REVIEW.md said what blocks and what does not, but never how to mark a finding,
so every review invented its own shape and none carried a severity. It now
names the three markers the hosted Code Review service uses - Important, Nit,
Pre-existing - and keys them to what the pull request did rather than to how
alarming a defect looks alone: a defect it introduces or worsens is Important,
one it merely brought into view is Pre-existing and cannot be a reason to hold
it. Pre-existing was missing entirely, and checking what this panel emits means
reading far outside the diff, so those findings had nowhere to go except a
wrong Important or silence.
The volume cap said how many and never which. It now collapses a nit repeated
across files into one finding, prefers a nit in code the pull request wrote
over one in code it only moved, caps pre-existing findings at three, and states
that Important findings are never capped - a section listing two caps otherwise
reads as licence to trim what matters. The review opens with a tally so the
author sees the shape before the detail.
Two contradictions went with it. The file told the reviewer to skip what CI
enforces and then to check that a new i18n key reaches all 13 locales, which
i18n-dead-keys.test.ts pins in both directions - the rule moves to "Do not
report" with the reason. "Anything CI already enforces: npm audit" overstated
what runs; CI audits production dependencies at high and above, so a
dev-dependency advisory is out of scope by design.
The reviewer could not read its own CI. Only postgres-durable-first runs
against PostgreSQL, and XRAY_E2E_BINARY and XUI_SCALE_TEST are set by no job,
so a dialect or migration change can carry a wall of green while the paths it
touches never executed. That belongs to the verification bar, next to the rule
that a behaviour claim needs a file:line citation, and "CI passed" now needs a
run actually read. Also names the two house choices no linter defends: neither
golangci-lint nor oxlint rejects a testify or Tailwind import.
Both kinds of claim rot on a rename, so a test pins them the way
repo-context.md's claims are already pinned - the CI jobs REVIEW.md names must
exist in ci.yml, the skip gates it calls unset must stay unset, and the locale
count must match the directory.
The review itself moves from high to max effort, and the prompt records why it
names REVIEW.md at all: the code-review skill reads CLAUDE.md on its own but
not REVIEW.md, so dropping that clause would silently stop the file applying.
Drops a CLAUDE.md reference to tools/seedperf/, which no longer exists - the
review reads that file as project context, so a stale path there misleads it.
Two failures from the same afternoon, both in the bot workflow.
The review of #6272 ran for 34 minutes across four subagents and posted
"No issues found. Checked for bugs and CLAUDE.md compliance." — three
lines for a 73-file diff. The agent had written a per-area coverage
summary in its own last turn and then dropped it on the floor, because
the code-review skill's comment template carries findings and nothing
else. A comment that cannot distinguish a thorough clean review from a
run that died early is not evidence, so REVIEW.md now states what the
posted comment must show and the system prompt points the run at it.
The same run logged 67 permission denials. Only the inline-comment MCP
tool was named in --allowedTools, so `gh api`, writing the diff to a
scratch file, and reading it back were all auto-denied: agents spent
turns hunting for a writable directory, and the openapi.json copy check
REVIEW.md calls blocking could not be run at all ("gh api was
unavailable in this sandbox"). Name the tools the review actually uses.
The conflict resolution on #6243 resolved both conflicted files
correctly and was then rejected by its own guard: "Edits outside the
conflicted set: CLAUDE.md". The agent never touched CLAUDE.md — it had
Edit rights on exactly two paths and no shell. claude-code-action
deletes and restores CLAUDE.md, .claude/, .mcp.json and friends from
the base branch before it runs, because the PR head is untrusted, and
that restore is what dirtied the tree. Name that set once, exclude it
from the stray-edit check, and hand back rather than resolve when a
conflict lands inside it — the restore would silently overwrite the
resolution and stage the base copy.
The four pull_request_target review jobs in claude-bot.yml (Senior
Developer / QA / Tester / Arbiter and their shared rubric) are replaced by
a single review job running the official code-review plugin - the same
skill behind Anthropic's hosted Code Review and the review workflow
/install-github-app generates. The hosted service needs a Team/Enterprise
organisation, so the plugin runs in CI on the maintainer's subscription
instead: inline findings on PR open and ready-for-review, plus manual
(re-)review when the owner or a collaborator comments "@claude review".
The official example triggers on pull_request, but GitHub withholds
secrets from fork runs and essentially every 3x-ui pull request is from a
fork, so the job keeps the lanes' pull_request_target posture: the
workspace is the base revision and nothing from the pull request is
checked out or executed.
What the lanes uniquely knew is distilled into REVIEW.md, handed to the
skill via --append-system-prompt and pinned by bot_context_test.go the way
repo-context.md is: the runtime.Runtime dispatch rule, migration and
upgrade safety, the four-step route contract chain including the unchecked
docs copy, the i18n rule, the three link implementations, and the
wire-format verification bar. The mention job now ignores "@claude
review" comments on pull requests so the review trigger does not also
wake the generic bot, and the lane-only rubric file goes with the lanes.
The remaining prompts also lose their tone micro-rules (no emoji, no
exclamation marks, no filler) and the workflow's comment banners are
removed.
