Skip to main content

verify-game

Test whether a game runs on the current engine and record the result in highball-db with honest provenance. Use when verifying a title, adding or updating a compatibility entry, writing a game recipe, choosing a renderer for a game, or answering whether a specific game works.

インストールへ移動

ソース情報

リポジトリ
gauthierpiarrette/highball
ソースの最終更新活動
2026年9月1日 19:46
検出された SKILL.md の言語
英語
スター
900
フォーク
32

インストール方法

デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。

ソースファイルを確認

インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。

SKILL.md を表示中

SKILL.md
ソースの指示 · 読み取り専用プレビュー
name
verify-game
description
Test whether a game runs on the current engine and record the result in highball-db with honest provenance. Use when verifying a title, adding or updating a compatibility entry, writing a game recipe, choosing a renderer for a game, or answering whether a specific game works.
# Verify a game and record the result The database is the product. An entry that overstates what was seen is worse than no entry, because every other claim inherits its credibility. Environment paths, commands, and traps: [../investigate-issue/reference.md](../investigate-issue/reference.md) ## 1. Check what is already known Read `../highball-db/db/games/<id>.json` and any existing recipe before testing. If the database already answers this, verify rather than rediscover, and say which claim you are re-testing. Check the anti-cheat data first. A kernel anti-cheat title cannot work under Wine on macOS; record that and stop rather than spending hours proving it. ## 2. Set the bottle up deliberately Decide the renderer and dependencies before launching, and change one variable at a time. After changing a bottle's renderer, restart the bottle. A running Steam client keeps the environment it started with, so the game inherits the old renderer and you test the wrong thing. ## 3. Reach the actual goal, not the first screen A menu is not a verdict. Launchers render while games fail; windows exist while nothing composites. Play far enough to prove the thing a player wants: - a level or match actually loads and runs - the frame rate holds after shader compilation settles - exiting is clean Record where it stops if it stops. ## 4. Capture evidence while it runs - frame rate from the Metal HUD (`bottle set <bottle> hud 1`), not an estimate - the log footer (`# exit=N … after Ns`) for the run - a screenshot when the result is visual ## 5. Test the renderers, and record the failures Try the plausible renderers rather than only the one that worked. Which renderer fails, and how, is the data nobody else publishes, and it is what makes an entry useful. ## 6. Repeat before believing a good result A single success that does not reproduce is not a verdict. Run it again from a clean start. If it only worked once, say so and mark the entry accordingly. ## 7. Write the entry with matching provenance Pick the status by what was actually observed: | status | means | |---|---| | `verified-local` | someone ran it here, start to finish | | `reported-upstream` | named in DXMT release notes as working or fixed; not verified here | | `community` | reported by users, unverified by us | | `blocked-anticheat` | cannot work, anti-cheat | There is deliberately no status for "we tested it here and it does not work". Do not stretch `reported-upstream` to mean that, and do not invent a status: record the failure in `notes` and `knownIssues`, leave the status at what is actually true, and raise the gap with the maintainer. `reported-upstream` is defined by the site build (`Scripts/build-site.py`), and changing what it means silently changes what every existing entry claims. `lastVerified` carries date, engine id, macOS version, and chip, because the answer differs across all four. Notes say what to do, what it costs, and what still fails. ## 8. Ship the fix as data If the game needs setup to work, put it in a recipe in `../highball-db`, never in Swift. Prefer steps that persist (config files the game reads) over launch arguments, and prove persistence by running once without the arguments. ## 9. Claim only what was seen Write "menus render, sessions crash" rather than "works". If the game was not played, say which part was exercised. If a number was not measured, leave it out.
GitHubで見る