Kiln Batch Dispatch Agent Skill

Run comparable asset trials

Use the user's requested harness and exact model identifier. Verify what the provider actually ran; do not silently substitute a default or older model.

• Create a fresh workspace for each candidate using clean-room evaluation. A named Kiln project is optional; use matching project configuration only when the trial calls for it. Keep the engine checkout and example collection outside the agent's task context.
• Give each candidate the same brief, installed skills, image budget, time limit, and starting asset where applicable. Record inherited tools, instructions, memory, and permissions.
• Verify tool access and actual image delivery with one small run before dispatching a batch. Model vision support and harness image forwarding are separate checks.
• Exercise the whole revision loop: render source once, use the returned programRef, read a source window, edit an exact anchor, inspect a part, and export the same revision. Check that later calls do not resend the program and that unrelated source survives the edit.
• Retain the brief, exact model, harness version, requested and independently confirmed thinking effort, transcript, source, images, package/skill hashes, and outcome. Leave unexposed effort unknown. Keep the original author and each later refiner distinct. Rebuild the saved source independently of the agent's success claim.
• Review shapes under consistent cameras and lighting. Separate provider outages or quota limits, harness failures, tool/schema failures, engine failures, and visual quality. An interrupted run is not a completed asset.

After a shared package or skill change, rerun the affected workflow with the final candidate. A successful earlier revision does not validate later instructions. Record manual repairs separately from model output.

Adding a reviewed asset to a gallery is a separate task. Do it only when that addition is requested, retaining model provenance and later edits.