본문으로 건너뛰기
cccskills
무료GitHub에 공개

build-submit-packaging

Use for Crewrift-specific build submit packaging recipes when optimizing a player.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

설치 방법 보기

포함된 파일(1)

  • SKILL.md4.0 KB

SKILL.md(원문)

설치하기 전에 에이전트에게 전달되는 지시 내용을 확인할 수 있습니다.

Build, package & submit — recipes (crewrift tier)

On-demand recipes (3). Trigger→action heuristics; pull the relevant one when its situation arises.

1. Crewrift configs and the nav bake pipeline ordering; use the full scoring config, verify games via the replay

crewrift

For a crewrift quality signal use the full Among Them config (8 players, 2 imposters, 8 tasks/player, 1200 kill-cooldown ticks, 600 vote-timer ticks, 10000 max ticks, ~120 role-reveal, ~360 game-over, mirrored from bitworld_among_them_8p_2imposters); smoke/certification configs that zero tasks and tick budgets are unusable for quality. Crewrift's certification (smoke) variant uses tasksPerPlayer 1, startWaitTicks 0, roleRevealTicks 0, gameOverTicks 1, maxTicks 300, voteTimerTicks 120, HUD arrows/bubbles off; use it for fast deterministic real-replay capture (pass --save-replay/--save-scores, synthesize episode metadata yourself — the game does not write it; a time-limit draw replay exercises the no-winner render branch). Among Them's canonical tournament variant runs 8 players with exactly 2 imposters and 8 tasks/player at 24 FPS (6000 ticks = 250s). The nav bake has STRICT prerequisite ordering that silently breaks if reordered: walk_mask.bin (derived from upstream assets) must exist before waypoint editing; the nav graph JSON the editor produces must exist before the path-baker computes per-edge costs and the packed nav_paths.bin; gate the re-bake on reporting zero failed/unreachable edges. League episodes can return empty game_stats/policy_results even though games ran, so confirm a real game by decompressing the replay (zlib-compressed as replay.json.z; zippy is a dependency) and checking it is a valid ~41 KB CREWRIFT replay. <sub>sources: claude-code:10f74020-e0e6-4515-8d0c-f071d41766ad, claude-code:25488cd9-61d5-4231-bd87-e2744b6db043, claude-code:48add98e-a83d-4348-97bf-0079c48c42d6, claude-code:63658f66-83c7-482e-a805-d58527e10975 (+6)</sub>

2. Pin CREWRIFT_REF centrally and keep it matched to the live league; record the engine version per submission

crewrift

Pin the crewrift game version centrally in crewrift_lab/tools/versions.env (CREWRIFT_REF), passed as a --build-arg to every Nim player build, and keep it matched to the game version running in the league you target: a player compiled against a different game version can skew from live behavior (the same version-skew that breaks expand_replay). CREWRIFT_REF cannot be auto-resolved (the platform exposes no commit and master runs ahead of :latest), so bump it deliberately when build_expand_replay starts hash-failing on fresh replays (the redeploy signal). Record the engine version a submission was built against, since smoke-config and protocol details (tick counts, frame format) are version-specific; building against a cached older among_them image is acceptable when the newer cannot be pulled (e.g. v0.1.20 used because v0.1.24 returned a ghcr.io 403). <sub>sources: personal_labs/crewrift_lab/docs/designs/building_players.md, coworlds/coworld-crewrift/README.md</sub>

3. The crewborg publish path: amd64 build, upload-policy, submit, live-resolved league id

crewrift · ⚠ session-derived, unverified · tool: coworld

The crewborg publish/submit path is: Docker build as linux/amd64, then coworld upload-policy, then coworld submit. Resolve the Crewrift league id LIVE (e.g. 'Crewrift Daily', or the current 'game-of-week') rather than trusting stale/memory-noted ids, and upload a new version under the EXISTING policy name (e.g. v4 over active v3). New submissions land in the Qualifiers division with status qualifying. Run the player's pytest suite and ruff check before uploading; the Dockerfile's FROM --platform=linux/amd64 triggers an expected lint warning that is not a build failure. <sub>sources: codex:019e8fc7-ebad-7b00-bc95-df1727641ce0, codex:019eaea8-4f78-7072-b4b0-539df963b7df</sub>

리뷰

아직 리뷰가 없습니다. 사용해 본 소감을 남겨 주세요.

댓글

로그인 없이 쓸 수 있습니다. 사용 후기, 질문, 작동하지 않은 환경 등을 편하게 남겨 주세요.

아직 댓글이 없습니다. 첫 댓글을 남겨 보세요.

보내면 바로 공개됩니다. 이메일 주소나 전화번호 같은 개인 정보는 쓰지 마세요. URL은 2개까지입니다.

0 / 1,000자

이 양식은 reCAPTCHA로 보호되며 Google 개인정보처리방침 및 서비스 약관이 적용됩니다.

같은 저장소의 스킬

Use for ab methodology recipes in scripted Coworld policy optimization.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Metta-AI/coworld172026년 10월 14일 업데이트

Use for ab methodology recipes when optimizing a Coworld player.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Metta-AI/coworld172026년 10월 14일 업데이트

Use for Crewrift-specific ab methodology recipes when optimizing a player.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Metta-AI/coworld172026년 10월 14일 업데이트

Use for reusable ab methodology recipes in software and coding-agent work.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Metta-AI/coworld172026년 10월 14일 업데이트

Use for reusable agent workflow collab recipes in software and coding-agent work.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Metta-AI/coworld172026년 10월 14일 업데이트

Use for agent workflow collab recipes when optimizing a Coworld player.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Metta-AI/coworld172026년 10월 14일 업데이트

Metta-AI의 스킬 모두 보기

이 스킬의 문제 신고