Custom reviewers
Add focused reviewers, scope specialists, and filter findings.
Each reviewer has a name, harness, and instructions. Declaring review replaces the default reviewer, so include every reviewer you want to run.
General review and a specialist
ellipsis:
kind: code_review
name: backend-review
description: []
review:
- name: correctness
harness:
type: claude_code
model: claude-opus-5
instructions: |
Find reproducible defects in the changed code. Check callers,
error handling, and tests. Report the failing input or state.
Skip style and formatting.
- name: authorization
harness:
type: claude_code
model: claude-opus-5
instructions: |
Check authorization on changed request handlers. Trace the
authenticated principal to each resource lookup. Report
cases where one account can read or modify another's data.
budget:
run: 12Both reviewers run against the same commit range. Their findings appear together in the posted review. Up to eight reviewers may run in parallel.
Scope a specialist
review:
- name: migrations
harness:
type: claude_code
model: claude-opus-5
pull_requests:
paths: ["migrations/**"]
instructions: |
Check migration ordering, compatibility with the previous
application version, and whether rollback preserves data.
Report concrete failure cases.This replaces the review stage with one migration specialist. It runs only when the pull request changes a matching path.
Reviewer filters accept paths, base, and head. A skipped reviewer uses no budget allocation.
Add a gatekeeper
ellipsis:
kind: code_review
name: verified-findings
filter:
name: verify-findings
harness:
type: claude_code
model: claude-opus-5
instructions: |
Verify each proposed finding against the code. Keep concrete
defects with a reproducible failure. Reject style opinions
and unsupported assumptions. Combine duplicate findings.
budget:
run: 12This keeps the built-in reviewer and adds a pass before posting. The dashboard retains rejected findings and the gatekeeper's reason.
The gatekeeper needs its own instructions. filter: [] disables it.
Run project checks
ellipsis:
kind: code_review
name: tested-review
environment:
image:
setup: |
cd /sandbox/api-repo
npm ci
review:
- name: test-regressions
harness:
type: claude_code
model: claude-opus-5
instructions: |
Review the changed code for regressions. Run the relevant
tests in api-repo. When reporting a failure, include the
command and observed error.
budget:
run: 10The pull request repository is cloned automatically. Environment setup installs dependencies before review starts.
Write reviewer instructions
Name the defects to look for, the evidence required, and what to omit. Ellipsis supplies the commit range and handles posting. Do not ask stage agents to submit GitHub reviews themselves.