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>
The previous policy was primarily written from a standpoint
that AI models are not very good and we didn't wanna waste any
time reviewing PRs generated by AI. Now that AI models have become
actually good and their output is just as good as regular contributions,
let's stop requiring the disclosure as its pointless to still have it,
it doesn't really matter anymore whether a patch was written with or without
AI. It's up to the author to make sure they're not wasting our time by
submitting unreviewed, untested code upstream, regardless of whether that
code is written by an AI or not.
The new policy is inspired by https://github.com/lxc/incus/pull/3506, with
various removals to be less adverse to the usage of AI.
Let's only keep instructions for stuff that we've seen the AI
mess up in practice rather than having a bunch of AI generated
text that it can figure out for itself these days (given it was
trained on systemd's source code in the first place).
Also add a rule to use git worktrees and check out PRs locally when
reviewing them, since I've seen it mess that up in practice.
This seems to be what all the tools are standardizing on, except
claude (https://github.com/anthropics/claude-code/issues/6235) so
add a symlink from CLAUDE.md to AGENTS.md for now until they support
it as well.
I also had claude extend the instructions a bit.
Co-developed-by: Claude <claude@anthropic.com>
Daan De Meyer
davidak