菜单
Implement one step from the build sequence in docs/TODO.md
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
A hierarchical GAN coder campaign. Run only when user mentions using Sol/Luna, claudex fleets, competing implementations, or architecture contests. The CMDR derives checkpointed operational phases from the ingested mission and advances them through fresh blind Opus supervisors; every code-bearing artifact is produced by claudex Sol/Luna Individual Units; supervisors only commission, judge, select, gate, and report.
Run the anopticengine FPS / GPU-pass benchmark and write the result to docs/benchmarks/. Use whenever asked to benchmark, run the sweep, measure frame throughput, or compare a shadow/render knob. Drives tools/perf/bench_fps_<platform>.py (win64 / linux / macos). NEVER invent a runner. Full methodology and every env knob live in tools/perf/bench_fps.md.
Create a test or benchmark in tests/ using the tests/templates primitives
Simplify comments and tighten naming in the given files
Drive an anoptic-engine change at its runtime surface — nix run to build+launch, screenshot-macos to capture, eval+substitution for Linux artifacts from a Mac.
Screenshot or visually inspect a running app's window on macOS (Darwin). ALWAYS use this whenever you need to see a native GUI on macOS — verify a rendered frame, confirm a visual change, capture any on-screen app. Runs .claude/tools/screenshot-macos (screencapture + CGWindowList). NEVER osascript / AppleScript / System Events.
| name | todostep |
| description | Implement one step from the build sequence in docs/TODO.md |
| argument-hint | [step number] |
Read docs/TODO.md, find Step $ARGUMENTS. Restate the acceptance criteria as a checklist. Present an implementation plan and WAIT for approval.
After approval: implement, test, run @tsan-runner if concurrent code was touched, then show git diff --stat and the checklist with pass/fail.
Anything out of scope goes in a "Deferred findings" list at the end.
