Export Review Proof for Glassbox

SkillMonitoring & ops

Export reviewable proof (screenshots, flows, logs) for Glassbox ground-truth review. Use this after implementing a change whose visible/behavioral/runtime effect should be reviewed — e.g. a UI change, a multi-step flow, or a computed result that needs evidence. Produces a version:2 manifest reviewable via `glassbox --ground-truth`.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Export Review Proof for Glassbox skill

What this skill tells your AI

The instructions your AI receives, as published by brianwestphal/glassbox in docs/skills/export-review-proof/SKILL.md and read by ahel’s review.

You are working in a project that uses Glassbox for review. When you finish a change whose effect should be seen, don't just describe it — export proof that a reviewer (or a later AI session) can inspect. Glassbox is consumer-only: it runs no capturer. You write proof to disk; Glassbox reviews it.

There are three kinds of proof. Pick by what you're proving.

1. Visual state — a screenshot

For a change to a screen, component, or rendered output:

  • Capture the actual state as a PNG/JPEG (your test runner's screenshot API — Playwright page.screenshot(), etc.). One file per state worth reviewing.
  • Pair it with an expected image: a design spec, a reference render, or the previous run's actual (a regression baseline). The expected may live outside the repo (specs often aren't committed).
  • Glassbox scores the perceptual difference, hides identical pairs, and sorts most-different-first.

2. Behavior over time — a flow

For a multi-step interaction (a journey, a sequence), choose one:

  • One animated SVG. Render the whole interaction to a single animated SVG (e.g. via domotion) and reference it as one step's actual. It animates live in the comparison view. Best when the motion is the point.
  • Per-frame still steps. One ordered PNG per step (NN-step.png), declared as a manifest set. Glassbox renders a per-step "Step k of N" navigator with per-step scoring. Best when each state needs its own pass.

3. Textual / runtime proof — a review note, NOT a screenshot

For logs, computed values, or request/response bodies: do not screenshot text. Write a line-anchored review-note text artifact with the Glassbox producer CLI, attached to the code it proves:

glassbox note add --file <file> --lines <A[-B]> --kind proof --body "What this proves" --artifact run.log

--kind is one of rationale, proof, assumption, alternative-considered, risk, test-evidence; --lines is 1-based (A or A-B); --body accepts - to read from stdin; --artifact (repeatable) attaches a committed proof file. Commit .pr-notes/ alongside the code. See the project's Glassbox doc 20 (or run glassbox note instructions) for the full contract.

Emit the manifest and hand it off

Write a predictable tree, then generate a version: 2 manifest over it:

proof/<area>/
  actual/   ...   # your captures (PNG/JPEG/SVG)
  expected/ ...   # parallel specs / baselines
  manifest.json
{
  "version": 2,
  "comparisons": [
    { "actual": "actual/login.png", "expected": "expected/login.png", "label": "Login", "expectedKind": "spec" }
  ],
  "sets": [
    { "label": "Checkout", "expectedKind": "spec", "steps": [
      { "actual": "actual/checkout/1-cart.png", "expected": "expected/checkout/1-cart.png", "label": "Cart" },
      { "actual": "actual/checkout/2-pay.png",  "expected": "expected/checkout/2-pay.png",  "label": "Pay" }
    ]}
  ]
}

Paths resolve relative to the manifest's directory. Then tell the reviewer:

glassbox --ground-truth proof/<area>/manifest.json

No git repo is required for the manifest mode.

Checklist before you finish

  • Every reviewable visual state has an actual + an expected.
  • Multi-step flows are either one animated SVG or an ordered manifest set.
  • Textual proof is a glassbox note artifact on the relevant line — never a screenshot of text.
  • A version: 2 manifest covers the captured tree, and you handed over the glassbox --ground-truth <manifest> command.

Notes

  • expectedKind is a display hint: spec (design spec), reference (reference render), or previous-actual (a regression baseline, shown as "Baseline").
  • For regression over a flow, point this run's expected at the prior run's actuals and rotate the baseline outside Glassbox (it stores no baseline).
  • Animated-SVG steps need no special handling — Glassbox serves them as raw image/svg+xml so they animate live.

Signals

GitHub stars
34
Forks
1
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
export-review-proof
Source
github.com/brianwestphal/glassbox