Menú
Verify ML correctness, library usage, formatting, tests, dead code, and client/backend integration.
Instalar con Codex o Claude Copia este prompt, pégalo en Codex, Claude u otro asistente, y deja que revise la página de la skill y la instale por ti.
Basado en la clasificación ocupacional SOC
Create a rigorous RL environment specification before implementation.
Add or modify a React/Three.js environment visualization.
Add or modify a supported Gymnasium/SB3 task in backend/mlagents.
Plan multi-agent and simulator-heavy environments using appropriate APIs instead of forcing them into single-agent SB3.
| name | scientific-verification |
| description | Verify ML correctness, library usage, formatting, tests, dead code, and client/backend integration. |
| argument-hint | [scope] |
| allowed-tools | Read, Write, Glob, Bash |
| context | repo |
| agent | three-mlagents-evaluation-scientist |
Use this skill before declaring work complete.
Run the broad backend checks unless the change is documentation-only:
uvx ruff check backend/examples backend/mlagents backend/main.py backend/services backend/tests
uvx ruff format --check backend/examples backend/mlagents backend/main.py backend/services backend/tests
uv run --project backend python -m unittest discover -s backend/tests
uv run --project backend three-mlagents list --trainable-only
For a changed trainable task, also inspect the task and run a short fixed-seed train/evaluate sanity pass. If full benchmark training is too expensive, state that clearly and do not claim final performance.
Run:
cd client && pnpm run build
For route or WebSocket changes, verify with live backend and frontend servers. Confirm that trainable tasks call real backend routes and roadmap tasks do not show fake training controls.
Check:
Search for:
rg "three_mlagents|backend/three_mlagents|onnxruntime|hand-rolled|fake policy|TODO" .
rg "console\\.log|debugger" client/src
git status --short
Do not remove unrelated user changes. Do not commit generated policies, caches, build output, TensorBoard runs, or dependency directories.
Smoke tests can support the evidence, but they do not satisfy a full implementation request by themselves. Completion requires code, docs, tests, and client/backend behavior to agree.