SKILL.md
Argos PR Review
Argos is a visual testing platform: each CI build captures screenshots and compares them to a baseline, producing per-snapshot diffs. A build with changes-detected is waiting for a human (or you) to decide whether each change is intentional, a regression, or a flaky capture.
Treat the Argos build as one input to the PR review, never the sole source of truth. Infer the intended UI change from the PR title, description, linked issue, and code diff first, then use Argos to confirm the rendered result matches.
Tooling & auth
Drive Argos through the argos CLI — load the argos-cli skill for the token model and flags. In short: build get / build snapshots need a project token (ARGOS_TOKEN/--token); submitting a review or comment needs a personal access token (--token / argos login). Always use --json when parsing, and never print token values. If no PAT is available, give the user your conclusion and evidence instead of posting — the CLI can't submit the review.
Workflow
- Inspect the build —
argos build get <ref> --json. Decide from status:
| Status | Meaning / next step |
|---|---|
accepted / no-changes |
Already approved / no visual diff — no review needed |
pending / progress |
Not ready — stop and report it can't be reviewed yet |
changes-detected |
Needs a decision — fetch snapshots |
rejected / error / aborted/ expired |
Don't approve until the cause is understood |
- Fetch what changed —
argos build snapshots <ref> --needs-review --json.
For each diff inspect url (diff mask), base.url (before), head.url (after), and head.metadata, plus the flakiness signals test.metrics and, on a change, change.occurrences / change.ignored (used in step 3).
- Judge each diff against the inferred intent:
- Intentional — matches the code change and renders cleanly. - Regression — broken layout/overlap, clipping, wrong state/theme/route, missing content, or a removed snapshot with no matching test removal. - Flaky — weigh two independent signals; when they agree, call it flaky with confidence: - Test history — test.metrics.flakiness (0 stable → 1 flaky) and change.occurrences (how many times this exact diff has recurred over the metrics period). A high flakiness score or a recurring change is strong evidence the diff is environmental noise, not this PR's work; stability (builds without a change), consistency (do changes repeat identically), and uniqueChanges explain why it scores that way. change.ignored: true is already known-flaky and auto-approved — never read it as a regression. - This capture — a spinner/skeleton, async content not yet loaded, mid-animation, drifting dynamic values, head.metadata.test.retry > 0, or identical score/head.url across browsers (both captured the same transient state). retries (the configured budget) is not itself a signal.
Conversely, a stable test (flakiness→0, high stability, a first-time change) that changed is more likely intentional or a real regression — don't dismiss it as flaky on the visuals alone. Tune the window with build snapshots --metrics-period <24h|3d|7d|30d|90d> (default 7d).
- Comment on specific diffs — the highest-value output of an agent review.
A binary approve/reject is cheap; specific, anchored feedback is what makes an agent review worth reading. For each problem diff, post a comment that names what's wrong and how to fix it, anchored to that snapshot:
``bash argos comment create <ref> --token <pat> --diff <screenshotDiffId> \ --body "Loader still visible — capture runs before data loads. Wait for settled content (or mark the loader aria-busy)." ``
- <screenshotDiffId> is the diff id from build snapshots --json. - Pin to a region with --anchor-lines <from,to> or --anchor-point <x,y> (normalized 0–1); reply in a thread with --reply-to <commentId>. - Use --draft to bundle comments into your pending review, then submit them together in the next step.
- Submit the review (or report if no PAT is available):
- Approve: argos review create <ref> --token <pat> --event approve - Request changes: argos review create <ref> --token <pat> --event reject --body "<summary>" - Neutral note only: --event comment --body "<summary>" - Add --project owner/project for build-number refs (not URLs). - Silence a confirmed recurring flake so it stops blocking future builds: argos change ignore <change.id> --token <pat> --project owner/project (reverse with change unignore). Only ignore flakes the metrics confirm — never to bypass a real diff; prefer a code fix ([references/flaky-fixes.md](references/flaky-fixes.md)) when one exists. - Lead with the inferred intent, the snapshots reviewed, and the evidence. In the PR, cite the build URL and affected snapshot names; for flakes, name the signal and recommend a fix.
References
- [references/baseline.md](references/baseline.md) — baseline selection and
orphan-build semantics (load when the review depends on which baseline was used).
- [references/flaky-fixes.md](references/flaky-fixes.md) — concrete code fixes for
flaky captures (aria-busy, data-visual-test, animation stabilization).