in CI) and
had an 89% false positive rate on the security-focused checks.
The workflow consumed significant CI resources (40+ minutes per run) and
the ongoing maintenance burden of triaging false positives outweighs the
marginal value of the 2-3 real findings it produced across its entire
lifetime.
I noticed looking at the logs that claude spends a lot of time re-checking
existing comments, so let's update the prompt to hopefully reduce
the amount of comments that it re-checks.
Let's stop using --json-schema and instead have claude write a JSON
file in the repo root which we pass around as an artifact similar to
how we pass around the input. This works around the bug where claude
receives task notifications after producing structured output which
breaks the structured output.
We often have a pattern where the same verb function is used for
multiple actions. This leads to an antipattern where we figure out what
action needs to be taken from argv[0] multiple times: often once in
arse_argv() to figure out what options are allowed, then once again
implicitly in dispatch_verb(), and then again in the action verb itself.
Let's allow passing a parameter into the verb to simplify this.
As it seems impossible to prevent claude from receiving notifications
about subagents finishing after it has produced structured output, which
breaks the structured output as it has to be the final reply, let's stop
using subagents and background tasks completely to avoid the issue.
claude keeps failing by its subagents completing after it has already
written the review for large prs. It seems to run out of turns, tries
to get the subagents to post partial reviews but doesn't seem to stop
them.
Let's insist that it waits for background tasks to stop but let's also
increase the max turns a bit so it doesn't run out as quickly.
claude wants to use python to access the JSON context so let's allow
it. Since python3 basically allows you to reimplement every other tool,
let's just enable all tools except the web related ones but enable network
isolation so it can't try to exfiltrate anything via python.
In CI we get spurious failures about unitialized variables with gcc
versions older then (depending on the case) 12, 13, or 14. Let's only
try to do this check with newer gcc which returns more useful results.
At the same time, do compile with both gcc and clang at -O2, just
disable the warning.
The old logic seems to have been confused. We compile with -Wall, at
least in some cases, which includes -Wmaybe-unitialized. So if we
_don't_ want it, we need to explicitly disable it.
The labelling approach introduced in 6089075265
means contributors can now trigger the workflow on their own when the label
is added by a maintainer and they update the PR. Hence we need to allow all
users to access the claude code action. This is safe because we already gate
the workflow ourselves to only the contributors that we want to allow.
Additionally, the claude code job has no permissions anymore except read access
to the repository and can execute very limited tools, so this should be safe.
Jörg Behrmann
Yu Watanabe
Luca Boccassi
Christian Brauner
Alyssa Ross
dependabot[bot]
Daan De Meyer
Michael Vogt
Zbigniew Jędrzejewski-Szmek