@opengamehq/cli 0.9.7 → 0.9.9

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (349) hide show
  1. package/AGENTS.md +37 -17
  2. package/CHANGELOG.md +18 -0
  3. package/README.md +7 -6
  4. package/dist/cli.mjs +144 -144
  5. package/dist/opengame-global.js +3 -3
  6. package/dist/preflight-worker.mjs +14 -14
  7. package/dist/runtime-builder.mjs +1 -1
  8. package/dist/skills/manifest.json +1406 -142
  9. package/dist/skills/opengame-accessibility/SKILL.md +48 -32
  10. package/dist/skills/opengame-all-in-one/SKILL.md +195 -0
  11. package/dist/skills/opengame-all-in-one/references/opengame-accessibility/README.md +102 -0
  12. package/dist/skills/opengame-all-in-one/references/opengame-api/README.md +107 -0
  13. package/dist/skills/opengame-all-in-one/references/opengame-assets/README.md +168 -0
  14. package/dist/skills/opengame-all-in-one/references/opengame-assets/references/pipeline.md +164 -0
  15. package/dist/skills/opengame-all-in-one/references/opengame-audio/README.md +238 -0
  16. package/dist/skills/opengame-all-in-one/references/opengame-audio/helpers/audio-chain.js +325 -0
  17. package/dist/skills/opengame-all-in-one/references/opengame-audio/references/mix-notes.md +125 -0
  18. package/dist/skills/opengame-all-in-one/references/opengame-audio/references/qa-without-ears.md +177 -0
  19. package/dist/skills/opengame-all-in-one/references/opengame-capture/README.md +128 -0
  20. package/dist/skills/opengame-all-in-one/references/opengame-capture/references/pipeline.md +114 -0
  21. package/dist/skills/opengame-all-in-one/references/opengame-capture/references/threejs.md +161 -0
  22. package/dist/skills/opengame-all-in-one/references/opengame-characters/README.md +180 -0
  23. package/dist/skills/opengame-all-in-one/references/opengame-characters/references/first-person-body.md +108 -0
  24. package/dist/skills/opengame-all-in-one/references/opengame-characters/references/integration.md +144 -0
  25. package/dist/skills/opengame-all-in-one/references/opengame-characters/references/player-og.md +160 -0
  26. package/dist/skills/opengame-all-in-one/references/opengame-cli/README.md +189 -0
  27. package/dist/skills/opengame-all-in-one/references/opengame-deploy/README.md +223 -0
  28. package/dist/skills/opengame-all-in-one/references/opengame-deploy/references/json-events.md +63 -0
  29. package/dist/skills/opengame-all-in-one/references/opengame-deploy/references/manifest.md +89 -0
  30. package/dist/skills/opengame-all-in-one/references/opengame-deploy/references/packaging.md +119 -0
  31. package/dist/skills/opengame-all-in-one/references/opengame-depth/README.md +154 -0
  32. package/dist/skills/opengame-all-in-one/references/opengame-enemies/README.md +155 -0
  33. package/dist/skills/opengame-all-in-one/references/opengame-enemies/references/pacing-and-bots.md +108 -0
  34. package/dist/skills/opengame-all-in-one/references/opengame-enemies/references/threejs.md +193 -0
  35. package/dist/skills/opengame-all-in-one/references/opengame-front-end/README.md +185 -0
  36. package/dist/skills/opengame-all-in-one/references/opengame-front-end/references/launch.md +50 -0
  37. package/dist/skills/opengame-all-in-one/references/opengame-front-end/references/menus.md +81 -0
  38. package/dist/skills/opengame-all-in-one/references/opengame-game-feel/README.md +181 -0
  39. package/dist/skills/opengame-all-in-one/references/opengame-game-feel/helpers/spring.js +146 -0
  40. package/dist/skills/opengame-all-in-one/references/opengame-game-feel/references/feel-numbers.md +142 -0
  41. package/dist/skills/opengame-all-in-one/references/opengame-game-feel/references/killcam.md +88 -0
  42. package/dist/skills/opengame-all-in-one/references/opengame-jam/README.md +114 -0
  43. package/dist/skills/opengame-all-in-one/references/opengame-lighting/README.md +141 -0
  44. package/dist/skills/opengame-all-in-one/references/opengame-lighting/helpers/light-pool.js +185 -0
  45. package/dist/skills/opengame-all-in-one/references/opengame-loading/README.md +135 -0
  46. package/dist/skills/opengame-all-in-one/references/opengame-loading/helpers/loading-count.mjs +115 -0
  47. package/dist/skills/opengame-all-in-one/references/opengame-loading/references/threejs.md +141 -0
  48. package/dist/skills/opengame-all-in-one/references/opengame-look/README.md +181 -0
  49. package/dist/skills/opengame-all-in-one/references/opengame-look/references/matching-a-reference.md +73 -0
  50. package/dist/skills/opengame-all-in-one/references/opengame-look/references/painterly.md +69 -0
  51. package/dist/skills/opengame-all-in-one/references/opengame-look/references/post-and-lens.md +73 -0
  52. package/dist/skills/opengame-all-in-one/references/opengame-maps/README.md +74 -0
  53. package/dist/skills/opengame-all-in-one/references/opengame-maps/references/map-spec.md +77 -0
  54. package/dist/skills/opengame-all-in-one/references/opengame-mobile/README.md +86 -0
  55. package/dist/skills/opengame-all-in-one/references/opengame-montage-make/README.md +183 -0
  56. package/dist/skills/opengame-all-in-one/references/opengame-montage-make/references/briefs.md +293 -0
  57. package/dist/skills/opengame-all-in-one/references/opengame-montage-make/references/day-one-questions.md +46 -0
  58. package/dist/skills/opengame-all-in-one/references/opengame-montage-make/references/edge-cases.md +27 -0
  59. package/dist/skills/opengame-all-in-one/references/opengame-montage-make/references/filmable.md +244 -0
  60. package/dist/skills/opengame-all-in-one/references/opengame-montage-make/references/finishing.md +126 -0
  61. package/dist/skills/opengame-all-in-one/references/opengame-montage-make/references/measure.md +282 -0
  62. package/dist/skills/opengame-all-in-one/references/opengame-montage-make/references/montage-capture.md +151 -0
  63. package/dist/skills/opengame-all-in-one/references/opengame-montage-make/references/montage-gates.md +84 -0
  64. package/dist/skills/opengame-all-in-one/references/opengame-montage-make/references/montage-genres.md +198 -0
  65. package/dist/skills/opengame-all-in-one/references/opengame-montage-make/references/montage-grammar.md +320 -0
  66. package/dist/skills/opengame-all-in-one/references/opengame-montage-make/references/montage-ops.md +127 -0
  67. package/dist/skills/opengame-all-in-one/references/opengame-montage-make/references/og-snipers-record.md +110 -0
  68. package/dist/skills/opengame-all-in-one/references/opengame-montage-make/references/owner-eye.md +214 -0
  69. package/dist/skills/opengame-all-in-one/references/opengame-montage-make/references/reference-study.md +293 -0
  70. package/dist/skills/opengame-all-in-one/references/opengame-montage-make/references/song.md +64 -0
  71. package/dist/skills/opengame-all-in-one/references/opengame-montage-plan/README.md +193 -0
  72. package/dist/skills/opengame-all-in-one/references/opengame-montage-plan/references/briefs.md +293 -0
  73. package/dist/skills/opengame-all-in-one/references/opengame-montage-plan/references/day-one-questions.md +46 -0
  74. package/dist/skills/opengame-all-in-one/references/opengame-montage-plan/references/filmable.md +244 -0
  75. package/dist/skills/opengame-all-in-one/references/opengame-montage-plan/references/measure.md +282 -0
  76. package/dist/skills/opengame-all-in-one/references/opengame-montage-plan/references/montage-capture.md +151 -0
  77. package/dist/skills/opengame-all-in-one/references/opengame-montage-plan/references/montage-gates.md +84 -0
  78. package/dist/skills/opengame-all-in-one/references/opengame-montage-plan/references/montage-genres.md +198 -0
  79. package/dist/skills/opengame-all-in-one/references/opengame-montage-plan/references/montage-grammar.md +320 -0
  80. package/dist/skills/opengame-all-in-one/references/opengame-montage-plan/references/montage-ops.md +127 -0
  81. package/dist/skills/opengame-all-in-one/references/opengame-montage-plan/references/og-snipers-record.md +110 -0
  82. package/dist/skills/opengame-all-in-one/references/opengame-montage-plan/references/owner-eye.md +214 -0
  83. package/dist/skills/opengame-all-in-one/references/opengame-montage-plan/references/reference-study.md +293 -0
  84. package/dist/skills/opengame-all-in-one/references/opengame-montage-plan/references/song.md +64 -0
  85. package/dist/skills/opengame-all-in-one/references/opengame-montage-review-and-improve/README.md +174 -0
  86. package/dist/skills/opengame-all-in-one/references/opengame-montage-review-and-improve/references/briefs.md +293 -0
  87. package/dist/skills/opengame-all-in-one/references/opengame-montage-review-and-improve/references/day-one-questions.md +46 -0
  88. package/dist/skills/opengame-all-in-one/references/opengame-montage-review-and-improve/references/filmable.md +244 -0
  89. package/dist/skills/opengame-all-in-one/references/opengame-montage-review-and-improve/references/finishing.md +126 -0
  90. package/dist/skills/opengame-all-in-one/references/opengame-montage-review-and-improve/references/measure.md +282 -0
  91. package/dist/skills/opengame-all-in-one/references/opengame-montage-review-and-improve/references/montage-capture.md +151 -0
  92. package/dist/skills/opengame-all-in-one/references/opengame-montage-review-and-improve/references/montage-gates.md +84 -0
  93. package/dist/skills/opengame-all-in-one/references/opengame-montage-review-and-improve/references/montage-genres.md +198 -0
  94. package/dist/skills/opengame-all-in-one/references/opengame-montage-review-and-improve/references/montage-grammar.md +320 -0
  95. package/dist/skills/opengame-all-in-one/references/opengame-montage-review-and-improve/references/montage-ops.md +127 -0
  96. package/dist/skills/opengame-all-in-one/references/opengame-montage-review-and-improve/references/og-snipers-record.md +110 -0
  97. package/dist/skills/opengame-all-in-one/references/opengame-montage-review-and-improve/references/owner-eye.md +214 -0
  98. package/dist/skills/opengame-all-in-one/references/opengame-montage-review-and-improve/references/reference-study.md +293 -0
  99. package/dist/skills/opengame-all-in-one/references/opengame-montage-review-and-improve/references/song.md +64 -0
  100. package/dist/skills/opengame-all-in-one/references/opengame-multiplayer/README.md +139 -0
  101. package/dist/skills/opengame-all-in-one/references/opengame-multiplayer/references/bot-lobbies.md +71 -0
  102. package/dist/skills/opengame-all-in-one/references/opengame-multiplayer/references/netcode.md +165 -0
  103. package/dist/skills/opengame-all-in-one/references/opengame-performance/README.md +177 -0
  104. package/dist/skills/opengame-all-in-one/references/opengame-performance/helpers/landing-pacer.mjs +86 -0
  105. package/dist/skills/opengame-all-in-one/references/opengame-performance/references/measuring.md +123 -0
  106. package/dist/skills/opengame-all-in-one/references/opengame-performance/references/webgpu.md +86 -0
  107. package/dist/skills/opengame-all-in-one/references/opengame-production/README.md +192 -0
  108. package/dist/skills/opengame-all-in-one/references/opengame-production/references/team.md +107 -0
  109. package/dist/skills/opengame-all-in-one/references/opengame-production/references/templates.md +153 -0
  110. package/dist/skills/opengame-all-in-one/references/opengame-quality-bar/README.md +170 -0
  111. package/dist/skills/opengame-all-in-one/references/opengame-quality-bar/references/cheap-tells.md +43 -0
  112. package/dist/skills/opengame-all-in-one/references/opengame-quality-bar/references/judges.md +95 -0
  113. package/dist/skills/opengame-all-in-one/references/opengame-quality-bar/references/review-prompt.md +32 -0
  114. package/dist/skills/opengame-all-in-one/references/opengame-rhythm/README.md +134 -0
  115. package/dist/skills/opengame-all-in-one/references/opengame-rhythm/references/charting.md +109 -0
  116. package/dist/skills/opengame-all-in-one/references/opengame-rhythm/references/timing.md +92 -0
  117. package/dist/skills/opengame-all-in-one/references/opengame-router/README.md +54 -0
  118. package/dist/skills/opengame-all-in-one/references/opengame-router/references/process.md +38 -0
  119. package/dist/skills/opengame-all-in-one/references/opengame-scene-audit/README.md +136 -0
  120. package/dist/skills/opengame-all-in-one/references/opengame-scene-audit/helpers/scene-audit.js +677 -0
  121. package/dist/skills/opengame-all-in-one/references/opengame-sdk/README.md +283 -0
  122. package/dist/skills/opengame-all-in-one/references/opengame-sdk/references/engines.md +93 -0
  123. package/dist/skills/opengame-all-in-one/references/opengame-sdk/references/frame.md +107 -0
  124. package/dist/skills/opengame-all-in-one/references/opengame-sdk/references/identity.md +111 -0
  125. package/dist/skills/opengame-all-in-one/references/opengame-sdk/references/leaderboards.md +115 -0
  126. package/dist/skills/opengame-all-in-one/references/opengame-sdk/references/radio.md +71 -0
  127. package/dist/skills/opengame-all-in-one/references/opengame-stability/README.md +112 -0
  128. package/dist/skills/opengame-all-in-one/references/opengame-trailer-make/README.md +160 -0
  129. package/dist/skills/opengame-all-in-one/references/opengame-trailer-make/references/archetypes.md +113 -0
  130. package/dist/skills/opengame-all-in-one/references/opengame-trailer-make/references/briefs.md +273 -0
  131. package/dist/skills/opengame-all-in-one/references/opengame-trailer-make/references/core.md +129 -0
  132. package/dist/skills/opengame-all-in-one/references/opengame-trailer-make/references/craft.md +299 -0
  133. package/dist/skills/opengame-all-in-one/references/opengame-trailer-make/references/edge-cases.md +27 -0
  134. package/dist/skills/opengame-all-in-one/references/opengame-trailer-make/references/filmable.md +244 -0
  135. package/dist/skills/opengame-all-in-one/references/opengame-trailer-make/references/finishing.md +126 -0
  136. package/dist/skills/opengame-all-in-one/references/opengame-trailer-make/references/gates.md +184 -0
  137. package/dist/skills/opengame-all-in-one/references/opengame-trailer-make/references/measure.md +282 -0
  138. package/dist/skills/opengame-all-in-one/references/opengame-trailer-make/references/music.md +148 -0
  139. package/dist/skills/opengame-all-in-one/references/opengame-trailer-make/references/og-clash-record.md +133 -0
  140. package/dist/skills/opengame-all-in-one/references/opengame-trailer-make/references/ops.md +154 -0
  141. package/dist/skills/opengame-all-in-one/references/opengame-trailer-make/references/owner-eye.md +214 -0
  142. package/dist/skills/opengame-all-in-one/references/opengame-trailer-make/references/review.md +141 -0
  143. package/dist/skills/opengame-all-in-one/references/opengame-trailer-make/references/title.md +92 -0
  144. package/dist/skills/opengame-all-in-one/references/opengame-trailer-plan/README.md +172 -0
  145. package/dist/skills/opengame-all-in-one/references/opengame-trailer-plan/references/archetypes.md +113 -0
  146. package/dist/skills/opengame-all-in-one/references/opengame-trailer-plan/references/briefs.md +273 -0
  147. package/dist/skills/opengame-all-in-one/references/opengame-trailer-plan/references/core.md +129 -0
  148. package/dist/skills/opengame-all-in-one/references/opengame-trailer-plan/references/craft.md +299 -0
  149. package/dist/skills/opengame-all-in-one/references/opengame-trailer-plan/references/edge-cases.md +27 -0
  150. package/dist/skills/opengame-all-in-one/references/opengame-trailer-plan/references/filmable.md +244 -0
  151. package/dist/skills/opengame-all-in-one/references/opengame-trailer-plan/references/finishing.md +126 -0
  152. package/dist/skills/opengame-all-in-one/references/opengame-trailer-plan/references/gates.md +184 -0
  153. package/dist/skills/opengame-all-in-one/references/opengame-trailer-plan/references/measure.md +282 -0
  154. package/dist/skills/opengame-all-in-one/references/opengame-trailer-plan/references/music.md +148 -0
  155. package/dist/skills/opengame-all-in-one/references/opengame-trailer-plan/references/og-clash-record.md +133 -0
  156. package/dist/skills/opengame-all-in-one/references/opengame-trailer-plan/references/ops.md +154 -0
  157. package/dist/skills/opengame-all-in-one/references/opengame-trailer-plan/references/owner-eye.md +214 -0
  158. package/dist/skills/opengame-all-in-one/references/opengame-trailer-plan/references/reference-set.md +92 -0
  159. package/dist/skills/opengame-all-in-one/references/opengame-trailer-plan/references/review.md +141 -0
  160. package/dist/skills/opengame-all-in-one/references/opengame-trailer-plan/references/title.md +92 -0
  161. package/dist/skills/opengame-all-in-one/references/opengame-trailer-review-and-improve/README.md +141 -0
  162. package/dist/skills/opengame-all-in-one/references/opengame-trailer-review-and-improve/references/archetypes.md +113 -0
  163. package/dist/skills/opengame-all-in-one/references/opengame-trailer-review-and-improve/references/briefs.md +273 -0
  164. package/dist/skills/opengame-all-in-one/references/opengame-trailer-review-and-improve/references/core.md +129 -0
  165. package/dist/skills/opengame-all-in-one/references/opengame-trailer-review-and-improve/references/craft.md +299 -0
  166. package/dist/skills/opengame-all-in-one/references/opengame-trailer-review-and-improve/references/edge-cases.md +27 -0
  167. package/dist/skills/opengame-all-in-one/references/opengame-trailer-review-and-improve/references/filmable.md +244 -0
  168. package/dist/skills/opengame-all-in-one/references/opengame-trailer-review-and-improve/references/finishing.md +126 -0
  169. package/dist/skills/opengame-all-in-one/references/opengame-trailer-review-and-improve/references/gates.md +184 -0
  170. package/dist/skills/opengame-all-in-one/references/opengame-trailer-review-and-improve/references/measure.md +282 -0
  171. package/dist/skills/opengame-all-in-one/references/opengame-trailer-review-and-improve/references/music.md +148 -0
  172. package/dist/skills/opengame-all-in-one/references/opengame-trailer-review-and-improve/references/og-clash-record.md +133 -0
  173. package/dist/skills/opengame-all-in-one/references/opengame-trailer-review-and-improve/references/ops.md +154 -0
  174. package/dist/skills/opengame-all-in-one/references/opengame-trailer-review-and-improve/references/owner-eye.md +214 -0
  175. package/dist/skills/opengame-all-in-one/references/opengame-trailer-review-and-improve/references/review.md +141 -0
  176. package/dist/skills/opengame-all-in-one/references/opengame-trailer-review-and-improve/references/title.md +92 -0
  177. package/dist/skills/opengame-all-in-one/references/opengame-verification/README.md +197 -0
  178. package/dist/skills/opengame-all-in-one/references/opengame-verification/helpers/browser-evidence.mjs +74 -0
  179. package/dist/skills/opengame-all-in-one/references/opengame-weapons/README.md +179 -0
  180. package/dist/skills/opengame-all-in-one/references/opengame-weapons/references/handling-numbers.md +114 -0
  181. package/dist/skills/opengame-all-in-one/references/opengame-weapons/references/reloads-and-hands.md +96 -0
  182. package/dist/skills/opengame-all-in-one/references/opengame-weapons/references/sights-and-scopes.md +98 -0
  183. package/dist/skills/opengame-all-in-one/references/opengame-weapons/references/threejs.md +240 -0
  184. package/dist/skills/opengame-all-in-one/references/opengame-world-physics/README.md +141 -0
  185. package/dist/skills/opengame-all-in-one/references/opengame-world-physics/references/measured-failures.md +79 -0
  186. package/dist/skills/opengame-api/SKILL.md +32 -6
  187. package/dist/skills/opengame-assets/SKILL.md +79 -31
  188. package/dist/skills/opengame-assets/references/pipeline.md +67 -0
  189. package/dist/skills/opengame-audio/SKILL.md +129 -41
  190. package/dist/skills/opengame-audio/references/mix-notes.md +44 -1
  191. package/dist/skills/opengame-audio/references/qa-without-ears.md +177 -0
  192. package/dist/skills/opengame-capture/SKILL.md +44 -20
  193. package/dist/skills/opengame-capture/references/pipeline.md +33 -1
  194. package/dist/skills/opengame-capture/references/threejs.md +17 -1
  195. package/dist/skills/opengame-characters/SKILL.md +94 -39
  196. package/dist/skills/opengame-characters/references/first-person-body.md +108 -0
  197. package/dist/skills/opengame-characters/references/player-og.md +1 -1
  198. package/dist/skills/opengame-cli/SKILL.md +4 -4
  199. package/dist/skills/opengame-deploy/SKILL.md +35 -15
  200. package/dist/skills/opengame-deploy/references/json-events.md +6 -5
  201. package/dist/skills/opengame-deploy/references/packaging.md +28 -26
  202. package/dist/skills/opengame-depth/SKILL.md +61 -33
  203. package/dist/skills/opengame-enemies/SKILL.md +96 -40
  204. package/dist/skills/opengame-enemies/references/pacing-and-bots.md +108 -0
  205. package/dist/skills/opengame-enemies/references/threejs.md +59 -0
  206. package/dist/skills/opengame-front-end/SKILL.md +85 -39
  207. package/dist/skills/opengame-front-end/references/menus.md +42 -8
  208. package/dist/skills/opengame-game-feel/SKILL.md +95 -47
  209. package/dist/skills/opengame-game-feel/references/feel-numbers.md +59 -0
  210. package/dist/skills/opengame-game-feel/references/killcam.md +88 -0
  211. package/dist/skills/opengame-jam/SKILL.md +42 -27
  212. package/dist/skills/opengame-lighting/SKILL.md +85 -34
  213. package/dist/skills/opengame-loading/SKILL.md +78 -38
  214. package/dist/skills/opengame-loading/helpers/loading-count.mjs +115 -0
  215. package/dist/skills/opengame-loading/references/threejs.md +32 -7
  216. package/dist/skills/opengame-look/SKILL.md +122 -51
  217. package/dist/skills/opengame-look/references/matching-a-reference.md +73 -0
  218. package/dist/skills/opengame-look/references/painterly.md +69 -0
  219. package/dist/skills/opengame-look/references/post-and-lens.md +10 -3
  220. package/dist/skills/opengame-maps/SKILL.md +78 -0
  221. package/dist/skills/opengame-maps/references/map-spec.md +77 -0
  222. package/dist/skills/opengame-mobile/SKILL.md +18 -18
  223. package/dist/skills/opengame-montage-make/SKILL.md +187 -0
  224. package/dist/skills/opengame-montage-make/references/briefs.md +293 -0
  225. package/dist/skills/opengame-montage-make/references/day-one-questions.md +46 -0
  226. package/dist/skills/opengame-montage-make/references/edge-cases.md +27 -0
  227. package/dist/skills/opengame-montage-make/references/filmable.md +244 -0
  228. package/dist/skills/opengame-montage-make/references/finishing.md +126 -0
  229. package/dist/skills/opengame-montage-make/references/measure.md +282 -0
  230. package/dist/skills/opengame-montage-make/references/montage-capture.md +151 -0
  231. package/dist/skills/opengame-montage-make/references/montage-gates.md +84 -0
  232. package/dist/skills/opengame-montage-make/references/montage-genres.md +198 -0
  233. package/dist/skills/opengame-montage-make/references/montage-grammar.md +320 -0
  234. package/dist/skills/opengame-montage-make/references/montage-ops.md +127 -0
  235. package/dist/skills/opengame-montage-make/references/og-snipers-record.md +110 -0
  236. package/dist/skills/opengame-montage-make/references/owner-eye.md +214 -0
  237. package/dist/skills/opengame-montage-make/references/reference-study.md +293 -0
  238. package/dist/skills/opengame-montage-make/references/song.md +64 -0
  239. package/dist/skills/opengame-montage-plan/SKILL.md +197 -0
  240. package/dist/skills/opengame-montage-plan/references/briefs.md +293 -0
  241. package/dist/skills/opengame-montage-plan/references/day-one-questions.md +46 -0
  242. package/dist/skills/opengame-montage-plan/references/filmable.md +244 -0
  243. package/dist/skills/opengame-montage-plan/references/measure.md +282 -0
  244. package/dist/skills/opengame-montage-plan/references/montage-capture.md +151 -0
  245. package/dist/skills/opengame-montage-plan/references/montage-gates.md +84 -0
  246. package/dist/skills/opengame-montage-plan/references/montage-genres.md +198 -0
  247. package/dist/skills/opengame-montage-plan/references/montage-grammar.md +320 -0
  248. package/dist/skills/opengame-montage-plan/references/montage-ops.md +127 -0
  249. package/dist/skills/opengame-montage-plan/references/og-snipers-record.md +110 -0
  250. package/dist/skills/opengame-montage-plan/references/owner-eye.md +214 -0
  251. package/dist/skills/opengame-montage-plan/references/reference-study.md +293 -0
  252. package/dist/skills/opengame-montage-plan/references/song.md +64 -0
  253. package/dist/skills/opengame-montage-review-and-improve/SKILL.md +178 -0
  254. package/dist/skills/opengame-montage-review-and-improve/references/briefs.md +293 -0
  255. package/dist/skills/opengame-montage-review-and-improve/references/day-one-questions.md +46 -0
  256. package/dist/skills/opengame-montage-review-and-improve/references/filmable.md +244 -0
  257. package/dist/skills/opengame-montage-review-and-improve/references/finishing.md +126 -0
  258. package/dist/skills/opengame-montage-review-and-improve/references/measure.md +282 -0
  259. package/dist/skills/opengame-montage-review-and-improve/references/montage-capture.md +151 -0
  260. package/dist/skills/opengame-montage-review-and-improve/references/montage-gates.md +84 -0
  261. package/dist/skills/opengame-montage-review-and-improve/references/montage-genres.md +198 -0
  262. package/dist/skills/opengame-montage-review-and-improve/references/montage-grammar.md +320 -0
  263. package/dist/skills/opengame-montage-review-and-improve/references/montage-ops.md +127 -0
  264. package/dist/skills/opengame-montage-review-and-improve/references/og-snipers-record.md +110 -0
  265. package/dist/skills/opengame-montage-review-and-improve/references/owner-eye.md +214 -0
  266. package/dist/skills/opengame-montage-review-and-improve/references/reference-study.md +293 -0
  267. package/dist/skills/opengame-montage-review-and-improve/references/song.md +64 -0
  268. package/dist/skills/opengame-multiplayer/SKILL.md +143 -0
  269. package/dist/skills/opengame-multiplayer/references/bot-lobbies.md +71 -0
  270. package/dist/skills/opengame-multiplayer/references/netcode.md +165 -0
  271. package/dist/skills/opengame-performance/SKILL.md +106 -65
  272. package/dist/skills/opengame-performance/helpers/landing-pacer.mjs +86 -0
  273. package/dist/skills/opengame-performance/references/measuring.md +51 -0
  274. package/dist/skills/opengame-performance/references/webgpu.md +86 -0
  275. package/dist/skills/opengame-production/SKILL.md +119 -54
  276. package/dist/skills/opengame-production/references/team.md +107 -0
  277. package/dist/skills/opengame-production/references/templates.md +106 -16
  278. package/dist/skills/opengame-quality-bar/SKILL.md +94 -42
  279. package/dist/skills/opengame-quality-bar/references/cheap-tells.md +14 -7
  280. package/dist/skills/opengame-quality-bar/references/judges.md +95 -0
  281. package/dist/skills/opengame-quality-bar/references/review-prompt.md +10 -5
  282. package/dist/skills/opengame-rhythm/SKILL.md +138 -0
  283. package/dist/skills/opengame-rhythm/references/charting.md +109 -0
  284. package/dist/skills/opengame-rhythm/references/timing.md +92 -0
  285. package/dist/skills/opengame-router/SKILL.md +10 -2
  286. package/dist/skills/opengame-scene-audit/SKILL.md +64 -47
  287. package/dist/skills/opengame-scene-audit/helpers/scene-audit.js +97 -19
  288. package/dist/skills/opengame-sdk/SKILL.md +26 -9
  289. package/dist/skills/opengame-sdk/references/engines.md +2 -2
  290. package/dist/skills/opengame-sdk/references/identity.md +10 -11
  291. package/dist/skills/opengame-stability/SKILL.md +59 -27
  292. package/dist/skills/opengame-trailer-make/SKILL.md +164 -0
  293. package/dist/skills/opengame-trailer-make/references/archetypes.md +113 -0
  294. package/dist/skills/opengame-trailer-make/references/briefs.md +273 -0
  295. package/dist/skills/opengame-trailer-make/references/core.md +129 -0
  296. package/dist/skills/opengame-trailer-make/references/craft.md +299 -0
  297. package/dist/skills/opengame-trailer-make/references/edge-cases.md +27 -0
  298. package/dist/skills/opengame-trailer-make/references/filmable.md +244 -0
  299. package/dist/skills/opengame-trailer-make/references/finishing.md +126 -0
  300. package/dist/skills/opengame-trailer-make/references/gates.md +184 -0
  301. package/dist/skills/opengame-trailer-make/references/measure.md +282 -0
  302. package/dist/skills/opengame-trailer-make/references/music.md +148 -0
  303. package/dist/skills/opengame-trailer-make/references/og-clash-record.md +133 -0
  304. package/dist/skills/opengame-trailer-make/references/ops.md +154 -0
  305. package/dist/skills/opengame-trailer-make/references/owner-eye.md +214 -0
  306. package/dist/skills/opengame-trailer-make/references/review.md +141 -0
  307. package/dist/skills/opengame-trailer-make/references/title.md +92 -0
  308. package/dist/skills/opengame-trailer-plan/SKILL.md +176 -0
  309. package/dist/skills/opengame-trailer-plan/references/archetypes.md +113 -0
  310. package/dist/skills/opengame-trailer-plan/references/briefs.md +273 -0
  311. package/dist/skills/opengame-trailer-plan/references/core.md +129 -0
  312. package/dist/skills/opengame-trailer-plan/references/craft.md +299 -0
  313. package/dist/skills/opengame-trailer-plan/references/edge-cases.md +27 -0
  314. package/dist/skills/opengame-trailer-plan/references/filmable.md +244 -0
  315. package/dist/skills/opengame-trailer-plan/references/finishing.md +126 -0
  316. package/dist/skills/opengame-trailer-plan/references/gates.md +184 -0
  317. package/dist/skills/opengame-trailer-plan/references/measure.md +282 -0
  318. package/dist/skills/opengame-trailer-plan/references/music.md +148 -0
  319. package/dist/skills/opengame-trailer-plan/references/og-clash-record.md +133 -0
  320. package/dist/skills/opengame-trailer-plan/references/ops.md +154 -0
  321. package/dist/skills/opengame-trailer-plan/references/owner-eye.md +214 -0
  322. package/dist/skills/opengame-trailer-plan/references/reference-set.md +92 -0
  323. package/dist/skills/opengame-trailer-plan/references/review.md +141 -0
  324. package/dist/skills/opengame-trailer-plan/references/title.md +92 -0
  325. package/dist/skills/opengame-trailer-review-and-improve/SKILL.md +145 -0
  326. package/dist/skills/opengame-trailer-review-and-improve/references/archetypes.md +113 -0
  327. package/dist/skills/opengame-trailer-review-and-improve/references/briefs.md +273 -0
  328. package/dist/skills/opengame-trailer-review-and-improve/references/core.md +129 -0
  329. package/dist/skills/opengame-trailer-review-and-improve/references/craft.md +299 -0
  330. package/dist/skills/opengame-trailer-review-and-improve/references/edge-cases.md +27 -0
  331. package/dist/skills/opengame-trailer-review-and-improve/references/filmable.md +244 -0
  332. package/dist/skills/opengame-trailer-review-and-improve/references/finishing.md +126 -0
  333. package/dist/skills/opengame-trailer-review-and-improve/references/gates.md +184 -0
  334. package/dist/skills/opengame-trailer-review-and-improve/references/measure.md +282 -0
  335. package/dist/skills/opengame-trailer-review-and-improve/references/music.md +148 -0
  336. package/dist/skills/opengame-trailer-review-and-improve/references/og-clash-record.md +133 -0
  337. package/dist/skills/opengame-trailer-review-and-improve/references/ops.md +154 -0
  338. package/dist/skills/opengame-trailer-review-and-improve/references/owner-eye.md +214 -0
  339. package/dist/skills/opengame-trailer-review-and-improve/references/review.md +141 -0
  340. package/dist/skills/opengame-trailer-review-and-improve/references/title.md +92 -0
  341. package/dist/skills/opengame-verification/SKILL.md +113 -48
  342. package/dist/skills/opengame-weapons/SKILL.md +129 -55
  343. package/dist/skills/opengame-weapons/references/handling-numbers.md +42 -0
  344. package/dist/skills/opengame-weapons/references/reloads-and-hands.md +96 -0
  345. package/dist/skills/opengame-weapons/references/sights-and-scopes.md +98 -0
  346. package/dist/skills/opengame-weapons/references/threejs.md +71 -0
  347. package/dist/skills/opengame-world-physics/SKILL.md +55 -24
  348. package/dist/skills/opengame-world-physics/references/measured-failures.md +28 -1
  349. package/package.json +1 -1
@@ -0,0 +1,177 @@
1
+ # Checking sound without ears
2
+
3
+ An agent cannot hear. It can record what the game actually sends to the speakers, line
4
+ that recording up with what the game did, and test stated relationships between sounds.
5
+ LC, OS and OH each built some of the tools below; together they found most of the sound
6
+ faults those games had before a person listened. None of them replaces the person, and
7
+ every report that uses them still opens with what nobody has heard.
8
+
9
+ ## The master-bus tap
10
+
11
+ Record the output, not the files. An AudioWorklet in parallel with the destination sees
12
+ exactly what the player would hear after every bus, duck and limiter.
13
+
14
+ ```js
15
+ // tap-worklet.js, loaded with audioContext.audioWorklet.addModule()
16
+ class MasterTap extends AudioWorkletProcessor {
17
+ constructor() { super(); this.blocks = []; this.frames = 0; this.start = null; }
18
+ process(inputs) {
19
+ const input = inputs[0];
20
+ // With nothing playing, an input can arrive with no channels. Record that quantum as
21
+ // silence, so a batch stays continuous and every later event still lines up.
22
+ const block = input && input.length
23
+ ? input.map((channel) => channel.slice())
24
+ : [new Float32Array(128), new Float32Array(128)];
25
+ if (this.start === null) this.start = currentTime; // the context time of this batch
26
+ this.blocks.push(block);
27
+ this.frames += block[0].length;
28
+ if (this.frames >= sampleRate / 4) { // post about four times a second
29
+ this.port.postMessage({ start: this.start, blocks: this.blocks });
30
+ this.blocks = []; this.frames = 0; this.start = null;
31
+ }
32
+ return true;
33
+ }
34
+ }
35
+ registerProcessor('master-tap', MasterTap);
36
+ ```
37
+
38
+ ```js
39
+ // in the game, behind a test flag that changes nothing else
40
+ await ctx.audioWorklet.addModule('tap-worklet.js');
41
+ const tap = new AudioWorkletNode(ctx, 'master-tap', {
42
+ channelCount: 2, channelCountMode: 'explicit', outputChannelCount: [2],
43
+ });
44
+ const silent = new GainNode(ctx, { gain: 0 });
45
+ master.connect(tap).connect(silent).connect(ctx.destination); // keeps the tap running, adds nothing
46
+ const batches = [];
47
+ tap.port.onmessage = (e) => batches.push(e.data);
48
+
49
+ const events = [];
50
+ // Call it with the same `when` you pass to `source.start(when)`, so the log holds the
51
+ // time the sound was placed on the audio clock, not the time the code ran.
52
+ export function logEvent(kind, when, detail = {}) {
53
+ events.push({ t: when, kind, ...detail });
54
+ }
55
+ ```
56
+
57
+ Place each batch at its `start` time; the worklet writes silent quanta itself, so batches
58
+ follow one another without gaps. Convert the samples to plain arrays or
59
+ base64 before returning them from the page, then write a WAV in Node:
60
+
61
+ ```js
62
+ import { writeFileSync } from 'node:fs';
63
+ // 32-bit float WAV from interleaved samples
64
+ export function writeWav(path, interleaved, channels, rate) {
65
+ const data = Buffer.from(interleaved.buffer, interleaved.byteOffset, interleaved.byteLength);
66
+ const h = Buffer.alloc(44);
67
+ h.write('RIFF', 0); h.writeUInt32LE(36 + data.length, 4); h.write('WAVE', 8);
68
+ h.write('fmt ', 12); h.writeUInt32LE(16, 16); h.writeUInt16LE(3, 20); // 3: IEEE float
69
+ h.writeUInt16LE(channels, 22); h.writeUInt32LE(rate, 24);
70
+ h.writeUInt32LE(rate * channels * 4, 28); h.writeUInt16LE(channels * 4, 32); h.writeUInt16LE(32, 34);
71
+ h.write('data', 36); h.writeUInt32LE(data.length, 40);
72
+ writeFileSync(path, Buffer.concat([h, data]));
73
+ }
74
+ ```
75
+
76
+ The event log and the samples share the audio clock, so every shot, step and line can be
77
+ found in the recording to well under a frame. OS's tap lined up with its engine log to
78
+ about 0.2 s, close enough to expose a 1.7 to 3.6 s skew that had been producing wrong
79
+ numbers for days.
80
+
81
+ ## The owner-path capture
82
+
83
+ Every repeated sound complaint LC had lived on a path nobody tested. Capture the path the
84
+ player takes:
85
+
86
+ - The built game, served the way it ships, with no autostart flag and no debug option
87
+ that changes behaviour.
88
+ - A real key press through the browser's input pipeline, not a call into the game.
89
+ - Two runs: the browser's default autoplay policy and the strict one
90
+ (`--autoplay-policy=user-gesture-required` in Chromium). Read the flags your test
91
+ launcher adds; many allow autoplay, which hides exactly this fault.
92
+ - The game's own count of refused `play()` calls and suspended-context recoveries, kept in
93
+ the shipped build too, read at the end. A wrapper the test injects into `play()` is a
94
+ tool changing the game.
95
+ - One tap per bus as well as the master, so a muted bed and a quiet bed read differently.
96
+
97
+ LC called this capture "the single best check" for sound that never starts: its rain was
98
+ silent under strict autoplay while every earlier capture, run with autostart, had it.
99
+
100
+ ## The offline mix render
101
+
102
+ For fast level work, build the game's real audio graph (the same factory functions, the
103
+ same buses) inside an `OfflineAudioContext`, schedule a scripted scene (a fight, a quiet
104
+ walk, a reload beside a crowd) and render it in well under real time. Measure integrated,
105
+ short-term and momentary loudness, true peak oversampled, band levels and duck depth. Wait
106
+ for streamed sources to load before rendering, and take the median of three renders for
107
+ anything streamed. LC's version fails the build if rain at rest reads under its floor in
108
+ its band.
109
+
110
+ ## Bands a laptop can play
111
+
112
+ Laptop and phone speakers give back little of the lowest octaves, so a cue that is all
113
+ low end can measure loud and be silent on the player's machine. Measure each bed and cue
114
+ in the band that carries it (LC: rain at rest in 200 Hz to 6 kHz at or over -33 dBFS;
115
+ enemy steps and voices in 100 Hz to 1 kHz against the rain). Give a low cue harmonics: a
116
+ parallel softly clipped path through the low mids made LC's heartbeat audible where it had
117
+ read -59 dBFS in 300 Hz to 4 kHz.
118
+
119
+ ## A gate table of relationships
120
+
121
+ Absolute levels drift with every change; relationships are what the player hears. LC's
122
+ table, read from a three-minute real run after every bank change:
123
+
124
+ | Relationship | LC bar |
125
+ | --- | --- |
126
+ | The player's shot over an enemy voice at arm's length | at least 4 LU |
127
+ | Footsteps over rain, most steps | at least 6 dB |
128
+ | A surge of enemies | heard before it is seen |
129
+ | A fight, integrated | -16 LUFS ±1.5 |
130
+ | A quiet night, integrated | -28 LUFS ±3 |
131
+ | Output true peak | under -2 dBTP |
132
+ | Mouth sounds in three minutes without a fight | zero |
133
+
134
+ OS's equivalent was a hierarchy: the player's shot 4.8 LU over anything in a fight, rivals
135
+ at least 6 dB under, reloads never within 6 LU, the announcer 2 to 3 LU under.
136
+
137
+ ## Spoken lines
138
+
139
+ Run every line through the game's final chain (radio, helmet, distance) into a local
140
+ speech recogniser, and compare it with the script. Recognisers lean towards dictionary
141
+ words, so add a phoneme comparison for names and invented words, and flag a line when the
142
+ two disagree with the text. OS's recogniser passed two invented place names that a
143
+ phoneme check caught, and could not tell /laɪv/ from /lɪv/, which the owner heard at once. LC reworded lines the recogniser misheard through the radio ("take an arm from
144
+ Iraq" 11 times in 12).
145
+
146
+ ## Spectrograms read as images
147
+
148
+ Render the shipped file first and each candidate after it on the same scale and range.
149
+ Look for horizontal lines (fixed tones and whines), vertical lines (clicks), tails cut
150
+ square, energy before the attack (codec pre-echo) and a top band that never moves. OS
151
+ found that its metrics alone preferred synthetic beeps to real takes; the picture with the
152
+ shipped file beside it did not.
153
+
154
+ ## Ledgers for cut-offs, delay and position
155
+
156
+ - **Stop ledger.** Every voice that ended before its file, with the reason: stolen, ended
157
+ by a rule, a queue overflow, never started. OS's showed 1,110 cut-offs in a ten-minute
158
+ fight that no level meter could see.
159
+ - **Delay probe.** The press, stamped from the input event and mapped to the audio clock
160
+ with `getOutputTimestamp()`, against the first sample over a threshold in the tap. OS's
161
+ found a 690 ms backlog; after the fix, press-to-output read 67 ms.
162
+ - **Position log.** Every world sound with its true distance and bearing and its level
163
+ against the ambience in its band. LC's found steps 32 times in 34 and a 2 degree bearing
164
+ error; OS's found rival reloads playing centred at full level.
165
+
166
+ ## A/B tapes for the person
167
+
168
+ Name each clip `<date>-<what>-{NEW|AS-IT-WAS|NOT-TAKEN}-<what to listen for>.mp3` and
169
+ render each pair from one boot, so the only difference is the change. Send two or three
170
+ pairs with one plain question each. Record the answer against the cue.
171
+
172
+ ## What none of this proves
173
+
174
+ Whether a sound is pleasant, whether it fits the world, whether a voice suits its
175
+ character, whether the game sounds finished. Those are a person's calls, made by
176
+ listening. Report the measurements as measurements, and keep the list of what nobody has
177
+ heard at the top.
@@ -1,8 +1,8 @@
1
1
  ---
2
2
  name: opengame-capture
3
- description: Make a browser game filmable and screenshot-testable. Covers a host-stepped clock, one time scale, seeded takes, a director camera whose cuts reset history, a clean feed with HUD masks and mute flags, an awaitable ready state, acknowledged capture requests, a state probe, stepped audio, 4K60 readback from the render surface and hooks kept out of public builds. Use when the person wants a trailer, gameplay footage, slow motion, a photo mode or screenshot tests, or says takes do not repeat.
3
+ description: Make a browser game filmable and screenshot-testable. Covers a host-stepped clock, one time scale, seeded takes, a director camera, a clean feed, a ready state, acknowledged requests, captures that prove their own build and state, motion filmed on the game clock at quarter speed, stepped audio, WebGPU readback, batched jobs and hooks absent from public builds. Use when the person wants a trailer, footage, slow motion or screenshot tests, or takes do not repeat or captures come out black.
4
4
  metadata:
5
- version: "1.0.0"
5
+ version: "1.1.0"
6
6
  ---
7
7
 
8
8
  # Capture: a game you can film and test
@@ -15,7 +15,10 @@ top. The owner's brief: "i do NOT want to have to keep watching shitty takes and
15
15
  feedbaxk thqt u should be able to know ill give u based on those skills".
16
16
 
17
17
  Build these into the game before anyone films it. The same hooks make screenshot tests
18
- repeatable, which is most of what `opengame-verification` asks of a capture.
18
+ repeatable, which is most of what `opengame-verification` asks of a capture. OS, LC and OH
19
+ added the testing half: judges scored captures that were not what their labels said,
20
+ WebGPU frames came back black, real-time films drew a handful of frames a second on a
21
+ loaded machine, and screencasts were read as real time.
19
22
 
20
23
  Rules carry five tags; `opengame-quality-bar` explains them and the game codes. The
21
24
  three.js way to do each is in [references/threejs.md](references/threejs.md); pipeline
@@ -25,17 +28,17 @@ numbers, the capability audit and process are in [references/pipeline.md](refere
25
28
 
26
29
  ### The host steps one clock
27
30
  - **Invariant:** In capture mode the host advances the game by a fixed simulation step per output frame, every system reads that one clock, and no game code reads wall time.
28
- - **Technique:** A capture mode where the host calls step, then render, then read back, and the game's own frame loop is off. Simulation, physics, animation, particles, weather, camera smoothing and audio all take their time from the step, and worker threads read the same clock. Slow motion is a smaller step, never frame interpolation. Gameplay never runs faster than real time; speed on screen comes from the camera.
29
- - **Reference configuration:** OC's time came from `performance.now()` in a worker, and page-level tricks failed: a fake page clock froze timers and virtual time never reached the workers. A shim replaced `performance.now` and `requestAnimationFrame` in the worker and retried a frame when the GPU slot was busy; its first version took 48-64 min and the full rig about 2 h. Real-time 4K ran at 25-31 fps, so 1 s of 4K60 cost about 5.2 s of wall time. Stepped, the rig ran 28.1 fps first person and 37.2 fps on a dolly at 4K on an idle M4 Max, 2.1-3.5 s of wall time per output second. A step of 1/60 is 1x and 1/240 is 0.25x; 240 frames covered 1.000 s of simulation with 0 skipped. The montage reused the rig unchanged on a new build in about an hour.
31
+ - **Technique:** A capture mode where the host calls step, then render, then read back, and the game's own frame loop is off. Simulation, physics, animation, particles, weather, camera smoothing and audio all take their time from the step, and worker threads read the same clock. Drive every scripted state by update count, never wall time. Slow motion is a smaller step, never frame interpolation. Gameplay never runs faster than real time; speed on screen comes from the camera.
32
+ - **Reference configuration:** OC's time came from `performance.now()` in a worker, and page-level tricks failed: a fake page clock froze timers and virtual time never reached the workers. A shim replaced `performance.now` and `requestAnimationFrame` in the worker and retried a frame when the GPU slot was busy; its first version took 48-64 min and the full rig about 2 h. Real-time 4K ran at 25-31 fps, so 1 s of 4K60 cost about 5.2 s of wall time. Stepped, the rig ran 28.1 fps first person and 37.2 fps on a dolly at 4K on an idle M4 Max, 2.1-3.5 s of wall time per output second. A step of 1/60 is 1x and 1/240 is 0.25x; 240 frames covered 1.000 s of simulation with 0 skipped. The montage reused the rig unchanged on a new build in about an hour. LC stepped 1/60 per drawn frame after real-time films on its loaded machine drew 8-25 fps ("nine pictures a second at best"), and recipes timed on the wall clock while the game clock was held made holds 3-4x short; the stepped burst was "the only film true under tonight's load".
30
33
  - **Applies when:** Every game that will be filmed, screenshot-tested or compared frame by frame.
31
34
  - **Verify:** Capture 240 frames at a 1/240 step: the probe's simulation time rises by exactly one step per frame and ends at 1 s. Repeat on a loaded machine and get the same frames.
32
35
 
33
36
  ### One time scale, and every timer obeys it
34
37
  - **Invariant:** Every timer in the game (hitstop, shake, camera lerps, particles, physics, corpse settling, emitters, flashes, audio) runs on scaled simulation time, so an effect at quarter speed looks like the same effect slowed down.
35
38
  - **Technique:** One time scale applied where the clock is read, never per system. Replace frame-count timers with time. Physics sub-steps or interpolates so a small step does not hold a pose and then jump. Motion blur takes its shutter from the step. A cinematic flag turns off hitstop and kill-cam time dilation. Review every effect at quarter speed before filming.
36
- - **Reference configuration:** OC: a 45 ms kill hitstop became about 9 near-frozen frames at 0.3x, and a 70 ms hammer hold about 12 frames at 0.25x. Corpses solved at a fixed 60 Hz, so at 0.25x a pose held for 4 frames, then jumped. A gib trail emitted every third frame, and per-frame emitters ran 4 times denser. A game-only time scale left particles on presentation time at full speed. Motion blur reached a 360-degree shutter at 1/240 until its strength was set to 60 times the step in seconds (0.25 at 1/240). At 0.25x a light re-firing every 0.09 s strobed at 10 Hz, a blood sprite covered about 40% of the frame, and a bloom clipped to white. The frozen-flip defect was predicted in an audit, still reached the owner about 28 h later, and cost about 3 h and reshoots for a 6-line fix.
39
+ - **Reference configuration:** OC: a 45 ms kill hitstop became about 9 near-frozen frames at 0.3x, and a 70 ms hammer hold about 12 frames at 0.25x. Corpses solved at a fixed 60 Hz, so at 0.25x a pose held for 4 frames, then jumped. A gib trail emitted every third frame, and per-frame emitters ran 4 times denser. A game-only time scale left particles on presentation time at full speed. Motion blur reached a 360-degree shutter at 1/240 until its strength was set to 60 times the step in seconds (0.25 at 1/240). At 0.25x a light re-firing every 0.09 s strobed at 10 Hz, a blood sprite covered about 40% of the frame, and a bloom clipped to white. The frozen-flip defect was predicted in an audit, still reached the owner about 28 h later, and cost about 3 h and reshoots for a 6-line fix. OS's dev time scale never reached the game, so every "slow-motion" animation capture ran at real speed until someone checked.
37
40
  - **Applies when:** Games with slow motion, hitstop, or any effect timed by frames.
38
- - **Verify:** Render each effect at 1x and at 0.25x and lay the frames side by side: the slow version has four times the frames of the same motion, with no holds, jumps or denser particles.
41
+ - **Verify:** Render each effect at 1x and at 0.25x and lay the frames side by side: the slow version has four times the frames of the same motion, with no holds, jumps or denser particles, and the probe's simulation time confirms the scale took.
39
42
 
40
43
  ### One seed makes a take repeat
41
44
  - **Invariant:** Two runs of the same take with the same seed, clock and scripted input produce the same frames.
@@ -53,25 +56,32 @@ numbers, the capability audit and process are in [references/pipeline.md](refere
53
56
 
54
57
  ### A clean feed, element by element
55
58
  - **Invariant:** Each HUD element and each world-space gameplay marker can be hidden on its own, music and voice can each be muted, and a clean take keeps whatever feedback the shot asks for.
56
- - **Technique:** A named flag for every interface element and world marker (hit markers, score pops, crosshair, plates, subtitles, set-piece UI, telegraphs, weak-point marks, rings) and a mask that picks which to draw. Tag world markers separately from art. One clean-feed switch sets the usual mask and mutes music and voice; prove the mute by loudness. Interface projected from the player's camera lines up only in first person, so hide it on other cameras. Dev overlays, page badges and cheat effects never reach the frame.
57
- - **Reference configuration:** OC's owner asked to hide "the H UD except hitmarkers and like the points juice". A HUD mask (score pops, hit marker, crosshair, points plate) plus flags for subtitles and set-piece UI took about 3 min to build. The old hide flag also hid boss marks and the letterbox, subtitles ignored it, and the mask drew over it, so cinematics had to clear both. Cyan rings, a weak-point capsule, floor telegraphs, glyphs and strike circles read as HUD on film. On the montage, dev-funded score jumps from 1,039 to 49,550 showed on screen.
59
+ - **Technique:** A named flag for every interface element and world marker (hit markers, score pops, crosshair, plates, subtitles, set-piece UI, telegraphs, weak-point marks, rings) and a mask that picks which to draw. Tag world markers separately from art. One clean-feed switch sets the usual mask and mutes music and voice; prove the mute by loudness. Interface projected from the player's camera lines up only in first person, so hide it on other cameras. Dev overlays, page badges and cheat effects never reach the frame. Give scoring tools the same mask, so a screen-fixed reticle or HUD is never counted as change in the scene.
60
+ - **Reference configuration:** OC's owner asked to hide "the H UD except hitmarkers and like the points juice". A HUD mask (score pops, hit marker, crosshair, points plate) plus flags for subtitles and set-piece UI took about 3 min to build. The old hide flag also hid boss marks and the letterbox, subtitles ignored it, and the mask drew over it, so cinematics had to clear both. Cyan rings, a weak-point capsule, floor telegraphs, glyphs and strike circles read as HUD on film. On the montage, dev-funded score jumps from 1,039 to 49,550 showed on screen. On OS the screen-fixed reticle made 75-91% of the spikes a flicker scorer counted in scoped shots until it was masked.
58
61
  - **Applies when:** Every game with a HUD that will be filmed or used for marketing stills.
59
62
  - **Verify:** A contact sheet of one frame with each flag on and off. On the clean feed, a reviewer marks every visible interface element; only the masked-in ones remain.
60
63
 
61
64
  ### Ready means resident
62
65
  - **Invariant:** The game exposes one state the capture script can await that means every model, texture, pipeline and sound the scene needs is resident and the quality setting is locked, and a load failure reports itself instead of crashing.
63
- - **Technique:** Count pending loads, compiles and decodes; resolve a promise or post a message when they reach zero after the scene is built and quality is fixed. Give each asset class a named failure and a timeout. In the unrecorded preroll, grant and show every item the take uses once, so late loads happen off camera. Lock quality: no dynamic resolution or automatic tier in capture mode.
64
- - **Reference configuration:** OC had no such state. Under load, 55 boot crashes paired with a renderer assert on a missing material; about 40 takes hit it and 5 of 7 attempts at one shot were lost. Boot waits grew from 180 to 420 s, and takes were capped at 2-3 min with 2-5 retries. Late guns drew black for about 10 s ("give every gun once before shooting"), and sound banks stayed silent for 11-20 s. The renderer started on Low and calibrated, so dynamic resolution popped lights until quality was fixed after the title; the rig waited for the log line "renderer quality: fixed at ultra".
66
+ - **Technique:** Count pending loads, compiles and decodes; resolve a promise or post a message when they reach zero after the scene is built and quality is fixed. Give each asset class a named failure and a timeout. In the unrecorded preroll, grant and show every item the take uses once, so late loads happen off camera. Lock quality: no dynamic resolution or automatic tier in capture mode. A capture that waits for ready never sees the seconds before it, so test the first seconds of play with a separate capture that starts at the press.
67
+ - **Reference configuration:** OC had no such state. Under load, 55 boot crashes paired with a renderer assert on a missing material; about 40 takes hit it and 5 of 7 attempts at one shot were lost. Boot waits grew from 180 to 420 s, and takes were capped at 2-3 min with 2-5 retries. Late guns drew black for about 10 s ("give every gun once before shooting"), and sound banks stayed silent for 11-20 s. The renderer started on Low and calibrated, so dynamic resolution popped lights until quality was fixed after the title; the rig waited for the log line "renderer quality: fixed at ultra". LC's captures waited for late assets and so never saw the white placeholder hand a player saw before the real hands arrived.
65
68
  - **Applies when:** Every game that streams assets or calibrates quality.
66
69
  - **Verify:** Await the state on a throttled connection and capture the next frame: every model is present and lit at final quality, and the first shot is audible. Block one asset: the state reports which one failed.
67
70
 
68
71
  ### Capture requests are addressed to frames and acknowledged one by one
69
72
  - **Invariant:** Every screenshot request and director command names the simulation frame it applies to, and the game acknowledges each one with the frame it ran on, so the capture script never guesses when something happened.
70
73
  - **Technique:** A request queue with an id and a target frame per request. The game runs it at that frame and replies with the id, the frame, the build id, the screen it was on and whether assets were loaded. The capture script sends the next request after the reply and treats a missing reply as a failure. Screenshots wait for the interface to settle. Document the limits: command size, deadlines, and what happens to a command sent before the scene exists.
71
- - **Reference configuration:** OC: real-time stage commands 0.2 s apart landed in the same frame and were lost under load, so staging moved to an unrecorded preroll. A shutter command sent only on change swallowed the frame-0 command. Commands sent before the scene existed were dropped, a scripted run aborted 10 s after its last command, and a command buffer over about 7.9 KB crashed the engine on load. Frame-addressed injection landed 59 of 59 on the right frame. Checking that each capture shows the screen it claims is in `opengame-verification`.
74
+ - **Reference configuration:** OC: real-time stage commands 0.2 s apart landed in the same frame and were lost under load, so staging moved to an unrecorded preroll. A shutter command sent only on change swallowed the frame-0 command. Commands sent before the scene existed were dropped, a scripted run aborted 10 s after its last command, and a command buffer over about 7.9 KB crashed the engine on load. Frame-addressed injection landed 59 of 59 on the right frame. OS's capture steps each waited for a log line the game printed, and every step that shoots shoots exactly once, after one boot took 184 s. Checking that each capture shows the screen it claims is in `opengame-verification`.
72
75
  - **Applies when:** Every game driven by a capture or test script.
73
76
  - **Verify:** Send 100 screenshot requests in a burst: 100 replies arrive, each with its own frame and the expected screen. Kill the game mid-burst: the capture script fails the run.
74
77
 
78
+ ### A capture proves its own build, state, camera and coverage
79
+ - **Invariant:** A frame reaches a judge, a metric or the owner only if its pixels and a log line from the same second agree with its label (the named build from committed code, the real interface path, the player alive in a first life, the requested camera outside geometry, quality and weather pinned), it was taken from a frozen snapshot nobody else can change while the run lasts, and the set covers every cell the rubric scores.
80
+ - **Technique:** Capture from a snapshot of the committed head plus your own files, served on a private port, so another agent's save cannot reload the page. Pin quality, weather and time of day, and write the renderer string, build hash and state line beside each image. Dump the camera before and after each still, ray-probe every judge camera against collision, and spawn with god mode off. Wait on game state, never on timers. Write each run to an immutable folder keyed by build hash. Quarantine black, low-variance and duplicate frames. Generate the capture plan from the rubric, so every scored cell has a frame. Label each clip for what it is: stepped or real time, its clock rate, brightened or not.
81
+ - **Reference configuration:** OS's first round had 6 of 10 maps at 1 HP, put the victory screen in all three scope slots, and sent judges four "gameplay" stills aimed at the sky; on a first pass 13 of one map's 39 checklist cameras sat inside walls, roofs or cliff tops. "Several smoke folders were rewritten while I was reading them." Capture "success" (files written, 0 errors, 4 of 4 shots) proved nothing four separate times. OH captured at 1600 x 900 from frozen snapshots of the committed head on private ports, quality pinned, a state line and renderer string beside each image, and still had a missing clip, a mid-animation title and a stale capture scored as game faults; its rule became a manifest of build hash and renderer, with no file older than the build. Checking a capture against its label before judging is in `opengame-verification`.
82
+ - **Applies when:** Any automated capture that feeds a judge, a gate, a metric or the owner.
83
+ - **Verify:** Kill the player partway through a run and move one camera into rock: the run exits non-zero and quarantines both. The manifest lists the build hash and renderer, and no file in the set is older than the build.
84
+
75
85
  ### A state probe in render space, checked on day one
76
86
  - **Invariant:** On request or every frame, the game reports where the camera, the player, each enemy's head and joints, bosses and objective states are, taken from the same snapshot that was rendered and in render space, and the report matches the picture.
77
87
  - **Technique:** A probe flag that emits one record per frame from post-skinning transforms in the rendered snapshot. Name items and tables in the record; never make anyone map indices by hand. On the first day, project probe points into a rendered frame and check each lands on the drawn head, joint or item. Export the per-frame camera and joints if CG will be composited.
@@ -79,24 +89,38 @@ numbers, the capability audit and process are in [references/pipeline.md](refere
79
89
  - **Applies when:** Games where a capture script aims, frames or checks shots from game state.
80
90
  - **Verify:** Draw probe points over 20 captured frames: each lands on its drawn feature within a few pixels.
81
91
 
92
+ ### Motion is filmed on the game clock and read at quarter speed, from the angle that shows it
93
+ - **Invariant:** Holds, reloads, gaits, deaths, effects, flicker and transitions are accepted only from a recording of every drawn frame on the game clock, played back slowed and stepped frame by frame, from the player's camera and from an angle that shows the motion, and also watched at full speed as a whole act by a person, since no agent can; an interface cut over the scene is judged from a recording that holds both.
94
+ - **Technique:** Record every drawn frame for a few seconds on the stepped clock and play it at quarter speed (drawn frames shown at a quarter of their rate), with a sheet of consecutive frames beside it. Film enemies from a side lens at chest height a few metres off as well as from the player's view. For flicker and camera feel, script a fixed still, turn, strafe, walk and aim sequence so before and after show the same poses. When the DOM and the canvas must be judged together, run a headed browser on its own virtual display and grab it at the display rate, with a small corner mark that changes every frame as the frame clock. Stamp any dark frame you brightened with its gain, and never pass anything on a brightened frame.
95
+ - **Reference configuration:** LC filmed every drawn frame and played it at 15 fps (quarter speed at 60) beside a 16-frame sheet, and filmed enemies through a side lens 3 m off at chest height with a 68-degree field of view, which exposed library "falls" as stunt breakfalls that the deformation audit had passed. LC's headed grab ran 34 fps at a machine load of 13 and 4-13 fps at 35-54, and showed the DOM changing 67-167 ms before the WebGPU picture. OS's game-clock flicker path held within about 5% run to run, against 10-30% for free-running captures. OH's judges found its board vanishing for single frames by stepping a clip, and an in-page probe counted 4 of 880.
96
+ - **Applies when:** Anything that moves, in any game whose motion an agent, a judge or the owner will review.
97
+ - **Verify:** The film's frame count equals game seconds times the step rate, the corner mark changes once per frame, and each motion has a sheet with named faults or "none", read at full size.
98
+
82
99
  ### Audio is captured at simulation rate, per bus
83
100
  - **Invariant:** A take's sound comes from the simulation at the take's own rate, music, voice, effects and ambience are available as separate stems, and each bus can be muted from the launch URL.
84
- - **Technique:** Mute and stem flags readable at launch. In capture mode, drain the mixer per output frame by exactly the samples one step covers, trim the output latency, and save a cue log with each cue's simulation time. Slow motion yields audio the length of simulated time, so the edit retimes or replaces it. Load every sound bank a take uses before recording.
85
- - **Reference configuration:** OC: music and voice were baked into every take until music and voice mute flags existed, found when 12 of 22 voice lines overlapped. The rig drained `sampleRate * step` samples per frame and trimmed 4,095 samples (85 ms) of latency; 13 of 145 cues needed re-alignment in slow-motion takes. On the montage, settings-menu mute clicks never landed and all 9 segments were reshot (`opengame-audio` has the numbers). The film mastered at -13.8 LUFS and -1.6 dBTP against a -14 LUFS target.
101
+ - **Technique:** Mute and stem flags readable at launch. In capture mode, drain the mixer per output frame by exactly the samples one step covers, trim the output latency, and save a cue log with each cue's simulation time. Slow motion yields audio the length of simulated time, so the edit retimes or replaces it. Load every sound bank a take uses before recording. Where a capture cannot carry sound, prove timing from logs on the game's clock and leave the listening to a person.
102
+ - **Reference configuration:** OC: music and voice were baked into every take until music and voice mute flags existed, found when 12 of 22 voice lines overlapped. The rig drained `sampleRate * step` samples per frame and trimmed 4,095 samples (85 ms) of latency; 13 of 145 cues needed re-alignment in slow-motion takes. On the montage, settings-menu mute clicks never landed and all 9 segments were reshot (`opengame-audio` has the numbers). The film mastered at -13.8 LUFS and -1.6 dBTP against a -14 LUFS target. Every clip OH's judges saw had no audio, so OH proved timing from logs on the song clock (drawn against heard under 4 ms). LC recorded the page's audio by routing everything sent to the output into a recorder while the browser stayed muted; OS recorded the exact WebAudio output aligned to the engine log to about 0.2 s, which exposed a 1.7-3.6 s skew behind wrong numbers.
86
103
  - **Applies when:** Every game with sound that will be filmed.
87
104
  - **Verify:** Capture the same shot at 1x and at 0.25x. At 1x each shot's sound lines up with its muzzle flash to within a frame. With music muted, the music stem measures silent.
88
105
 
89
106
  ### Every frame comes from the render surface
90
- - **Invariant:** Delivered footage holds every frame at the delivery resolution and frame rate, read from the render surface with correct colour tags, however loaded the machine is.
91
- - **Technique:** After each stepped frame, read back the render surface, convert colour on the GPU to the encoder's format and tag it, and encode with a hardware encoder through WebCodecs. Page screenshots and screencasts are for stills and debugging, never for motion. Render at native resolution with a device pixel ratio of 1, lock the quality preset, and stop the run if the browser fell back to a software rasteriser. Check output by counting unique frames and measuring colour, never by frame count alone.
92
- - **Reference configuration:** OC: `page.screenshot` at 4K took 2,516 ms; a CDP screencast gave 12-19 fps, lossy; CPU x264 encoded 1.7-4.2 fps at 1080p. A swapchain copy with `copyTextureToBuffer` and `mapAsync` took 30.4 ms at 4K. WebCodecs tagged frames BT.601, so a GPU pass to NV12 BT.709 fixed it (Y 68.7 dB against ffmpeg, against 41.3 dB for BT.601). HEVC at 300 Mbps banded least on the storm sky (H.264 at 160 Mbps flattened 51% of blocks, HEVC 14.5%); the 4K master is 2.25 GB per output minute. Supersampling at dpr 1.5 cost 3.4x (11.0 against 37.2 fps) for no visible gain. Dropped frames were 0 regardless of load, because frames were stepped, not timed. Frame-count QA passed a 3,720-frame cut whose frames 1-402 were frozen.
107
+ - **Invariant:** Delivered footage and test captures hold every frame at the intended resolution and rate, read from the render surface with correct colour tags, however loaded the machine is, and a capture whose first picture is near black fails.
108
+ - **Technique:** After each stepped frame, read back the render surface, convert colour on the GPU to the encoder's format and tag it, and encode with a hardware encoder through WebCodecs. Page screenshots and screencasts are for stills and debugging, never for motion, and a screencast's timing and frame count are not the game's. Read a WebGPU canvas inside the task that drew it, from a declared after-frame hook that changes nothing else (see the hooks rule below), with `canvas.toBlob` or a copy of the presented texture; depending on the browser's mode, copying the canvas into another canvas, `captureStream`, a screencast or a page screenshot can return black, so prove the path on one frame first and check that the first frame's mean clears a near-black floor. Render at native resolution with a device pixel ratio of 1, lock the quality preset, and stop the run if the browser fell back to a software rasteriser. Check output by counting unique frames and measuring colour, never by frame count alone.
109
+ - **Reference configuration:** OC: `page.screenshot` at 4K took 2,516 ms; a CDP screencast gave 12-19 fps, lossy; CPU x264 encoded 1.7-4.2 fps at 1080p. A swapchain copy with `copyTextureToBuffer` and `mapAsync` took 30.4 ms at 4K. WebCodecs tagged frames BT.601, so a GPU pass to NV12 BT.709 fixed it (Y 68.7 dB against ffmpeg, against 41.3 dB for BT.601). HEVC at 300 Mbps banded least on the storm sky (H.264 at 160 Mbps flattened 51% of blocks, HEVC 14.5%); the 4K master is 2.25 GB per output minute. Supersampling at dpr 1.5 cost 3.4x (11.0 against 37.2 fps) for no visible gain. Dropped frames were 0 regardless of load, because frames were stepped, not timed. Frame-count QA passed a 3,720-frame cut whose frames 1-402 were frozen. On LC, headless, copying a WebGPU canvas into another canvas, `captureStream()`, the CDP screencast and `page.screenshot` all gave black frames (79 of 79), and LC fails a capture whose first frame's mean is under 1-2 of 255. OH's 25 fps screencasts sat 1-4.6 s off the page, and judges who timed effects by counting their frames were wrong.
93
110
  - **Applies when:** Every game filmed for a trailer, montage or store page.
94
111
  - **Verify:** `ffprobe` reports the expected size, rate and BT.709 tags; the count of unique frames equals the count of frames; one frame matches a direct readback of the same step.
95
112
 
113
+ ### Batch the shots for one look into one job, and prove the tool on one frame first
114
+ - **Invariant:** Every shot needed for one look is taken in one job on your machine's job queue, with no more browsers than the GPU holds, the job is watched until its result is read, and a new or changed capture tool has been proved on one frame before a batch spends a slot.
115
+ - **Technique:** Write the batch as a spec (scenes, cameras, states, frames) and run it in one slot from one build. Run a dry pass that boots no browser and checks names and paths, then one shot, then the batch. Wait for the job inside your own turn with a bounded wait on its end mark; a job left running after its launcher's turn ends reaches nobody. Keep file names short and encoder tile widths even.
116
+ - **Reference configuration:** LC kept at most three browsers in one slot after five lost the WebGPU device; an odd tile width gave 0-frame clips, and under load a kept picture cost about 1 s. Learning the black-frame rule cost one LC stream five slot jobs and every stream about one. OH kept every shot for a judge packet inside one slot grant of under 8 minutes, never nested. OS grouped steps into slot jobs of under 20 minutes and spaced screenshots at least 1.2 s apart. How to share the slots between agents is in `opengame-production`.
117
+ - **Applies when:** Every capture run on a machine whose browsers, GPU or job slots are shared.
118
+ - **Verify:** The job log shows one slot per look, the dry pass and the single-shot proof before the batch, and no capture job still running after its launcher's turn ended.
119
+
96
120
  ### Capture hooks are off by default and absent from public builds
97
- - **Invariant:** Capture hooks live in the shipping game's code, each off by default behind its own flag, change nothing while off, and are absent or locked in the public build.
98
- - **Technique:** One flag per hook, listed in one file with what it does. A parity frame proves flags-off equals the release. Strip hooks at build time for the public build, or gate them behind a key whose secret is not in the bundle. Record the build hash and every flag a take used in the take's spec. Write each film fix as a flag the game can adopt.
99
- - **Reference configuration:** OC added about 25-30 default-off hooks in a day, each diffable as a hooks list and a patch, versioned film builds by wasm hash with receipts, and proved parity with the release round on a frame. The fork had costs: the final film mixes selects from at least 5 builds and no single build reproduces it, and one take depended on environment variables nobody recorded. Production disabled dev URL switches with a dev key stored only as a SHA-256 hash. Film fixes flowed back: the film's lightning became the game default (flat-white share 10.47% to 0.58%).
121
+ - **Invariant:** Capture hooks live in the shipping game's code, each off by default behind its own flag, change nothing while off, change nothing but what they declare while on, and are absent or locked in the public build.
122
+ - **Technique:** One flag per hook, listed in one file with what it does. No game code reads whether a hook is attached except through that declaration. A parity frame proves flags-off equals the release. Strip hooks at build time for the public build, or gate them behind a key whose secret is not in the bundle. Record the build hash and every flag a take used in the take's spec. Write each film fix as a flag the game can adopt.
123
+ - **Reference configuration:** OC added about 25-30 default-off hooks in a day, each diffable as a hooks list and a patch, versioned film builds by wasm hash with receipts, and proved parity with the release round on a frame. The fork had costs: the final film mixes selects from at least 5 builds and no single build reproduces it, and one take depended on environment variables nobody recorded. Production disabled dev URL switches with a dev key stored only as a SHA-256 hash. Film fixes flowed back: the film's lightning became the game default (flat-white share 10.47% to 0.58%). LC's engine skipped its frame-pacing rule whenever an after-frame hook was set, so every capture sat a different game from the owner's; `opengame-verification` has the rule.
100
124
  - **Applies when:** Every game with capture or test hooks and a public build.
101
125
  - **Verify:** Grep the public bundle for every hook name; open the public URL with each flag set and see no change; diff a flags-off frame from the capture build against the release.
102
126
 
@@ -1,7 +1,8 @@
1
1
  # Capture pipeline, audit and process
2
2
 
3
3
  Numbers from OC's launch film (a 62 s 4K60 trailer) and its gameplay montage, filmed on an
4
- M4 Max with 16 cores.
4
+ M4 Max with 16 cores, and from the test captures of OS, LC and OH, taken on shared Linux
5
+ machines with one GPU.
5
6
 
6
7
  ## Audit before planning shots
7
8
 
@@ -50,6 +51,32 @@ what is missing. OC's film team found most of these one at a time, mid-productio
50
51
  passed a cut whose first 402 frames were frozen, because two directors shared one
51
52
  minute-stamped work folder.
52
53
 
54
+ ## Test captures and judge packets
55
+
56
+ - Capture from a snapshot of the committed head plus your own files, never the shared
57
+ working tree. LC's slot tool read the tree when the slot was granted, not when the job
58
+ was queued, so an edit made during the wait killed jobs or filmed someone else's
59
+ half-written file.
60
+ - Write a manifest per packet: build hash, renderer string, quality, weather, camera per
61
+ frame, the state line, and whether each clip is stepped or real time. OH's judges scored
62
+ a missing clip, a mid-animation title and a stale capture as game faults, and OH's
63
+ rule became a manifest with no file older than the build.
64
+ - Generate the capture list from the rubric. OS's council scored more than 20 "gaps" that
65
+ were built items no camera showed.
66
+ - Space stills when the game settles UI after input: OS waited at least 1.2 s between
67
+ screenshots and A/B'd inside one boot, because "Numbers across boots move about 20 %".
68
+ - Software rasterisers prove function only. Refuse one for anything judged on looks or
69
+ timing, and print the adapter on every run (LC and OH both did).
70
+
71
+ ## Filming the interface with the scene
72
+
73
+ Stepped capture films the canvas. When a cut, a menu or a HUD change must be judged
74
+ against the scene at speed, LC ran a headed browser on the job's own virtual display and
75
+ grabbed it with `ffmpeg -f x11grab -framerate 60`, with a 6-pixel white corner mark that
76
+ changed every frame as the frame clock. It ran 34 fps at a load of 13 and 4-13 fps at
77
+ 35-54, so read it for order and offsets, never for frame time. It showed the DOM changing
78
+ 67-167 ms before the WebGPU picture.
79
+
53
80
  ## Contention
54
81
 
55
82
  Capture is the heaviest job on the machine, and it slows down when anything else runs.
@@ -66,6 +93,11 @@ Capture is the heaviest job on the machine, and it slows down when anything else
66
93
  10% other work only after more than 40 h of contention.
67
94
  - Private build trees as full disk clones cost about 26 GB for 3 flags. Keep a disk floor
68
95
  (OC's was 4 GB).
96
+ - Job slots do not govern the GPU. OH held two of six slots while another project's
97
+ headless browsers kept the card 14-100% busy, so every OH frame-time number is
98
+ contaminated. Record GPU load beside every frame number.
99
+ - On LC, at most three browsers fit in one slot (five lost the WebGPU device), and real-time
100
+ films on the loaded machine drew 8-25 fps, which is why motion is filmed stepped.
69
101
 
70
102
  ## Process with the game team
71
103
 
@@ -94,8 +94,24 @@ pixels with the drawing buffer size. Draw those points over the captured frame o
94
94
  target, or build a `VideoFrame` from the canvas right after rendering. If you configure the
95
95
  canvas context yourself, add `GPUTextureUsage.COPY_SRC` to its usage; OC read its swapchain
96
96
  with `copyTextureToBuffer` and `mapAsync` in 30.4 ms at 4K.
97
+ - Headless WebGPU: read the canvas with `canvas.toBlob()` in the same task that rendered
98
+ it, one copy per frame, and hand frames to the host in batches (LC returned 20 at a
99
+ time). On LC, `drawImage` of the WebGPU canvas into a 2D canvas, `captureStream()`, the
100
+ CDP screencast and `page.screenshot` all returned black (79 of 79), so prove the path on
101
+ one frame and check its mean before a batch:
102
+
103
+ ```js
104
+ async function captureFrame(dt) { // capture mode only, behind its flag
105
+ step(dt)
106
+ renderer.render(scene, camera) // the read below must be in this task
107
+ const blob = await new Promise((r) => renderer.domElement.toBlob(r, 'image/png'))
108
+ return blob // check the first one is not near black
109
+ }
110
+ ```
111
+
97
112
  - `canvas.captureStream(0)` with `track.requestFrame()` after each render feeds a
98
- `MediaRecorder`. Its timestamps are wall time, so use it for real-time previews only.
113
+ `MediaRecorder`. Its timestamps are wall time, so use it for real-time previews only,
114
+ and check it is not black on a WebGPU canvas.
99
115
  - Refuse a software rasteriser. WebGL: read `UNMASKED_RENDERER_WEBGL` through
100
116
  `WEBGL_debug_renderer_info` and stop on SwiftShader or llvmpipe. WebGPU: inspect
101
117
  `adapter.info`, including `isFallbackAdapter` where the browser reports it.