rvst-validation

RVST validation discipline

The rule

You cannot see the window. Every claim about visual or layout behavior has to be backed by a tool output. No eyeball, no vibe check, no "looks good."

This document exists because the failure mode is real and repeats: agent makes a change, launches the app, sees "something," declares victory, moves on. Meanwhile the user sees a white screen, or text overflowing, or a button that doesn't actually click. The discipline below prevents that.

The four levels

Every RVST change should clear as many of these as apply:

1. Unit test

Does the logic compile and its pure behavior round-trip? For parsers, converters, op handlers — this is where drift gets caught cheapest. Run: cargo test --release

2. Snapshot test (--snapshot / --dump)

Does the resolved DOM + layout look right? For layout-sized changes, this is the strongest non-interactive signal. Run: target/blitz/release/<bin> --snapshot app.js app.css | jq ...

3. Interactive test (--test + assertions)

Does the interaction produce the expected state change? Click a button, dispatch an event, verify the DOM or a JS global updated. Run: command pipeline into --test

4. Windowed check (last resort)

User launches the app and confirms it looks right. This is NOT sufficient on its own — the user doesn't see what you think you changed; they see whatever happened after. Use it to confirm after 1–3 pass, not to skip them.

The five-step change template

Every non-trivial RVST change follows these five steps. Skip any and you are guessing.

Step 1 — HYPOTHESIS
  Write down, in words, what measurable thing will be different after the
  change. Example: "Node #95 will have height >= 300px (currently 0)."

Step 2 — BASELINE
  Run the tool that measures the thing. Save output.
  Example: `--dump | grep 'node-95'` → "420x0"

Step 3 — CHANGE
  Make the smallest possible code change. One file ideally.

Step 4 — REMEASURE
  Rerun the same tool on the rebuilt binary. Compare to baseline.
  Example: "420x362" — hypothesis confirmed.

Step 5 — COMMIT or ITERATE
  If confirmed: commit with the hypothesis/measurement in the message.
  If not: the hypothesis was wrong — update your understanding, don't
  keep fiddling.

Anti-patterns that produce bad work

❌ "Let me just launch the app to see"

When the fix concerns layout or computed styles, the windowed app is the least informative surface. You can't inspect computed boxes. You can't compare to the prior state. Use --dump or --snapshot.

❌ "I made a change, it seems to work"

"Seems" is not a measurement. If the baseline was "canvas is 0px tall" and you can't produce a tool output showing "canvas is 362px tall," you haven't confirmed the change.

❌ "Let me disable the broken thing"

Gating off a code path is a workaround, not a fix. If you have to do it (pre-existing bug, out of scope), file a beads issue, leave a comment that says "gated off because , tracked in ," and don't describe it as "fixed."

❌ "The log says the event fired, so it worked"

An eprintln! firing means the Rust code ran. It does not mean the Svelte handler got called, the DOM mutated, the paint surface updated, or the user would see the change. Confirm downstream via --test + diff.

❌ "The test harness doesn't have a command for this"

It almost certainly does, or can be extended in 10 lines. Check rvst-cli/SKILL.md first; test_harness.rs is ~500 lines and easy to add to.

When you genuinely need a windowed check

Some cases only make sense in a real window:

• Vello GPU paint correctness at target DPI
• Animation / frame pacing
• Font metrics under hidpi
• winit window decoration / OS chrome

For those: ask the user to launch and report back. Do not guess.

The recovery path when things go sideways

If an RVST change produces unexpected behavior:

1. Don't undo. Keep the change; the information is more useful than the fear of breaking more.
2. Snapshot the broken state. --dump and --snapshot both — layout
   • structure.
3. Form a specific hypothesis about why the observed behavior differs from expected. Which tool output first shows the deviation?
4. Bisect — if you made three edits, revert two, confirm which one is responsible. Don't "fix in aggregate."
5. Write the learning into AGENTS.md (or the relevant skill) so the next agent doesn't make the same mistake.

Worked example — the canvas-was-0px-tall bug

(This is a real bug we hit during HMR work. Shipped as the fix pattern.)

1. Hypothesis: <canvas> elements with no intrinsic size inside a flex column should resolve to parent height via height:100%.
2. Baseline: --dump showed node #103 as <div> 420×0 — not canvas, and zero height.
3. First finding: Svelte template <canvas> was being mapped to NodeType::View → rendered as <div> by the bridge. Fix: add NodeType::Canvas.
4. Remeasure: <canvas> now. Still 420×0.
5. Second hypothesis: Blitz's replaced_measure_function was doing aspect_ratio = 0 / 0 = NaN and propagating it into the declared size, zeroing the height.
6. Fix: Fall back to 1:1 aspect when inherent is zero.
7. Remeasure: 420×362. Confirmed via --dump and --ascii.

None of those steps involved opening a window. Every hypothesis was confirmed or disproven against a tool output.

Pointer to the tool reference

rvst-cli/SKILL.md — every --test command, every flag, every canonical workflow.