A Claude Code skill that reviews a change with several parallel agents and consolidates their findings into one report.
The lenses are separated by evidence source, not by model. Two agents reading the same diff produce correlated findings even on different models; two agents with different evidence produce independent findings even on the same model. Independence is what parallelism buys, so each lens gets something the others do not have: the whole files and their callers, the project's conventions and git history, a deterministic sweep of the repository, or the results of actually running tests and diagnostics.
The corollary is the point of the whole design: an agent that only sees the diff can only review style. Most findings worth having come from files the diff never touched.
- Acquires the target — a local branch, a PR number, or a PR URL in a repository you have not cloned (it clones shallowly and materializes the PR branch, so lenses can read neighbouring code).
- Triages a tier from
gitalone: effective lines changed, whether any touched path is sensitive, whether a public contract moved. The tier and the signal that chose it are printed. - Dispatches the lenses for that tier in parallel.
- Grades every finding by how it was verified — V1 verified with a quoted command output, V2 read off cited code, V3 hypothesis — and requires cheap verification to be run rather than assumed.
- Audits the report: the orchestrator checks that every citation exists and every grade holds, but does not review the code itself or invent findings.
- Writes the report and prints a short summary.
Work is dosed by a deterministic triage computed from git before anything is dispatched: trivial
changes cost nothing, risky ones get every lens. The thresholds and the sensitive-path list live in
references/triage.md, and a project can extend them with a committed .local-ultrareview.yml.
--tier=N always overrides.
The skill is SKILL.md plus a references/ directory read on demand.
# for every project
mkdir -p ~/.claude/skills/local-ultrareview
cp -r SKILL.md references ~/.claude/skills/local-ultrareview/
# or scoped to one repository
mkdir -p <repo>/.claude/skills/local-ultrareview
cp -r SKILL.md references <repo>/.claude/skills/local-ultrareview/Install it in one place only — two copies drift apart, and the one you are reading stops being the one that runs.
| Command | Scope |
|---|---|
/local-ultrareview |
current branch vs the default branch, plus uncommitted changes |
/local-ultrareview 123 |
PR #123 of the current repository |
/local-ultrareview https://github.com/org/repo/pull/123 |
a PR in any repository, cloned or not |
/local-ultrareview src/foo/ |
pending changes under a path |
/local-ultrareview --all-staged |
staged changes only |
/local-ultrareview --tier=3 |
force the full treatment |
git. That is the only hard requirement.gh(or an equivalent platform CLI) optional — needed for PR targets and for the concurrent-change collision check. Without it, those steps skip with one line and the local review continues.- Model hints (
opus/sonnet/haiku) are suggestions. If an alias is unavailable the lens runs on the default model and the report says so. A run never fails over a model hint.
.local-ultrareview/report-<UTC-timestamp>.md at the repo root, gitignored — except when the
repository is not yours (a temporary clone, another organization's code), where nothing is written
inside it and the report goes to the session's working directory instead. The review artifact
belongs to the reviewer, not to the reviewed repository.
references/calibration.md records why the design is what it is — why the evidence axis replaced the
model axis, why consensus stopped being the headline signal, and what observation would falsify the
current design. Read it before changing anything structural.
MIT. See LICENSE.