mlx-bun 0.4.0 → 1.0.0

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 (601) hide show
  1. package/README.md +1453 -88
  2. package/THIRD_PARTY_NOTICES.md +882 -0
  3. package/bin/mlx-bun.convert.mjs +4 -0
  4. package/bin/mlx-bun.fuse.mjs +4 -0
  5. package/bin/mlx-bun.generate.mjs +4 -0
  6. package/bin/mlx-bun.lora.mjs +4 -0
  7. package/bin/mlx-bun.mjs +13 -21
  8. package/bin/mlx-bun.server.mjs +4 -0
  9. package/bin/mlx-bun.upload.mjs +4 -0
  10. package/dist/web/app.js +9584 -0
  11. package/package.json +50 -53
  12. package/src/cli/ane-install.ts +76 -0
  13. package/src/cli/args.ts +179 -0
  14. package/src/cli/engine-entry.ts +68 -0
  15. package/src/cli/inference.ts +214 -0
  16. package/src/cli/job-entry.ts +97 -0
  17. package/src/cli/library-host.ts +117 -0
  18. package/src/cli/main.ts +57 -0
  19. package/src/cli/memory-engine.ts +163 -0
  20. package/src/cli/memory.ts +36 -0
  21. package/src/cli/mlx-lm-aliases.ts +309 -0
  22. package/src/cli/model-choice.ts +33 -0
  23. package/src/cli/model-selection.ts +84 -0
  24. package/src/cli/module-host.ts +11 -0
  25. package/src/cli/module-verbs.ts +50 -0
  26. package/src/cli/numerical-policy.ts +12 -0
  27. package/src/cli/perplexity.ts +89 -0
  28. package/src/cli/serve-host.ts +423 -0
  29. package/src/cli/serve-isolated.ts +217 -0
  30. package/src/cli/serve-options.ts +106 -0
  31. package/src/cli/serve-state.ts +205 -0
  32. package/src/cli/serve.ts +500 -0
  33. package/src/cli/served-model-host.ts +102 -0
  34. package/src/cli/server-entry.ts +148 -0
  35. package/src/cli/serving-unit.ts +177 -0
  36. package/src/{tui.ts → cli/terminal.ts} +13 -8
  37. package/src/cli/worker-adapters.ts +34 -0
  38. package/src/cli/worker-entry.ts +272 -0
  39. package/src/cli/worker-unit.ts +179 -0
  40. package/src/engine/adapter-operation.ts +26 -0
  41. package/src/engine/cache-services.ts +114 -0
  42. package/src/engine/completion.ts +39 -0
  43. package/src/engine/generation-gateway.ts +358 -0
  44. package/src/engine/index.ts +59 -0
  45. package/src/engine/kv-budget.ts +120 -0
  46. package/src/engine/media-preparation.ts +232 -0
  47. package/src/engine/model-binding.ts +24 -0
  48. package/src/engine/model-host.ts +513 -0
  49. package/src/engine/model-serving.ts +89 -0
  50. package/src/{serve → engine}/preparation.ts +10 -8
  51. package/src/{serve → engine}/session-completion-engine.ts +6 -6
  52. package/src/engine/telemetry.ts +109 -0
  53. package/src/hub/downloads.ts +116 -0
  54. package/src/jobs/db.ts +16 -7
  55. package/src/jobs/events.ts +29 -0
  56. package/src/jobs/executable.ts +5 -0
  57. package/src/jobs/host.ts +51 -0
  58. package/src/jobs/in-process.ts +55 -0
  59. package/src/{contracts/jobs.ts → jobs/protocol.ts} +7 -9
  60. package/src/jobs/runner.ts +150 -231
  61. package/src/jobs/service.ts +75 -0
  62. package/src/jobs/sse.ts +59 -23
  63. package/src/jobs/worker-memory.ts +46 -0
  64. package/src/jobs/worker-process.ts +163 -0
  65. package/src/jobs/worker-supervisor.ts +258 -0
  66. package/src/modules.ts +58 -0
  67. package/src/publishing/catalog-hub.ts +18 -0
  68. package/src/publishing/credentials.ts +31 -0
  69. package/src/publishing/upload.ts +37 -0
  70. package/src/residency/leased-adapters.ts +21 -0
  71. package/src/residency/local-records.ts +16 -0
  72. package/src/residency/model-residency.ts +405 -0
  73. package/src/residency/record-index.ts +28 -0
  74. package/src/residency/resident-estimate.ts +50 -0
  75. package/src/{anthropic.ts → server/anthropic.ts} +3 -81
  76. package/src/server/cache-routes.ts +36 -0
  77. package/src/{serve → server}/chat-request.ts +12 -10
  78. package/src/{serve → server}/chat-stage.ts +27 -22
  79. package/src/{client.ts → server/client.ts} +28 -2
  80. package/src/{serve → server}/completion-executor.ts +42 -42
  81. package/src/server/decision-body.ts +55 -0
  82. package/src/server/decision-routes.ts +134 -0
  83. package/src/server/decision-wire.ts +228 -0
  84. package/src/{serve → server}/discovery-routes.ts +42 -83
  85. package/src/server/generated-token-history.ts +57 -0
  86. package/src/{serve → server}/http.ts +25 -13
  87. package/src/{serve → server}/inference-request.ts +5 -5
  88. package/src/server/job-routes.ts +20 -0
  89. package/src/server/media-prompt.ts +63 -0
  90. package/src/server/memory-completion-client.ts +46 -0
  91. package/src/server/model-routes.ts +192 -0
  92. package/src/{serve → server}/openai-wire.ts +3 -2
  93. package/src/{serve → server}/pipeline.ts +1 -1
  94. package/src/server/prompt-builder.ts +13 -0
  95. package/src/server/prompt-contracts.ts +17 -0
  96. package/src/server/proxy-routes.ts +187 -0
  97. package/src/server/publishing-routes.ts +41 -0
  98. package/src/{serve → server}/request-plan.ts +8 -8
  99. package/src/{serve → server}/request-prep.ts +82 -68
  100. package/src/{serve → server}/responses-client.ts +30 -13
  101. package/src/{responses.ts → server/responses.ts} +12 -93
  102. package/src/server/routes.ts +205 -0
  103. package/src/server/start.ts +72 -0
  104. package/src/server/status-routes.ts +108 -0
  105. package/src/{serve → server}/text-completion-stage.ts +25 -4
  106. package/src/server/text-prompt.ts +12 -0
  107. package/src/{serve → server}/token-streams.ts +36 -38
  108. package/src/server/worker-routes.ts +311 -0
  109. package/src/storage/paths.ts +61 -0
  110. package/src/web/assets.ts +37 -0
  111. package/src/web/{src → browser}/api.ts +5 -23
  112. package/src/web/browser/chat-host.ts +17 -0
  113. package/src/web/browser/installed-panels.ts +6 -0
  114. package/src/web/browser/main.ts +55 -0
  115. package/src/web/browser/markdown.ts +279 -0
  116. package/src/web/browser/memory-panel.ts +9 -0
  117. package/src/web/browser/palette.ts +64 -0
  118. package/src/web/browser/protocol.ts +13 -0
  119. package/src/web/browser/shell.ts +467 -0
  120. package/src/web/{src → browser}/status.ts +32 -36
  121. package/src/web/build.ts +51 -0
  122. package/src/web/public/app.html +1054 -0
  123. package/src/web/{sw.js → public/sw.js} +4 -3
  124. package/src/web/{vendor → public/vendor}/hljs-theme.css +2 -2
  125. package/src/web/{vendor → public/vendor}/hljs.js +2 -2
  126. package/THIRD_PARTY_LICENSES.md +0 -51
  127. package/docs/README.md +0 -94
  128. package/docs/archive/hlg-sampling.md +0 -350
  129. package/docs/archive/investigations/audio-input-plan.md +0 -356
  130. package/docs/archive/investigations/chunk-finetune-experiment.md +0 -84
  131. package/docs/archive/investigations/colibri-glm52-port.md +0 -1486
  132. package/docs/archive/investigations/colibri-oracle-pin.md +0 -1284
  133. package/docs/archive/investigations/curve-bisector-routing.md +0 -123
  134. package/docs/archive/investigations/decode-roofline-lookagain.md +0 -348
  135. package/docs/archive/investigations/diffusion-gemma-port.md +0 -602
  136. package/docs/archive/investigations/dspark-handoff.md +0 -224
  137. package/docs/archive/investigations/expert-offload-single-user-moe.md +0 -292
  138. package/docs/archive/investigations/faithful-l1-consolidation.md +0 -310
  139. package/docs/archive/investigations/hlg-sampling-investigation.md +0 -970
  140. package/docs/archive/investigations/indexshare-performance-spike.md +0 -357
  141. package/docs/archive/investigations/kernel-perf-review-2026-07.md +0 -171
  142. package/docs/archive/investigations/lab-build-journal.md +0 -157
  143. package/docs/archive/investigations/minicpm5-decode-megakernel.md +0 -489
  144. package/docs/archive/investigations/openjarvis-fit.md +0 -198
  145. package/docs/archive/investigations/openwiki-evaluation.md +0 -249
  146. package/docs/archive/investigations/optimization_plan.md +0 -346
  147. package/docs/archive/investigations/orpo-base-uf-experiment-and-directions.md +0 -120
  148. package/docs/archive/investigations/orpo-flash-cce-pin-leak.md +0 -167
  149. package/docs/archive/investigations/orpo-uf-testing-handoff.md +0 -126
  150. package/docs/archive/investigations/pareto-specialized-runtime-findings.md +0 -92
  151. package/docs/archive/investigations/pi-builtin-investigation.md +0 -308
  152. package/docs/archive/investigations/pre-colibri-stabilization.md +0 -329
  153. package/docs/archive/investigations/qwen38-27b-campaign-log.md +0 -1092
  154. package/docs/archive/investigations/segmented-backward-handoff.md +0 -148
  155. package/docs/archive/investigations/starter-model-port-handoff.md +0 -342
  156. package/docs/archive/investigations/steel-flash-cce-handoff.md +0 -489
  157. package/docs/archive/investigations/the-dreaming-handoff.md +0 -141
  158. package/docs/archive/investigations/trainer-validation-experiment.md +0 -173
  159. package/docs/archive/planning/chat-ui-vision-memory-first.md +0 -199
  160. package/docs/archive/planning/chat-ui-vision.md +0 -50
  161. package/docs/archive/planning/chunk-task-roadmap.md +0 -77
  162. package/docs/archive/planning/curve-sampler-research-plan.md +0 -331
  163. package/docs/archive/planning/journal.md +0 -316
  164. package/docs/archive/planning/memory-docs-and-dag-plan.md +0 -194
  165. package/docs/archive/planning/pi-integration-review.md +0 -178
  166. package/docs/archive/planning/release-notes-v0.0.10.md +0 -57
  167. package/docs/archive/planning/release-notes-v0.0.11.md +0 -239
  168. package/docs/archive/planning/release-notes-v0.0.12.md +0 -80
  169. package/docs/archive/planning/release-notes-v0.0.13.md +0 -77
  170. package/docs/archive/planning/release-notes-v0.0.9.md +0 -122
  171. package/docs/archive/planning/release-notes-v0.2.0.md +0 -49
  172. package/docs/archive/planning/release-notes-v0.3.0.md +0 -35
  173. package/docs/archive/planning/web-ui-pass-plan.md +0 -213
  174. package/docs/archive/planning/website-readme-pass-plan.md +0 -224
  175. package/docs/design/batching.md +0 -1841
  176. package/docs/design/decode-speed-program.md +0 -2349
  177. package/docs/design/dreaming-nightly-pipeline.md +0 -684
  178. package/docs/design/generic-model-support.md +0 -705
  179. package/docs/design/kv-cache.md +0 -824
  180. package/docs/design/orpo-training.md +0 -680
  181. package/docs/design/speculative-decoding.md +0 -1990
  182. package/docs/design/turboquant.md +0 -2212
  183. package/docs/design/unified-engine-frontier-plan.md +0 -1880
  184. package/docs/design/web-chat-redesign.md +0 -695
  185. package/docs/planning/PRODUCT_ROADMAP.md +0 -594
  186. package/docs/planning/ResearchTopics.md +0 -119
  187. package/docs/planning/release-notes-v0.4.0.md +0 -45
  188. package/docs/planning/video-series.md +0 -421
  189. package/docs/reference/benchmarks.md +0 -3613
  190. package/docs/reference/cli.md +0 -585
  191. package/docs/reference/distribution.md +0 -355
  192. package/docs/reference/environment.md +0 -398
  193. package/docs/reference/glossary.md +0 -39
  194. package/docs/reference/library-api.md +0 -674
  195. package/docs/reference/memory.md +0 -260
  196. package/docs/reference/models.md +0 -334
  197. package/docs/reference/server-api.md +0 -1775
  198. package/docs/reference/server-config.md +0 -846
  199. package/docs/reference/training.md +0 -462
  200. package/docs/reference/troubleshooting.md +0 -158
  201. package/src/admission-plan.ts +0 -18
  202. package/src/audio/conformer.ts +0 -647
  203. package/src/audio/decode.ts +0 -133
  204. package/src/audio/features.ts +0 -202
  205. package/src/audio/transcode.ts +0 -67
  206. package/src/backends/mlx/assistant-target.ts +0 -46
  207. package/src/backends/mlx/autoregressive-draft-rows.ts +0 -120
  208. package/src/backends/mlx/autoregressive.ts +0 -108
  209. package/src/backends/mlx/batch-group.ts +0 -1790
  210. package/src/backends/mlx/cache-attachment.ts +0 -67
  211. package/src/backends/mlx/cache-layout.ts +0 -31
  212. package/src/backends/mlx/checkpoint-state.ts +0 -53
  213. package/src/backends/mlx/continuation-persistence.ts +0 -106
  214. package/src/backends/mlx/continuation-request.ts +0 -43
  215. package/src/backends/mlx/continuation.ts +0 -77
  216. package/src/backends/mlx/diffusion.ts +0 -39
  217. package/src/backends/mlx/draft-target.ts +0 -32
  218. package/src/backends/mlx/evaluation.ts +0 -105
  219. package/src/backends/mlx/fill-append.ts +0 -45
  220. package/src/backends/mlx/full-kv-row-append.ts +0 -32
  221. package/src/backends/mlx/full-kv-row-donor.ts +0 -54
  222. package/src/backends/mlx/full-transitioning-kv-rows.ts +0 -65
  223. package/src/backends/mlx/gateway-binding.ts +0 -130
  224. package/src/backends/mlx/graph.ts +0 -38
  225. package/src/backends/mlx/kv-codec.ts +0 -9
  226. package/src/backends/mlx/kv-maintenance.ts +0 -133
  227. package/src/backends/mlx/kv-tensor-rows.ts +0 -212
  228. package/src/backends/mlx/memory-client.ts +0 -189
  229. package/src/backends/mlx/methods.ts +0 -90
  230. package/src/backends/mlx/model-prompt.ts +0 -164
  231. package/src/backends/mlx/model-serving.ts +0 -79
  232. package/src/backends/mlx/prefill-cohort.ts +0 -89
  233. package/src/backends/mlx/prefill-rows.ts +0 -200
  234. package/src/backends/mlx/prefill.ts +0 -33
  235. package/src/backends/mlx/projected-context-rows.ts +0 -72
  236. package/src/backends/mlx/request-policy.ts +0 -20
  237. package/src/backends/mlx/request-state-policy.ts +0 -41
  238. package/src/backends/mlx/rollback.ts +0 -89
  239. package/src/backends/mlx/row-sampling.ts +0 -49
  240. package/src/backends/mlx/serial-executor.ts +0 -242
  241. package/src/backends/mlx/speculative-group.ts +0 -359
  242. package/src/backends/mlx/speculative-round.ts +0 -161
  243. package/src/backends/mlx/speculative.ts +0 -114
  244. package/src/backends/mlx/state-rows.ts +0 -104
  245. package/src/backends/mlx/state-views.ts +0 -46
  246. package/src/backends/mlx/target-layout-capability.ts +0 -35
  247. package/src/backends/mlx/text-engine.ts +0 -73
  248. package/src/backends/mlx/transitioning-kv-rows.ts +0 -165
  249. package/src/chat-template.ts +0 -382
  250. package/src/cli.ts +0 -3168
  251. package/src/config.ts +0 -399
  252. package/src/contracts/completion.ts +0 -16
  253. package/src/contracts/execution.ts +0 -60
  254. package/src/contracts/generation.ts +0 -103
  255. package/src/contracts/host.ts +0 -7
  256. package/src/contracts/pi-provider.ts +0 -30
  257. package/src/contracts/pi.ts +0 -259
  258. package/src/contracts/prefix-cache.ts +0 -23
  259. package/src/contracts/resources.ts +0 -25
  260. package/src/contracts/sampling.ts +0 -23
  261. package/src/contracts/scheduling.ts +0 -45
  262. package/src/contracts/tsconfig.json +0 -14
  263. package/src/dataset/generators.ts +0 -1087
  264. package/src/dataset/index.ts +0 -12
  265. package/src/dataset/job.ts +0 -81
  266. package/src/dataset/llm.ts +0 -71
  267. package/src/dataset/registry.ts +0 -528
  268. package/src/diffusion/diffusion-generate.ts +0 -507
  269. package/src/download.ts +0 -505
  270. package/src/embed.ts +0 -77
  271. package/src/engine/admission.ts +0 -59
  272. package/src/engine/cancellation.ts +0 -38
  273. package/src/engine/engine.ts +0 -56
  274. package/src/engine/execution-coordinator.ts +0 -70
  275. package/src/engine/execution-plan.ts +0 -53
  276. package/src/engine/resources.ts +0 -53
  277. package/src/engine/scheduler.ts +0 -72
  278. package/src/engine/session.ts +0 -247
  279. package/src/eval/README.md +0 -130
  280. package/src/eval/ifeval.ts +0 -81
  281. package/src/eval/kl-prompts.ts +0 -29
  282. package/src/eval/kl.ts +0 -590
  283. package/src/eval/perplexity.ts +0 -164
  284. package/src/eval/quality-db.ts +0 -92
  285. package/src/eval/reference-logits.ts +0 -207
  286. package/src/eval/runner.ts +0 -124
  287. package/src/eval/score.ts +0 -49
  288. package/src/eval/tasks/bfcl.ts +0 -535
  289. package/src/eval/tasks/gsm8k.ts +0 -98
  290. package/src/eval/tasks/hashhop.ts +0 -139
  291. package/src/eval/tasks/humaneval.ts +0 -275
  292. package/src/eval/tasks/ifeval.ts +0 -525
  293. package/src/eval/tasks/mmlu.ts +0 -157
  294. package/src/eval/ultrafeedback.ts +0 -60
  295. package/src/evaldb.ts +0 -199
  296. package/src/expert-io.ts +0 -393
  297. package/src/expert-offload-build.ts +0 -77
  298. package/src/expert-offload.ts +0 -76
  299. package/src/expert-residency.ts +0 -1134
  300. package/src/expert-usage.ts +0 -415
  301. package/src/fill/echo-index.ts +0 -188
  302. package/src/fill/fill-session.ts +0 -452
  303. package/src/fill/proposal.ts +0 -68
  304. package/src/fill/schema-rows.ts +0 -506
  305. package/src/fill/tool-boundary.ts +0 -103
  306. package/src/fit.ts +0 -249
  307. package/src/generate.ts +0 -1346
  308. package/src/grammar.ts +0 -591
  309. package/src/harness-pi.ts +0 -184
  310. package/src/hf-push.ts +0 -628
  311. package/src/hub-rest.ts +0 -242
  312. package/src/index.ts +0 -63
  313. package/src/inference/denoising.ts +0 -17
  314. package/src/inference/draft-acceptance.ts +0 -41
  315. package/src/inference/draft-output.ts +0 -49
  316. package/src/inference/graph.ts +0 -33
  317. package/src/inference/prefill.ts +0 -39
  318. package/src/inference/rollback.ts +0 -7
  319. package/src/jobs/index.ts +0 -31
  320. package/src/jobs/job-entry.ts +0 -107
  321. package/src/jobs/task-client.ts +0 -19
  322. package/src/jobs/types.ts +0 -2
  323. package/src/kv-scheme.ts +0 -257
  324. package/src/kv-store.ts +0 -1099
  325. package/src/lab/README.md +0 -18
  326. package/src/lab/curve/curve-designer.html +0 -303
  327. package/src/lab/curve/curve-sampler.ts +0 -122
  328. package/src/lab/expert-trace/expert-trace.ts +0 -69
  329. package/src/lab/paged-kv/paged-kv-rows.ts +0 -83
  330. package/src/lab/paged-kv/paged-kv.ts +0 -270
  331. package/src/library.ts +0 -40
  332. package/src/lora.ts +0 -408
  333. package/src/md-modules.d.ts +0 -49
  334. package/src/media-fetch.ts +0 -442
  335. package/src/memory/article.ts +0 -558
  336. package/src/memory/chunk-validation.ts +0 -150
  337. package/src/memory/chunk.ts +0 -318
  338. package/src/memory/cluster.ts +0 -351
  339. package/src/memory/coverage.ts +0 -134
  340. package/src/memory/crosslink.ts +0 -0
  341. package/src/memory/db.ts +0 -253
  342. package/src/memory/entity.ts +0 -203
  343. package/src/memory/gate.ts +0 -187
  344. package/src/memory/history.ts +0 -152
  345. package/src/memory/ingest.ts +0 -335
  346. package/src/memory/model.ts +0 -183
  347. package/src/memory/normalize.ts +0 -671
  348. package/src/memory/parse.ts +0 -22
  349. package/src/memory/pipeline.ts +0 -275
  350. package/src/memory/prompts.ts +0 -65
  351. package/src/memory/query.ts +0 -334
  352. package/src/memory/reconcile.ts +0 -734
  353. package/src/memory/redirect.ts +0 -108
  354. package/src/memory/reindex.ts +0 -325
  355. package/src/memory/resolve.ts +0 -307
  356. package/src/memory/rest.ts +0 -245
  357. package/src/memory/route.ts +0 -272
  358. package/src/memory/schedule.ts +0 -165
  359. package/src/memory/stages.ts +0 -777
  360. package/src/memory/synthesize.ts +0 -1638
  361. package/src/memory/tools.ts +0 -749
  362. package/src/memory/vault.ts +0 -1194
  363. package/src/memory/wikify.ts +0 -1064
  364. package/src/mlx/array.ts +0 -388
  365. package/src/mlx/autograd.ts +0 -304
  366. package/src/mlx/checkpoint.ts +0 -149
  367. package/src/mlx/compile.ts +0 -168
  368. package/src/mlx/custom-vjp.ts +0 -188
  369. package/src/mlx/ffi.ts +0 -512
  370. package/src/mlx/materialize.ts +0 -23
  371. package/src/mlx/metal-kernel.ts +0 -159
  372. package/src/mlx/ops.ts +0 -1021
  373. package/src/mlx/quantized-rows.ts +0 -36
  374. package/src/mlx/token-bitmask.ts +0 -52
  375. package/src/mlx/turboquant-kv-decode.ts +0 -82
  376. package/src/mlx/turboquant-kv-inverse.ts +0 -122
  377. package/src/mlx/turboquant-ops.ts +0 -861
  378. package/src/mlx/turboquant-tables.ts +0 -237
  379. package/src/mmap.ts +0 -90
  380. package/src/model/artifact-identity.ts +0 -73
  381. package/src/model/batched-kv.ts +0 -235
  382. package/src/model/batched-mask.ts +0 -279
  383. package/src/model/batched-quant.ts +0 -212
  384. package/src/model/batched-quantized-kv.ts +0 -144
  385. package/src/model/batched-rotating-quant.ts +0 -353
  386. package/src/model/batched-rotating-state.ts +0 -149
  387. package/src/model/batched-rotating.ts +0 -338
  388. package/src/model/batched-row-storage.ts +0 -179
  389. package/src/model/batched-ssm.ts +0 -52
  390. package/src/model/batched-turboquant-kv.ts +0 -93
  391. package/src/model/compiled-decode.ts +0 -712
  392. package/src/model/decoded-kv-donor.ts +0 -12
  393. package/src/model/delayed-quantized-kv.ts +0 -72
  394. package/src/model/delayed-rotating-quantized-kv.ts +0 -93
  395. package/src/model/delayed-turboquant-kv.ts +0 -55
  396. package/src/model/diffusion-gemma.ts +0 -1026
  397. package/src/model/factory.ts +0 -189
  398. package/src/model/fingerprint.ts +0 -33
  399. package/src/model/flash-attention.ts +0 -694
  400. package/src/model/full-prefill-padding.ts +0 -56
  401. package/src/model/full-prefill-row.ts +0 -105
  402. package/src/model/gemma4-base.ts +0 -2351
  403. package/src/model/gemma4.ts +0 -1315
  404. package/src/model/generated/gemma4-12b.ts +0 -429
  405. package/src/model/generated/gemma4-26b.ts +0 -448
  406. package/src/model/generated/gemma4-e4b.ts +0 -717
  407. package/src/model/generated/index.ts +0 -18
  408. package/src/model/glm52-atlas.ts +0 -511
  409. package/src/model/glm52-cache.ts +0 -792
  410. package/src/model/glm52-config.ts +0 -262
  411. package/src/model/glm52-container.ts +0 -305
  412. package/src/model/glm52-coupling.ts +0 -388
  413. package/src/model/glm52-dsa.ts +0 -606
  414. package/src/model/glm52-expert-layout.ts +0 -253
  415. package/src/model/glm52-layout.ts +0 -152
  416. package/src/model/glm52-memory.ts +0 -500
  417. package/src/model/glm52-mla.ts +0 -792
  418. package/src/model/glm52-moe.ts +0 -223
  419. package/src/model/glm52-pilot.ts +0 -537
  420. package/src/model/glm52-reference.ts +0 -445
  421. package/src/model/glm52-residency.ts +0 -557
  422. package/src/model/glm52-resident-weights.ts +0 -135
  423. package/src/model/glm52-streamed-experts.ts +0 -453
  424. package/src/model/glm52-streamed-metal.ts +0 -610
  425. package/src/model/glm52-weights.ts +0 -503
  426. package/src/model/glm52.ts +0 -1132
  427. package/src/model/implementation.ts +0 -69
  428. package/src/model/kv-attention-view.ts +0 -67
  429. package/src/model/minicpm5.ts +0 -376
  430. package/src/model/profile.ts +0 -443
  431. package/src/model/qwen-append.ts +0 -12
  432. package/src/model/qwen-conv.ts +0 -52
  433. package/src/model/qwen3-delta-state.ts +0 -74
  434. package/src/model/qwen3-delta.ts +0 -451
  435. package/src/model/qwen3-moe.ts +0 -404
  436. package/src/model/qwen3-mrope.ts +0 -252
  437. package/src/model/qwen3.ts +0 -277
  438. package/src/model/qwen3_5-checkpoint.ts +0 -164
  439. package/src/model/qwen3_5.ts +0 -819
  440. package/src/model/rotating-kv-donor.ts +0 -35
  441. package/src/model/rotating-kv-layout.ts +0 -262
  442. package/src/model/rotating-row-transaction.ts +0 -35
  443. package/src/model/speculative-rotating-kv.ts +0 -101
  444. package/src/model/ssm-prefill-padding.ts +0 -39
  445. package/src/model/support.ts +0 -102
  446. package/src/model/trellis-balanced-scatter.ts +0 -51
  447. package/src/model/trellis-linear.ts +0 -672
  448. package/src/model/trellis-shared-m.ts +0 -151
  449. package/src/model/trellis-shared-scatter.ts +0 -144
  450. package/src/model/trellis-splitk-prefill.ts +0 -158
  451. package/src/model/trellis-tiled-prefill.ts +0 -136
  452. package/src/model/trellis-vector-expand.ts +0 -62
  453. package/src/model/trellis-wide-prefill.ts +0 -94
  454. package/src/model/turboquant-codec.ts +0 -83
  455. package/src/model/universal/archs.ts +0 -332
  456. package/src/model/universal/dense.ts +0 -612
  457. package/src/model/universal/modules.ts +0 -170
  458. package/src/model/universal/rope.ts +0 -381
  459. package/src/native/expert_io.c +0 -523
  460. package/src/native/frame_extract.swift +0 -140
  461. package/src/native-pack.ts +0 -162
  462. package/src/pi-provider.ts +0 -53
  463. package/src/pi-session.ts +0 -81
  464. package/src/pi-terminal.ts +0 -274
  465. package/src/pi-web.ts +0 -1932
  466. package/src/preflight.ts +0 -158
  467. package/src/prompt-cache.ts +0 -460
  468. package/src/quantize/allocator.ts +0 -627
  469. package/src/quantize/atomic-output.ts +0 -38
  470. package/src/quantize/calibration.ts +0 -181
  471. package/src/quantize/config-writer.ts +0 -145
  472. package/src/quantize/index.ts +0 -84
  473. package/src/quantize/job.ts +0 -162
  474. package/src/quantize/quantizer.ts +0 -666
  475. package/src/quantize/rotate.ts +0 -603
  476. package/src/quantize/safetensors-writer.ts +0 -249
  477. package/src/quantize/sensitivity.ts +0 -216
  478. package/src/quantize/trellis.ts +0 -586
  479. package/src/quantize/weight-transform.ts +0 -292
  480. package/src/registry.ts +0 -602
  481. package/src/runtime-config.ts +0 -81
  482. package/src/safetensors.ts +0 -158
  483. package/src/sampler.ts +0 -1011
  484. package/src/serve/admin-routes.ts +0 -184
  485. package/src/serve/assets/curve-terrain.html +0 -105
  486. package/src/serve/assets/training-inference-map.html +0 -575
  487. package/src/serve/aux-routes.ts +0 -132
  488. package/src/serve/batch-scheduler.ts +0 -8
  489. package/src/serve/checkpoint-identity.ts +0 -64
  490. package/src/serve/generation-gateway.ts +0 -496
  491. package/src/serve/isolate.ts +0 -609
  492. package/src/serve/kv-budget.ts +0 -14
  493. package/src/serve/lab-routes.ts +0 -341
  494. package/src/serve/lane-registry.ts +0 -45
  495. package/src/serve/model-admin-routes.ts +0 -155
  496. package/src/serve/model-binding.ts +0 -43
  497. package/src/serve/model-host.ts +0 -511
  498. package/src/serve/parent-application.ts +0 -56
  499. package/src/serve/prompt-response-trace.ts +0 -147
  500. package/src/serve/response-lease.ts +0 -24
  501. package/src/serve/session-search.ts +0 -326
  502. package/src/serve/static-routes.ts +0 -101
  503. package/src/server.ts +0 -1314
  504. package/src/spec/assistant-rows.ts +0 -92
  505. package/src/spec/assistant-source.ts +0 -126
  506. package/src/spec/cache-identity.ts +0 -22
  507. package/src/spec/deepspec-rows.ts +0 -29
  508. package/src/spec/deepspec-source.ts +0 -145
  509. package/src/spec/dflash-source.ts +0 -149
  510. package/src/spec/drafter.ts +0 -304
  511. package/src/spec/dspark/ab-stats.ts +0 -136
  512. package/src/spec/dspark/calibration.ts +0 -112
  513. package/src/spec/dspark/data-dflash.ts +0 -156
  514. package/src/spec/dspark/data.ts +0 -201
  515. package/src/spec/dspark/deepspec-module.ts +0 -776
  516. package/src/spec/dspark/generate-dflash.ts +0 -245
  517. package/src/spec/dspark/generate.ts +0 -290
  518. package/src/spec/dspark/loader.ts +0 -23
  519. package/src/spec/dspark/loss.ts +0 -213
  520. package/src/spec/dspark/module-dflash.ts +0 -560
  521. package/src/spec/dspark/module.ts +0 -540
  522. package/src/spec/dspark/quantize-drafter.ts +0 -91
  523. package/src/spec/dspark/sample.ts +0 -112
  524. package/src/spec/generate.ts +0 -276
  525. package/src/spec/glm52-mtp-graph.ts +0 -29
  526. package/src/spec/glm52-mtp-rows.ts +0 -132
  527. package/src/spec/glm52-mtp-source.ts +0 -317
  528. package/src/spec/glm52-mtp-state.ts +0 -33
  529. package/src/spec/ngram-source.ts +0 -204
  530. package/src/spec/projected-draft-rows.ts +0 -73
  531. package/src/spec/qwen-mtp-module.ts +0 -171
  532. package/src/spec/qwen-mtp-rows.ts +0 -219
  533. package/src/spec/qwen-mtp-source.ts +0 -383
  534. package/src/spec/qwen-mtp-state.ts +0 -30
  535. package/src/spec/serve-loop.ts +0 -585
  536. package/src/spec/source.ts +0 -215
  537. package/src/spec/two-model-rows.ts +0 -61
  538. package/src/spec/two-model.ts +0 -113
  539. package/src/ssd-cache.ts +0 -372
  540. package/src/ssd-durability.ts +0 -211
  541. package/src/tokenizer.ts +0 -335
  542. package/src/tool-approvals.ts +0 -131
  543. package/src/tool-call.ts +0 -695
  544. package/src/train/dataset.ts +0 -484
  545. package/src/train/diffusion-lora.ts +0 -216
  546. package/src/train/export.ts +0 -32
  547. package/src/train/flash-cce.ts +0 -1234
  548. package/src/train/forward.ts +0 -266
  549. package/src/train/fuse.ts +0 -243
  550. package/src/train/index.ts +0 -46
  551. package/src/train/job.ts +0 -205
  552. package/src/train/lora-params.ts +0 -254
  553. package/src/train/loss.ts +0 -1591
  554. package/src/train/merge.ts +0 -195
  555. package/src/train/optimizer.ts +0 -183
  556. package/src/train/prefix-shared.ts +0 -490
  557. package/src/train/rank.ts +0 -137
  558. package/src/train/segmented.ts +0 -1828
  559. package/src/train/steel-qmm-header.ts +0 -1461
  560. package/src/train/trainer.ts +0 -1221
  561. package/src/train/watch.ts +0 -542
  562. package/src/vision/diffusion-vision.ts +0 -620
  563. package/src/vision/embedder.ts +0 -168
  564. package/src/vision/preprocess.ts +0 -247
  565. package/src/vision/prompt.ts +0 -373
  566. package/src/vision/qwen3vl-preprocess.ts +0 -409
  567. package/src/vision/qwen3vl-prompt.ts +0 -196
  568. package/src/vision/qwen3vl-tower.ts +0 -555
  569. package/src/vision/siglip.ts +0 -599
  570. package/src/vision/video-frames.ts +0 -148
  571. package/src/web/app.html +0 -2143
  572. package/src/web/app.js +0 -5703
  573. package/src/web/skills/memory/SKILL.md +0 -120
  574. package/src/web/skills/web-research/SKILL.md +0 -46
  575. package/src/web/skills.ts +0 -97
  576. package/src/web/src/adapters-panel.ts +0 -288
  577. package/src/web/src/assistant.ts +0 -451
  578. package/src/web/src/chat.ts +0 -1057
  579. package/src/web/src/composer.ts +0 -959
  580. package/src/web/src/dataset.ts +0 -132
  581. package/src/web/src/finetune.ts +0 -267
  582. package/src/web/src/hub.ts +0 -428
  583. package/src/web/src/main.ts +0 -98
  584. package/src/web/src/markdown.ts +0 -0
  585. package/src/web/src/memory-panel.ts +0 -572
  586. package/src/web/src/model-picker.ts +0 -159
  587. package/src/web/src/palette.ts +0 -346
  588. package/src/web/src/protocol.ts +0 -26
  589. package/src/web/src/quantize.ts +0 -167
  590. package/src/web/src/rag.ts +0 -274
  591. package/src/web/src/sessions.ts +0 -469
  592. package/src/web/src/shell.ts +0 -804
  593. package/src/web/src/ui-catalog.ts +0 -78
  594. package/src/web/vendor/README.md +0 -61
  595. package/src/web-assets.ts +0 -47
  596. package/src/web-tools.ts +0 -478
  597. package/src/weights.ts +0 -255
  598. /package/src/{serve → server}/completion-sink.ts +0 -0
  599. /package/src/web/{icon.svg → public/icon.svg} +0 -0
  600. /package/src/web/{manifest.webmanifest → public/manifest.webmanifest} +0 -0
  601. /package/src/web/{vendor → public/vendor}/hljs-LICENSE +0 -0
@@ -1,777 +0,0 @@
1
- // mlx-bun memory — "The Dreaming" STAGE WORKERS (pipeline decomposition).
2
- //
3
- // The monolithic runPipeline (segment-ALL → extract-ALL → route → create, one
4
- // blocking call) is decomposed here into four INDEPENDENT, DB-state-keyed stage
5
- // workers — exactly the split Lucien made between chunk-recent / cluster-assign /
6
- // synthesize. Each worker:
7
- // • pulls its eligible work from the DB *by state* (a watermark / the absence
8
- // of a downstream row), so it is RESUMABLE — an interrupted run continues;
9
- // • processes a bounded batch and PERSISTS, so two of them can run as separate
10
- // concurrent processes on different slices;
11
- // • is individually runnable (`mlx-bun memory segment|extract|route|synthesize`).
12
- //
13
- // CHRONOLOGY IS LOAD-BEARING. Every stage walks oldest-conversation-first
14
- // (conversations.updated_at ASC, then intra-conversation chunks.start ASC).
15
- // `messages` carry only a position, so conv-level updated_at + start is the
16
- // chronological key. This is what makes self-healing correct: the synthesis chunk
17
- // feed for an entity is chronological, so the user's LATEST statement dominates a
18
- // contradicted claim — never "whichever conversation UUID sorts last wins" (the
19
- // old ORDER BY chunk-id == conv-UUID, which is random w.r.t. time).
20
- //
21
- // runPipeline (pipeline.ts) now simply drives these four in sequence, so the full
22
- // DAG and the existing tests still pass.
23
-
24
- import { toEntityKind, type EntityKind } from "./article";
25
- import { chunkConversations, type ChunkCall, type ChunkResult } from "./chunk";
26
- import {
27
- firstSentences,
28
- routeSections,
29
- type SectionCall,
30
- type SectionRouteArticle,
31
- type SectionRouteChunk,
32
- } from "./cluster";
33
- import type { MemoryStore } from "./db";
34
- import { buildEntityPrompt, extractEntities, type ExtractCall } from "./entity";
35
- import { callLocalBatch, MAX_OUTPUT_TOKENS } from "./model";
36
- import type { SynthesisEvent } from "./pipeline";
37
- import { loadMetaPolicy } from "./prompts";
38
- import {
39
- EntityResolver,
40
- goldSeeds,
41
- loadDreamingGold,
42
- type DreamingGold,
43
- } from "./resolve";
44
- import {
45
- RouteAccumulator,
46
- engagesAsSubject,
47
- CAPTURED_BUCKET,
48
- type CreateStats,
49
- type RouteAction,
50
- } from "./route";
51
- import { reconcileArticle } from "./reconcile";
52
- import { registerRedirect, relatedArticleStem } from "./redirect";
53
- import {
54
- entityStem,
55
- synthesizeCreate,
56
- synthesizeNewSection,
57
- synthesizePatch,
58
- type SynthesisCall,
59
- } from "./synthesize";
60
- import { commitVault, listArticles, readArticle, vaultRoot } from "./vault";
61
-
62
- /** Default CREATE cap per synthesize pass (bounds a run's GPU cost). */
63
- export const DEFAULT_ARTICLE_CAP = 20;
64
-
65
- // ---- chronological selectors (shared) --------------------------------------
66
-
67
- /**
68
- * Every chunk id in CHRONOLOGICAL order — conversations.updated_at ASC, then
69
- * chunks.start ASC — optionally restricted to `convIds`. Also returns the
70
- * distinct conversation count. This replaces the old `ORDER BY chunks.id` (chunk
71
- * id == conv UUID, which sorts RANDOMLY w.r.t. time).
72
- */
73
- export function chronoChunkIds(
74
- store: MemoryStore,
75
- convIds?: string[],
76
- ): { convs: number; chunkIds: string[] } {
77
- const base =
78
- "SELECT chunks.id AS id, chunks.conv AS conv FROM chunks " +
79
- "JOIN conversations c ON c.conv = chunks.conv";
80
- const order = " ORDER BY c.updated_at ASC, chunks.start ASC";
81
- let rows: { id: string; conv: string }[];
82
- if (convIds && convIds.length) {
83
- const ph = convIds.map(() => "?").join(",");
84
- rows = store.db.query(`${base} WHERE chunks.conv IN (${ph})${order}`).all(...convIds) as {
85
- id: string;
86
- conv: string;
87
- }[];
88
- } else {
89
- rows = store.db.query(`${base}${order}`).all() as { id: string; conv: string }[];
90
- }
91
- return { convs: new Set(rows.map((r) => r.conv)).size, chunkIds: rows.map((r) => r.id) };
92
- }
93
-
94
- /**
95
- * Routed chunk ids for ONE entity in CHRONOLOGICAL order (conv.updated_at ASC,
96
- * chunks.start ASC), optionally restricted to a run's chunk set. The synthesis
97
- * feed rides this order so the latest statement dominates.
98
- */
99
- export function entityChunkIdsChrono(
100
- store: MemoryStore,
101
- entity: string,
102
- runChunks?: Set<string>,
103
- ): string[] {
104
- const rows = store.db
105
- .query(
106
- "SELECT ce.chunk_id AS id FROM chunk_entities ce " +
107
- "JOIN chunks ch ON ch.id = ce.chunk_id " +
108
- "JOIN conversations c ON c.conv = ch.conv " +
109
- "WHERE ce.entity_name = ? ORDER BY c.updated_at ASC, ch.start ASC",
110
- )
111
- .all(entity) as { id: string }[];
112
- const ids = rows.map((r) => r.id);
113
- return runChunks ? ids.filter((id) => runChunks.has(id)) : ids;
114
- }
115
-
116
- /** The earliest source-conversation updated_at across an entity's in-scope
117
- * chunks — the chronological sort key for ordering entities oldest-first. */
118
- function entityEarliestAt(store: MemoryStore, entity: string, runChunks?: Set<string>): number {
119
- const ids = entityChunkIdsChrono(store, entity, runChunks);
120
- if (ids.length === 0) return Number.POSITIVE_INFINITY;
121
- const row = store.db
122
- .query(
123
- "SELECT c.updated_at AS at FROM chunks ch JOIN conversations c ON c.conv = ch.conv WHERE ch.id = ?",
124
- )
125
- .get(ids[0]!) as { at: number | null } | null;
126
- return row?.at ?? Number.POSITIVE_INFINITY;
127
- }
128
-
129
- /** A chunk's label (the SECTION-ROUTE note label), or null. */
130
- function chunkLabel(store: MemoryStore, chunkId: string): string | null {
131
- const row = store.db.query("SELECT label FROM chunks WHERE id = ?").get(chunkId) as
132
- | { label: string | null }
133
- | null;
134
- return row?.label ?? null;
135
- }
136
-
137
- /** `loadMetaPolicy` that never throws — a missing Meta page yields "" so a
138
- * fixture/smoke vault without seeded Meta still runs. */
139
- function loadMetaPolicyQuiet(names: string[]): string {
140
- try {
141
- return loadMetaPolicy(names);
142
- } catch {
143
- return "";
144
- }
145
- }
146
-
147
- // ---- gold-derived entity metadata (kind + aliases) -------------------------
148
-
149
- /** Map every gold canonical → its declared kind + alias set, so a created
150
- * article carries the right `kind:` and seeds its `aliases:`. Non-gold entities
151
- * default to `thing`. (Lives here so the synthesize stage can stand alone.) */
152
- export interface EntityMeta {
153
- kindByCanonical: Map<string, EntityKind>;
154
- aliasesByCanonical: Map<string, string[]>;
155
- }
156
-
157
- export function buildEntityMeta(gold: DreamingGold = loadDreamingGold()): EntityMeta {
158
- const kindByCanonical = new Map<string, EntityKind>();
159
- const aliasesByCanonical = new Map<string, string[]>();
160
- const addAliases = (canonical: string, aliases: string[]): void => {
161
- const set = new Set(aliasesByCanonical.get(canonical) ?? []);
162
- for (const a of aliases) set.add(a);
163
- aliasesByCanonical.set(canonical, [...set]);
164
- };
165
- for (const g of gold.variantGroups) {
166
- if (!kindByCanonical.has(g.canonical)) kindByCanonical.set(g.canonical, toEntityKind(g.kind));
167
- addAliases(g.canonical, g.variants);
168
- }
169
- for (const n of gold.notableEntities) {
170
- if (!kindByCanonical.has(n.name)) kindByCanonical.set(n.name, toEntityKind(n.kind));
171
- addAliases(n.name, n.aliases);
172
- }
173
- return { kindByCanonical, aliasesByCanonical };
174
- }
175
-
176
- // ---- CREATE-target selection (pure) ----------------------------------------
177
-
178
- export interface RouteDecision {
179
- entity: string;
180
- action: RouteAction;
181
- stats: CreateStats;
182
- }
183
-
184
- /**
185
- * Pick which `create` decisions actually become articles this run: rank by
186
- * routed-chunk count (then name) and cap, but ALWAYS keep the `mustInclude`
187
- * entities when they earned a `create`. Pure. Retained for callers/tests that
188
- * want the routed-count ranking; the chronological synthesize stage orders by
189
- * earliest-chunk time instead.
190
- */
191
- export function selectCreateTargets(
192
- decisions: RouteDecision[],
193
- cap: number,
194
- mustInclude: Iterable<string> = [],
195
- ): string[] {
196
- const creatable = decisions.filter((d) => d.action === "create");
197
- const byEntity = new Map(creatable.map((d) => [d.entity, d]));
198
- const must = [...mustInclude].filter((e) => byEntity.has(e));
199
- const ranked = creatable
200
- .slice()
201
- .sort((a, b) => b.stats.routedChunks - a.stats.routedChunks || a.entity.localeCompare(b.entity))
202
- .map((d) => d.entity);
203
- const out: string[] = [];
204
- const seen = new Set<string>();
205
- for (const e of [...must, ...ranked]) {
206
- if (seen.has(e)) continue;
207
- seen.add(e);
208
- out.push(e);
209
- if (out.length >= cap) break;
210
- }
211
- return out;
212
- }
213
-
214
- // ---- result shapes (shared with pipeline.ts) -------------------------------
215
-
216
- export interface PipelineCreated {
217
- stem: string;
218
- hasInfobox: boolean;
219
- citedSections: number;
220
- chunkEdges: number;
221
- }
222
-
223
- export interface PipelinePatched {
224
- stem: string;
225
- anchor: string;
226
- chunkId: string;
227
- footnote: number | null;
228
- }
229
-
230
- // ===========================================================================
231
- // STAGE 1 — SEGMENT
232
- // ===========================================================================
233
-
234
- export interface SegmentStageOptions {
235
- /** Restrict to these conversation ids (default: every eligible conversation). */
236
- convIds?: string[];
237
- /** Cap the number of conversations (re)segmented this pass. */
238
- limit?: number;
239
- /** Generation budget per conversation. */
240
- maxTokens?: number;
241
- /** Model-call seam (default `callLocal("chunk", …)`). Tests inject a fake. */
242
- call?: ChunkCall;
243
- onEvent?: (e: SynthesisEvent) => void;
244
- }
245
-
246
- /**
247
- * SEGMENT — (re)chunk every conversation whose `chunked_at` is NULL or stale
248
- * (`updated_at > chunked_at`), oldest-first, DELETE+rewrite per conv with OUR
249
- * adapter, and set `chunked_at`. Resumable by the watermark (a re-run is a
250
- * no-op). This is a thin, named wrapper over `chunkConversations` so SEGMENT is
251
- * an independently-runnable worker.
252
- */
253
- export async function runSegmentStage(
254
- store: MemoryStore,
255
- opts: SegmentStageOptions = {},
256
- ): Promise<ChunkResult> {
257
- const result = await chunkConversations(
258
- store,
259
- { convs: opts.convIds, limit: opts.limit, maxTokens: opts.maxTokens, call: opts.call },
260
- opts.onEvent,
261
- );
262
- opts.onEvent?.({
263
- type: "stage",
264
- stage: "segment",
265
- message: `segment: ${result.valid} conversation(s) segmented, ${result.chunks} chunk(s) written (${result.skipped} skipped, ${result.errored} errored)`,
266
- });
267
- return result;
268
- }
269
-
270
- // ===========================================================================
271
- // STAGE 2 — ENTITY-EXTRACT
272
- // ===========================================================================
273
-
274
- export interface ExtractStageOptions {
275
- /** Restrict to these conversation ids (default: all). */
276
- convIds?: string[];
277
- /** Cap the number of chunks extracted this pass (default: all pending). */
278
- limit?: number;
279
- /** Pre-rendered Entities policy (cached across passes); default loads it. */
280
- policy?: string;
281
- /** Model-call seam (default the local `entity` stage). Tests inject a fake. */
282
- call?: ExtractCall;
283
- onEvent?: (e: SynthesisEvent) => void;
284
- }
285
-
286
- export interface ExtractStageResult {
287
- /** Chunks extracted this pass. */
288
- extracted: number;
289
- /** Pending chunks still un-extracted after the cap (0 ⇒ stage drained). */
290
- remaining: number;
291
- }
292
-
293
- /** Chunks that have NO `chunk_entities` row yet, CHRONOLOGICAL — the extract
294
- * work queue (state-keyed: presence of a downstream row ⇒ done). */
295
- function pendingExtractChunkIds(store: MemoryStore, convIds?: string[]): string[] {
296
- const base =
297
- "SELECT chunks.id AS id FROM chunks JOIN conversations c ON c.conv = chunks.conv " +
298
- "WHERE NOT EXISTS (SELECT 1 FROM chunk_entities ce WHERE ce.chunk_id = chunks.id)";
299
- const order = " ORDER BY c.updated_at ASC, chunks.start ASC";
300
- if (convIds && convIds.length) {
301
- const ph = convIds.map(() => "?").join(",");
302
- return (store.db.query(`${base} AND chunks.conv IN (${ph})${order}`).all(...convIds) as {
303
- id: string;
304
- }[]).map((r) => r.id);
305
- }
306
- return (store.db.query(`${base}${order}`).all() as { id: string }[]).map((r) => r.id);
307
- }
308
-
309
- /**
310
- * ENTITY-EXTRACT — read each not-yet-extracted chunk (CHRONOLOGICAL: oldest
311
- * conversation first, then by position) and persist the entities it is about via
312
- * the store-seeded + gold-seeded resolver, so canonicals fold across the pass.
313
- * Resumable: a chunk with `chunk_entities` rows is skipped. Returns how many were
314
- * extracted and how many remain (so a bounded pass can be repeated to drain).
315
- */
316
- export async function runExtractStage(
317
- store: MemoryStore,
318
- opts: ExtractStageOptions = {},
319
- ): Promise<ExtractStageResult> {
320
- const policy = opts.policy ?? loadMetaPolicyQuiet(["Entities"]);
321
- const resolver = EntityResolver.fromStore(store);
322
- for (const s of goldSeeds()) resolver.register(s.canonical, s.aliases);
323
-
324
- const pending = pendingExtractChunkIds(store, opts.convIds);
325
- const cap = opts.limit ?? pending.length;
326
- const target = Math.min(cap, pending.length);
327
- opts.onEvent?.({
328
- type: "stage",
329
- stage: "extract",
330
- message: `extract: ${target} of ${pending.length} pending chunk(s) for the entities they are about`,
331
- });
332
-
333
- const targetIds = pending.slice(0, target);
334
-
335
- // BATCHED model path (the prime uniform-prompt win: one 128-tok call per chunk,
336
- // independent). Only when no `call` is injected (tests inject a string-matching
337
- // fake → keep the serial seam) — callLocalBatch itself falls back to a serial
338
- // bit-exact loop when batching is disabled, so this stays correct either way.
339
- // The model calls run in a batch; PARSE + RESOLVE + PERSIST stay sequential so
340
- // the shared resolver folds canonicals in chronological order.
341
- if (!opts.call && targetIds.length > 1) {
342
- const inputs = targetIds.map((id) => ({ user: buildEntityPrompt(store.chunkText(id), policy) }));
343
- const outputs = await callLocalBatch("entity", inputs, { maxTokens: MAX_OUTPUT_TOKENS });
344
- let extracted = 0;
345
- for (let i = 0; i < targetIds.length; i++) {
346
- const out = outputs[i] ?? "";
347
- await extractEntities(store, targetIds[i]!, { resolver, policy, call: async () => out });
348
- extracted++;
349
- if (extracted % 10 === 0) {
350
- opts.onEvent?.({ type: "log", message: ` extracted ${extracted}/${target} chunks` });
351
- }
352
- }
353
- return { extracted, remaining: pending.length - extracted };
354
- }
355
-
356
- // SERIAL fallback (injected `call`, single chunk, or batching disabled).
357
- let extracted = 0;
358
- for (const id of targetIds) {
359
- // Pass the (possibly empty) policy through: `|| undefined` made extractEntities
360
- // re-load Meta/Entities.md and THROW on a vault without it, while the batched
361
- // path above proceeds with "" — the serial path must match.
362
- await extractEntities(store, id, { resolver, policy, call: opts.call });
363
- extracted++;
364
- if (extracted % 10 === 0) {
365
- opts.onEvent?.({ type: "log", message: ` extracted ${extracted}/${target} chunks` });
366
- }
367
- }
368
- return { extracted, remaining: pending.length - extracted };
369
- }
370
-
371
- // ===========================================================================
372
- // STAGE 3 — ROUTE
373
- // ===========================================================================
374
-
375
- export interface RouteStageOptions {
376
- /** Restrict the routed chunk set to these conversation ids (default: all). */
377
- convIds?: string[];
378
- /** Run the subject-engagement model gate to promote a thin (single-chunk)
379
- * genuine-subject candidate to a `create` stub. Off by default (a recurring
380
- * subject already creates without it). */
381
- useSubjectGate?: boolean;
382
- onEvent?: (e: SynthesisEvent) => void;
383
- }
384
-
385
- export interface RouteStageResult {
386
- /** Per-entity create/capture/routed decisions (name-sorted). */
387
- decisions: RouteDecision[];
388
- /** Count of entities the gate marked create-eligible. */
389
- createEligible: number;
390
- /** Subjects captured (not yet articled) — their chunks land in `_captured`,
391
- * still retrievable by search, never dropped. */
392
- captured: string[];
393
- }
394
-
395
- /**
396
- * ROUTE — recompute the create/capture decisions over `chunk_entities` with the
397
- * deterministic `RouteAccumulator` + the CREATE gate (Bucketing.md), then PERSIST
398
- * the outcome so SYNTHESIZE can read it without recomputing: `entities.notable =
399
- * 1` for a create-eligible (or already-homed) entity, `0` for a captured one, and
400
- * the captured subjects' chunks into the reserved `_captured` bucket (still
401
- * searchable — capture is not a drop). Deterministic/CPU (only the optional
402
- * subject gate touches the GPU). Resumable/idempotent — re-running rewrites the
403
- * same flags.
404
- */
405
- export async function runRouteStage(
406
- store: MemoryStore,
407
- opts: RouteStageOptions = {},
408
- ): Promise<RouteStageResult> {
409
- const { chunkIds } = chronoChunkIds(store, opts.convIds);
410
- const chunkSet = new Set(chunkIds);
411
-
412
- const known = (
413
- store.db.query("SELECT name FROM entities WHERE article_stem IS NOT NULL").all() as {
414
- name: string;
415
- }[]
416
- ).map((r) => r.name);
417
- const acc = new RouteAccumulator(known);
418
-
419
- // Feed the fan-out CHRONOLOGICALLY (oldest conversation first).
420
- const rows = store.db
421
- .query(
422
- "SELECT ce.chunk_id AS chunk_id, ce.entity_name AS entity_name FROM chunk_entities ce " +
423
- "JOIN chunks ch ON ch.id = ce.chunk_id JOIN conversations c ON c.conv = ch.conv " +
424
- "ORDER BY c.updated_at ASC, ch.start ASC, ce.entity_name ASC",
425
- )
426
- .all() as { chunk_id: string; entity_name: string }[];
427
- for (const r of rows) {
428
- if (!chunkSet.has(r.chunk_id)) continue;
429
- acc.enqueue(r.entity_name, r.chunk_id);
430
- }
431
-
432
- // Optional model gate: promote a thin genuine-subject candidate to a `create`
433
- // stub (chronological scan). NO ownership test — pure subject engagement.
434
- if (opts.useSubjectGate) {
435
- for (const d of acc.decisions()) {
436
- if (d.action !== "capture") continue;
437
- for (const cid of entityChunkIdsChrono(store, d.entity, chunkSet)) {
438
- if (await engagesAsSubject(d.entity, store.chunkText(cid))) {
439
- acc.enqueue(d.entity, cid, { subjectEngagement: true });
440
- break;
441
- }
442
- }
443
- }
444
- }
445
-
446
- const decisions = acc.decisions() as RouteDecision[];
447
- const captured = decisions.filter((d) => d.action === "capture").map((d) => d.entity);
448
-
449
- // PERSIST: notable flag + capture chunks into the reserved _captured bucket
450
- // (still searchable — capture is not a drop).
451
- const setNotable = store.db.query("UPDATE entities SET notable = ? WHERE name = ?");
452
- const ensureBucket = store.db.query(
453
- "INSERT OR IGNORE INTO buckets (name, article_stem, description) VALUES (?, '', ?)",
454
- );
455
- const insChunkBucket = store.db.query(
456
- "INSERT OR IGNORE INTO chunk_buckets (chunk_id, bucket) VALUES (?, ?)",
457
- );
458
- const capturedChunks = new Map<string, string[]>();
459
- for (const e of captured) capturedChunks.set(e, entityChunkIdsChrono(store, e, chunkSet));
460
- const tx = store.db.transaction(() => {
461
- for (const d of decisions) {
462
- const notable = d.action === "create" || d.action === "routed" ? 1 : 0;
463
- setNotable.run(notable, d.entity);
464
- }
465
- if (captured.length) {
466
- ensureBucket.run(CAPTURED_BUCKET, "subjects captured but not yet articled — still retrievable by search");
467
- for (const [, cids] of capturedChunks) for (const cid of cids) insChunkBucket.run(cid, CAPTURED_BUCKET);
468
- }
469
- });
470
- tx();
471
-
472
- // REDIRECTS: every discussed subject must resolve to a home, even when it never
473
- // earns its own article. A subject FOLDED into an existing article (action
474
- // "routed") registers its surfaces → that article; a CAPTURED subject registers
475
- // its surfaces → the article it most co-occurs with (its neighbourhood), so a
476
- // later resolveName reaches its home instead of NOT FOUND. Idempotent; reindex
477
- // projects these into entity_aliases. Purely additive — does not affect routing.
478
- const stemOf = store.db.query("SELECT article_stem AS stem FROM entities WHERE name = ?");
479
- const surfacesOf = (entity: string): string[] =>
480
- (
481
- store.db
482
- .query("SELECT DISTINCT surface_form AS s FROM chunk_entities WHERE entity_name = ? AND surface_form IS NOT NULL")
483
- .all(entity) as { s: string }[]
484
- ).map((r) => r.s);
485
- const registerHome = (entity: string, home: string, kind: "fold" | "capture"): void => {
486
- if (!home) return;
487
- registerRedirect(store, entity, home, kind);
488
- for (const s of surfacesOf(entity)) registerRedirect(store, s, home, kind);
489
- };
490
- for (const d of decisions) {
491
- if (d.action === "routed") {
492
- const home = (stemOf.get(d.entity) as { stem: string | null } | null)?.stem;
493
- if (home) registerHome(d.entity, home, "fold");
494
- } else if (d.action === "capture") {
495
- const home = relatedArticleStem(store, d.entity, chunkSet);
496
- if (home) registerHome(d.entity, home, "capture");
497
- }
498
- }
499
-
500
- const createEligible = decisions.filter((d) => d.action === "create").length;
501
- opts.onEvent?.({
502
- type: "stage",
503
- stage: "route",
504
- message: `route: ${decisions.length} entities — ${createEligible} create, ${captured.length} captured`,
505
- });
506
- return { decisions, createEligible, captured };
507
- }
508
-
509
- // ===========================================================================
510
- // STAGE 4 — SYNTHESIZE (CREATE + PATCH)
511
- // ===========================================================================
512
-
513
- export interface SynthesizeStageOptions {
514
- /** Vault root (honors MLX_BUN_WIKI); defaults to `vaultRoot()`. */
515
- root?: string;
516
- /** Max articles to CREATE this pass (defaults to DEFAULT_ARTICLE_CAP). */
517
- limit?: number;
518
- /** Restrict the run to these conversation ids (default: all). */
519
- convIds?: string[];
520
- /** Entities to force into the CREATE set when create-eligible. */
521
- mustCreate?: string[];
522
- /** Model-call seam for CREATE drafting AND PATCH integration. */
523
- call?: SynthesisCall;
524
- /** Model-call seam for SECTION-ROUTE (which existing section a patch folds into). */
525
- sectionCall?: SectionCall;
526
- /** Skip the final git commit (tests). */
527
- commit?: boolean;
528
- now?: number;
529
- onEvent?: (e: SynthesisEvent) => void;
530
- }
531
-
532
- export interface SynthesizeStageResult {
533
- created: PipelineCreated[];
534
- patched: PipelinePatched[];
535
- skippedByGate: string[];
536
- /** Article stems a reconcile pass updated to stay consistent with the latest
537
- * position (infobox field refresh and/or stale sibling-section rewrite). */
538
- reconciled: string[];
539
- }
540
-
541
- /**
542
- * SYNTHESIZE — the write stage, CHRONOLOGICAL throughout.
543
- *
544
- * (a) CREATE: for each create-eligible entity (`notable = 1`) with NO article
545
- * file yet, draft via `synthesizeCreate` FEEDING ITS CHUNKS OLDEST-FIRST so
546
- * the latest statement dominates; entities are processed in order of their
547
- * EARLIEST chunk (oldest entity first), capped by `limit`.
548
- * (b) PATCH: for each chunk routed to an EXISTING article and not already in the
549
- * `synthesized_chunk_sections` ledger, SECTION-ROUTE then fold via
550
- * `synthesizePatch`, oldest chunk first (so a newer correction overwrites an
551
- * older take, not the reverse).
552
- * (c) RECONCILE: after the PATCH loop, for every article it touched, run
553
- * `reconcileArticle` so a fold that flipped the verdict also updates whatever
554
- * it made stale (infobox relationship fields + sibling sections still asserting
555
- * the old value). Bounded + idempotent — a consistent article is a NO-OP.
556
- *
557
- * Resumable via article presence + the ledger. Commits once at the end.
558
- */
559
- export async function runSynthesizeStage(
560
- store: MemoryStore,
561
- opts: SynthesizeStageOptions = {},
562
- ): Promise<SynthesizeStageResult> {
563
- const root = opts.root ?? vaultRoot();
564
- const cap = opts.limit ?? DEFAULT_ARTICLE_CAP;
565
- const emit = opts.onEvent ?? (() => {});
566
- const { chunkIds } = chronoChunkIds(store, opts.convIds);
567
- const runChunks = new Set(chunkIds);
568
-
569
- // Snapshot the EXISTING article files BEFORE drafting, so the create-vs-patch
570
- // split is stable (an entity created THIS run is not also patched this run).
571
- const existingStems = new Set(await listArticles(root));
572
- const dbStemByEntity = new Map(
573
- (
574
- store.db.query("SELECT name, article_stem FROM entities WHERE article_stem IS NOT NULL").all() as {
575
- name: string;
576
- article_stem: string;
577
- }[]
578
- ).map((r) => [r.name, r.article_stem] as const),
579
- );
580
- const existingArticleStem = (entity: string): string | null => {
581
- const dbStem = dbStemByEntity.get(entity);
582
- if (dbStem && existingStems.has(dbStem)) return dbStem;
583
- const s = entityStem(entity);
584
- return existingStems.has(s) ? s : null;
585
- };
586
-
587
- const meta = buildEntityMeta();
588
- const created: PipelineCreated[] = [];
589
- const skippedByGate: string[] = [];
590
-
591
- // ---- CREATE: notable=1, no article file, in scope, oldest entity first ----
592
- const convFilter = opts.convIds?.length ? `AND ch.conv IN (${opts.convIds.map(() => "?").join(",")})` : "";
593
- const createRows = store.db
594
- .query(
595
- "SELECT e.name AS name, MIN(c.updated_at) AS first_at FROM entities e " +
596
- "JOIN chunk_entities ce ON ce.entity_name = e.name " +
597
- "JOIN chunks ch ON ch.id = ce.chunk_id " +
598
- "JOIN conversations c ON c.conv = ch.conv " +
599
- `WHERE e.notable = 1 ${convFilter} ` +
600
- "GROUP BY e.name ORDER BY first_at ASC, e.name ASC",
601
- )
602
- .all(...(opts.convIds ?? [])) as { name: string; first_at: number }[];
603
- const eligible = createRows.map((r) => r.name).filter((n) => existingArticleStem(n) == null);
604
- const must = (opts.mustCreate ?? []).filter((n) => eligible.includes(n));
605
- const targets: string[] = [];
606
- const seen = new Set<string>();
607
- for (const e of [...must, ...eligible]) {
608
- if (seen.has(e)) continue;
609
- seen.add(e);
610
- targets.push(e);
611
- if (targets.length >= cap) break;
612
- }
613
- emit({ type: "stage", stage: "create", message: `create: drafting ${targets.length} entity article(s) (cap ${cap})` });
614
-
615
- for (const entity of targets) {
616
- const cids = entityChunkIdsChrono(store, entity, runChunks); // chronological feed
617
- const kind = meta.kindByCanonical.get(entity) ?? "thing";
618
- let res;
619
- try {
620
- res = await synthesizeCreate(store, {
621
- entity,
622
- kind,
623
- chunkIds: cids,
624
- root,
625
- aliases: meta.aliasesByCanonical.get(entity) ?? [],
626
- call: opts.call,
627
- commit: false,
628
- now: opts.now,
629
- });
630
- } catch (err) {
631
- skippedByGate.push(entity);
632
- emit({ type: "log", message: ` skipped ${entity} (error: ${String(err)})` });
633
- continue;
634
- }
635
- if (res.created) {
636
- created.push({ stem: res.stem, hasInfobox: res.hasInfobox, citedSections: res.citedSections, chunkEdges: cids.length });
637
- emit({ type: "log", message: ` created ${res.stem} (infobox=${res.hasInfobox}, cited=${res.citedSections})` });
638
- } else if (res.skippedByGate) {
639
- skippedByGate.push(res.stem);
640
- emit({ type: "log", message: ` skipped ${res.stem} (gate: ${res.reason})` });
641
- } else {
642
- emit({ type: "log", message: ` skipped ${res.stem} (${res.reason})` });
643
- }
644
- }
645
-
646
- // ---- SECTION-ROUTE → PATCH: existing-article entities, oldest chunk first ----
647
- const patchEntities = (
648
- opts.convIds?.length
649
- ? (store.db
650
- .query(
651
- `SELECT DISTINCT ce.entity_name AS name FROM chunk_entities ce JOIN chunks ch ON ch.id = ce.chunk_id WHERE ch.conv IN (${opts.convIds
652
- .map(() => "?")
653
- .join(",")})`,
654
- )
655
- .all(...opts.convIds) as { name: string }[])
656
- : (store.db.query("SELECT DISTINCT entity_name AS name FROM chunk_entities").all() as { name: string }[])
657
- ).map((r) => r.name);
658
-
659
- const patchTargets = patchEntities
660
- .map((entity) => ({ entity, stem: existingArticleStem(entity) }))
661
- .filter((t): t is { entity: string; stem: string } => t.stem != null)
662
- .sort(
663
- (a, b) =>
664
- entityEarliestAt(store, a.entity, runChunks) - entityEarliestAt(store, b.entity, runChunks) ||
665
- a.entity.localeCompare(b.entity),
666
- );
667
- emit({
668
- type: "stage",
669
- stage: "section-route",
670
- message: `section-route: ${patchTargets.length} existing-article entit${patchTargets.length === 1 ? "y" : "ies"} to fold into`,
671
- });
672
-
673
- const patched: PipelinePatched[] = [];
674
- let patchNoops = 0;
675
- for (const { entity, stem } of patchTargets) {
676
- for (const cid of entityChunkIdsChrono(store, entity, runChunks)) {
677
- let article: SectionRouteArticle;
678
- try {
679
- const { content } = await readArticle(root, stem);
680
- article = { stem, content };
681
- } catch {
682
- break; // article vanished mid-run
683
- }
684
- const chunk: SectionRouteChunk = {
685
- id: cid,
686
- label: chunkLabel(store, cid),
687
- gist: firstSentences(store.chunkText(cid), 2),
688
- };
689
- const routeRes = await routeSections(chunk, article, { call: opts.sectionCall });
690
- for (const anchor of routeRes.matchedAnchors) {
691
- const pr = await synthesizePatch(store, {
692
- stem,
693
- anchor,
694
- chunkId: cid,
695
- root,
696
- call: opts.call,
697
- commit: false,
698
- now: opts.now,
699
- });
700
- if (pr.patched) {
701
- patched.push({ stem, anchor, chunkId: cid, footnote: pr.footnote });
702
- emit({ type: "log", message: ` patched ${stem} #${anchor} (+chunk ${cid}, [^${pr.footnote}])` });
703
- } else if (!pr.alreadyIntegrated) {
704
- patchNoops++;
705
- emit({ type: "log", message: ` patch NO-OP ${stem} #${anchor} (${pr.reason})` });
706
- }
707
- }
708
- // No existing section fit but the chunk is substantive ⇒ SECTION-ROUTE
709
- // named a NEW section. HONOR it (mint heading + fold the chunk) instead of
710
- // dropping the chunk — silent loss is the self-healing bug we are fixing.
711
- if (routeRes.newSection && routeRes.matchedAnchors.length === 0) {
712
- const { title, anchor } = routeRes.newSection;
713
- const nr = await synthesizeNewSection(store, {
714
- stem,
715
- title,
716
- anchor,
717
- chunkId: cid,
718
- root,
719
- call: opts.call,
720
- commit: false,
721
- now: opts.now,
722
- });
723
- if (nr.patched) {
724
- patched.push({ stem, anchor, chunkId: cid, footnote: nr.footnote });
725
- emit({ type: "log", message: ` new section ${stem} #${anchor} (+chunk ${cid}, [^${nr.footnote}])` });
726
- } else if (!nr.alreadyIntegrated) {
727
- patchNoops++;
728
- emit({ type: "log", message: ` new-section NO-OP ${stem} #${anchor} (${nr.reason})` });
729
- }
730
- }
731
- }
732
- }
733
- emit({ type: "stage", stage: "patch", message: `patch: ${patched.length} section fold(s), ${patchNoops} NO-OP(s)` });
734
-
735
- // ---- RECONCILE: after a fold flips an article's verdict, update everything it
736
- // made stale (infobox relationship fields + sibling sections still asserting the
737
- // old value) so the article ends consistent with the latest position. Runs once
738
- // per article touched this run; a consistent article is a NO-OP.
739
- const reconciled: string[] = [];
740
- const touchedStems = [...new Set(patched.map((p) => p.stem))];
741
- for (const stem of touchedStems) {
742
- let rr;
743
- try {
744
- rr = await reconcileArticle(store, stem, { root, call: opts.call, commit: false, now: opts.now });
745
- } catch (err) {
746
- emit({ type: "log", message: ` reconcile error ${stem}: ${String(err)}` });
747
- continue;
748
- }
749
- if (rr.reconciled) {
750
- reconciled.push(stem);
751
- emit({
752
- type: "log",
753
- message: ` reconciled ${stem} (fields: ${rr.refreshedFields.join(", ") || "none"}; sections: ${rr.rewrittenSections.join(", ") || "none"}${rr.retried ? "; retried" : ""})`,
754
- });
755
- } else if (rr.gateVetoed) {
756
- emit({ type: "log", message: ` reconcile NO-OP ${stem} (gate: ${rr.reason})` });
757
- }
758
- if (rr.unresolved.length) {
759
- emit({
760
- type: "log",
761
- message: ` reconcile UNRESOLVED ${stem} — base-model left a stale present-tense assertion in: ${[...new Set(rr.unresolved)].join(", ")} (a memory-synthesis LoRA / P10-T1 would close it)`,
762
- });
763
- }
764
- }
765
- emit({ type: "stage", stage: "reconcile", message: `reconcile: ${reconciled.length} article(s) made consistent` });
766
-
767
- const wrote = created.length + patched.length + reconciled.length;
768
- if (opts.commit !== false && wrote) {
769
- await commitVault(
770
- root,
771
- `memory: synthesis (${created.length} created, ${patched.length} patched, ${reconciled.length} reconciled)`,
772
- );
773
- }
774
- emit({ type: "stage", stage: "commit", message: `commit: ${created.length} created + ${patched.length} patched + ${reconciled.length} reconciled written to ${root}` });
775
-
776
- return { created, patched, skippedByGate, reconciled };
777
- }