DisDat capture → delivery v2 — executable spec

Spec'd 2026-06-18 (via /gstack-spec, adapted: spec drives the build, not a GH issue). Decisions locked with the founder. Source of the gaps: DISDAT-FIDELITY-ROADMAP.md.

Why

Capture is greedy-cheap to build (AI codes it) but the consuming agent pays runtime tokens to ingest a big blob. v1 stuffs everything into one bundle + a long ground-truth wall on the clipboard. v2: capture MORE (close the audit's capture gaps), deliver it as a menu of separately-downloadable artifacts so the agent fetches only the depth its task needs, and shrink the paste to ≤20 lines of codespeak + the full transcript + the menu. North-star: the paste reads like a brief handed to a 10x engineer who will go see the computer-stuff to rebuild the architecture + feel + outcomes the user reacted to.

Locked decisions

• D1 — 5 grouped artifacts (not 8, not 2): core · dom · screenshots · assets · sources. core = codespeak + FULL transcript + elements + network + state + timeline (small, always-grab). The heavy/optional layers are their own files. ~5 uploads/capture.
• D2 — skip cookies + auth-ish keys. Capture localStorage/sessionStorage only; skip cookies entirely, and drop any key matching /token|auth|session|secret|key|jwt|bearer|password|credential/i AND any value that looks like a JWT / long high-entropy secret. Nothing secret reaches the public URL.

What v2 delivers

Delivery (A1/A2) — the menu

1. At Stop, build 5 artifacts; gzip + upload each non-empty one (bg, parallel, the existing 3-host retry per file). core is mandatory; if core upload fails, fall back to the v1 inline ground truth.
2. The clipboard = ≤20 lines: the codespeak prose + a one-line transcript pointer + the MENU (one row per artifact: name · size · one-line "grab-when" · URL) + a per-intent recommendation line.
3. Menu row format: core 18KB codespeak+transcript+elements+API+state [always] <url>. Recommendation (static, by intent): _"clone the look → core + assets + screenshots; rebuild behaviour →
   • sources + dom; wire the data → core (has network+state)."_
4. The FULL transcript lives in core (keeps the paste tight); the paste shows only the crunched codespeak
   • "full transcript + N elements in core".

Capture-more (B*) — each becomes/feeds an artifact

• B1 state → core.state: localStorage + sessionStorage, D2-redacted.
• B2 assets → assets{}: re-host image (<img> currentSrc + CSS background-image) + font (woff2 from document.fonts/loaded faces) + external-SVG bytes as base64, via bg (host perms, past CORS). Capped (per-file 2MB, total 6MB), only first-party + the page's asset CDN, skip analytics. Mirrors pullSource().
• B3 dom → dom: a bounded, script/style-stripped serialization of document.documentElement (node-capped, e.g. ≤4000 nodes / ≤600KB), so structure/order/un-engaged regions are present.
• B4 head-meta → core.head: title, meta (description/viewport/theme-color), OG/Twitter, canonical, JSON-LD. Cheap document.head read.
• B5 soft-nav → hook history.pushState/replaceState so SPA route changes are recorded (today only hard loads re-inject).

Deferred to a v2.1 follow-up (trickier, own pass): rAF/motion sampler (played-timeline), global-store snapshots (Redux/Zustand/Apollo). Noted so they're not silently dropped.

Acceptance criteria

1. The clipboard paste is ≤20 lines for a typical capture (measured), contains the codespeak + the menu
   • a recommendation, and references the full transcript (which is present in core).
2. A capture produces ≥2 artifact URLs (core + at least one heavy) when the page has assets/screenshots; each URL fetches a valid gzipped JSON of just that layer.
3. core JSON contains: prose, transcript (full), elements[], network[], state{}, head{}, timeline.
4. state{} never contains a JWT/bearer/api-key/cookie — verified by re-capturing a page with an auth token in localStorage and confirming it's absent/redacted in the artifact.
5. assets{} carries real image/font BYTES (base64) for first-party assets, capped at 6MB total, with a logged count of what was dropped over budget (no silent truncation).
6. dom carries the full document structure (not just engaged els), node-capped, truncation noted.
7. If core upload fails, the paste falls back to the v1 inline ground truth (never non-functional).
8. Versions stay unified (manifest = bundle-meta = doc).

Files

Out of scope (this pass)

rAF/motion played-timeline, global-store snapshots, IndexedDB, WebGPU, worker-realm capture, video frame grabs — tracked in the roadmap, not built here.

Source document ↗