Author a cinematic recipe pack

Slopcamera's extension surface is declarative: coding agents extend behavior by authoring bounded JSON documents, never by registering code. The operation registry is closed, and the cinematic-world workflow composes only registered operations. A recipe pack is the complete input contract for that loop.

Documents an agent authors

1. Scene (slopcamera.spatial-scene, schemaVersion 1) — the admitted world. Usually already admitted through slopcamera scene commands; the pack binds it by digest, it is never embedded twice.
2. Direction (slopcamera.spatial-direction, schemaVersion 1) — semantic beats, actions, camera coverage, and look intents. projectDigest must equal the scene's canonical SHA-256. Character and subject references are semantic ids (hero), which compile against entity_hero-style scene entities.
3. Recipe pack (slopcamera.spatial-recipe-pack, schemaVersion 1) — the inert plan document:

{
  "kind": "slopcamera.spatial-recipe-pack",
  "schemaVersion": 1,
  "packId": "recipe_<slug>",
  "sceneSha256": "<canonical scene sha-256>",
  "direction": { "<the direction document>" },
  "cameraId": "camera_main",
  "axes": ["camera", "lighting"],
  "previews": [{
    "name": "reel",
    "request": {
      "cameraId": "camera_main",
      "mode": { "kind": "beauty" },
      "selection": { "kind": "frame", "timeUs": 0 }
    }
  }],
  "effects": {
    "renderPlan": {
      "kind": "slopcamera.spatial-render-plan",
      "schemaVersion": 1,
      "quality": {
        "outputBytes": 8000000, "particleCount": 0, "pixelBudget": 2073600,
        "simulationSteps": 0, "texturePixelBudget": 2073600, "tier": "preview"
      }
    },
    "particleSystems": [],
    "simulationBakes": []
  },
  "temporalAudit": { "timesUs": [0, 2000000, 4000000] }
}

Boundaries the pack enforces

• axes — 1 to 6 unique values from performance | camera | lighting | materials | effects | sequence.
• previews — at most 4 named render requests; each request is revalidated against the spatial render-request schema at workflow admission.
• effects — optional; when present it requires renderPlan (a declarative slopcamera.spatial-render-plan document with a quality tier and optional post-process stack) plus bounded particleSystems/simulationBakes. Each preview render first plans an integrity-bound effects document (scene.effects.plan) and renders with it.
• temporalAudit — optional; requires cameraId; timesUs bounds to 64 samples.
• packId — recipe_ prefix, at most 64 slug characters.

Invoking the workflow

The workflow input is { "pack": <pack>, "scene": <scene>, "source": { "path": "<repo-relative scene path>" } }. pack.sceneSha256 must equal the canonical scene digest, so compute it first:

slopcamera scene inspect <scene.json> --json   # read sceneSha256

Plan is effect-free — it compiles the graph and reports requirements only:

slopcamera workflows plan cinematic-world --input input.json --json

Run executes the graph: inspection, direction check, compilation, one gallery plan per axis, per-preview effects binding and render, and the temporal audit.

slopcamera workflows run cinematic-world --input input.json --json

What the workflow never does

• It never selects or promotes a candidate. spatial.project.select-candidate and iteration.select are separate explicit operations a caller invokes after reviewing the plan outputs.
• Compiled direction stays verified: false; it is a proposal document, not applied state.
• Planning and graph compilation execute no operations, read no credentials, contact no providers, and mutate no project state.