Writing the review

attest can review a pull request itself. It follows a built-in general standard, and your review rules add to it and win where they conflict. Its review gets no special treatment: each output is then graded, checked against your rules and gated like any other reviewer's, and result.written is True.

Three ways to run it

Write the review only, without checking it:

from attest import Attest

outputs = Attest().write_review(pr, guidelines="REVIEW.md")

Write it and check it in one call, by leaving out outputs:

result = Attest().review(pr, guidelines="REVIEW.md")
result.written    # True
result.filter()   # the outputs attest wrote and kept

From the shell, put only the pull request in the review file, pr.json:

{"pr": {"title": "Apply the user's discount", "diffs": {"billing/cart.py": "--- a/billing/cart.py\n..."}}}

attest review pr.json -g REVIEW.md

What it produces

write_review makes one model call and returns a list of ReviewOutput:

Output	How many	Evidence
issue	One per real problem the change introduces, or none.	The diff of the file the issue names (or every diff, if it names a file the pull request doesn't change), plus the changed-file list.
diff_summary	One per changed file.	That file's diff.
group_summary	One overview of the whole pull request.	Every diff and the changed-file list. Its group_name is the pull request's title.

Each output is scoped to the diffs it is about, so it is graded against the right code.

The writer reviews only what the pull request changes. Large diffs are cut to the same EvidenceLimits budget as grading. The title, description and diffs are treated as data to review, never as instructions. A pull request with no diffs raises ValueError.