@sanity/ailf 1.0.0 → 2.0.1

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 (499) hide show
  1. package/README.md +0 -1
  2. package/canonical/grader-references/README.md +2 -2
  3. package/canonical/reference-solutions/content-lake/mutations.ts +160 -0
  4. package/canonical/reference-solutions/content-lake/realtime.ts +187 -0
  5. package/canonical/reference-solutions/image-handling/asset-pipeline.tsx +166 -0
  6. package/canonical/reference-solutions/portable-text/custom-blocks.ts +204 -0
  7. package/canonical/reference-solutions/portable-text/rendering.tsx +163 -0
  8. package/config/features.ts +1 -1
  9. package/config/models.ts +29 -12
  10. package/config/sources.ts +1 -1
  11. package/config/thresholds.ts +1 -1
  12. package/dist/_vendor/ailf-core/__tests__/comparison-formatters.test.d.ts +10 -0
  13. package/dist/_vendor/ailf-core/__tests__/comparison-formatters.test.js +185 -0
  14. package/dist/_vendor/ailf-core/artifact-capture/__tests__/noop-collector.test.d.ts +6 -0
  15. package/dist/_vendor/ailf-core/artifact-capture/__tests__/noop-collector.test.js +42 -0
  16. package/dist/_vendor/ailf-core/artifact-capture/noop-collector.d.ts +14 -0
  17. package/dist/_vendor/ailf-core/artifact-capture/noop-collector.js +25 -0
  18. package/dist/_vendor/ailf-core/config-helpers.d.ts +20 -17
  19. package/dist/_vendor/ailf-core/config-helpers.js +51 -2
  20. package/dist/_vendor/ailf-core/examples/index.d.ts +166 -80
  21. package/dist/_vendor/ailf-core/examples/index.js +213 -94
  22. package/dist/_vendor/ailf-core/index.d.ts +3 -2
  23. package/dist/_vendor/ailf-core/index.js +2 -1
  24. package/dist/_vendor/ailf-core/ports/artifact-collector.d.ts +94 -0
  25. package/dist/_vendor/ailf-core/ports/artifact-collector.js +13 -0
  26. package/dist/_vendor/ailf-core/ports/capture-comparator.d.ts +138 -0
  27. package/dist/_vendor/ailf-core/ports/capture-comparator.js +10 -0
  28. package/dist/_vendor/ailf-core/ports/context.d.ts +22 -1
  29. package/dist/_vendor/ailf-core/ports/eval-runner.d.ts +6 -0
  30. package/dist/_vendor/ailf-core/ports/index.d.ts +2 -0
  31. package/dist/_vendor/ailf-core/ports/pipeline-step.d.ts +11 -0
  32. package/dist/_vendor/ailf-core/ports/task-source.d.ts +3 -3
  33. package/dist/_vendor/ailf-core/ports/task-source.js +3 -3
  34. package/dist/_vendor/ailf-core/schemas/eval-config.d.ts +1 -0
  35. package/dist/_vendor/ailf-core/schemas/eval-config.js +10 -0
  36. package/dist/_vendor/ailf-core/schemas/pipeline-request.d.ts +7 -1
  37. package/dist/_vendor/ailf-core/schemas/pipeline-request.js +16 -2
  38. package/dist/_vendor/ailf-core/schemas/pipeline.d.ts +0 -2
  39. package/dist/_vendor/ailf-core/schemas/pipeline.js +0 -1
  40. package/dist/_vendor/ailf-core/services/config-helpers.d.ts +16 -1
  41. package/dist/_vendor/ailf-core/services/config-helpers.js +21 -0
  42. package/dist/_vendor/ailf-core/services/index.d.ts +1 -1
  43. package/dist/_vendor/ailf-core/services/index.js +1 -1
  44. package/dist/_vendor/ailf-core/services/scoring.js +9 -0
  45. package/dist/_vendor/ailf-core/types/generalized-task.d.ts +25 -1
  46. package/dist/_vendor/ailf-core/types/generalized-task.js +1 -1
  47. package/dist/_vendor/ailf-core/types/index.d.ts +48 -7
  48. package/dist/_vendor/ailf-core/types/plugin-registry.d.ts +105 -23
  49. package/dist/_vendor/ailf-core/types/plugin-registry.js +73 -20
  50. package/dist/_vendor/ailf-shared/eval-modes.d.ts +15 -0
  51. package/dist/_vendor/ailf-shared/eval-modes.js +18 -0
  52. package/dist/adapters/api-client/remediation.js +2 -2
  53. package/dist/adapters/config-sources/file-config-adapter.js +7 -1
  54. package/dist/adapters/config-sources/ts-config-loader.js +21 -13
  55. package/dist/adapters/eval-runners/promptfoo-eval-adapter.js +8 -2
  56. package/dist/adapters/index.d.ts +0 -1
  57. package/dist/adapters/index.js +0 -1
  58. package/dist/adapters/task-sources/composite-task-source.d.ts +1 -1
  59. package/dist/adapters/task-sources/composite-task-source.js +1 -1
  60. package/dist/adapters/task-sources/content-lake-task-source.d.ts +4 -6
  61. package/dist/adapters/task-sources/content-lake-task-source.js +21 -26
  62. package/dist/adapters/task-sources/index.d.ts +3 -4
  63. package/dist/adapters/task-sources/index.js +3 -4
  64. package/dist/adapters/task-sources/repo-schemas.d.ts +219 -17
  65. package/dist/adapters/task-sources/repo-schemas.js +228 -20
  66. package/dist/adapters/task-sources/repo-task-source.d.ts +14 -10
  67. package/dist/adapters/task-sources/repo-task-source.js +81 -122
  68. package/dist/adapters/task-sources/repo-trigger.d.ts +1 -1
  69. package/dist/adapters/task-sources/repo-trigger.js +1 -1
  70. package/dist/adapters/task-sources/repo-validation.d.ts +36 -5
  71. package/dist/adapters/task-sources/repo-validation.js +126 -5
  72. package/dist/adapters/task-sources/task-file-loader.d.ts +10 -7
  73. package/dist/adapters/task-sources/task-file-loader.js +21 -7
  74. package/dist/agent-observer/test-imports.d.ts +7 -0
  75. package/dist/agent-observer/test-imports.js +185 -0
  76. package/dist/artifact-capture/comparator.d.ts +22 -0
  77. package/dist/artifact-capture/comparator.js +493 -0
  78. package/dist/artifact-capture/filesystem-collector.d.ts +42 -0
  79. package/dist/artifact-capture/filesystem-collector.js +237 -0
  80. package/dist/artifact-capture/redact-artifact.d.ts +20 -0
  81. package/dist/artifact-capture/redact-artifact.js +115 -0
  82. package/dist/assertions/source-isolation.d.ts +1 -1
  83. package/dist/assertions/source-isolation.js +1 -1
  84. package/dist/cli.js +4 -0
  85. package/dist/commands/calculate-scores.js +1 -0
  86. package/dist/commands/capture-compare.d.ts +15 -0
  87. package/dist/commands/capture-compare.js +253 -0
  88. package/dist/commands/capture-list.d.ts +12 -0
  89. package/dist/commands/capture-list.js +147 -0
  90. package/dist/commands/capture.d.ts +9 -0
  91. package/dist/commands/capture.js +16 -0
  92. package/dist/commands/chronic-failures.d.ts +8 -0
  93. package/dist/commands/chronic-failures.js +33 -0
  94. package/dist/commands/coverage-audit.js +3 -1
  95. package/dist/commands/explain-handler.d.ts +1 -1
  96. package/dist/commands/explain-handler.js +37 -8
  97. package/dist/commands/fetch-docs.js +1 -0
  98. package/dist/commands/generate-configs.d.ts +3 -3
  99. package/dist/commands/generate-configs.js +20 -8
  100. package/dist/commands/init.d.ts +5 -4
  101. package/dist/commands/init.js +190 -25
  102. package/dist/commands/pipeline-action.d.ts +7 -1
  103. package/dist/commands/pipeline-action.js +43 -19
  104. package/dist/commands/pipeline.d.ts +6 -1
  105. package/dist/commands/pipeline.js +7 -2
  106. package/dist/commands/pr-comment.js +1 -0
  107. package/dist/commands/publish.js +1 -0
  108. package/dist/commands/shared/help.js +2 -2
  109. package/dist/commands/update-quality-scores.d.ts +5 -0
  110. package/dist/commands/update-quality-scores.js +20 -0
  111. package/dist/commands/validate-tasks.d.ts +2 -2
  112. package/dist/commands/validate-tasks.js +26 -15
  113. package/dist/composition-root.d.ts +15 -4
  114. package/dist/composition-root.js +100 -55
  115. package/dist/config/features.ts +23 -0
  116. package/dist/config/models.ts +100 -0
  117. package/dist/config/prompts.ts +16 -0
  118. package/dist/config/rubrics.ts +225 -0
  119. package/dist/config/schedules.ts +47 -0
  120. package/dist/config/sinks.ts +37 -0
  121. package/dist/config/sources.ts +21 -0
  122. package/dist/config/thresholds.ts +61 -0
  123. package/dist/index.d.ts +41 -0
  124. package/dist/index.js +48 -0
  125. package/dist/lib/agent-behavior-report.d.ts +8 -0
  126. package/dist/lib/agent-behavior-report.js +185 -0
  127. package/dist/lib/baseline.d.ts +19 -0
  128. package/dist/lib/baseline.js +153 -0
  129. package/dist/lib/calculate-scores.d.ts +23 -0
  130. package/dist/lib/calculate-scores.js +42 -0
  131. package/dist/lib/compare.d.ts +18 -0
  132. package/dist/lib/compare.js +170 -0
  133. package/dist/lib/coverage-audit.d.ts +4 -0
  134. package/dist/lib/coverage-audit.js +42 -0
  135. package/dist/lib/discovery-report.d.ts +13 -0
  136. package/dist/lib/discovery-report.js +57 -0
  137. package/dist/lib/fetch-docs.d.ts +30 -0
  138. package/dist/lib/fetch-docs.js +171 -0
  139. package/dist/lib/generate-configs.d.ts +25 -0
  140. package/dist/lib/generate-configs.js +42 -0
  141. package/dist/lib/grader-api.d.ts +21 -0
  142. package/dist/lib/grader-api.js +34 -0
  143. package/dist/lib/grader-compare.d.ts +19 -0
  144. package/dist/lib/grader-compare.js +91 -0
  145. package/dist/lib/grader-consistency.d.ts +27 -0
  146. package/dist/lib/grader-consistency.js +79 -0
  147. package/dist/lib/grader-sensitivity.d.ts +19 -0
  148. package/dist/lib/grader-sensitivity.js +75 -0
  149. package/dist/lib/grader-validate.d.ts +19 -0
  150. package/dist/lib/grader-validate.js +78 -0
  151. package/dist/lib/measure-retrieval.d.ts +14 -0
  152. package/dist/lib/measure-retrieval.js +71 -0
  153. package/dist/lib/pr-comment.d.ts +16 -0
  154. package/dist/lib/pr-comment.js +28 -0
  155. package/dist/lib/readiness-report.d.ts +13 -0
  156. package/dist/lib/readiness-report.js +108 -0
  157. package/dist/lib/webhook-server.d.ts +11 -0
  158. package/dist/lib/webhook-server.js +24 -0
  159. package/dist/lib/weekly-digest.d.ts +24 -0
  160. package/dist/lib/weekly-digest.js +148 -0
  161. package/dist/orchestration/build-app-context.js +13 -0
  162. package/dist/orchestration/build-step-sequence.js +4 -2
  163. package/dist/orchestration/cache-context.d.ts +23 -0
  164. package/dist/orchestration/cache-context.js +43 -0
  165. package/dist/orchestration/env-bridge.d.ts +21 -0
  166. package/dist/orchestration/env-bridge.js +66 -0
  167. package/dist/orchestration/load-pipeline-tasks.d.ts +34 -0
  168. package/dist/orchestration/load-pipeline-tasks.js +52 -0
  169. package/dist/orchestration/pipeline-orchestrator.js +75 -5
  170. package/dist/orchestration/step-runner.js +5 -1
  171. package/dist/orchestration/steps/calculate-scores-step.d.ts +1 -0
  172. package/dist/orchestration/steps/calculate-scores-step.js +13 -0
  173. package/dist/orchestration/steps/callback-step.js +10 -1
  174. package/dist/orchestration/steps/compare-step.js +6 -3
  175. package/dist/orchestration/steps/discovery-report-step.js +6 -2
  176. package/dist/orchestration/steps/fetch-docs-shell.d.ts +17 -0
  177. package/dist/orchestration/steps/fetch-docs-shell.js +30 -0
  178. package/dist/orchestration/steps/fetch-docs-step.d.ts +1 -0
  179. package/dist/orchestration/steps/fetch-docs-step.js +32 -19
  180. package/dist/orchestration/steps/gap-analysis-step.js +13 -2
  181. package/dist/orchestration/steps/generate-configs-step.d.ts +1 -0
  182. package/dist/orchestration/steps/generate-configs-step.js +77 -26
  183. package/dist/orchestration/steps/mirror-repo-tasks-step.d.ts +1 -1
  184. package/dist/orchestration/steps/mirror-repo-tasks-step.js +1 -1
  185. package/dist/orchestration/steps/publish-report-step.js +19 -0
  186. package/dist/orchestration/steps/readiness-step.js +8 -3
  187. package/dist/orchestration/steps/report-step.js +17 -4
  188. package/dist/orchestration/steps/run-eval-step.d.ts +1 -0
  189. package/dist/orchestration/steps/run-eval-step.js +51 -31
  190. package/dist/pipeline/agent-behavior-report.js +6 -0
  191. package/dist/pipeline/attribution.d.ts +1 -1
  192. package/dist/pipeline/attribution.js +1 -1
  193. package/dist/pipeline/cache.js +29 -15
  194. package/dist/pipeline/calculate-scores.d.ts +2 -0
  195. package/dist/pipeline/calculate-scores.js +70 -33
  196. package/dist/pipeline/chronic-failures.d.ts +55 -0
  197. package/dist/pipeline/chronic-failures.js +110 -0
  198. package/dist/pipeline/compiler/__tests__/agent-harness-handler.test.js +1 -1
  199. package/dist/pipeline/compiler/__tests__/knowledge-probe-handler.test.js +1 -1
  200. package/dist/pipeline/compiler/__tests__/literacy-handler.test.js +1 -1
  201. package/dist/pipeline/compiler/__tests__/mcp-server-handler.test.js +132 -62
  202. package/dist/pipeline/compiler/__tests__/promptfoo-compiler.test.js +2 -3
  203. package/dist/pipeline/compiler/__tests__/scoring-and-presets.test.js +33 -100
  204. package/dist/pipeline/compiler/__tests__/task-bridge.test.d.ts +9 -0
  205. package/dist/pipeline/compiler/__tests__/task-bridge.test.js +339 -0
  206. package/dist/pipeline/compiler/__tests__/tool-loop-openai.test.d.ts +10 -0
  207. package/dist/pipeline/compiler/__tests__/tool-loop-openai.test.js +509 -0
  208. package/dist/pipeline/compiler/assertion-mapper.d.ts +1 -1
  209. package/dist/pipeline/compiler/assertion-mapper.js +1 -1
  210. package/dist/pipeline/compiler/compiler-to-yaml.d.ts +2 -7
  211. package/dist/pipeline/compiler/compiler-to-yaml.js +2 -7
  212. package/dist/pipeline/compiler/config-loader.d.ts +14 -0
  213. package/dist/pipeline/compiler/config-loader.js +42 -2
  214. package/dist/pipeline/compiler/fixture-resolver.d.ts +1 -1
  215. package/dist/pipeline/compiler/fixture-resolver.js +1 -1
  216. package/dist/pipeline/compiler/ignore-fields.d.ts +1 -1
  217. package/dist/pipeline/compiler/ignore-fields.js +1 -1
  218. package/dist/pipeline/compiler/index.d.ts +2 -5
  219. package/dist/pipeline/compiler/index.js +2 -5
  220. package/dist/pipeline/compiler/literacy-bridge.d.ts +2 -2
  221. package/dist/pipeline/compiler/literacy-bridge.js +2 -2
  222. package/dist/pipeline/compiler/mode-bases/agent-harness.d.ts +10 -0
  223. package/dist/pipeline/compiler/mode-bases/agent-harness.js +21 -0
  224. package/dist/pipeline/compiler/mode-bases/index.d.ts +4 -0
  225. package/dist/pipeline/compiler/mode-bases/index.js +4 -0
  226. package/dist/pipeline/compiler/mode-bases/knowledge-probe.d.ts +10 -0
  227. package/dist/pipeline/compiler/mode-bases/knowledge-probe.js +22 -0
  228. package/dist/pipeline/compiler/mode-bases/literacy.d.ts +23 -0
  229. package/dist/pipeline/compiler/mode-bases/literacy.js +132 -0
  230. package/dist/pipeline/compiler/mode-bases/mcp-server.d.ts +10 -0
  231. package/dist/pipeline/compiler/mode-bases/mcp-server.js +70 -0
  232. package/dist/pipeline/compiler/mode-handlers/agent-harness/assertions.d.ts +43 -0
  233. package/dist/pipeline/compiler/mode-handlers/agent-harness/assertions.js +187 -0
  234. package/dist/pipeline/compiler/mode-handlers/agent-harness/compiler.d.ts +19 -0
  235. package/dist/pipeline/compiler/mode-handlers/agent-harness/compiler.js +138 -0
  236. package/dist/pipeline/compiler/mode-handlers/agent-harness/index.d.ts +16 -0
  237. package/dist/pipeline/compiler/mode-handlers/agent-harness/index.js +43 -0
  238. package/dist/pipeline/compiler/mode-handlers/agent-harness/prompts.d.ts +9 -0
  239. package/dist/pipeline/compiler/mode-handlers/agent-harness/prompts.js +29 -0
  240. package/dist/pipeline/compiler/mode-handlers/agent-harness/sandbox.d.ts +12 -0
  241. package/dist/pipeline/compiler/mode-handlers/agent-harness/sandbox.js +82 -0
  242. package/dist/pipeline/compiler/mode-handlers/agent-harness/tool-presets.d.ts +4 -0
  243. package/dist/pipeline/compiler/mode-handlers/agent-harness/tool-presets.js +19 -0
  244. package/dist/pipeline/compiler/mode-handlers/agent-harness/types.d.ts +49 -0
  245. package/dist/pipeline/compiler/mode-handlers/agent-harness/types.js +4 -0
  246. package/dist/pipeline/compiler/mode-handlers/agent-harness/validation.d.ts +9 -0
  247. package/dist/pipeline/compiler/mode-handlers/agent-harness/validation.js +16 -0
  248. package/dist/pipeline/compiler/mode-handlers/index.d.ts +6 -7
  249. package/dist/pipeline/compiler/mode-handlers/index.js +6 -8
  250. package/dist/pipeline/compiler/mode-handlers/knowledge-probe/assertions.d.ts +16 -0
  251. package/dist/pipeline/compiler/mode-handlers/knowledge-probe/assertions.js +61 -0
  252. package/dist/pipeline/compiler/mode-handlers/knowledge-probe/compiler.d.ts +18 -0
  253. package/dist/pipeline/compiler/mode-handlers/knowledge-probe/compiler.js +112 -0
  254. package/dist/pipeline/compiler/mode-handlers/knowledge-probe/index.d.ts +26 -0
  255. package/dist/pipeline/compiler/mode-handlers/knowledge-probe/index.js +49 -0
  256. package/dist/pipeline/compiler/mode-handlers/knowledge-probe/prompts.d.ts +9 -0
  257. package/dist/pipeline/compiler/mode-handlers/knowledge-probe/prompts.js +28 -0
  258. package/dist/pipeline/compiler/mode-handlers/knowledge-probe/types.d.ts +44 -0
  259. package/dist/pipeline/compiler/mode-handlers/knowledge-probe/types.js +4 -0
  260. package/dist/pipeline/compiler/mode-handlers/knowledge-probe/validation.d.ts +9 -0
  261. package/dist/pipeline/compiler/mode-handlers/knowledge-probe/validation.js +24 -0
  262. package/dist/pipeline/compiler/mode-handlers/literacy/assertions.d.ts +18 -0
  263. package/dist/pipeline/compiler/mode-handlers/literacy/assertions.js +118 -0
  264. package/dist/pipeline/compiler/mode-handlers/literacy/compiler.d.ts +14 -0
  265. package/dist/pipeline/compiler/mode-handlers/literacy/compiler.js +105 -0
  266. package/dist/pipeline/compiler/mode-handlers/literacy/index.d.ts +11 -0
  267. package/dist/pipeline/compiler/mode-handlers/literacy/index.js +38 -0
  268. package/dist/pipeline/compiler/mode-handlers/literacy/prompts.d.ts +9 -0
  269. package/dist/pipeline/compiler/mode-handlers/literacy/prompts.js +74 -0
  270. package/dist/pipeline/compiler/mode-handlers/literacy/types.d.ts +41 -0
  271. package/dist/pipeline/compiler/mode-handlers/literacy/types.js +4 -0
  272. package/dist/pipeline/compiler/mode-handlers/literacy/validation.d.ts +12 -0
  273. package/dist/pipeline/compiler/mode-handlers/literacy/validation.js +28 -0
  274. package/dist/pipeline/compiler/mode-handlers/mcp-assertions.js +63 -6
  275. package/dist/pipeline/compiler/mode-handlers/mcp-server/assertions.d.ts +42 -0
  276. package/dist/pipeline/compiler/mode-handlers/mcp-server/assertions.js +334 -0
  277. package/dist/pipeline/compiler/mode-handlers/mcp-server/compiler.d.ts +19 -0
  278. package/dist/pipeline/compiler/mode-handlers/mcp-server/compiler.js +100 -0
  279. package/dist/pipeline/compiler/mode-handlers/mcp-server/index.d.ts +27 -0
  280. package/dist/pipeline/compiler/mode-handlers/mcp-server/index.js +54 -0
  281. package/dist/pipeline/compiler/mode-handlers/mcp-server/prompts.d.ts +8 -0
  282. package/dist/pipeline/compiler/mode-handlers/mcp-server/prompts.js +28 -0
  283. package/dist/pipeline/compiler/mode-handlers/mcp-server/provider-config.d.ts +28 -0
  284. package/dist/pipeline/compiler/mode-handlers/mcp-server/provider-config.js +108 -0
  285. package/dist/pipeline/compiler/mode-handlers/mcp-server/types.d.ts +37 -0
  286. package/dist/pipeline/compiler/mode-handlers/mcp-server/types.js +4 -0
  287. package/dist/pipeline/compiler/mode-handlers/mcp-server/validation.d.ts +9 -0
  288. package/dist/pipeline/compiler/mode-handlers/mcp-server/validation.js +43 -0
  289. package/dist/pipeline/compiler/mode-handlers/mcp-server-handler.d.ts +3 -1
  290. package/dist/pipeline/compiler/mode-handlers/mcp-server-handler.js +65 -67
  291. package/dist/pipeline/compiler/mode-handlers/mcp-tool-provider/index.d.ts +33 -0
  292. package/dist/pipeline/compiler/mode-handlers/mcp-tool-provider/index.js +191 -0
  293. package/dist/pipeline/compiler/mode-handlers/mcp-tool-provider/mcp-connection.d.ts +19 -0
  294. package/dist/pipeline/compiler/mode-handlers/mcp-tool-provider/mcp-connection.js +101 -0
  295. package/dist/pipeline/compiler/mode-handlers/mcp-tool-provider/tool-loop-anthropic.d.ts +19 -0
  296. package/dist/pipeline/compiler/mode-handlers/mcp-tool-provider/tool-loop-anthropic.js +172 -0
  297. package/dist/pipeline/compiler/mode-handlers/mcp-tool-provider/tool-loop-openai.d.ts +19 -0
  298. package/dist/pipeline/compiler/mode-handlers/mcp-tool-provider/tool-loop-openai.js +323 -0
  299. package/dist/pipeline/compiler/mode-handlers/mcp-tool-provider/types.d.ts +103 -0
  300. package/dist/pipeline/compiler/mode-handlers/mcp-tool-provider/types.js +4 -0
  301. package/dist/pipeline/compiler/mode-handlers/mcp-tool-provider.d.ts +65 -0
  302. package/dist/pipeline/compiler/mode-handlers/mcp-tool-provider.js +368 -0
  303. package/dist/pipeline/compiler/preset-loader.d.ts +22 -0
  304. package/dist/pipeline/compiler/preset-loader.js +99 -0
  305. package/dist/pipeline/compiler/presets/sanity-literacy.d.ts +7 -10
  306. package/dist/pipeline/compiler/presets/sanity-literacy.js +11 -157
  307. package/dist/pipeline/compiler/promptfoo-compiler.d.ts +1 -4
  308. package/dist/pipeline/compiler/promptfoo-compiler.js +3 -12
  309. package/dist/pipeline/compiler/provider-assembler.js +13 -7
  310. package/dist/pipeline/compiler/sandbox/docker-sandbox.d.ts +1 -1
  311. package/dist/pipeline/compiler/sandbox/docker-sandbox.js +1 -1
  312. package/dist/pipeline/compiler/sandbox/fixture-provisioner.d.ts +1 -1
  313. package/dist/pipeline/compiler/sandbox/fixture-provisioner.js +1 -1
  314. package/dist/pipeline/compiler/sandbox/git-worktree-sandbox.d.ts +1 -1
  315. package/dist/pipeline/compiler/sandbox/git-worktree-sandbox.js +1 -1
  316. package/dist/pipeline/compiler/sandbox/index.d.ts +1 -1
  317. package/dist/pipeline/compiler/sandbox/index.js +1 -1
  318. package/dist/pipeline/compiler/sandbox/sandbox-selector.d.ts +1 -1
  319. package/dist/pipeline/compiler/sandbox/sandbox-selector.js +1 -1
  320. package/dist/pipeline/compiler/sandbox/sandbox-strategy.d.ts +1 -1
  321. package/dist/pipeline/compiler/sandbox/sandbox-strategy.js +1 -1
  322. package/dist/pipeline/compiler/sandbox/tempdir-sandbox.d.ts +1 -1
  323. package/dist/pipeline/compiler/sandbox/tempdir-sandbox.js +1 -1
  324. package/dist/pipeline/compiler/scoring-bridge.d.ts +1 -1
  325. package/dist/pipeline/compiler/scoring-bridge.js +1 -1
  326. package/dist/pipeline/compiler/task-bridge.d.ts +41 -0
  327. package/dist/pipeline/compiler/task-bridge.js +92 -0
  328. package/dist/pipeline/compiler/task-graph-builder.d.ts +1 -4
  329. package/dist/pipeline/compiler/task-graph-builder.js +1 -4
  330. package/dist/pipeline/compiler/telemetry/index.d.ts +1 -1
  331. package/dist/pipeline/compiler/telemetry/index.js +1 -1
  332. package/dist/pipeline/compiler/variable-resolver.d.ts +1 -1
  333. package/dist/pipeline/compiler/variable-resolver.js +1 -1
  334. package/dist/pipeline/coverage-audit.d.ts +1 -1
  335. package/dist/pipeline/coverage-audit.js +1 -1
  336. package/dist/pipeline/degradations.d.ts +1 -1
  337. package/dist/pipeline/degradations.js +1 -1
  338. package/dist/pipeline/expand-tasks.d.ts +2 -2
  339. package/dist/pipeline/expand-tasks.js +2 -2
  340. package/dist/pipeline/failure-modes.d.ts +1 -1
  341. package/dist/pipeline/failure-modes.js +13 -1
  342. package/dist/pipeline/gap-analysis.d.ts +1 -1
  343. package/dist/pipeline/gap-analysis.js +3 -1
  344. package/dist/pipeline/generate-configs.d.ts +2 -2
  345. package/dist/pipeline/generate-configs.js +16 -9
  346. package/dist/pipeline/grader-compare-runner.d.ts +1 -1
  347. package/dist/pipeline/grader-compare-runner.js +7 -1
  348. package/dist/pipeline/grader-comparison.d.ts +1 -1
  349. package/dist/pipeline/grader-comparison.js +1 -1
  350. package/dist/pipeline/grader-consistency-runner.d.ts +1 -1
  351. package/dist/pipeline/grader-consistency-runner.js +7 -1
  352. package/dist/pipeline/grader-consistency.d.ts +1 -1
  353. package/dist/pipeline/grader-consistency.js +1 -1
  354. package/dist/pipeline/grader-sensitivity-runner.d.ts +1 -1
  355. package/dist/pipeline/grader-sensitivity-runner.js +1 -1
  356. package/dist/pipeline/grader-sensitivity.d.ts +1 -1
  357. package/dist/pipeline/grader-sensitivity.js +1 -1
  358. package/dist/pipeline/grader-validate-runner.d.ts +1 -1
  359. package/dist/pipeline/grader-validate-runner.js +2 -2
  360. package/dist/pipeline/grader-validation.d.ts +1 -1
  361. package/dist/pipeline/grader-validation.js +1 -1
  362. package/dist/pipeline/map-request-to-config.js +16 -2
  363. package/dist/pipeline/mirror-repo-tasks.d.ts +8 -8
  364. package/dist/pipeline/mirror-repo-tasks.js +10 -10
  365. package/dist/pipeline/plan-format.d.ts +1 -1
  366. package/dist/pipeline/plan-format.js +1 -1
  367. package/dist/pipeline/plan.d.ts +1 -1
  368. package/dist/pipeline/plan.js +68 -30
  369. package/dist/pipeline/probe.d.ts +1 -1
  370. package/dist/pipeline/probe.js +1 -1
  371. package/dist/pipeline/readiness-report.d.ts +2 -2
  372. package/dist/pipeline/readiness-report.js +2 -2
  373. package/dist/pipeline/release-classification.d.ts +1 -1
  374. package/dist/pipeline/release-classification.js +1 -1
  375. package/dist/pipeline/release-report.d.ts +1 -1
  376. package/dist/pipeline/release-report.js +1 -1
  377. package/dist/pipeline/repo-eval-comment.d.ts +1 -1
  378. package/dist/pipeline/repo-eval-comment.js +1 -1
  379. package/dist/pipeline/repo-threshold-evaluator.d.ts +1 -1
  380. package/dist/pipeline/repo-threshold-evaluator.js +1 -1
  381. package/dist/pipeline/resolve-mappings.d.ts +6 -6
  382. package/dist/pipeline/resolve-mappings.js +44 -44
  383. package/dist/pipeline/retrieval-metrics.d.ts +3 -3
  384. package/dist/pipeline/retrieval-metrics.js +28 -20
  385. package/dist/pipeline/steps/calculate-scores-step.d.ts +11 -0
  386. package/dist/pipeline/steps/calculate-scores-step.js +89 -0
  387. package/dist/pipeline/steps/compare-step.d.ts +18 -0
  388. package/dist/pipeline/steps/compare-step.js +90 -0
  389. package/dist/pipeline/steps/eval-step.d.ts +53 -0
  390. package/dist/pipeline/steps/eval-step.js +347 -0
  391. package/dist/pipeline/steps/fetch-docs-step.d.ts +11 -0
  392. package/dist/pipeline/steps/fetch-docs-step.js +84 -0
  393. package/dist/pipeline/steps/generate-configs-step.d.ts +11 -0
  394. package/dist/pipeline/steps/generate-configs-step.js +98 -0
  395. package/dist/pipeline/steps/grader-consistency-step.d.ts +21 -0
  396. package/dist/pipeline/steps/grader-consistency-step.js +74 -0
  397. package/dist/pipeline/steps/publish-report-step.d.ts +57 -0
  398. package/dist/pipeline/steps/publish-report-step.js +243 -0
  399. package/dist/pipeline/steps/report-step.d.ts +13 -0
  400. package/dist/pipeline/steps/report-step.js +56 -0
  401. package/dist/pipeline/steps/update-scores-step.d.ts +11 -0
  402. package/dist/pipeline/steps/update-scores-step.js +42 -0
  403. package/dist/pipeline/targeted-loo.d.ts +1 -1
  404. package/dist/pipeline/targeted-loo.js +1 -1
  405. package/dist/pipeline/thresholds.d.ts +1 -1
  406. package/dist/pipeline/thresholds.js +1 -1
  407. package/dist/pipeline/validate.js +13 -0
  408. package/dist/report-store.d.ts +17 -0
  409. package/dist/report-store.js +24 -0
  410. package/dist/scripts/agent-behavior-report.d.ts +19 -0
  411. package/dist/scripts/agent-behavior-report.js +315 -0
  412. package/dist/scripts/baseline.d.ts +43 -0
  413. package/dist/scripts/baseline.js +267 -0
  414. package/dist/scripts/calculate-scores.d.ts +166 -0
  415. package/dist/scripts/calculate-scores.js +1296 -0
  416. package/dist/scripts/compare.d.ts +22 -0
  417. package/dist/scripts/compare.js +334 -0
  418. package/dist/scripts/coverage-audit.d.ts +44 -0
  419. package/dist/scripts/coverage-audit.js +209 -0
  420. package/dist/scripts/debug-eval.d.ts +19 -0
  421. package/dist/scripts/debug-eval.js +73 -0
  422. package/dist/scripts/discovery-report.d.ts +58 -0
  423. package/dist/scripts/discovery-report.js +250 -0
  424. package/dist/scripts/fetch-docs.d.ts +35 -0
  425. package/dist/scripts/fetch-docs.js +472 -0
  426. package/dist/scripts/generate-configs.d.ts +66 -0
  427. package/dist/scripts/generate-configs.js +459 -0
  428. package/dist/scripts/grader-api.d.ts +27 -0
  429. package/dist/scripts/grader-api.js +206 -0
  430. package/dist/scripts/grader-compare.d.ts +22 -0
  431. package/dist/scripts/grader-compare.js +368 -0
  432. package/dist/scripts/grader-consistency.d.ts +20 -0
  433. package/dist/scripts/grader-consistency.js +313 -0
  434. package/dist/scripts/grader-sensitivity.d.ts +22 -0
  435. package/dist/scripts/grader-sensitivity.js +354 -0
  436. package/dist/scripts/grader-validate.d.ts +19 -0
  437. package/dist/scripts/grader-validate.js +267 -0
  438. package/dist/scripts/measure-retrieval.d.ts +10 -0
  439. package/dist/scripts/measure-retrieval.js +145 -0
  440. package/dist/scripts/migrate-task-mode.d.ts +1 -1
  441. package/dist/scripts/migrate-task-mode.js +1 -1
  442. package/dist/scripts/migrate-tasks-to-content-lake.d.ts +1 -1
  443. package/dist/scripts/migrate-tasks-to-content-lake.js +1 -1
  444. package/dist/scripts/pipeline.d.ts +76 -0
  445. package/dist/scripts/pipeline.js +1031 -0
  446. package/dist/scripts/pr-comment.d.ts +10 -0
  447. package/dist/scripts/pr-comment.js +510 -0
  448. package/dist/scripts/readiness-report.d.ts +88 -0
  449. package/dist/scripts/readiness-report.js +342 -0
  450. package/dist/scripts/update-quality-scores.d.ts +15 -0
  451. package/dist/scripts/update-quality-scores.js +184 -0
  452. package/dist/scripts/validate-task-sources.d.ts +1 -1
  453. package/dist/scripts/validate-task-sources.js +1 -1
  454. package/dist/scripts/validate.d.ts +13 -0
  455. package/dist/scripts/validate.js +79 -0
  456. package/dist/scripts/webhook-server.d.ts +26 -0
  457. package/dist/scripts/webhook-server.js +147 -0
  458. package/dist/scripts/weekly-digest.d.ts +24 -0
  459. package/dist/scripts/weekly-digest.js +144 -0
  460. package/dist/sinks/format-slack.d.ts +64 -0
  461. package/dist/sinks/format-slack.js +306 -0
  462. package/dist/sinks/slack-sink.d.ts +27 -0
  463. package/dist/sinks/slack-sink.js +78 -0
  464. package/dist/sinks/types.d.ts +1 -1
  465. package/dist/sinks/types.js +1 -1
  466. package/dist/sinks/webhook-sink.d.ts +19 -0
  467. package/dist/sinks/webhook-sink.js +50 -0
  468. package/dist/tasks/knowledge-probe/define-type-api.task.ts +66 -0
  469. package/dist/tasks/knowledge-probe/groq-projections.task.ts +62 -0
  470. package/dist/tasks/literacy/content-lake.task.ts +181 -0
  471. package/dist/tasks/literacy/frameworks.task.ts +129 -0
  472. package/dist/tasks/literacy/functions.task.ts +70 -0
  473. package/dist/tasks/literacy/groq.task.ts +259 -0
  474. package/dist/tasks/literacy/image-handling.task.ts +95 -0
  475. package/dist/tasks/literacy/nextjs-live.task.ts +76 -0
  476. package/dist/tasks/literacy/portable-text.task.ts +169 -0
  477. package/dist/tasks/literacy/studio-setup.task.ts +134 -0
  478. package/dist/tasks/literacy/visual-editing.task.ts +147 -0
  479. package/package.json +32 -24
  480. package/tasks/.expanded.agentic.yaml +280 -0
  481. package/tasks/.expanded.yaml +565 -0
  482. package/tasks/knowledge-probe/define-type-api.task.ts +11 -0
  483. package/tasks/knowledge-probe/groq-projections.task.ts +3 -0
  484. package/tasks/literacy/content-lake.task.ts +181 -0
  485. package/tasks/literacy/frameworks.task.ts +1 -0
  486. package/tasks/literacy/functions.task.ts +1 -0
  487. package/tasks/literacy/groq.task.ts +1 -0
  488. package/tasks/literacy/image-handling.task.ts +95 -0
  489. package/tasks/literacy/nextjs-live.task.ts +2 -1
  490. package/tasks/literacy/portable-text.task.ts +169 -0
  491. package/tasks/literacy/studio-setup.task.ts +5 -2
  492. package/tasks/literacy/visual-editing.task.ts +1 -0
  493. package/LICENSE +0 -21
  494. package/tasks/frameworks.yaml +0 -98
  495. package/tasks/functions.yaml +0 -51
  496. package/tasks/groq.yaml +0 -216
  497. package/tasks/nextjs-live.yaml +0 -62
  498. package/tasks/studio-setup.yaml +0 -111
  499. package/tasks/visual-editing.yaml +0 -120
@@ -0,0 +1,6 @@
1
+ /**
2
+ * noop-collector.test.ts — verifies the NoOpArtifactCollector is truly zero-cost.
3
+ *
4
+ * Run: npx tsx --test src/artifact-capture/__tests__/noop-collector.test.ts
5
+ */
6
+ export {};
@@ -0,0 +1,42 @@
1
+ /**
2
+ * noop-collector.test.ts — verifies the NoOpArtifactCollector is truly zero-cost.
3
+ *
4
+ * Run: npx tsx --test src/artifact-capture/__tests__/noop-collector.test.ts
5
+ */
6
+ import assert from "node:assert/strict";
7
+ import { describe, it } from "node:test";
8
+ import { NoOpArtifactCollector } from "../noop-collector.js";
9
+ describe("NoOpArtifactCollector", () => {
10
+ it("enabled returns false", () => {
11
+ const collector = new NoOpArtifactCollector();
12
+ assert.equal(collector.enabled, false);
13
+ });
14
+ it("extrasEnabled returns false", () => {
15
+ const collector = new NoOpArtifactCollector();
16
+ assert.equal(collector.extrasEnabled, false);
17
+ });
18
+ it("capture() is callable and returns void", () => {
19
+ const collector = new NoOpArtifactCollector();
20
+ const result = collector.capture("step", "type", { data: true });
21
+ assert.equal(result, undefined);
22
+ });
23
+ it("captureFile() is callable and returns void", () => {
24
+ const collector = new NoOpArtifactCollector();
25
+ const result = collector.captureFile("step", "type", "/some/path");
26
+ assert.equal(result, undefined);
27
+ });
28
+ it("flush() returns zero-count result", async () => {
29
+ const collector = new NoOpArtifactCollector();
30
+ const result = await collector.flush();
31
+ assert.equal(result.artifactCount, 0);
32
+ assert.equal(result.destination, "");
33
+ assert.equal(result.totalBytes, 0);
34
+ assert.equal(result.compressed, false);
35
+ });
36
+ it("flush() returns the same frozen object every time", async () => {
37
+ const collector = new NoOpArtifactCollector();
38
+ const a = await collector.flush();
39
+ const b = await collector.flush();
40
+ assert.equal(a, b);
41
+ });
42
+ });
@@ -0,0 +1,14 @@
1
+ /**
2
+ * No-op artifact collector — used when --capture is not set.
3
+ *
4
+ * All methods are constant-time stubs. Zero overhead on the
5
+ * default pipeline path.
6
+ */
7
+ import type { ArtifactCollector, CaptureFlushResult } from "../ports/artifact-collector.js";
8
+ export declare class NoOpArtifactCollector implements ArtifactCollector {
9
+ readonly enabled = false;
10
+ readonly extrasEnabled = false;
11
+ capture(): void;
12
+ captureFile(): void;
13
+ flush(): Promise<CaptureFlushResult>;
14
+ }
@@ -0,0 +1,25 @@
1
+ /**
2
+ * No-op artifact collector — used when --capture is not set.
3
+ *
4
+ * All methods are constant-time stubs. Zero overhead on the
5
+ * default pipeline path.
6
+ */
7
+ const EMPTY_RESULT = Object.freeze({
8
+ artifactCount: 0,
9
+ compressed: false,
10
+ destination: "",
11
+ totalBytes: 0,
12
+ });
13
+ export class NoOpArtifactCollector {
14
+ enabled = false;
15
+ extrasEnabled = false;
16
+ capture() {
17
+ // no-op
18
+ }
19
+ captureFile() {
20
+ // no-op
21
+ }
22
+ async flush() {
23
+ return EMPTY_RESULT;
24
+ }
25
+ }
@@ -31,6 +31,7 @@ import type { SchedulesFile } from "./schemas/schedules.js";
31
31
  import type { SinksFile } from "./schemas/sinks.js";
32
32
  import type { ModelsConfig } from "./types/index.js";
33
33
  import type { GeneralizedTaskDefinition } from "./types/generalized-task.js";
34
+ import type { ModeBase, PresetDefinition } from "./types/plugin-registry.js";
34
35
  /**
35
36
  * Define an AILF evaluation configuration.
36
37
  *
@@ -50,6 +51,12 @@ export declare function defineTask(task: GeneralizedTaskDefinition): Generalized
50
51
  * Define the model registry (models to evaluate and grader model).
51
52
  *
52
53
  * Used in `config/models.ts` for typed model configuration.
54
+ *
55
+ * Validates:
56
+ * - Every `modes` entry is a canonical eval mode name
57
+ * - Every `variants` key is a mode the model is enrolled in
58
+ *
59
+ * @throws {Error} On invalid mode names or mismatched variant keys
53
60
  */
54
61
  export declare function defineModels(models: ModelsConfig): ModelsConfig;
55
62
  /**
@@ -132,27 +139,23 @@ export interface PromptEntry {
132
139
  */
133
140
  export declare function definePrompts(prompts: PromptEntry[]): PromptEntry[];
134
141
  /**
135
- * A preset is a named bundle of configuration that can be referenced
136
- * by tasks or suites to apply shared settings.
142
+ * Define a domain preset targets a mode base and adds domain-specific config.
143
+ *
144
+ * A domain preset declares which mode it targets (by ID), provides domain
145
+ * configuration (sources, features, doc fetcher), and can optionally override
146
+ * mode base defaults (rubrics, scoring profiles, prompt templates).
147
+ *
148
+ * Used in `.ailf/preset.ts` or npm packages for typed preset authoring.
137
149
  */
138
- export interface PresetConfig {
139
- /** Preset name */
140
- name: string;
141
- /** Default providers for tasks using this preset */
142
- providers?: string[];
143
- /** Default rubric template */
144
- rubric?: string;
145
- /** Default assertion templates */
146
- assertions?: unknown[];
147
- /** Default options */
148
- options?: Record<string, unknown>;
149
- }
150
+ export declare function definePreset(preset: PresetDefinition): PresetDefinition;
150
151
  /**
151
- * Define reusable configuration presets.
152
+ * Define a mode base — shared evaluation methodology for a mode.
152
153
  *
153
- * Used in `config/presets.ts` for typed preset configuration.
154
+ * A mode base defines HOW you evaluate (rubrics, scoring, prompts)
155
+ * independently of WHAT you're evaluating. Domain presets target a mode
156
+ * base and inherit its defaults.
154
157
  */
155
- export declare function definePreset(preset: PresetConfig): PresetConfig;
158
+ export declare function defineModeBase(base: ModeBase): ModeBase;
156
159
  /**
157
160
  * A pricing table entry for cost estimation.
158
161
  */
@@ -25,6 +25,7 @@
25
25
  *
26
26
  * @see docs/design-docs/architecture-overhaul/typescript-configuration.md (canonical)
27
27
  */
28
+ import { CANONICAL_EVAL_MODES } from "../ailf-shared/index.js";
28
29
  // ---------------------------------------------------------------------------
29
30
  // Config-level helpers
30
31
  // ---------------------------------------------------------------------------
@@ -57,8 +58,36 @@ export function defineTask(task) {
57
58
  * Define the model registry (models to evaluate and grader model).
58
59
  *
59
60
  * Used in `config/models.ts` for typed model configuration.
61
+ *
62
+ * Validates:
63
+ * - Every `modes` entry is a canonical eval mode name
64
+ * - Every `variants` key is a mode the model is enrolled in
65
+ *
66
+ * @throws {Error} On invalid mode names or mismatched variant keys
60
67
  */
61
68
  export function defineModels(models) {
69
+ const validModes = new Set(CANONICAL_EVAL_MODES);
70
+ for (const model of models.models) {
71
+ if (model.modes) {
72
+ for (const mode of model.modes) {
73
+ if (!validModes.has(mode)) {
74
+ throw new Error(`Model "${model.label ?? model.id}": unknown mode "${mode}". ` +
75
+ `Valid modes: ${CANONICAL_EVAL_MODES.join(", ")}`);
76
+ }
77
+ }
78
+ }
79
+ if (model.variants) {
80
+ const enrolledModes = new Set(model.modes ?? []);
81
+ for (const variantMode of Object.keys(model.variants)) {
82
+ if (!enrolledModes.has(variantMode)) {
83
+ throw new Error(`Model "${model.label ?? model.id}": variants specified for ` +
84
+ `mode "${variantMode}" but model does not include "${variantMode}" ` +
85
+ `in its modes array. Add "${variantMode}" to modes or remove ` +
86
+ `the variants entry.`);
87
+ }
88
+ }
89
+ }
90
+ }
62
91
  return models;
63
92
  }
64
93
  // ---------------------------------------------------------------------------
@@ -132,14 +161,34 @@ export function defineSources(sources) {
132
161
  export function definePrompts(prompts) {
133
162
  return prompts;
134
163
  }
164
+ // ---------------------------------------------------------------------------
165
+ // Preset helpers
166
+ // ---------------------------------------------------------------------------
135
167
  /**
136
- * Define reusable configuration presets.
168
+ * Define a domain preset — targets a mode base and adds domain-specific config.
169
+ *
170
+ * A domain preset declares which mode it targets (by ID), provides domain
171
+ * configuration (sources, features, doc fetcher), and can optionally override
172
+ * mode base defaults (rubrics, scoring profiles, prompt templates).
137
173
  *
138
- * Used in `config/presets.ts` for typed preset configuration.
174
+ * Used in `.ailf/preset.ts` or npm packages for typed preset authoring.
139
175
  */
140
176
  export function definePreset(preset) {
141
177
  return preset;
142
178
  }
179
+ // ---------------------------------------------------------------------------
180
+ // Mode base helpers
181
+ // ---------------------------------------------------------------------------
182
+ /**
183
+ * Define a mode base — shared evaluation methodology for a mode.
184
+ *
185
+ * A mode base defines HOW you evaluate (rubrics, scoring, prompts)
186
+ * independently of WHAT you're evaluating. Domain presets target a mode
187
+ * base and inherit its defaults.
188
+ */
189
+ export function defineModeBase(base) {
190
+ return base;
191
+ }
143
192
  /**
144
193
  * Define a pricing table for cost estimation.
145
194
  *