Gate assets for review

Gate assets for review

The chokepoint between "an agent built something" and "a human reviews it". Standards come from golden.json and .hologram/checks.py — the gate applies them, it never defines them.

The one rule: you may stamp pending-review or gate-failed. You must NEVER mark an asset approved or rejected — approval is the human's move.

What to do

1. Resolve the target. An asset id, a .glb path, a category, or all. A bare name → match against manifest id stems; nothing matches → stop and say so.

2. Run the gate checks. hologram check --json, filtered to the target assets. With a golden.json present this includes the tri budgets and golden-thumbnail drift automatically. An error-severity finding fails the gate for that asset; warnings pass with notes.

3. Stamp the manifest. For each gated asset, edit its record in exports/manifest.json (agent-written — Hologram itself never writes):

   "review": {"status": "pending-review" | "gate-failed", "version": <record version>, "at": "<ISO-8601Z>"}
   

   • A human decision (approved / rejected) on the same version stands — do not overwrite it. A version bump re-opens the asset: stamp normally.
   • Preserve every other key in the record; edit surgically.

4. Log the gate. Append one line per run to <export_root>/activity.jsonl:

   {"ts": "<ISO-8601Z>", "actor": "claude", "skill": "asset-gate", "target": "<target>", "note": "P passed, F failed"}
   

5. Report. Per asset: id, pass/fail, resulting status, and for failures the exact checks that missed with their numbers:

   Gated 3 assets (target: props)
     ✓ props/market_stall   → pending-review
     ✗ props/crane_arm      → gate-failed (1 error)
         - tri budget: 44,120 tris > props budget 30,000
   

On failures: report and STOP

The fix belongs in the generating script or source file, not the exported GLB — and the user may want to make that call. Do not auto-fix unless explicitly asked. Never touch golden.json: agents build against the goldens; only the human moves them.