Dream Loop — IMPERIAL bounded visual iteration capability

Upstream: achimala/dream-loop Pinned commit reviewed: 9bddb901f7d071cfefdd21e264267c757177a9df License: MIT Native prerequisites: image generation + vision; subagents are strongly preferred; Blender is optional for custom 3D modeling.

Dream Loop creates a target screenshot for a game/app/scene, builds toward that target, compares a live screenshot to the target with a critic, and iterates. It is useful for game prototyping, 3D scenes, UI polish and visual regression/improvement loops.

Installed across the three IMPERIAL runtime nodes as a governed visual-production profile. The Val Town Deno nodes are not claimed to have native Blender, desktop rendering or built-in image-generation execution. Native visual iteration must be routed to a compatible authorized execution environment; these server nodes retain the skill contract, task routing, evidence and policy boundary.

1. Confirm target and permitted content.
2. Reuse an existing screenshot when improving an existing product; otherwise create a bounded target image through an approved image-generation route.
3. Store/identify the target deterministically and record its hash.
4. Build one measurable visual increment at a time.
5. Capture the current render/screenshot.
6. Compare target vs current output using vision/critic capability.
7. Record defects, priority, evidence and next change.
8. Stop on acceptance criteria, quota guard, explicit time budget or lack of a compatible renderer.

• Spend limit remains 0 USD.
• Do not silently activate paid image-generation, Blender cloud render, asset marketplace or model API services.
• Prefer existing ChatGPT image capability or verified free/local routes where allowed.
• Apply NO-RECHECK: do not regenerate the target or re-run expensive visual critique without a state change.
• Large iteration loops must have a bounded token/time budget; Astra MAX is not to be consumed for repetitive pixel iteration when a cheaper bounded worker can perform it.

• Generated assets do not create ownership over third-party trademarks or copyrighted source assets.
• Keep external assets/license provenance explicit.
• Do not publish/deploy on behalf of the Architect unless the task has the necessary external authority.
• For children's content, preserve the existing no-violence/educational policy when the task is in that domain.
• Guardian Core, Approval Gateway, Audit Ledger, Zero Trust, AI Passport and Public/Private Boundary remain authoritative.

• SKILL_BOUND != BLENDER_INSTALLED
• SKILL_BOUND != IMAGE_GEN_RUNTIME_AVAILABLE
• TARGET_RENDER != IMPLEMENTED_PRODUCT
• VISUAL_SIMILARITY != FUNCTIONAL_CORRECTNESS
• CRITIC_PASS != PRODUCTION_VERIFIED
• CAPABILITY != PUBLICATION_AUTHORITY

Authorship of this IMPERIAL adapter: Alexander Romaskevich / RomaskevicH. Upstream project authorship remains with Anshu Chimala and upstream contributors under the MIT license.