Ultra Plan Mode

Ultra_Plan_Mode — human-in-the-loop decision by HTML

When there is a lot to review (N items, each with context) and you want a human to decide without reading a wall of text: hand them a single self-contained HTML file where they scan cards, mark each in two clicks, and hand back their selection. It closes the loop — their filtering trains the agent's criteria for the next pass. Atomic: one tool = one clickable decision HTML.

When to use it

• Any review / approval / filtering / triage flow over complex or long items: comms threads, signals/discrepancies, candidates, decisions, queue items, recommendations.
• Auditing a project or area → use the "master questions" preset (below).
• If it's short (1–3 items) → you don't need this; just answer inline.

Procedure

1. Digest the items (never raw): per item → a short title/ID · 1–2 lines of "what it is / what you read" · the key datum (amount, decision, who, date) · a deep link ↗ to the source.
2. Generate the self-contained HTML (no deps/CDN → opens offline) from a data dict via a small script (see examples/generate.py), not by hand, so it scales to N cards. Or edit the REPORT object in template.html directly for a handful of items.
3. Per card: radios ✅ keep / ⚠️ review / ❌ drop + a free-text note for corrections. Optionally pre-load a preset verdict per card; the reviewer confirms or overrides.
4. "Copy selection" button: collects id — mark — note for every touched card into a visible <textarea> + navigator.clipboard. The reviewer pastes it back.
5. Save the .html where the reviewer can open it (a reports folder, or ~/Downloads/).
6. Validate before delivering: balanced tags · zero href="#" (all deep links resolved) · extract the <script> and node --check it.
7. Close the loop: when they paste their selection back → apply the corrections to the data/criteria; if the flow repeats, feed what you learned into the next run.

REPORT schema (for the generator / template)

const REPORT = {
  title, subtitle /* HTML */,
  meta: { date: "YYYY-MM-DD", generator: "ultra-plan-mode", target: "…" },
  readonly: true,
  legend: [{ c: "keep|review|drop", t: "label" }],
  sections: [{
    title,
    items: [{
      id,               // short stable id (printed in the export)
      title,            // card heading
      desc,             // 1–2 digested lines
      why,              // optional: why you lean a way (the key datum)
      link,             // optional deep link (http(s):// or file://)
      preset            // optional: "keep" | "review" | "drop"
    }]
  }]
};

"Master questions" preset (audit a project/area) — READ-ONLY

For each project/area, answer these five; the recommendations that come out become the clickable cards:

1. What's here? — map the structure.
2. Live or legacy? — file count + most-recent mtime per folder → active / dormant / legacy.
3. Curated or orphan? — is there a README / guide / owner? curated vs draft vs no guidance.
4. Where can an agent go wrong? — stale copies of canonical files, mis-named folders, PII, deprecated paths, missing guidance.
5. What to improve / merge / archive? — concrete recommendations, each with a severity.

Guardrails

• Self-contained (zero CDN) · deep links resolved (no link → can't be audited) · digested data, not raw · clipboard with try/catch + <textarea> fallback.
• In audits: READ-ONLY — propose, don't execute. Any destructive reorg (move/delete/ rename) needs the owner's explicit approval.
• Rule #1: name PII as a risk, never copy it into the report.

Output

An .html to open + the return contract: the reviewer pastes back id: ✅ keep / ⚠️ review / ❌ drop — note. That is the training signal for the agent's criteria.