At high effort, Opus 5.5 uses about half as many thinking tokens per review as
Opus 5 did. The lenses return after about five of their twenty minutes. Use
xhigh effort, and tell the lenses to go through every hunk of the diff instead
of stopping at the first findings.
Pull requests from members of downstream organizations such as amutable
get no label that shows which organization they come from. Add a
downstream/<org> label when the author is a member of the organization.
The workflow token is not a member of the downstream organizations.
GitHub only shows public memberships to the token, so a member has to
make their membership public to get the label.
The review runs on Claude Opus 5. In Anthropic's testing, Claude Opus 5.5 scores
higher than Claude Opus 5 on coding evaluations. Switch the review to Claude
Opus 5.5.
With CLAUDE_CODE_SUBPROCESS_ENV_SCRUB set, Claude Code 2.1.220 lets the
sandbox write to /home. Every sandboxed Bash call then fails because bwrap
cannot create a placeholder at /home/.mcp.json. Bump to 2.1.288, which no
longer creates that placeholder. A sandboxed shell can also write to
$RUNNER_TEMP. There it can forge the results of other lenses, or set the
environment of later steps through $GITHUB_ENV. Deny sandbox writes to
$RUNNER_TEMP. Lenses are also killed in their final turn because the timeout
fires three minutes after the deadline. Give them eight minutes instead. The
summary says "No issues found" when lenses have failed. Make it say that the
review is incomplete. A lens also counts as done when all of its Bash calls
failed. Count it as failed instead. Agents cannot run git diff without a
working shell, so write the diff since the last review to a file.
Syncing a fork's main branch triggers the push workflows if actions are
enabled, which then run the full CI matrix and use up the fork owner's
Actions minutes for no benefit. Only skip push events so that PRs within
a fork still run CI, which is useful for testing actions.
Signed-off-by: Paul Meyer <katexochen0@gmail.com>
Let's maintain the list of subsystem maintainers
in CODEOWNERS, which github knows about and can do
useful stuff with like automatically adding those
usernames as reviewers.
* 117d382486 nura: use modern url for Alpine binary repos
* 7872fd58f7 nura: drop x86
* 3533864541 gitignore: add mkosi.pkgcache
* 580c4d94d7 Rename postmarketOS --> Nura
* ba96aba91e Bump version to 28~devel
* ecf7f5702b Release 27.1
* ca08e2ccef docs: add notes on nsresourced in execution flow
* 146b5dd4de docs: add a few notes on package manager specific behaviour
* 344e42ee20 Update NEWS
* dde1bf020f debian: fix debian-security/debug url when using a mirror
* 699aa3b565 mkosi: upgrade debug message about fallback path to info
* cb6da3f3ce mkosi: move fallback environment variable check to bypass nsresourced fully
* 88c716a6b9 config: add delegate_ranges
* f4de9a102a mkosi: use the foreign UID range only when requested when running as root
* d091fed134 mkosi: only delegate UID ranges if we're not already root
* 165bc99675 sandbox: rename delegate to delegate_ranges
* bba53c50d0 sandbox: use equality check instead of boolean not for delegate
* a033d03f59 summary: add foreign uid range
* f13e96eb95 user: drop unused variable SUBRANGE
* 847d113863 mkosi-obs: Only configure MakeScriptsExecutable in the main image
* 568b1bce56 dnf: Only read repositories and variables from the sandbox tree
* 938d6a1c21 obs: use policy signing data when available
* 19d045e5e9 Rename sysupdate configuration files to .transfer
* 9539f1771a action: bump fallback apk version to 3.0.8
* 8de4d25596 sandbox: add force_fallback to acquire_privileges and wire it up via env
* af97d6ffef opensuse: tolerate missing RPM keyring
* 0f2b7e0656 initrd: don't fail when encountering directories among ucode files
* 5240191893 Allow using SignInitrdPCRs with SplitArtifacts=pcrs
* ce90009686 Bump ukify required version from 261.999 to 262~ for SignInitrdPCRs
* 58bbd63aca news: fix v27 `DiskType=` setting typo
* 1da3299aec tests: skip test_fork_and_wait if on rc 1
* 5658e7abc9 Bump version to 28~devel
* 4736cd8361 Release 27
* 97f305bf0a action: make shell style consistent
* dd7140cbae action: only use apk version from API if non-empty
* 2cfe0cefb8 news: update NEWS
* 7bca598fa4 action: support container-based self-hosted runners
* 08c86a25e1 action: bump fallback APK version
* cc0fcef38f action: reduce timeouts and number of API calls
* 211993d5d2 action: download latest version of apk
| * f33cc45db3 Add ToolsTreeSnapshot= for pinning the default tools tree
| * ddfc5a2cb5 tests: join options and their values with f-strings
| * a497ff0251 doc: add missing matchers
* | 589226cfd4 apk: disable grub's update-grub trigger during package installation
/
| * 595451bb4b debian: add new split-out systemd packages
| * 7ce1851901 Revert "mkosi-tools: Install systemd-tpm on Debian"
| * 50a1137be5 ci: pin resolved release for integration tests
| * e30b0a0ba4 tests: drop obsolete @ prefix
| * e7f4d57fac news: update NEWS
|/
* bf2addb2ec build(deps): bump aws-actions/configure-aws-credentials
* a58ded6a2a build(deps): bump actions/checkout from 6.0.2 to 7.0.1
* 9b7d87a170 add support for prebuilt UKIs
* 75fdb9be15 mkosi-tools: Install systemd-tpm on Debian
* d5ff0d0d98 Fall back to unprivileged user namespace on polkit denial
* 0e4d5546ac Bump 261~devel version checks to 261
* 180d4a5329 Use systemd nightly OBS version in signing PCR policies auto detection
* f7401bdc8d Add support for signing PCR policies for the initrd
* 33d17b2b92 sandbox: Don't create shared propagation peer group
* a078088f8e sandbox: Declare MS_PRIVATE as well
* e1e9eafc87 kmod: eliminate duplication in calls to modinfo
* df68b27e8e tools: add mount to Debian tools tree
* 205dcedd45 util: don't drop single-character variable names in read_env_file
* 35a9d38ea0 distribution: report the actual architecture in unsupported-arch errors
* d4f44d212c sandbox: handle case where dev node for tty doesn't exist better
* acf08c5eca build(deps): bump actions/upload-artifact from 7.0.0 to 7.0.1
* b3ec53743a tests: Pass --output-directory to vm() and boot()
* 9bc430a7fa config: Isolate build history per output directory
* fd3d9bfe57 tests: Add regression test for build history lookup via configured output dir
| * 55ab00bb13 Fix tar exclude patterns for apivfs paths
|/
* 699da862f8 sandbox: Split a chroot() context manager out of chase()
* 70ce10d132 sandbox: Default FSOperation.execute oldroot/newroot to /
* 610b38538f Forbid AI attribution in commit messages
* b7425bbe42 ci: Sync claude-review workflow with systemd
* b9d4c06239 action: Make apk installation configurable and reliable again
actions/labeler applies path based labels by reading the pull request's full
label set and writing it back with setLabels(). It re-reads the labels just
before writing, but the window is still open.
Match the configured patterns in the workflow instead and apply them with
addLabels() and removeLabel(), which only ever touch the labels the step
decided on. The configuration moves to .github/labeler.json so parsing needs
nothing github-script does not already provide, and "*" now means any run of
characters, "/" included. That also makes the patterns say what they looked
like they said: "src/home*" matched nothing below src/home/, so the homed
label was dead.
Crossing the too-many-labels limit in either direction takes the other
outcome's labels back off, so a pull request cannot end up carrying tree-wide
and a handful of subsystem labels at the same time.
The pull_request trigger is there to check a configuration change against the
head copy, so it no longer writes anything: a fork only ever got a read-only
token, and pull_request_target runs for the same pull request, so the two would
fight over the labels. Drop its paths-ignore in turn, because it only kept
pull_request_target away when the pull request changed nothing else, and a
configuration change bundled with other files left nobody doing the writing.
The hook that enforces the review deadline used bash's [[ test, but hook
commands run under /bin/sh, which is dash on the runners. Every tool call was
denied with "[[: not found", so all six lenses returned empty findings and
the run reported a clean review in 73 seconds. Use a POSIX test.
An empty result from an agent that never got a tool call through looks exactly
like a clean review, so require at least one successful tool call before a lens
counts as completed. A run like that now fails every lens, which also stops it
from raising the review count.
The "Pull Request Labeler" workflow read the full label set of a pull
request and wrote it back with setLabels(), which is a full replace. Other
workflows label the same pull request at the same time: claude-review.yml
adds the claude-review label on the pull_request_target "opened" event,
which is the very same event that starts the labeler. Whenever that add
landed between the labeler's listLabelsOnIssue() and its setLabels(), the
label was dropped again and the review never ran.
Only ever add or remove the individual labels the labeler owns, so
concurrent changes by other workflows survive. removeLabel() returns 404
when the label is already gone, which is exactly what a racing run causes,
so swallow that status.
Reviewers ignored the tool call budget in their prompt, and the budget was a
poor stand-in for what matters anyway: per-call latency varied from 25 to
85 seconds between lenses, so the same count meant 7 minutes for one and 20
for another. Two lenses hit the timeout and lost everything they had found.
Enforce the deadline itself. run_claude exports it per invocation and a
PreToolUse hook denies further tool calls once it has passed, telling the
model to return what it has. The structured output call stays allowed, so
a slow lens still produces findings. The timeout, three minutes later, only
catches a hang. Give reviewers 22 minutes and the validator 12.
AI agents prepend a notice to README.md that the author removes once they
have reviewed the changes. Until that happens nobody else should look at
the PR: convert it to a draft, don't add the please-review label, and
don't start a Claude review. The labeler and the review workflow check
README.md at the PR head for the notice. Marking the PR ready for review
after removing the notice triggers the normal labeling again.
A triple backtick inside prose made the fence count odd, so a body with an
inline marker plus a valid suggestion block got cut at the block's closing
fence and left it open. Only fences at the start of a line delimit blocks,
so count and strip those alone.
Reviews keep hitting the 50 minute timeout. The coordinator agent that was
supposed to launch reviewers, watch the clock, stop stragglers, validate and
write the result does none of that reliably: reviewers get stopped and lose
their findings, foreground validators can't be stopped, deadlines get
overshot by minutes, and one leftover agent keeps the process alive until
the job fails. Piling more rules into its prompt hasn't helped.
Do the orchestration in the shell instead. Run one bounded claude invocation
per review lens in parallel under timeout(1), each returning its findings as
structured output, merge them with jq and run one bounded validator over the
candidates. A lens that times out is just reported as failed. The post job
builds the tracking comment from what it actually posted, so the model no
longer authors the summary or the review bookkeeping. Agents run from an
empty directory with all inputs outside it, so the sandbox keeps both the
checkout and the inputs read-only.
The Claude review workflow posted long, padded inline comments, and every
force push produced a fresh batch of new findings on code that had not
changed. Both make authors stop reading the reviews.
Add a comment style section to the prompt that limits each comment to the
problem and the obvious fix, and truncate anything over 1000 characters in
the post job. Track the number of completed reviews and the last reviewed
head in a hidden state line in the tracking comment. Each successful review
narrows the severities that may be posted: everything on the first review,
no nits on the second, and only must-fix findings from the third on. The
post job enforces the policy so it holds even if the model ignores the
prompt, and reviewers are told to range-diff against the previous head and
skip non-must-fix findings on unchanged code. Commenting "@claude review
full" resets the count for that run.
Running the review step from /tmp/claude-review-sandbox put the directory the
review result has to be written to outside the agent's working directory. The
Write tool refuses to touch anything outside that directory, no matter what the
permission rules say, so every attempt to write review-result.json was denied.
Runs only passed because the agent fell back to writing the file from a
sandboxed shell command instead, and the last one hit the 50 minute step
timeout before it got that far, leaving nothing for the post job to upload.
The separate working directory bought little on top of the write allow rule,
which already confines the agent to claude-review-work/, so drop it and run in
the workspace again.
Replace opaque built-in review behavior with explicit parallel review lenses for
correctness, concurrency, security, architecture, coding style, and regression
coverage, plus bounded domain-specific reviewers and independent validation.
Prefetch pull request discussion in the setup job while letting the read-only
review job fetch and inspect the full commit history without per-commit
worktrees.
Keep GitHub writes isolated to the setup and posting jobs, pin the Claude models
and CLI, and bound execution below the AWS credential lifetime. Materialize
review instructions from the trusted base revision and scope agent writes to the
structured result directory so untrusted pull request content cannot widen the
review authority.
Treat the posting job as the authoritative trust boundary: validate all
model-selected commits, paths, lines, and thread IDs, normalize degenerate line
ranges, bound sanitized output, and surface GitHub API failures in the tracking
comment before performing writes. Retain daandemeyer/systemd in the repository
allowlist so the workflow can be exercised end to end on the personal fork.
Manual runs against fork PR #15 covered OIDC authentication, sandboxing, result
artifact handoff, tracking-comment updates, and validation of posted output.
Signed-off-by: Daan De Meyer <daan@amutable.com>
Recent Ubuntu 24.04 GHA images have /etc and /usr owned by runner
instead of root, which breaks some of our tests. This has been filed to
GH as https://github.com/actions/runner-images/issues/14477, so let's
work around this in our jobs until it's fixed.
1f76654f94 added support to build .standalone
variants of most binaries. test-dlopen-note.py is hooked in the test suite for
everything that can do .standalone, but this means that the .standalone
variants become a prerequisite for tests. This means that meson will build the
.standalone variants for all binaries before running tests, which takes quite a
bit of time in some situations. (Arguably, the dlopen test suite was excluded
by default, so maybe meson could be smart and figure out that those tests are
not enabled by default. But it doesn't seem to make this distinction.)
We often forget to run this command when updating manpages, and notice
only much later. Add a step in the CI to flag it, as we already do for
the D-Bus docs.
Enable the `dlopen` unit test suite in GitHub Actions, except for the
following configurations where linker garbage collection (`--gc-sections`)
or dead-code elimination fails to drop unused dlopen symbols:
- ppc64le (both GCC and Clang)
- s390x (GCC only)
- Sanitizer-enabled setups
* f7762b7143 sandbox: Preserve net caps across user namespace before unsharing net
* 582eadee34 Revert "Put build history into the output directory"
* 5ef262bc53 action: don't fail if apk cannot be downloaded
* bdd341ff9b Lock the package cache during package manager invocations
* da49fe976c Put build history into the output directory
* 1c392f1918 tests: Use unique machine names
* e4f4026e30 tests: Reduce VM RAM size
* de41a5e03e Don't leak gpg-agent when signing with gpg
* 1bc5d61e1d ci: Pin openSUSE to second-to-last Tumbleweed snapshot
* c4d565a009 test: Use the main build's snapshot for extension builds
* 718b06c866 tests: ignore masked units in check-and-shutdown
* 0dc5ecbc02 ci: enable postmarketOS in integration testing
* d4c6761ad3 action: install apk to /usr/bin
* 9980f31309 mkosi-vm: add systemd-efistub to postmarketOS config
* 5640ace38f mkosi.conf: add grub to postmarketOS
* 6741b440c0 mkosi-initrd: add sulogin, device-mapper to postmarketOS initrd
* c3575c035c mkosi-tools: add missing packages to postmarketOS tools tree
* 0774bc2498 mkosi-tools: add apk-tools to tools trees for Arch and OpenSuSE
| * bb87e48401 curl: Retry on failures
|/
* 41fea1dd8d dnf: Work around librepo rejecting valid repomd signatures cross-distro
* 647e3b610b dnf: Proper repository metadata signature requirement
* 46d907cce2 dnf: Don't skip unavailable repositories during makecache
* a91e89c3b7 run_locale_gen: noop if output_format is confext
* 30329e401b tests: Make integration tests runnable locally
* be549f04db config: Don't propagate $MKOSI_DNF when using a tools tree
* 42ed648981 build(deps): bump actions/upload-artifact from 7.0.0 to 7.0.1
* fd5eedd62b build(deps): bump aws-actions/configure-aws-credentials
* 86733c703d tree: check for root when copying SELinux attributes as well
* de2256f8fe Skip security.ima xattrs when copying tree as non-root
| * 08ebf6d678 vmspawn: Exclude secure-boot unless requested
|/
* 1d3c51e36d obs workflow: do not build aarch64/i586
Daan De Meyer