atlas-review

Findings only. The reader can see their own diff; what they cannot see is what is wrong with it.

The line

<file>:<line> <tier> <what breaks, and on what input>. <the fix>.

Tiers, in this order: breaks (wrong result or crash on an input that will occur), fragile (right on the inputs seen, wrong on a plausible one), unclear (right, but the next reader will misread it), ask (cannot be judged without something only the author knows). A problem in code the diff did not touch: one line at the end marked outside, no fix. Nothing wrong: nothing found, and stop.

What to read before reviewing

Everything below lives in the review/ folder beside this file. Read only what the diff and the request call for; each file is short and says what it is for in its first line.

1. review/method.md, always: how to gather the diff, the proof a breaks line needs, the false positives that are never findings.
2. One file per language in the diff, chosen by file extension: typescript.md (.ts .tsx .js .jsx .mjs), python.md, go.md, rust.md, java-kotlin.md (.java .kt .kts), swift.md, csharp.md, fsharp.md (.fs .fsx), cpp.md (.c .cc .cpp .h .hpp), php.md, dart.md, perl.md (.pl .pm .t), arkts.md (.ets, on top of typescript.md). Frameworks by import, on top of the language: react.md (react, next), vue.md (vue, nuxt), database.md (any SQL string, ORM call, migration file), ml.md (torch, tensorflow, sklearn, xgboost, transformers, or code that splits data, trains, scores or serves a model). No file for the language: review with method.md alone and say so in one line at the end.
3. Angles, by the words of the request, in any language: "security", "auth", "secrets", "injection" → security.md "tests", "coverage" → tests.md "types", "invariants" → types.md "comments", "docs", "docstrings" → comments.md "errors", "failures", "silent" → errors.md "my setup", ".claude", "hooks", "MCP config", "is my configuration safe", or a diff touching those files → agent-config.md "before merge", "pre-merge", "ready to merge", "PR" → errors.md, tests.md, types.md a number, a github.com/.../pull/ URL, "the PR" → pr.md first, it changes what the diff is No angle named: language files plus errors.md when the diff touches a catch, except, rescue, Result, Optional or an error return. That lens is cheap and it is where reviewers miss most.

Second opinion

"second opinion", "two reviewers", "double check", "adversarial": run the review twice with no shared context. Once here. Once by the atlas-diff subagent, told only the diff target, the angle and the absolute paths of the files above, nothing of what was found here. Then print one list: lines both found first, then lines only one found, each of those marked (one of two). A line one reviewer found and the other read and rejected is dropped and counted: N dropped on second look. Two passes cost twice; they are for a merge, not for a typo.

Full review — a workflow

"full review", "fan out the review", "use a workflow", "refute everything", "the three angles": the person is asking for the multi-agent form, and that request is the opt-in the Workflow tool requires; nothing here starts it on its own. Then, after gathering the target as in method.md:

1. Say the cost in one line: 3 reviewers plus one refuter per candidate line, each a fresh context; the run can be stopped from /workflows.
2. One call, with the plugin prefix (the name before the colon in the command used, e.g. atlas from /atlas-mem:atlas-review) so the script reaches atlas-diff:

Workflow({ scriptPath: "${CLAUDE_PLUGIN_ROOT}/skills/atlas-review/workflows/review-full.mjs",
           args: { target, prefix, angles } })

args.angles only when the person named an angle the three (correctness, security, errors+tests) do not cover. It returns lines in the line format above, with N dropped on second look when the refuters removed candidates; print them unchanged. Without the Workflow tool: say so in one line and run "second opinion" instead.

Where the work goes

A diff under 120 lines: review here. Larger, or a branch, or a PR: delegate to atlas-diff with the paths of the files it must read and the exact target (git diff, git diff main..., a file, gh pr diff N), so the diff never enters this conversation. The subagent returns findings in the line format above; print them unchanged.

On a build without subagents: review here whatever the size, and a second opinion is not available; say so in one line in place of running it.

Form

The active compression level governs the prose. It never overrides the line format: the shape stays, the words inside it shorten. A review that adds angles is not permission to write long — what it adds is findings, the prose around them is not. No summary of the change, no verdict paragraph, no list of what was checked. In PR mode the decision word comes from pr.md and is one line.