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
@@ -1,78 +0,0 @@
1
- // GENERATED-ADJACENT source module — part of the src/web/src/* split (plan
2
- // §7/§9 Phase 3, §6.6 "app-aware assistant"). Built into src/web/app.js by
3
- // scripts/build-web.ts.
4
- //
5
- // Shared route IDs keep browser navigation and server validation in agreement.
6
- // Spotlight selectors and overlay IDs belong to the browser.
7
-
8
- /** A hash-routed page (shell.ts's Route union, minus "routes" — the DAG
9
- * diagram tab isn't a useful assistant destination and is itself
10
- * feature-detected/hidden when unavailable). */
11
- import { APP_ROUTE_IDS as ROUTE_IDS } from "../../contracts/pi";
12
- export { ROUTE_IDS };
13
- export type RouteId = (typeof ROUTE_IDS)[number];
14
-
15
- /** A non-route overlay/panel the assistant can navigate to or spotlight
16
- * within. Opening one of these does not change location.hash. */
17
- export const VIEW_IDS = ["memory-panel", "hub-panel", "settings", "adapters-panel"] as const;
18
- export type ViewId = (typeof VIEW_IDS)[number];
19
-
20
- export function isRouteId(v: string): v is RouteId {
21
- return (ROUTE_IDS as readonly string[]).includes(v);
22
- }
23
- export function isViewId(v: string): v is ViewId {
24
- return (VIEW_IDS as readonly string[]).includes(v);
25
- }
26
-
27
- /** Human labels for routes, used in the "unknown route" tool-error message
28
- * and nowhere else — kept tiny on purpose. */
29
- export const ROUTE_LABELS: Record<RouteId, string> = {
30
- chat: "Chat",
31
- quantize: "Quantize",
32
- finetune: "Fine-tune",
33
- dataset: "Build Dataset",
34
- status: "Status",
35
- };
36
-
37
- /** A curated spotlight target: a stable id -> the selector to highlight +
38
- * a human label for the popover title, plus which route/view it lives on
39
- * (spotlight_ui navigates there first if the browser isn't already on it).
40
- * This is the "catalog target" resolution path in resolve-spotlight.ts's
41
- * terms — the fallback when ref/label/selector don't resolve from the live
42
- * DOM or the last uiSnapshot. Anchors added to app.html via data-spotlight
43
- * in this same wave; ids are stable and boring (element id minus "chat-"
44
- * prefix noise where it reads fine either way). */
45
- export interface SpotlightTarget {
46
- selector: string;
47
- label: string;
48
- route?: RouteId;
49
- view?: ViewId;
50
- }
51
-
52
- export const SPOTLIGHT_TARGETS: Record<string, SpotlightTarget> = {
53
- composer: { selector: "#chat-box", label: "Message box", route: "chat" },
54
- send: { selector: "#chat-send", label: "Send", route: "chat" },
55
- "new-chat": { selector: "#chat-new", label: "New chat", route: "chat" },
56
- "session-search": { selector: "#chat-sess-search", label: "Search chats", route: "chat" },
57
- "sampling-pill": { selector: "#chat-sampling", label: "Sampling", route: "chat" },
58
- "prompt-pill": { selector: "#chat-sysprompt", label: "System prompt", route: "chat" },
59
- "adapter-select": { selector: "#chat-adapter", label: "LoRA adapter", route: "chat" },
60
- "memory-entry": { selector: "#chat-memory-entry", label: "Memory", route: "chat" },
61
- "developer-toggle": { selector: "#nav-developer", label: "Developer toggle" },
62
- "model-picker": { selector: "#nav-model", label: "Active model" },
63
- "hub-browse": { selector: "#model-pop-browse", label: "Browse models", view: "hub-panel" },
64
- "quantize-source": { selector: "#q-model", label: "Source model", route: "quantize" },
65
- "finetune-base": { selector: "#f-model", label: "Base model", route: "finetune" },
66
- };
67
-
68
- export function getSpotlightTarget(id: string): SpotlightTarget | null {
69
- return SPOTLIGHT_TARGETS[id] ?? null;
70
- }
71
-
72
- /** Resolve a free-form route/page string (either a bare route id or a
73
- * "#/route" hash) the way navigate_app's tool param is described to the
74
- * model. Returns null for anything not in ROUTE_IDS. */
75
- export function resolveRouteId(routeOrPage: string): RouteId | null {
76
- const bare = routeOrPage.trim().replace(/^#\/?/, "");
77
- return isRouteId(bare) ? bare : null;
78
- }
@@ -1,61 +0,0 @@
1
- # src/web/vendor/ — vendored, no-CDN web assets
2
-
3
- Assets here are bundled locally and served same-origin by `src/server.ts`
4
- (`GET /assets/<name>`, `with { type: "text" }` import — same pattern as
5
- `src/web/app.html` itself). Nothing in this app loads a script or
6
- stylesheet from a CDN; this directory is where that tradeoff lives.
7
-
8
- ## hljs.js — syntax highlighting
9
-
10
- `hljs.js` is [highlight.js](https://github.com/highlightjs/highlight.js)
11
- (BSD-3-Clause, license text in `hljs-LICENSE`) core engine + a curated
12
- language set: javascript/typescript, python, bash/shell, json, html/xml,
13
- css, c/cpp, rust, go, sql, yaml, markdown, diff, plaintext. Registered
14
- languages expose themselves as `class="language-<name>"` on `<code>`
15
- elements — see `mdCodeBlock()` and `highlightIn()` in `app.html`.
16
-
17
- To rebuild (e.g. to bump the version or change the language set):
18
-
19
- ```sh
20
- mkdir /tmp/hljs-build && cd /tmp/hljs-build
21
- npm init -y
22
- npm install highlight.js@<version> --no-save
23
- ```
24
-
25
- Write an `entry.js` that imports `highlight.js/lib/core` and
26
- `highlight.js/lib/languages/<lang>` for each language you want, calling
27
- `hljs.registerLanguage(name, lang)` for each (see mlx-bun's own build log /
28
- commit that introduced this file for the exact entry.js used), then:
29
-
30
- ```sh
31
- globalThis.hljs = hljs; # inside entry.js, after registering
32
- bun build ./entry.js --minify --outfile=hljs.bundle.js
33
- ```
34
-
35
- Prepend a short license/provenance header comment (BSD-3-Clause,
36
- upstream URL, rebuild instructions — copy the one already at the top of
37
- `hljs.js`) and copy the result over `src/web/vendor/hljs.js`.
38
-
39
- **Size discipline:** the hygiene gate (`scripts/check-hygiene.ts`) fails
40
- any tracked file over 1 MB unless it's on the explicit allowlist. The
41
- current bundle (core + 14 languages) is ~83 KB — comfortably under the
42
- cap. If you add enough languages to bust 1 MB, either trim the language
43
- list or add a justified, size-capped `ALLOW` entry in
44
- `scripts/check-hygiene.ts` (see the existing entries there for the
45
- pattern) — don't silently let it grow past the cap.
46
-
47
- `hljs-theme.css` is NOT generated from highlight.js's stock themes — it's
48
- hand-written, mapping the standard `hljs-*` token classes onto mlx-bun's
49
- own design tokens (see the file's own header comment) so code blocks match
50
- the app's palette in both light and dark.
51
-
52
- ## KaTeX (math) — deliberately not vendored
53
-
54
- Investigated for the same treatment; skipped. KaTeX's CSS depends on its
55
- own bundled math font (`dist/fonts/`, ~60 files / ~1.1 MB of woff2/woff/ttf)
56
- for correct glyph metrics — there's no clean "system font fallback" path
57
- that renders math correctly. Vendoring it would mean adding ~1 MB of
58
- binary font files, which the hygiene gate is specifically designed to keep
59
- out of this repo (see `scripts/check-hygiene.ts`'s header). Needs the
60
- build-step/asset-hosting decision from `docs/design/web-chat-redesign.md`
61
- §7 before it's worth revisiting.
package/src/web-assets.ts DELETED
@@ -1,47 +0,0 @@
1
- // Embedded web app — the unified SPA (chat / quantize / finetune /
2
- // dataset / status). `with { type: "text" }` inlines the file in both
3
- // `bun run` and the compiled single binary. bun-types types *.html
4
- // imports as HTMLBundle (the html loader), but the text attribute makes
5
- // the runtime value a string — hence the double cast.
6
- import appHtml from "./web/app.html" with { type: "text" };
7
- import curveDesignerHtml from "./lab/curve/curve-designer.html" with { type: "text" };
8
- // Vendored, no-CDN static assets referenced by app.html (convention: any
9
- // self-contained JS/CSS too big to inline gets `with { type: "text" }`
10
- // imported here and served under /assets/<name>; see src/web/vendor/README
11
- // for how hljs.js was built). Add new vendored assets the same way.
12
- import hljsJs from "./web/vendor/hljs.js" with { type: "text" };
13
- import hljsCss from "./web/vendor/hljs-theme.css" with { type: "text" };
14
- // The frontend bundle (plan §7/§9 Phase 2 module split): GENERATED from
15
- // src/web/src/*.ts by `bun scripts/build-web.ts` — see that file's header
16
- // and tests/using/web-build.test.ts (the freshness gate). Same
17
- // with { type: "text" } + /assets/<name> pattern as the vendored assets
18
- // above; app.html's <script defer src="/assets/app.js"> loads it.
19
- import appJs from "./web/app.js" with { type: "text" };
20
- // PWA installability (plan §9 Phase 3, beat-matrix Axis 10): a manifest +
21
- // a single inline SVG icon (no binary PNGs — the hygiene gate forbids
22
- // tracked binary files; some browsers won't show an SVG app icon, which is
23
- // an accepted tradeoff, see docs/reference/server-config.md) + a shell-
24
- // only service worker. Same with { type: "text" } + /assets/<name>-shaped
25
- // pattern as the vendored assets above; see src/web/sw.js's header for why
26
- // it deliberately does NOT cache API/WS traffic.
27
- import manifestWebmanifest from "./web/manifest.webmanifest" with { type: "text" };
28
- import iconSvg from "./web/icon.svg" with { type: "text" };
29
- import swJs from "./web/sw.js" with { type: "text" };
30
- const APP_PAGE = appHtml as unknown as string;
31
- const HLJS_JS = hljsJs as unknown as string;
32
- const HLJS_CSS = hljsCss as unknown as string;
33
- const APP_JS = appJs as unknown as string;
34
- const MANIFEST_WEBMANIFEST = manifestWebmanifest as unknown as string;
35
- const ICON_SVG = iconSvg as unknown as string;
36
- const SW_JS = swJs as unknown as string;
37
- const CURVE_PAGE = curveDesignerHtml as unknown as string;
38
- export const STATIC_ROUTE_ASSETS = {
39
- appPage: APP_PAGE,
40
- appJs: APP_JS,
41
- hljsJs: HLJS_JS,
42
- hljsCss: HLJS_CSS,
43
- manifest: MANIFEST_WEBMANIFEST,
44
- iconSvg: ICON_SVG,
45
- serviceWorker: SW_JS,
46
- curvePage: CURVE_PAGE,
47
- };
package/src/web-tools.ts DELETED
@@ -1,478 +0,0 @@
1
- // mlx-bun web-tools — outward-facing capabilities for the web chat agent:
2
- // web_search, web_fetch, and weather. These are pi custom tools (defineTool)
3
- // registered alongside the built-in coding tools so the local model can pull
4
- // in current, real-world information instead of only inspecting the machine.
5
- //
6
- // Design notes:
7
- // - Keyless by default. weather uses Open-Meteo (no key, no anti-bot),
8
- // web_fetch is a plain HTTP GET, and web_search defaults to DuckDuckGo's
9
- // keyless HTML endpoint. Setting TAVILY_API_KEY or BRAVE_API_KEY upgrades
10
- // web_search to that provider's JSON API (more reliable, ranked results).
11
- // - Non-mutating: none of these touch the user's filesystem, so the web
12
- // chat's approval gate auto-allows them (they're not in GATED_TOOLS).
13
- // They DO make outbound network requests — that's the one thing that
14
- // leaves the machine, and only when the user asks for live info.
15
- // - execute() never throws on a network/upstream failure: it returns the
16
- // error as tool text so the model can read it and adapt (retry, tell the
17
- // user, fall back) rather than the turn aborting.
18
- //
19
- // The pure helpers (htmlToText, wmoCodeToText, parseDuckDuckGoHtml, the
20
- // formatters) are exported and unit-tested without touching the network.
21
-
22
- import { defineTool, type ToolDefinition } from "@earendil-works/pi-coding-agent";
23
- import { Type } from "typebox";
24
- import {
25
- fetchRestrictedHttpBytes,
26
- type MediaFetchPolicy,
27
- type RestrictedFetchDependencies,
28
- } from "./media-fetch";
29
-
30
- /** Tool names, exported so pi-web can add them to its tool allowlist. */
31
- export const WEB_TOOL_NAMES = ["web_search", "web_fetch", "weather"] as const;
32
-
33
- /** Browser-ish UA: some endpoints (DuckDuckGo) serve a challenge to blanks. */
34
- const USER_AGENT =
35
- "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 " +
36
- "(KHTML, like Gecko) Chrome/124.0 Safari/537.36";
37
-
38
- const FETCH_TIMEOUT_MS = 20_000;
39
- const MAX_FETCH_BYTES = 1024 * 1024;
40
- /** Cap on text returned to the model from a fetched page. */
41
- const MAX_FETCH_CHARS = 12_000;
42
- const DEFAULT_SEARCH_RESULTS = 5;
43
- const MAX_SEARCH_RESULTS = 10;
44
-
45
- // ---- low-level fetch helper ------------------------------------------
46
-
47
- /** Combine the tool's abort signal with a wall-clock timeout. */
48
- function withTimeout(signal: AbortSignal | undefined, ms: number): AbortSignal {
49
- const timeout = AbortSignal.timeout(ms);
50
- return signal ? AbortSignal.any([signal, timeout]) : timeout;
51
- }
52
-
53
- export interface FetchResult {
54
- ok: boolean;
55
- status: number;
56
- finalUrl: string;
57
- contentType: string;
58
- text: string;
59
- }
60
-
61
- export interface WebFetchOptions {
62
- policy?: MediaFetchPolicy;
63
- dependencies?: Partial<RestrictedFetchDependencies>;
64
- }
65
-
66
- const WEB_FETCH_POLICY: MediaFetchPolicy = {
67
- allowPrivate: false,
68
- privateDestinationHint:
69
- "private/loopback destinations are never available to the model-driven web_fetch tool",
70
- timeoutMs: FETCH_TIMEOUT_MS,
71
- maxBytes: MAX_FETCH_BYTES,
72
- maxRedirects: 5,
73
- };
74
-
75
- /** GET/POST a URL and read the body as text, with timeout + abort wiring. */
76
- async function httpText(
77
- url: string,
78
- opts: { method?: string; headers?: Record<string, string>; body?: string; signal?: AbortSignal; timeoutMs?: number },
79
- ): Promise<FetchResult> {
80
- const res = await fetch(url, {
81
- method: opts.method ?? "GET",
82
- headers: { "user-agent": USER_AGENT, ...opts.headers },
83
- body: opts.body,
84
- redirect: "follow",
85
- signal: withTimeout(opts.signal, opts.timeoutMs ?? FETCH_TIMEOUT_MS),
86
- });
87
- const contentType = res.headers.get("content-type") ?? "";
88
- const text = await res.text();
89
- return { ok: res.ok, status: res.status, finalUrl: res.url || url, contentType, text };
90
- }
91
-
92
- /**
93
- * Fetch a model-selected URL through the shared restricted-destination
94
- * transport. Exported with dependency injection for deterministic,
95
- * model-free security tests.
96
- */
97
- export async function fetchWebText(
98
- url: string,
99
- signal?: AbortSignal,
100
- options: WebFetchOptions = {},
101
- ): Promise<FetchResult> {
102
- const result = await fetchRestrictedHttpBytes(
103
- url,
104
- "web",
105
- options.policy ?? WEB_FETCH_POLICY,
106
- {
107
- signal,
108
- dependencies: options.dependencies,
109
- headers: {
110
- "user-agent": USER_AGENT,
111
- accept: "text/html,application/json,text/plain,*/*",
112
- },
113
- },
114
- );
115
- return {
116
- ok: result.ok,
117
- status: result.status,
118
- finalUrl: result.finalUrl,
119
- contentType: result.contentType,
120
- text: new TextDecoder().decode(result.bytes),
121
- };
122
- }
123
-
124
- /** Normalize anything thrown by fetch (timeout, DNS, abort) into a message. */
125
- function describeFetchError(err: unknown): string {
126
- if (err instanceof Error) {
127
- if (err.name === "TimeoutError") return "request timed out";
128
- if (err.name === "AbortError") return "request was cancelled";
129
- return err.message;
130
- }
131
- return String(err);
132
- }
133
-
134
- // ---- pure helpers (unit-tested) --------------------------------------
135
-
136
- const NAMED_ENTITIES: Record<string, string> = {
137
- amp: "&", lt: "<", gt: ">", quot: '"', apos: "'", nbsp: " ", "#39": "'", "#34": '"',
138
- };
139
-
140
- /** Decode the handful of HTML entities that show up in scraped text. */
141
- export function decodeEntities(s: string): string {
142
- return s.replace(/&(#x?[0-9a-fA-F]+|[a-zA-Z]+);/g, (whole, body: string) => {
143
- if (body[0] === "#") {
144
- const code = body[1] === "x" || body[1] === "X" ? parseInt(body.slice(2), 16) : parseInt(body.slice(1), 10);
145
- return Number.isFinite(code) ? String.fromCodePoint(code) : whole;
146
- }
147
- return NAMED_ENTITIES[body] ?? whole;
148
- });
149
- }
150
-
151
- /** Strip all HTML tags from a fragment and decode entities (single-line). */
152
- export function stripTags(html: string): string {
153
- return decodeEntities(html.replace(/<[^>]*>/g, "")).replace(/\s+/g, " ").trim();
154
- }
155
-
156
- /** Extract the <title> of an HTML document, if any. */
157
- export function extractTitle(html: string): string | undefined {
158
- const m = html.match(/<title[^>]*>([\s\S]*?)<\/title>/i);
159
- return m ? stripTags(m[1] ?? "") : undefined;
160
- }
161
-
162
- /**
163
- * Reduce an HTML document to readable plain text: drop script/style/head,
164
- * turn block-level boundaries into newlines, strip remaining tags, decode
165
- * entities, and collapse runaway whitespace.
166
- */
167
- export function htmlToText(html: string): string {
168
- let s = html
169
- .replace(/<!--[\s\S]*?-->/g, "")
170
- .replace(/<(script|style|noscript|template|svg)[^>]*>[\s\S]*?<\/\1>/gi, "")
171
- .replace(/<head[^>]*>[\s\S]*?<\/head>/gi, "");
172
- // Block-level close/break tags become newlines so structure survives.
173
- s = s.replace(/<\/(p|div|section|article|li|ul|ol|tr|h[1-6]|header|footer|nav|blockquote)>/gi, "\n");
174
- s = s.replace(/<br\s*\/?>/gi, "\n");
175
- s = s.replace(/<[^>]+>/g, "");
176
- s = decodeEntities(s);
177
- // Collapse spaces/tabs, trim each line, squeeze blank-line runs.
178
- s = s.replace(/[ \t\f\v]+/g, " ");
179
- s = s.split("\n").map((line) => line.trim()).join("\n");
180
- s = s.replace(/\n{3,}/g, "\n\n");
181
- return s.trim();
182
- }
183
-
184
- /** Clamp text to a max length, appending a clear truncation marker. */
185
- export function truncate(text: string, max: number): string {
186
- if (text.length <= max) return text;
187
- return text.slice(0, max) + `\n\n…[truncated ${text.length - max} more characters]`;
188
- }
189
-
190
- /** WMO weather interpretation codes (Open-Meteo) → human description. */
191
- export function wmoCodeToText(code: number): string {
192
- const map: Record<number, string> = {
193
- 0: "clear sky", 1: "mainly clear", 2: "partly cloudy", 3: "overcast",
194
- 45: "fog", 48: "depositing rime fog",
195
- 51: "light drizzle", 53: "moderate drizzle", 55: "dense drizzle",
196
- 56: "light freezing drizzle", 57: "dense freezing drizzle",
197
- 61: "slight rain", 63: "moderate rain", 65: "heavy rain",
198
- 66: "light freezing rain", 67: "heavy freezing rain",
199
- 71: "slight snowfall", 73: "moderate snowfall", 75: "heavy snowfall",
200
- 77: "snow grains",
201
- 80: "slight rain showers", 81: "moderate rain showers", 82: "violent rain showers",
202
- 85: "slight snow showers", 86: "heavy snow showers",
203
- 95: "thunderstorm", 96: "thunderstorm with slight hail", 99: "thunderstorm with heavy hail",
204
- };
205
- return map[code] ?? `unknown (code ${code})`;
206
- }
207
-
208
- export interface SearchResult {
209
- title: string;
210
- url: string;
211
- snippet: string;
212
- }
213
-
214
- /** Resolve a DuckDuckGo redirect href (//duckduckgo.com/l/?uddg=…) to the real URL. */
215
- export function resolveDdgHref(href: string): string {
216
- const decoded = decodeEntities(href);
217
- const m = decoded.match(/[?&]uddg=([^&]+)/);
218
- if (m && m[1]) {
219
- try {
220
- return decodeURIComponent(m[1]);
221
- } catch {
222
- return decoded;
223
- }
224
- }
225
- return decoded.startsWith("//") ? "https:" + decoded : decoded;
226
- }
227
-
228
- /**
229
- * Parse the result list out of html.duckduckgo.com/html/ output. Pairs each
230
- * result anchor (title + href) with the following snippet anchor by order.
231
- * Returns [] when the page carries no results (e.g. a challenge/empty page).
232
- */
233
- export function parseDuckDuckGoHtml(html: string, limit: number): SearchResult[] {
234
- const results: SearchResult[] = [];
235
- const linkRe = /<a[^>]*class="[^"]*result__a[^"]*"[^>]*href="([^"]+)"[^>]*>([\s\S]*?)<\/a>/gi;
236
- const snippetRe = /<a[^>]*class="[^"]*result__snippet[^"]*"[^>]*>([\s\S]*?)<\/a>/gi;
237
- const snippets: string[] = [];
238
- let sm: RegExpExecArray | null;
239
- while ((sm = snippetRe.exec(html)) !== null) snippets.push(stripTags(sm[1] ?? ""));
240
- let lm: RegExpExecArray | null;
241
- let i = 0;
242
- while ((lm = linkRe.exec(html)) !== null && results.length < limit) {
243
- const url = resolveDdgHref(lm[1] ?? "");
244
- const title = stripTags(lm[2] ?? "");
245
- if (!title || !url) {
246
- i++;
247
- continue;
248
- }
249
- results.push({ title, url, snippet: snippets[i] ?? "" });
250
- i++;
251
- }
252
- return results;
253
- }
254
-
255
- /** Render a result list as numbered Markdown for the model. */
256
- export function formatSearchResults(query: string, results: SearchResult[], provider: string): string {
257
- if (results.length === 0) {
258
- return `No results found for "${query}" (via ${provider}).`;
259
- }
260
- const lines = results.map((r, idx) => {
261
- const head = `${idx + 1}. ${r.title}\n ${r.url}`;
262
- return r.snippet ? `${head}\n ${r.snippet}` : head;
263
- });
264
- return `Search results for "${query}" (via ${provider}):\n\n${lines.join("\n\n")}`;
265
- }
266
-
267
- // ---- weather: Open-Meteo geocode + forecast --------------------------
268
-
269
- interface GeoPlace {
270
- name: string;
271
- latitude: number;
272
- longitude: number;
273
- country?: string;
274
- admin1?: string;
275
- }
276
-
277
- /** Format an Open-Meteo forecast response into a compact human summary. */
278
- export function formatWeather(place: GeoPlace, data: unknown, fahrenheit: boolean): string {
279
- const d = data as {
280
- current?: {
281
- temperature_2m?: number;
282
- apparent_temperature?: number;
283
- relative_humidity_2m?: number;
284
- wind_speed_10m?: number;
285
- weather_code?: number;
286
- };
287
- daily?: { time?: string[]; weather_code?: number[]; temperature_2m_max?: number[]; temperature_2m_min?: number[]; precipitation_probability_max?: number[] };
288
- };
289
- const tUnit = fahrenheit ? "°F" : "°C";
290
- const wUnit = fahrenheit ? "mph" : "km/h";
291
- const where = [place.name, place.admin1, place.country].filter(Boolean).join(", ");
292
- const lines: string[] = [`Weather for ${where}:`];
293
-
294
- const cur = d.current;
295
- if (cur?.temperature_2m != null) {
296
- const desc = cur.weather_code != null ? `, ${wmoCodeToText(cur.weather_code)}` : "";
297
- const parts = [`${Math.round(cur.temperature_2m)}${tUnit}${desc}`];
298
- if (cur.apparent_temperature != null) parts.push(`feels like ${Math.round(cur.apparent_temperature)}${tUnit}`);
299
- if (cur.relative_humidity_2m != null) parts.push(`humidity ${Math.round(cur.relative_humidity_2m)}%`);
300
- if (cur.wind_speed_10m != null) parts.push(`wind ${Math.round(cur.wind_speed_10m)} ${wUnit}`);
301
- lines.push(`Now: ${parts.join(", ")}.`);
302
- }
303
-
304
- const daily = d.daily;
305
- if (daily?.time?.length) {
306
- lines.push("");
307
- lines.push("Forecast:");
308
- for (let i = 0; i < daily.time.length; i++) {
309
- const code = daily.weather_code?.[i];
310
- const hi = daily.temperature_2m_max?.[i];
311
- const lo = daily.temperature_2m_min?.[i];
312
- const pop = daily.precipitation_probability_max?.[i];
313
- const desc = code != null ? wmoCodeToText(code) : "";
314
- const range = hi != null && lo != null ? `${Math.round(lo)}–${Math.round(hi)}${tUnit}` : "";
315
- const rain = pop != null ? `, precip ${pop}%` : "";
316
- lines.push(`- ${daily.time[i]}: ${[range, desc].filter(Boolean).join(", ")}${rain}`);
317
- }
318
- }
319
- return lines.join("\n");
320
- }
321
-
322
- async function geocode(name: string, signal: AbortSignal | undefined): Promise<GeoPlace | undefined> {
323
- const url = `https://geocoding-api.open-meteo.com/v1/search?name=${encodeURIComponent(name)}&count=1&language=en&format=json`;
324
- const res = await httpText(url, { signal });
325
- if (!res.ok) return undefined;
326
- const json = JSON.parse(res.text) as { results?: GeoPlace[] };
327
- return json.results?.[0];
328
- }
329
-
330
- // ---- web_search providers --------------------------------------------
331
-
332
- async function searchTavily(query: string, count: number, signal: AbortSignal | undefined, apiKey: string): Promise<SearchResult[]> {
333
- const res = await httpText("https://api.tavily.com/search", {
334
- method: "POST",
335
- headers: { "content-type": "application/json" },
336
- body: JSON.stringify({ api_key: apiKey, query, max_results: count, search_depth: "basic" }),
337
- signal,
338
- });
339
- if (!res.ok) throw new Error(`Tavily returned HTTP ${res.status}`);
340
- const json = JSON.parse(res.text) as { results?: Array<{ title?: string; url?: string; content?: string }> };
341
- return (json.results ?? []).slice(0, count).map((r) => ({ title: r.title ?? r.url ?? "", url: r.url ?? "", snippet: r.content ?? "" }));
342
- }
343
-
344
- async function searchBrave(query: string, count: number, signal: AbortSignal | undefined, apiKey: string): Promise<SearchResult[]> {
345
- const url = `https://api.search.brave.com/res/v1/web/search?q=${encodeURIComponent(query)}&count=${count}`;
346
- const res = await httpText(url, { headers: { "x-subscription-token": apiKey, accept: "application/json" }, signal });
347
- if (!res.ok) throw new Error(`Brave returned HTTP ${res.status}`);
348
- const json = JSON.parse(res.text) as { web?: { results?: Array<{ title?: string; url?: string; description?: string }> } };
349
- return (json.web?.results ?? []).slice(0, count).map((r) => ({ title: r.title ?? r.url ?? "", url: r.url ?? "", snippet: stripTags(r.description ?? "") }));
350
- }
351
-
352
- async function searchDuckDuckGo(query: string, count: number, signal: AbortSignal | undefined): Promise<SearchResult[]> {
353
- const url = `https://html.duckduckgo.com/html/?q=${encodeURIComponent(query)}`;
354
- const res = await httpText(url, { headers: { accept: "text/html" }, signal });
355
- if (!res.ok) throw new Error(`DuckDuckGo returned HTTP ${res.status}`);
356
- return parseDuckDuckGoHtml(res.text, count);
357
- }
358
-
359
- /** Pick the search backend from env; returns the provider label too. */
360
- function resolveSearchProvider(): { name: string; run: (q: string, n: number, s: AbortSignal | undefined) => Promise<SearchResult[]> } {
361
- const tavily = process.env.TAVILY_API_KEY;
362
- if (tavily) return { name: "Tavily", run: (q, n, s) => searchTavily(q, n, s, tavily) };
363
- const brave = process.env.BRAVE_API_KEY;
364
- if (brave) return { name: "Brave", run: (q, n, s) => searchBrave(q, n, s, brave) };
365
- return { name: "DuckDuckGo", run: searchDuckDuckGo };
366
- }
367
-
368
- // ---- tool definitions ------------------------------------------------
369
-
370
- function textResult(text: string): { content: [{ type: "text"; text: string }]; details: Record<string, never> } {
371
- return { content: [{ type: "text", text }], details: {} };
372
- }
373
-
374
- const webSearchTool = defineTool({
375
- name: "web_search",
376
- label: "Web Search",
377
- description:
378
- "Search the web for current, real-world information (news, facts, docs, anything that changes over time or that you don't already know). " +
379
- "Returns a ranked list of result titles, URLs, and snippets. Follow up with web_fetch to read a result in full.",
380
- parameters: Type.Object({
381
- query: Type.String({ description: "The search query." }),
382
- count: Type.Optional(Type.Number({ description: `Number of results to return (default ${DEFAULT_SEARCH_RESULTS}, max ${MAX_SEARCH_RESULTS}).` })),
383
- }),
384
- execute: async (_id, params, signal) => {
385
- const query = params.query.trim();
386
- if (!query) return textResult("web_search needs a non-empty query.");
387
- const count = Math.max(1, Math.min(MAX_SEARCH_RESULTS, Math.round(params.count ?? DEFAULT_SEARCH_RESULTS)));
388
- const provider = resolveSearchProvider();
389
- try {
390
- const results = await provider.run(query, count, signal);
391
- if (results.length === 0 && provider.name === "DuckDuckGo") {
392
- return textResult(
393
- `No results for "${query}". DuckDuckGo (the keyless default) may be rate-limiting; ` +
394
- `set TAVILY_API_KEY or BRAVE_API_KEY for a reliable search backend.`,
395
- );
396
- }
397
- return textResult(formatSearchResults(query, results, provider.name));
398
- } catch (err) {
399
- return textResult(`Web search failed (${provider.name}): ${describeFetchError(err)}`);
400
- }
401
- },
402
- });
403
-
404
- const webFetchTool = defineTool({
405
- name: "web_fetch",
406
- label: "Fetch URL",
407
- description:
408
- "Fetch a URL and return its contents as readable text (HTML pages are stripped to plain text). " +
409
- "Use this to read a web page, article, or API/JSON endpoint — e.g. after web_search, or when the user gives you a link.",
410
- parameters: Type.Object({
411
- url: Type.String({ description: "The absolute http(s) URL to fetch." }),
412
- }),
413
- execute: async (_id, params, signal) => {
414
- let url = params.url.trim();
415
- if (!/^https?:\/\//i.test(url)) url = "https://" + url;
416
- try {
417
- const res = await fetchWebText(url, signal);
418
- const header = `Fetched ${res.finalUrl} (HTTP ${res.status})`;
419
- if (!res.ok) return textResult(`${header}\n\n${truncate(htmlToText(res.text), 2_000)}`);
420
- const isHtml = /text\/html|application\/xhtml/i.test(res.contentType);
421
- if (isHtml) {
422
- const title = extractTitle(res.text);
423
- const body = truncate(htmlToText(res.text), MAX_FETCH_CHARS);
424
- return textResult(`${header}${title ? `\nTitle: ${title}` : ""}\n\n${body}`);
425
- }
426
- const isTextual = /text\/|application\/(json|xml|javascript|x-ndjson)|\+json|\+xml/i.test(res.contentType) || res.contentType === "";
427
- if (!isTextual) {
428
- return textResult(`${header}\n\nContent-Type ${res.contentType || "unknown"} is not text; not displaying ${res.text.length} bytes of binary data.`);
429
- }
430
- return textResult(`${header}\n\n${truncate(res.text, MAX_FETCH_CHARS)}`);
431
- } catch (err) {
432
- return textResult(`Failed to fetch ${url}: ${describeFetchError(err)}`);
433
- }
434
- },
435
- });
436
-
437
- const weatherTool = defineTool({
438
- name: "weather",
439
- label: "Weather",
440
- description:
441
- "Get current weather conditions and a short forecast for a place. Accepts a city/place name (e.g. 'Austin', 'Paris, France', 'Tokyo'). " +
442
- "No API key required.",
443
- parameters: Type.Object({
444
- location: Type.String({ description: "City or place name to look up." }),
445
- unit: Type.Optional(Type.String({ description: "Temperature unit: 'celsius' or 'fahrenheit' (default celsius)." })),
446
- }),
447
- execute: async (_id, params, signal) => {
448
- const location = params.location.trim();
449
- if (!location) return textResult("weather needs a location.");
450
- const fahrenheit = params.unit === "fahrenheit";
451
- try {
452
- const place = await geocode(location, signal);
453
- if (!place) return textResult(`Couldn't find a place named "${location}".`);
454
- const params2 = new URLSearchParams({
455
- latitude: String(place.latitude),
456
- longitude: String(place.longitude),
457
- current: "temperature_2m,relative_humidity_2m,apparent_temperature,precipitation,weather_code,wind_speed_10m",
458
- daily: "weather_code,temperature_2m_max,temperature_2m_min,precipitation_probability_max",
459
- timezone: "auto",
460
- forecast_days: "3",
461
- });
462
- if (fahrenheit) {
463
- params2.set("temperature_unit", "fahrenheit");
464
- params2.set("wind_speed_unit", "mph");
465
- }
466
- const res = await httpText(`https://api.open-meteo.com/v1/forecast?${params2.toString()}`, { signal });
467
- if (!res.ok) return textResult(`Weather lookup failed: Open-Meteo returned HTTP ${res.status}.`);
468
- return textResult(formatWeather(place, JSON.parse(res.text), fahrenheit));
469
- } catch (err) {
470
- return textResult(`Weather lookup failed: ${describeFetchError(err)}`);
471
- }
472
- },
473
- });
474
-
475
- /** All web tools, ready to pass to createAgentSession({ customTools }). */
476
- export function createWebTools(): ToolDefinition[] {
477
- return [webSearchTool, webFetchTool, weatherTool];
478
- }