Differ GitHub issue

vibecode
{"vibecode": {
    "doc": "differ",
    "status": "brainstorm — not part of any version target",
    "purpose": "Online diff viewer that runs both sides through the viewer's Caspian formatter before diffing, so only semantic changes surface",
    "modes": ["paste two scripts", "walk a file's GitHub history"]
}}

Brainstorm for a Caspian-aware diff service.

Why GitHub issue

Caspian's formatter is explicitly personal, not project policy (formatting.md). That makes ad-hoc collaboration work — but two contributors editing the same file can easily produce a diff that's 80% style noise. Differ normalizes both sides to the viewer's style.json before diffing, so the result reflects what actually changed.

It's the "run the formatter before complaining about formatting" social contract, implemented as a viewing surface instead of an editing one.

Paste mode GitHub issue

Two text areas, one diff pane. User pastes two scripts, hits compare, gets a unified or side-by-side diff with both sides reformatted to their personal style first.

GitHub mode GitHub issue

Point Differ at a github.com/user/repo/blob/.../file.casp URL and it walks the file's commit history, normalizes each version, and lets you step through clean version-to-version diffs — useful for "what actually changed in this file over the last week" without GitHub's own style-noisy blob diffs.

Style source GitHub issue

Differ uses the viewer's personal style to normalize both sides before diffing. Two paths in, one policy:

Policy: anonymous default equals canonical default GitHub issue

The style applied to an anonymous user must be the exact same file as the puck.uno canonical default — the same file a logged-in user sees when they haven't customized anything yet. The canonical default is the floor; user preferences are deltas on top.

Rationale: a logged-in user with an unconfigured profile should see identical results to an anonymous user. No surprise difference based on "did you make an account but skip the style step." First-contact and second-contact converge on the same baseline.

This makes the model boringly predictable: there is one default, used everywhere, and everything else is a delta from it.

How it works (and why to trust it) GitHub issue

Differ never parses Caspian text directly. Both sides go through the canonical Caspian→CJS transpiler, then a code generator renders each side back to Caspian using the viewer's style. The diff runs on those rendered outputs.

This is the same architecture as the engine itself — the interpreter consumes CJS, never Caspian source. Extended outward, every tool that reads Caspian goes through the one parser. Formatters, linters, refactors, and Differ are all CJS-in / Caspian-out generators on top.

The trust requirement collapses to a single sharper property: the transpiler is lossless. Lossless here means CJS preserves everything semantically meaningful — code, comments, %vibecode heredocs. If the transpiler round-trips correctly (verifiable against a corpus), Differ can't silently hide a real change. A "show raw text diff" toggle is still worth offering as a sanity-check escape hatch — belt-and-suspenders, not the primary safety mechanism.

This pushes a dependency back onto CJS: comments and %vibecode blocks must be first-class CJS nodes. If they aren't yet, that's the spec change that has to land before any formatter or Differ work begins.

Open questions GitHub issue

Out of scope (for now) GitHub issue

© 2026 Puck.uno