Browser playground with BETH operator console

[buchk]

What

A zero-install, zero-backend playground deployed to GitHub Pages (https://buchk.github.io/TypeFerence/) on every push to main: edit typed source in the browser and watch compiled Codex/Copilot/Cursor/neutral artifacts, diagnostics, the embedding graph, and the determinism digest update live — plus an Equivalence tab that walks the full BETH loop without ever asking for a credential.

How

• go/cmd/typeference-wasm — a js/wasm entry point over the existing internal/ packages, exposing compile and pack. The compiler is not forked, wrapped, or reimplemented; the wasm runs the same compile.Validate → compile.Build → compile.HashDirectory and eval.Pack pipelines as the CLI.
• web/playground/memfs.js — an in-memory implementation of the Node-style fs API that Go's js/wasm syscall layer expects, so the compiler's ordinary os calls work unchanged inside the tab.
• web/playground/ — dependency-free HTML/CSS/JS: syntax-colored editor, artifact browser, embedding graph (computed structural-interface satisfaction shown as dashed edges), resolved-bundle view, shareable gzip+base64url links, light/dark theme, CSP pinning connect-src to the page's own origin.
• BETH operator console (beth.js, ADR-0011) — packs scenario × surface cells in-browser with the real equivalence pack code (examples ship scenarios); cells render as copy-prompt/paste-response cards, which is exactly BETH's attested-operator collection model (ADR-0009); the assembled run exports as a deterministic ustar+gzip archive (fixed mtime); scoring happens locally (typeference equivalence score <run> --live — the judge key stays in your terminal); dropping scorecard.json back onto the page renders adherence per surface, cross-surface agreement, and every divergence with the judge's reasoning.
• make playground builds the gitignored assets; wasm_exec.js is copied from the building toolchain's GOROOT so it can never skew from the wasm binary. CI builds the bridge and packer to prevent bitrot.

Determinism, made interactive

The in-memory source dir is named after the loaded example and the ARD publisher domain matches the repo's own build commands, so the Helio example reproduces the committed dist/ digest exactly (a22f5410ba5f8e17…) and the maintainer example recompiles the repository-root AGENTS.md byte for byte (8b21fee0…) — browser, Windows CLI, and CI's Linux build all agree on the bytes.

No secrets in the browser (ADR-0011)

A bring-your-own-key run surface was built first, worked, and was rejected anyway: asking visitors to paste provider keys into a hosted page is the anti-pattern this project's trust posture argues against. The ADR records the decision and preserves the verified transport knowledge (e.g. OpenAI only answers browser CORS on the Responses API). The console needs no keys at any step that touches this page.

Verified end to end

• gofmt/go vet/go test ./... clean; wasm bridge and packer build in CI.
• All three examples compile in-browser with digests matching CLI runs; error paths, graph, bundle, share links, stale-pack marker exercised in a real browser session.
• Full BETH round trip with the real toolchain: browser-packed starter run → real download → extracted with system tar → scored by the CLI with zero workspace drift (the CLI's SHA-256 digest check verified the browser-written bytes exactly) → judged via seeded judge-response.json including a deliberate divergence → the CLI's scorecard.json rendered back in the browser (fail banner, 3/3·3/3·3/3·2/3 adherence, 2/3 agreement, divergence with all four verdicts).

Also

• ADR-0010 (playground), ADR-0011 (equivalence console), and a design note capturing the typed-context-shapes idea for a future ADR (docs/design-notes/typed-context-shapes.md).
• Pages must be enabled once with source "GitHub Actions" (the workflow passes enablement: true, which usually handles this on first deploy).

🤖 Generated with Claude Code