Unity expert (technical director and agent protocol)

Expert Unity work is a loop run against numbers: set the budget, build one stage, measure it, look at it, fix it, then advance. An agent without a mouse gets there by choosing the right channel for each action, editing through serialization with explicit saves, and never trusting its own eyes alone. This skill is that protocol, the shared toolkit every scenario-unity-* skill imports (scripts/), and the map of the team. Target: Unity 6000.3.21f1 (6.3 LTS), URP 17.3, macOS Apple Silicon. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

Status (2026-09-24, after blind grading and the refactor): every toolkit function ran in Unity 6000.3.21f1 here (tests/code/unity-expert/run_all.sh --live --builds; results in archive/tests/unity-expert/live_results.jsonl). Every call from before the refactor is unchanged.

Stance (the expert delta)

1. One editor per project, and the live editor wins. "You can't open a project in batch mode while the Editor has the same project open" (6.3 Manual). When an editor holds the project, drive it (ut_live), never hand-edit .unity/.prefab/.asset YAML: wrong fileIDs, invisible until reimport, easy to hit the wrong scene (Unity's unity-cli skill).
2. A job is only as honest as its result line. Pass on the AGENT_RESULT envelope and the log scan, not on "it exited". A compile error in ANY assembly aborts every batch job with "Scripts have compiler errors" (observed; the live editor's twin is Safe Mode, which also blocks com.unity.pipeline).
3. Numbers before pixels, milliseconds not fps. 16.67 ms at 60 fps, 33.33 at 30; mobile keeps about 65% for thermals (Unity profiling e-book). Editor Play mode is for iteration; the verdict comes from a development player on the lowest target device.
4. Look AND check, and know the capture's lies. Every frame goes through ut_review.image_checks and a contact sheet you open. Epic's lighting agent approved an all-white frame (scenario-unreal-expert sources); here batch captures showed first-frame white materials, culled 2D lights, bind poses, no Overlay UI and black metals (observed by the team): anything wrong only in an Editor capture is re-checked in a player first. Agent-built levels come out "90%" done, a shelf in front of a door (Code Monkey, ppFshgFOgXU [00:10:52]).
5. Fix the platform at launch. Target switches inside a batch job silently do nothing (6.3 Manual); -buildTarget/-activeBuildProfile at launch, one process per target. The editor reopens on the last active platform (observed): pin it.
6. Named jobs over eval. eval to explore, a named job for anything repeated: "A written command turns that workflow into something deterministic and repeatable" (DgNrgZeJOxQ [00:07:41]); generic run-command tools drift (git-amend, VPjo-M6mPkE [00:03:54]).
7. Tests are the contract. Give work a failing test and a fence ("It can only touch the card data", DgNrgZeJOxQ [00:05:31]); -runTests never with -quit.
8. Be honest about visual editors. Shader Graph, VFX Graph, the Animator window, UI Builder, Tile Palette painting and terrain sculpting have no authoring API: use the substitute (HLSL, templates with exposed properties, AnimatorController API, UXML/USS text, TerrainData) or name the human path.

Channels