How the Claude Code team does code review (Anthropic, 2026-10-15)
id
resource
title
pr-checker
.gitea/workflows/pr-checker.yml
PR Check workflow
Decision
Nothing decided yet. This file records candidate improvements for the PR
pipeline so the analysis survives until they are picked up. Total effort if
all are done: roughly half a day (3-6 h), all config/markdown work on the
existing pipeline, no new infrastructure.
Context
The existing three-layer pipeline (pre-commit, Gitea Actions, AI review)
already matches the core of Anthropic's model: cost-ordered layers, blocking
deterministic checks, severity-tagged findings, review-scope discipline
(other jobs' findings are not repeated; docs-only diffs get one line), and
fixing repeated findings at the source. The backlog below closes the gaps
the post exposed; the machinery at their scale is deliberately skipped.
Candidates
#
Item
Effort
Complexity
Lands in
1
pnpm build CI job
~30 min
Low
New job in the workflow, copied from the lint job block. Closes the biggest gap: nothing currently proves the app compiles or renders. Only unknown: build-time env/runtime config
2
PR intent into the review
~30 min
Low
Review job step: write github.event.pull_request.title/body to a file (node, as for the comment JSON), one prompt line. Reviewer then asks "does the change solve what it claims", like the post's reviewer
3
No-repeat re-review
1-2 h
Medium-low
Review job: fetch the previous review comment via the Gitea API (token scopes already sufficient) and pass it into the prompt as "already reported, do not repeat". Motivation: Anthropic measured that 59% of re-review findings were pre-existing code
4
PR template
~10 min
Trivial
.gitea/PULL_REQUEST_TEMPLATE.md with "How tested / Risk / How to revert"
5
Reviewer eval set
1-2 h
Low-medium
Note with 2-3 past PR diffs and their known findings; the reviewer prompt is rerun against them whenever prompt or model change. Seed exists: PR #5 findings in AI review setup and the failure-modes table in PR pipeline layering
Order and cost notes
Suggested order: 1 first (catches the most), then 2 and 3 together (both
touch the review step), 4 and 5 whenever.
Items 2 and 3 cost a paid review per test run. Test them in one batch on a
draft-to-ready PR, not iteratively, or the Go plan's usage window pays for
the tuning.
Item 1 needs no review budget - it is a deterministic job.
Deliberately not adopted
Scale-dependent machinery from the post; meaningless or wasteful for a
single-author repository at a few PRs per week:
Merge queue, feature flags, staged latest/stable channels - downtime
is accepted for v1 anyway; the user is the only user.
Risk-tier classifiers, trusted-author lists, two-person approvals - the
one-person equivalent: never let an agent-authored change to
.gitea/workflows/ merge unread (every pipeline regression so far lived
there: the lost chmod +x, the empty hashFiles).
Multi-agent finding-voting (2-3x review cost) - verifying a BLOCKER
manually before fixing achieves the same at solo scale, without touching
the $3/5 h usage window.
Measuring human-approval metrics (approvals per PR, time waiting on a
person) - there is no second reviewer to wait for.
Revisit when
An item is picked up - it moves from this backlog into its own decision
file, and this file's table gets a "done" link like open questions.