Review Video With Pingfusi

Workflow

1. Run pingfusi doctor. If the review login is missing, stop and have the user run pingfusi setup; there is no offline substitute for a human verdict.

2. Assemble the review context before rendering anything final: every prompt in authored order (active / replaced / context), the distilled current_brief, and requirements with prompt provenance. The complete context caps at 250 KB.

3. Render the MP4 and publish it through Pingfusi hosting by default:

   pingfusi publish <render.mp4> --name <name>-round-1 \
     --record .pingfusi/video/<name>/round-1.json --json

   The command creates the player wrapper, uploads immutable bytes, and returns a direct asset_url; use that value as video_url. Pingfusi serves it with 206 and Content-Range, so the native player can scrub. The current hosted-video cap is 25 MB per render. If a render cannot fit after reasonable encoding, use another long-lived public host that serves Range requests; do not introduce a live-site tunnel for a file.

4. File the round against a caller-owned state file:

   Matching a render against a brief you wrote — the iteration loop:

   const core = require("pingfusi/packages/core");
   const { ping_id } = await core.review.file(stateFile, {
     media_type: "video",
     video_url,
     current_brief,
     prompt_history,  // [{ id, text, state: "active"|"replaced"|"context", replaced_by? }]
     requirements,    // [{ id, text, prompt_ids }]
     n_target: 1,
     approve_verdicts: ["Matches the prompt"], // local bookkeeping, stripped before the wire
   });

   Judging any other video — no brief, your own questions:

   const { ping_id } = await core.review.file(stateFile, {
     media_type: "video",
     video_url,
     title: "Launch video, third cut",          // PUBLIC — on the browsing row
     video_headline: "Does this ad land?",      // private, top of the reviewer's panel
     video_intro:
       "Watch it once at full speed, then scrub back to anything that made you hesitate.",
     steps: [                                    // private; your questions
       { text: "Would you keep watching past the first five seconds?", options: ["Yes", "No"] },
       { text: "What is this selling, in your words?" },
     ],
     verdict_options: ["Ready to run", "Needs another cut"],   // private
     n_target: 3,
     approve_verdicts: ["Ready to run"],
   });

   Mix them freely: keep requirements without a prompt_history when you want timestamped notes linked to specific claims, or send a current_brief with your own steps. The one rule is that the reviewer must have something to answer — a brief, requirements, steps, or an intro. url and draft_url must be absent; video mode refuses them.

5. The filing command automatically chains client-safe wait legs until feedback. If a raw MCP leg returns pending, immediately call pingfusi_wait again; never return pending to the user or file a duplicate. Each leg renews the short idle lease; passive result/verify reads do not (a lapse only pulls the round from the feed for new claims; a reviewer mid-review can still finish). When results land, read the envelope: comments arrive sorted by video_anchor.time_ms, drawn annotations in normalized frame coordinates (0 = left/top, 1 = right/bottom), questionnaire answers attached to their questions. Fix every noted moment in the source, re-render, publish the NEW file under a new receipt/URL, and refile with the same context — update current_brief/requirements only if the user's ask actually changed.

6. Repeat until core.review.verify(stateFile) returns ok === true on your declared approving verdict — Matches the prompt by default, or whichever of your own verdict_options you named in approve_verdicts. Record the receipt; stop only on approval or when the user says stop.