메뉴
Repo-specific validation and PR instructions for mage-bench.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Debug a failing golden test from GitHub Actions logs and fix the root cause without papering over nondeterminism.
Investigate a game from raw logs, trace bugs to source code, and file detailed issues. Use for deeper debugging than export-only analysis.
Find Magic cards for compact golden-test scenarios using existing repo goldens, curated references, and Scryfall search recipes.
Analyze exported game files to assess gameplay, model decisions, and likely bugs without raw logs. Use for quick game triage.
Write or update low-noise golden prompt tests in tests/, including minimal decks, replay scripts, and golden regeneration.
Create a new game export schema version, migration module, and related tests.
| name | create-pr-local |
| description | Repo-specific validation and PR instructions for mage-bench. |
This file provides repo-specific instructions for the global /create-pr skill.
If the diff changes prompt rendering, bridge responses, MCP tool output, replay behavior, or exported game data, search existing goldens for the affected behavior and regenerate every stale prompt/export now. Do not assume only newly added tests need updates, and do not wait for CI to remind you.
make check # Full lint + typecheck + tests
make check VERBOSE=1 over launching a second blind make check.scripts/checks/quiet_check.py) can stay silent for minutes while long subchecks run. If it looks hung, inspect the process tree (e.g. pstree -ap <quiet_check_pid>) to see which subcheck is active before assuming it's stuck.make check fails in website-related targets with npm ERR! EEXIST symlink errors, or verify-schema-types claims website/src/types/game-export.d.ts is stale on an otherwise clean tree, check whether parallel npm install runs are racing inside the website targets before assuming the generated types actually need regeneration.Check website/package-lock.json before pushing. make check / website tooling can add incidental "peer": true lockfile churn even when you did not intentionally change website dependencies; drop unrelated lockfile noise so the PR stays scoped.