← ClaudeAtlas

astra-reviewlisted

Review a Builder's change against three axes, each in its own context: correctness (the built-in code-review), the ticket's intent and criteria, and the repo's coding standards; then verify every finding at its line, give it one verdict with evidence, and route it. Use after the tests and the UI pass, before simplify, on a full ticket. A quick ticket runs the built-in review alone.
thientranhung/astragentic · ★ 4 · Code & Development · score 70
Install: claude install-skill thientranhung/astragentic
# Review, three axes, one triage ## 1. Pin the change The fixed point is the ticket's base (`Base:` in the brief); the diff is `git diff <base>...HEAD`, untracked files included, staged to a temp file outside the repo. Empty diff: stop. The **intent** is the ticket verbatim: What to build, Criteria, Must not change, and for UI Glance and States; plus the spec FRs the criteria cite. The **standards** are every file that documents how code is written here (`CODING_STANDARDS.md`, `CONTRIBUTING.md`, `AGENTS.md`/`CLAUDE.md` at root and in touched directories) plus the smell baseline below. Skip anything a linter already enforces. ## 2. Three lenses, launched together before any result is read - **Correctness**: `Skill(skill: "code-review")`, the built-in, on the staged diff. It hunts bugs without the intent, which is the point: what is missing, not only what is wrong. - **Intent** (sub-agent, read-only, given the intent and the diff, nothing else): per criterion, the test that proves it and whether it ran; criteria missing or partial; behaviour nobody asked for; criteria that look met but are met at the wrong surface (the intent's expectation lives at surface A, the diff and its tests exercise surface B); each Must-not-change line, still true or not. **Verification gap**: for each changed consumer, would any test fail if this broke? Read the test before saying what it covers; search for the symbol before saying no test exists; a source-text, no-throw, snapshot-on