Atlas check
This compares what was built with what was intended and writes the result down, so the reviewer starts from a list instead of a blank page. It does not decide whether the change is good, safe or complete: a person does.
Read WORKING-AGREEMENT.md and the "Work artifacts" section of AGENTS.md first.
Steps
- Identify the inputs: the story (its numbered acceptance criteria), the brief that declares the expected
paths:, and the change: a PR (--pr <n> --repo owner/name), a git range in a registered clone (--range base..head --source <repo-id>), or a list of files. Pull requests are read-only: never comment on, approve, merge or push to one.
- Scope:
node scripts/scope-check.mjs work/<ID>/brief-<role>.md <the change options>, once per brief if there are several and their paths: differ (and say which you used). Note the files outside the declared paths, declared paths nothing touched, and the kb/ pages that rest on touched files. This compares paths only; say so.
- Scaffold:
node scripts/work.mjs new check <ID> --owner "<name>" --origin agent --parent work/<ID>/story.md.
- Criteria section: one numbered line per acceptance criterion of the story, in the same order and count, each beginning with a bold verdict:
**met** or **partly met**: cite the lines of the change or its tests that show it, as code at the commit ([repo@sha] path:lines). Read what the code does (conditions, defaults, error paths); a comment is what someone wrote, not proof of behaviour, so a criterion shown only by a comment is at most partly met.
**not met**: cite the code that does something else, or say what is absent.
**no evidence**: when there is nothing to cite. This is the honest answer for a criterion nobody tested or implemented visibly (for example behaviour on a restart, or under load that no test exercises). Do not mark it met because it is plausible.
You judge what the cited lines show, not what the author probably meant, and not whether the criterion is a good one.
- Scope section: what the change touched outside the brief, what the brief covers that the change did not, and the
kb/ pages resting on touched files (recommend atlas-refresh after merge; do not edit those pages here). Cite each fact.
- File findings for mismatches where both sides can be cited (
node scripts/findings.mjs new ...; node scripts/findings.mjs list first to avoid duplicates): a changed file outside the brief; a change that contradicts the story or the knowledge base; a story criterion the code contradicts. Point to each from Open questions. A gap with only one side is an open question instead.
- Check it:
node scripts/lint-work.mjs (a "met" with no citation is an error), node scripts/sync-status.mjs, then the atlas-verify skill on the check, because its verdicts are claims about code.
- Tell the user the counts per verdict, the criteria with no evidence, the out-of-scope files, the findings raised, and what you could not check (you did not run the code or the tests unless you say you did; a verdict is about what the lines show). The check is not approved by anyone.
You never mark a story, check or finding as approved, decided or closed, and you never describe a change as ready to merge. The reviewer does that.
1---2name: atlas-check3description: Compares a change (a pull request, a branch range or a set of files) with the story's acceptance criteria and the brief's declared scope, writes a check artifact with a cited verdict per criterion (met, partly met, not met or no evidence), lists files outside the scope and knowledge pages the change touches, and files a finding for each mismatch. Use when someone asks whether a change does what the story says, to review a PR against its story, to verify or test against acceptance criteria, or before saying a piece of work is done.4disable-model-invocation: true5metadata:6 source: https://skillmd.com/skills/chauthehienp3c1/atlas-check7---89# Atlas check1011This compares what was built with what was intended and writes the result down, so the reviewer starts from a list instead of a blank page. It does **not** decide whether the change is good, safe or complete: a person does.12Read `WORKING-AGREEMENT.md` and the "Work artifacts" section of `AGENTS.md` first.1314## Steps15161. **Identify the inputs**: the story (its numbered acceptance criteria), the brief that declares the expected `paths:`, and the change: a PR (`--pr <n> --repo owner/name`), a git range in a registered clone (`--range base..head --source <repo-id>`), or a list of files. Pull requests are read-only: never comment on, approve, merge or push to one.172. **Scope**: `node scripts/scope-check.mjs work/<ID>/brief-<role>.md <the change options>`, once per brief if there are several and their `paths:` differ (and say which you used). Note the files outside the declared paths, declared paths nothing touched, and the `kb/` pages that rest on touched files. This compares paths only; say so.183. **Scaffold**: `node scripts/work.mjs new check <ID> --owner "<name>" --origin agent --parent work/<ID>/story.md`.194. **Criteria section**: one numbered line per acceptance criterion of the story, **in the same order and count**, each beginning with a bold verdict:20 - `**met**` or `**partly met**`: cite the lines of the change or its tests that show it, as code at the commit (`[repo@sha] path:lines`). Read what the code does (conditions, defaults, error paths); a comment is what someone wrote, not proof of behaviour, so a criterion shown only by a comment is at most partly met.21 - `**not met**`: cite the code that does something else, or say what is absent.22 - `**no evidence**`: when there is nothing to cite. This is the honest answer for a criterion nobody tested or implemented visibly (for example behaviour on a restart, or under load that no test exercises). Do not mark it met because it is plausible.23 You judge what the cited lines show, not what the author probably meant, and not whether the criterion is a good one.245. **Scope section**: what the change touched outside the brief, what the brief covers that the change did not, and the `kb/` pages resting on touched files (recommend `atlas-refresh` after merge; do not edit those pages here). Cite each fact.256. **File findings** for mismatches where both sides can be cited (`node scripts/findings.mjs new ...`; `node scripts/findings.mjs list` first to avoid duplicates): a changed file outside the brief; a change that contradicts the story or the knowledge base; a story criterion the code contradicts. Point to each from Open questions. A gap with only one side is an open question instead.267. **Check it**: `node scripts/lint-work.mjs` (a "met" with no citation is an error), `node scripts/sync-status.mjs`, then the `atlas-verify` skill on the check, because its verdicts are claims about code.278. **Tell the user** the counts per verdict, the criteria with no evidence, the out-of-scope files, the findings raised, and what you could not check (you did not run the code or the tests unless you say you did; a verdict is about what the lines show). The check is not approved by anyone.2829You never mark a story, check or finding as approved, decided or closed, and you never describe a change as ready to merge. The reviewer does that.