@r1z4x/tezgah 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 (582) hide show
  1. package/CHANGELOG.md +738 -0
  2. package/LICENSE +21 -0
  3. package/MANIFEST +753 -0
  4. package/README.de.md +169 -0
  5. package/README.es.md +168 -0
  6. package/README.fr.md +166 -0
  7. package/README.md +161 -0
  8. package/README.tr.md +163 -0
  9. package/agents/tezgah-reviewer.md +70 -0
  10. package/bin/codegen +337 -0
  11. package/bin/consult +797 -0
  12. package/bin/tezgah-adhd +50 -0
  13. package/bin/tezgah-agents +45 -0
  14. package/bin/tezgah-capture +66 -0
  15. package/bin/tezgah-context +96 -0
  16. package/bin/tezgah-decisions +213 -0
  17. package/bin/tezgah-design +953 -0
  18. package/bin/tezgah-docs +1199 -0
  19. package/bin/tezgah-doctor +497 -0
  20. package/bin/tezgah-dsh +37 -0
  21. package/bin/tezgah-dsh.cmd +70 -0
  22. package/bin/tezgah-gate +75 -0
  23. package/bin/tezgah-import-ai-research +439 -0
  24. package/bin/tezgah-index +32 -0
  25. package/bin/tezgah-lessons +148 -0
  26. package/bin/tezgah-mcp +322 -0
  27. package/bin/tezgah-migrate +324 -0
  28. package/bin/tezgah-pony +50 -0
  29. package/bin/tezgah-research +772 -0
  30. package/bin/tezgah-rollback +93 -0
  31. package/bin/tezgah-route +124 -0
  32. package/bin/tezgah-setup +5798 -0
  33. package/bin/tezgah-status +383 -0
  34. package/bin/tezgah-task +607 -0
  35. package/bin/tezgah-taste +423 -0
  36. package/bin/tezgah-triage +485 -0
  37. package/bin/tezgah.js +36 -0
  38. package/commands/adhd.md +21 -0
  39. package/commands/ponytail.md +21 -0
  40. package/docs/README.md +112 -0
  41. package/docs/architecture.md +226 -0
  42. package/docs/citations-baseline.json +59 -0
  43. package/docs/clarity-baseline.json +116 -0
  44. package/docs/contract.md +416 -0
  45. package/docs/environment.md +141 -0
  46. package/docs/evidence.md +831 -0
  47. package/docs/feature-audit.md +139 -0
  48. package/docs/gate.md +407 -0
  49. package/docs/glossary.md +201 -0
  50. package/docs/hosts.md +400 -0
  51. package/docs/index.json +471 -0
  52. package/docs/judge.md +193 -0
  53. package/docs/layers.md +189 -0
  54. package/docs/models.md +160 -0
  55. package/docs/operations.md +666 -0
  56. package/docs/research.md +809 -0
  57. package/docs/results/paired-outcome-2026-10.jsonl +254 -0
  58. package/docs/results/paired-outcome-2026-10.md +204 -0
  59. package/docs/skills.md +236 -0
  60. package/docs/status-line.md +315 -0
  61. package/docs/testing.md +287 -0
  62. package/docs/vision.md +117 -0
  63. package/hooks/hooks.json +33 -0
  64. package/hooks/projects-auto-init.py +98 -0
  65. package/hooks/projects-posttooluse.py +149 -0
  66. package/hooks/projects-pretooluse.py +40 -0
  67. package/hooks/projects-stop.py +60 -0
  68. package/hooks/tezgah_agents.py +817 -0
  69. package/hooks/tezgah_apps.py +266 -0
  70. package/hooks/tezgah_attest.py +255 -0
  71. package/hooks/tezgah_clarity.py +171 -0
  72. package/hooks/tezgah_components.py +140 -0
  73. package/hooks/tezgah_context.py +2953 -0
  74. package/hooks/tezgah_embed.py +765 -0
  75. package/hooks/tezgah_gate.py +2736 -0
  76. package/hooks/tezgah_guard.py +177 -0
  77. package/hooks/tezgah_index.py +124 -0
  78. package/hooks/tezgah_integrity.py +5141 -0
  79. package/hooks/tezgah_judge.py +487 -0
  80. package/hooks/tezgah_lang.py +230 -0
  81. package/hooks/tezgah_lessons.py +210 -0
  82. package/hooks/tezgah_models.py +855 -0
  83. package/hooks/tezgah_paths.py +1132 -0
  84. package/hooks/tezgah_policy.py +1039 -0
  85. package/hooks/tezgah_rank.py +77 -0
  86. package/hooks/tezgah_replay.py +1109 -0
  87. package/hooks/tezgah_research.py +5176 -0
  88. package/hooks/tezgah_shapes.py +305 -0
  89. package/hooks/tezgah_skill_pick.py +227 -0
  90. package/hooks/tezgah_snapshot.py +490 -0
  91. package/hooks/tezgah_stopspec.py +320 -0
  92. package/hooks/tezgah_task.py +458 -0
  93. package/hooks/tezgah_taste.py +201 -0
  94. package/hooks/tezgah_untrusted.py +112 -0
  95. package/hooks/tezgah_update.py +430 -0
  96. package/hosts/codex/hook.py +268 -0
  97. package/hosts/codex/hooks.json +26 -0
  98. package/hosts/cursor/hook.py +461 -0
  99. package/hosts/cursor/hooks.json +44 -0
  100. package/hosts/dsh/hooks.json +21 -0
  101. package/hosts/dsh/statusline/lib/client.js +138 -0
  102. package/hosts/dsh/statusline/lib/index.js +104 -0
  103. package/hosts/dsh/statusline/package.json +28 -0
  104. package/hosts/omp/hook.py +277 -0
  105. package/hosts/omp/tezgah-hook.ts.in +701 -0
  106. package/hosts/opencode/plugins/tezgah.js +2742 -0
  107. package/hosts/opencode/tui/tezgah-tui.tsx +151 -0
  108. package/output-styles/tezgah.md +142 -0
  109. package/package.json +32 -0
  110. package/packaging/build.sh +342 -0
  111. package/packaging/install.ps1 +166 -0
  112. package/packaging/install.sh +56 -0
  113. package/packaging/upgrade.sh +142 -0
  114. package/skills/ai-research/0-autoresearch-skill/SKILL.md +411 -0
  115. package/skills/ai-research/0-autoresearch-skill/references/agent-continuity.md +95 -0
  116. package/skills/ai-research/0-autoresearch-skill/references/progress-reporting.md +166 -0
  117. package/skills/ai-research/0-autoresearch-skill/references/skill-routing.md +219 -0
  118. package/skills/ai-research/0-autoresearch-skill/templates/findings.md +44 -0
  119. package/skills/ai-research/0-autoresearch-skill/templates/research-log.md +41 -0
  120. package/skills/ai-research/01-model-architecture/litgpt/SKILL.md +469 -0
  121. package/skills/ai-research/01-model-architecture/litgpt/references/custom-models.md +569 -0
  122. package/skills/ai-research/01-model-architecture/litgpt/references/distributed-training.md +452 -0
  123. package/skills/ai-research/01-model-architecture/litgpt/references/supported-models.md +337 -0
  124. package/skills/ai-research/01-model-architecture/litgpt/references/training-recipes.md +620 -0
  125. package/skills/ai-research/01-model-architecture/mamba/SKILL.md +260 -0
  126. package/skills/ai-research/01-model-architecture/mamba/references/architecture-details.md +207 -0
  127. package/skills/ai-research/01-model-architecture/mamba/references/benchmarks.md +256 -0
  128. package/skills/ai-research/01-model-architecture/mamba/references/training-guide.md +389 -0
  129. package/skills/ai-research/01-model-architecture/nanogpt/SKILL.md +290 -0
  130. package/skills/ai-research/01-model-architecture/nanogpt/references/architecture.md +383 -0
  131. package/skills/ai-research/01-model-architecture/nanogpt/references/data.md +477 -0
  132. package/skills/ai-research/01-model-architecture/nanogpt/references/training.md +565 -0
  133. package/skills/ai-research/01-model-architecture/rwkv/SKILL.md +260 -0
  134. package/skills/ai-research/01-model-architecture/rwkv/references/architecture-details.md +345 -0
  135. package/skills/ai-research/01-model-architecture/rwkv/references/rwkv7.md +387 -0
  136. package/skills/ai-research/01-model-architecture/rwkv/references/state-management.md +370 -0
  137. package/skills/ai-research/01-model-architecture/torchtitan/SKILL.md +358 -0
  138. package/skills/ai-research/01-model-architecture/torchtitan/references/checkpoint.md +182 -0
  139. package/skills/ai-research/01-model-architecture/torchtitan/references/custom-models.md +259 -0
  140. package/skills/ai-research/01-model-architecture/torchtitan/references/float8.md +134 -0
  141. package/skills/ai-research/01-model-architecture/torchtitan/references/fsdp.md +127 -0
  142. package/skills/ai-research/02-tokenization/huggingface-tokenizers/SKILL.md +516 -0
  143. package/skills/ai-research/02-tokenization/huggingface-tokenizers/references/algorithms.md +654 -0
  144. package/skills/ai-research/02-tokenization/huggingface-tokenizers/references/integration.md +638 -0
  145. package/skills/ai-research/02-tokenization/huggingface-tokenizers/references/pipeline.md +724 -0
  146. package/skills/ai-research/02-tokenization/huggingface-tokenizers/references/training.md +566 -0
  147. package/skills/ai-research/02-tokenization/sentencepiece/SKILL.md +235 -0
  148. package/skills/ai-research/02-tokenization/sentencepiece/references/algorithms.md +201 -0
  149. package/skills/ai-research/02-tokenization/sentencepiece/references/training.md +305 -0
  150. package/skills/ai-research/03-fine-tuning/axolotl/SKILL.md +158 -0
  151. package/skills/ai-research/03-fine-tuning/axolotl/references/api.md +5549 -0
  152. package/skills/ai-research/03-fine-tuning/axolotl/references/dataset-formats.md +1030 -0
  153. package/skills/ai-research/03-fine-tuning/axolotl/references/index.md +16 -0
  154. package/skills/ai-research/03-fine-tuning/axolotl/references/other.md +3564 -0
  155. package/skills/ai-research/03-fine-tuning/llama-factory/SKILL.md +80 -0
  156. package/skills/ai-research/03-fine-tuning/llama-factory/references/_images.md +24 -0
  157. package/skills/ai-research/03-fine-tuning/llama-factory/references/advanced.md +1056 -0
  158. package/skills/ai-research/03-fine-tuning/llama-factory/references/getting_started.md +350 -0
  159. package/skills/ai-research/03-fine-tuning/llama-factory/references/index.md +20 -0
  160. package/skills/ai-research/03-fine-tuning/llama-factory/references/other.md +32 -0
  161. package/skills/ai-research/03-fine-tuning/peft/SKILL.md +431 -0
  162. package/skills/ai-research/03-fine-tuning/peft/references/advanced-usage.md +515 -0
  163. package/skills/ai-research/03-fine-tuning/peft/references/troubleshooting.md +481 -0
  164. package/skills/ai-research/03-fine-tuning/unsloth/SKILL.md +80 -0
  165. package/skills/ai-research/03-fine-tuning/unsloth/references/index.md +8 -0
  166. package/skills/ai-research/03-fine-tuning/unsloth/references/llms.md +83 -0
  167. package/skills/ai-research/04-mechanistic-interpretability/nnsight/SKILL.md +436 -0
  168. package/skills/ai-research/04-mechanistic-interpretability/nnsight/references/README.md +79 -0
  169. package/skills/ai-research/04-mechanistic-interpretability/nnsight/references/api.md +345 -0
  170. package/skills/ai-research/04-mechanistic-interpretability/nnsight/references/tutorials.md +301 -0
  171. package/skills/ai-research/04-mechanistic-interpretability/pyvene/SKILL.md +473 -0
  172. package/skills/ai-research/04-mechanistic-interpretability/pyvene/references/README.md +74 -0
  173. package/skills/ai-research/04-mechanistic-interpretability/pyvene/references/api.md +384 -0
  174. package/skills/ai-research/04-mechanistic-interpretability/pyvene/references/tutorials.md +377 -0
  175. package/skills/ai-research/04-mechanistic-interpretability/saelens/SKILL.md +386 -0
  176. package/skills/ai-research/04-mechanistic-interpretability/saelens/references/README.md +71 -0
  177. package/skills/ai-research/04-mechanistic-interpretability/saelens/references/api.md +334 -0
  178. package/skills/ai-research/04-mechanistic-interpretability/saelens/references/tutorials.md +319 -0
  179. package/skills/ai-research/04-mechanistic-interpretability/transformer-lens/SKILL.md +346 -0
  180. package/skills/ai-research/04-mechanistic-interpretability/transformer-lens/references/README.md +55 -0
  181. package/skills/ai-research/04-mechanistic-interpretability/transformer-lens/references/api.md +363 -0
  182. package/skills/ai-research/04-mechanistic-interpretability/transformer-lens/references/tutorials.md +340 -0
  183. package/skills/ai-research/05-data-processing/nemo-curator/SKILL.md +383 -0
  184. package/skills/ai-research/05-data-processing/nemo-curator/references/deduplication.md +88 -0
  185. package/skills/ai-research/05-data-processing/nemo-curator/references/filtering.md +103 -0
  186. package/skills/ai-research/05-data-processing/ray-data/SKILL.md +326 -0
  187. package/skills/ai-research/05-data-processing/ray-data/references/integration.md +83 -0
  188. package/skills/ai-research/05-data-processing/ray-data/references/transformations.md +84 -0
  189. package/skills/ai-research/06-post-training/grpo-rl-training/README.md +98 -0
  190. package/skills/ai-research/06-post-training/grpo-rl-training/SKILL.md +572 -0
  191. package/skills/ai-research/06-post-training/miles/SKILL.md +315 -0
  192. package/skills/ai-research/06-post-training/miles/references/api-reference.md +142 -0
  193. package/skills/ai-research/06-post-training/miles/references/troubleshooting.md +353 -0
  194. package/skills/ai-research/06-post-training/openrlhf/SKILL.md +249 -0
  195. package/skills/ai-research/06-post-training/openrlhf/references/algorithm-comparison.md +405 -0
  196. package/skills/ai-research/06-post-training/openrlhf/references/custom-rewards.md +531 -0
  197. package/skills/ai-research/06-post-training/openrlhf/references/hybrid-engine.md +288 -0
  198. package/skills/ai-research/06-post-training/openrlhf/references/multi-node-training.md +455 -0
  199. package/skills/ai-research/06-post-training/simpo/SKILL.md +219 -0
  200. package/skills/ai-research/06-post-training/simpo/references/datasets.md +479 -0
  201. package/skills/ai-research/06-post-training/simpo/references/hyperparameters.md +453 -0
  202. package/skills/ai-research/06-post-training/simpo/references/loss-functions.md +351 -0
  203. package/skills/ai-research/06-post-training/slime/SKILL.md +464 -0
  204. package/skills/ai-research/06-post-training/slime/references/api-reference.md +393 -0
  205. package/skills/ai-research/06-post-training/slime/references/troubleshooting.md +387 -0
  206. package/skills/ai-research/06-post-training/torchforge/SKILL.md +433 -0
  207. package/skills/ai-research/06-post-training/torchforge/references/api-reference.md +328 -0
  208. package/skills/ai-research/06-post-training/torchforge/references/troubleshooting.md +410 -0
  209. package/skills/ai-research/06-post-training/trl-fine-tuning/SKILL.md +455 -0
  210. package/skills/ai-research/06-post-training/trl-fine-tuning/references/dpo-variants.md +228 -0
  211. package/skills/ai-research/06-post-training/trl-fine-tuning/references/online-rl.md +83 -0
  212. package/skills/ai-research/06-post-training/trl-fine-tuning/references/reward-modeling.md +123 -0
  213. package/skills/ai-research/06-post-training/trl-fine-tuning/references/sft-training.md +169 -0
  214. package/skills/ai-research/06-post-training/verl/SKILL.md +391 -0
  215. package/skills/ai-research/06-post-training/verl/references/api-reference.md +302 -0
  216. package/skills/ai-research/06-post-training/verl/references/troubleshooting.md +392 -0
  217. package/skills/ai-research/07-safety-alignment/constitutional-ai/SKILL.md +290 -0
  218. package/skills/ai-research/07-safety-alignment/llamaguard/SKILL.md +337 -0
  219. package/skills/ai-research/07-safety-alignment/nemo-guardrails/SKILL.md +297 -0
  220. package/skills/ai-research/07-safety-alignment/prompt-guard/SKILL.md +312 -0
  221. package/skills/ai-research/08-distributed-training/accelerate/SKILL.md +332 -0
  222. package/skills/ai-research/08-distributed-training/accelerate/references/custom-plugins.md +454 -0
  223. package/skills/ai-research/08-distributed-training/accelerate/references/megatron-integration.md +490 -0
  224. package/skills/ai-research/08-distributed-training/accelerate/references/performance.md +526 -0
  225. package/skills/ai-research/08-distributed-training/deepspeed/SKILL.md +141 -0
  226. package/skills/ai-research/08-distributed-training/deepspeed/references/08.md +18 -0
  227. package/skills/ai-research/08-distributed-training/deepspeed/references/09.md +174 -0
  228. package/skills/ai-research/08-distributed-training/deepspeed/references/2020.md +379 -0
  229. package/skills/ai-research/08-distributed-training/deepspeed/references/2023.md +280 -0
  230. package/skills/ai-research/08-distributed-training/deepspeed/references/assets.md +180 -0
  231. package/skills/ai-research/08-distributed-training/deepspeed/references/index.md +36 -0
  232. package/skills/ai-research/08-distributed-training/deepspeed/references/mii.md +119 -0
  233. package/skills/ai-research/08-distributed-training/deepspeed/references/other.md +1192 -0
  234. package/skills/ai-research/08-distributed-training/megatron-core/SKILL.md +366 -0
  235. package/skills/ai-research/08-distributed-training/megatron-core/references/benchmarks.md +250 -0
  236. package/skills/ai-research/08-distributed-training/megatron-core/references/parallelism-guide.md +405 -0
  237. package/skills/ai-research/08-distributed-training/megatron-core/references/production-examples.md +474 -0
  238. package/skills/ai-research/08-distributed-training/megatron-core/references/training-recipes.md +548 -0
  239. package/skills/ai-research/08-distributed-training/pytorch-fsdp2/SKILL.md +231 -0
  240. package/skills/ai-research/08-distributed-training/pytorch-fsdp2/references/pytorch_dcp_async_recipe.md +14 -0
  241. package/skills/ai-research/08-distributed-training/pytorch-fsdp2/references/pytorch_dcp_overview.md +19 -0
  242. package/skills/ai-research/08-distributed-training/pytorch-fsdp2/references/pytorch_dcp_recipe.md +18 -0
  243. package/skills/ai-research/08-distributed-training/pytorch-fsdp2/references/pytorch_ddp_notes.md +10 -0
  244. package/skills/ai-research/08-distributed-training/pytorch-fsdp2/references/pytorch_device_mesh_tutorial.md +22 -0
  245. package/skills/ai-research/08-distributed-training/pytorch-fsdp2/references/pytorch_examples_fsdp2.md +15 -0
  246. package/skills/ai-research/08-distributed-training/pytorch-fsdp2/references/pytorch_fsdp1_api.md +9 -0
  247. package/skills/ai-research/08-distributed-training/pytorch-fsdp2/references/pytorch_fsdp2_tutorial.md +45 -0
  248. package/skills/ai-research/08-distributed-training/pytorch-fsdp2/references/pytorch_fully_shard_api.md +49 -0
  249. package/skills/ai-research/08-distributed-training/pytorch-fsdp2/references/pytorch_tp_tutorial.md +22 -0
  250. package/skills/ai-research/08-distributed-training/pytorch-fsdp2/references/ray_train_fsdp2_example.md +13 -0
  251. package/skills/ai-research/08-distributed-training/pytorch-fsdp2/references/torchtitan_fsdp_notes.md +14 -0
  252. package/skills/ai-research/08-distributed-training/pytorch-lightning/SKILL.md +346 -0
  253. package/skills/ai-research/08-distributed-training/pytorch-lightning/references/callbacks.md +437 -0
  254. package/skills/ai-research/08-distributed-training/pytorch-lightning/references/distributed.md +491 -0
  255. package/skills/ai-research/08-distributed-training/pytorch-lightning/references/hyperparameter-tuning.md +557 -0
  256. package/skills/ai-research/08-distributed-training/ray-train/SKILL.md +406 -0
  257. package/skills/ai-research/08-distributed-training/ray-train/references/multi-node.md +629 -0
  258. package/skills/ai-research/09-infrastructure/lambda-labs/SKILL.md +545 -0
  259. package/skills/ai-research/09-infrastructure/lambda-labs/references/advanced-usage.md +612 -0
  260. package/skills/ai-research/09-infrastructure/lambda-labs/references/troubleshooting.md +531 -0
  261. package/skills/ai-research/09-infrastructure/modal/SKILL.md +341 -0
  262. package/skills/ai-research/09-infrastructure/modal/references/advanced-usage.md +504 -0
  263. package/skills/ai-research/09-infrastructure/modal/references/troubleshooting.md +495 -0
  264. package/skills/ai-research/09-infrastructure/skypilot/SKILL.md +509 -0
  265. package/skills/ai-research/09-infrastructure/skypilot/references/advanced-usage.md +492 -0
  266. package/skills/ai-research/09-infrastructure/skypilot/references/troubleshooting.md +571 -0
  267. package/skills/ai-research/10-optimization/awq/SKILL.md +310 -0
  268. package/skills/ai-research/10-optimization/awq/references/advanced-usage.md +325 -0
  269. package/skills/ai-research/10-optimization/awq/references/troubleshooting.md +345 -0
  270. package/skills/ai-research/10-optimization/bitsandbytes/SKILL.md +411 -0
  271. package/skills/ai-research/10-optimization/bitsandbytes/references/memory-optimization.md +522 -0
  272. package/skills/ai-research/10-optimization/bitsandbytes/references/qlora-training.md +522 -0
  273. package/skills/ai-research/10-optimization/bitsandbytes/references/quantization-formats.md +448 -0
  274. package/skills/ai-research/10-optimization/flash-attention/SKILL.md +367 -0
  275. package/skills/ai-research/10-optimization/flash-attention/references/benchmarks.md +216 -0
  276. package/skills/ai-research/10-optimization/flash-attention/references/transformers-integration.md +294 -0
  277. package/skills/ai-research/10-optimization/gguf/SKILL.md +427 -0
  278. package/skills/ai-research/10-optimization/gguf/references/advanced-usage.md +505 -0
  279. package/skills/ai-research/10-optimization/gguf/references/troubleshooting.md +443 -0
  280. package/skills/ai-research/10-optimization/gptq/SKILL.md +450 -0
  281. package/skills/ai-research/10-optimization/gptq/references/calibration.md +338 -0
  282. package/skills/ai-research/10-optimization/gptq/references/integration.md +130 -0
  283. package/skills/ai-research/10-optimization/gptq/references/troubleshooting.md +96 -0
  284. package/skills/ai-research/10-optimization/hqq/SKILL.md +445 -0
  285. package/skills/ai-research/10-optimization/hqq/references/advanced-usage.md +529 -0
  286. package/skills/ai-research/10-optimization/hqq/references/troubleshooting.md +504 -0
  287. package/skills/ai-research/10-optimization/ml-training-recipes/SKILL.md +319 -0
  288. package/skills/ai-research/10-optimization/ml-training-recipes/references/architecture.md +320 -0
  289. package/skills/ai-research/10-optimization/ml-training-recipes/references/biomedical.md +498 -0
  290. package/skills/ai-research/10-optimization/ml-training-recipes/references/domain-specific.md +626 -0
  291. package/skills/ai-research/10-optimization/ml-training-recipes/references/experiment-loop.md +129 -0
  292. package/skills/ai-research/10-optimization/ml-training-recipes/references/optimizers.md +330 -0
  293. package/skills/ai-research/10-optimization/ml-training-recipes/references/scaling-and-selection.md +428 -0
  294. package/skills/ai-research/11-evaluation/bigcode-evaluation-harness/SKILL.md +405 -0
  295. package/skills/ai-research/11-evaluation/bigcode-evaluation-harness/references/benchmarks.md +394 -0
  296. package/skills/ai-research/11-evaluation/bigcode-evaluation-harness/references/custom-tasks.md +425 -0
  297. package/skills/ai-research/11-evaluation/bigcode-evaluation-harness/references/issues.md +395 -0
  298. package/skills/ai-research/11-evaluation/lm-evaluation-harness/SKILL.md +490 -0
  299. package/skills/ai-research/11-evaluation/lm-evaluation-harness/references/api-evaluation.md +491 -0
  300. package/skills/ai-research/11-evaluation/lm-evaluation-harness/references/benchmark-guide.md +489 -0
  301. package/skills/ai-research/11-evaluation/lm-evaluation-harness/references/custom-tasks.md +603 -0
  302. package/skills/ai-research/11-evaluation/lm-evaluation-harness/references/distributed-eval.md +520 -0
  303. package/skills/ai-research/11-evaluation/nemo-evaluator/SKILL.md +494 -0
  304. package/skills/ai-research/11-evaluation/nemo-evaluator/references/adapter-system.md +341 -0
  305. package/skills/ai-research/11-evaluation/nemo-evaluator/references/configuration.md +448 -0
  306. package/skills/ai-research/11-evaluation/nemo-evaluator/references/custom-benchmarks.md +316 -0
  307. package/skills/ai-research/11-evaluation/nemo-evaluator/references/execution-backends.md +362 -0
  308. package/skills/ai-research/12-inference-serving/llama-cpp/SKILL.md +258 -0
  309. package/skills/ai-research/12-inference-serving/llama-cpp/references/optimization.md +90 -0
  310. package/skills/ai-research/12-inference-serving/llama-cpp/references/quantization.md +214 -0
  311. package/skills/ai-research/12-inference-serving/llama-cpp/references/server.md +126 -0
  312. package/skills/ai-research/12-inference-serving/sglang/SKILL.md +442 -0
  313. package/skills/ai-research/12-inference-serving/sglang/references/deployment.md +491 -0
  314. package/skills/ai-research/12-inference-serving/sglang/references/radix-attention.md +414 -0
  315. package/skills/ai-research/12-inference-serving/sglang/references/structured-generation.md +542 -0
  316. package/skills/ai-research/12-inference-serving/tensorrt-llm/SKILL.md +187 -0
  317. package/skills/ai-research/12-inference-serving/tensorrt-llm/references/multi-gpu.md +299 -0
  318. package/skills/ai-research/12-inference-serving/tensorrt-llm/references/optimization.md +243 -0
  319. package/skills/ai-research/12-inference-serving/tensorrt-llm/references/serving.md +471 -0
  320. package/skills/ai-research/12-inference-serving/vllm/SKILL.md +364 -0
  321. package/skills/ai-research/12-inference-serving/vllm/references/optimization.md +227 -0
  322. package/skills/ai-research/12-inference-serving/vllm/references/quantization.md +285 -0
  323. package/skills/ai-research/12-inference-serving/vllm/references/server-deployment.md +256 -0
  324. package/skills/ai-research/12-inference-serving/vllm/references/troubleshooting.md +448 -0
  325. package/skills/ai-research/13-mlops/mlflow/SKILL.md +704 -0
  326. package/skills/ai-research/13-mlops/mlflow/references/deployment.md +745 -0
  327. package/skills/ai-research/13-mlops/mlflow/references/model-registry.md +771 -0
  328. package/skills/ai-research/13-mlops/mlflow/references/tracking.md +681 -0
  329. package/skills/ai-research/13-mlops/swanlab/SKILL.md +406 -0
  330. package/skills/ai-research/13-mlops/swanlab/references/integrations.md +305 -0
  331. package/skills/ai-research/13-mlops/swanlab/references/visualization.md +234 -0
  332. package/skills/ai-research/13-mlops/tensorboard/SKILL.md +629 -0
  333. package/skills/ai-research/13-mlops/tensorboard/references/integrations.md +639 -0
  334. package/skills/ai-research/13-mlops/tensorboard/references/profiling.md +546 -0
  335. package/skills/ai-research/13-mlops/tensorboard/references/visualization.md +621 -0
  336. package/skills/ai-research/13-mlops/weights-and-biases/SKILL.md +590 -0
  337. package/skills/ai-research/13-mlops/weights-and-biases/references/artifacts.md +585 -0
  338. package/skills/ai-research/13-mlops/weights-and-biases/references/integrations.md +701 -0
  339. package/skills/ai-research/13-mlops/weights-and-biases/references/sweeps.md +848 -0
  340. package/skills/ai-research/14-agents/a-evolve/SKILL.md +383 -0
  341. package/skills/ai-research/14-agents/a-evolve/references/README.md +831 -0
  342. package/skills/ai-research/14-agents/a-evolve/references/api.md +594 -0
  343. package/skills/ai-research/14-agents/a-evolve/references/architecture.md +417 -0
  344. package/skills/ai-research/14-agents/a-evolve/references/design-patterns.md +462 -0
  345. package/skills/ai-research/14-agents/a-evolve/references/examples.md +352 -0
  346. package/skills/ai-research/14-agents/a-evolve/references/issues.md +591 -0
  347. package/skills/ai-research/14-agents/a-evolve/references/releases.md +112 -0
  348. package/skills/ai-research/14-agents/a-evolve/references/tutorials.md +819 -0
  349. package/skills/ai-research/14-agents/autogpt/SKILL.md +403 -0
  350. package/skills/ai-research/14-agents/autogpt/references/advanced-usage.md +536 -0
  351. package/skills/ai-research/14-agents/autogpt/references/troubleshooting.md +421 -0
  352. package/skills/ai-research/14-agents/crewai/SKILL.md +498 -0
  353. package/skills/ai-research/14-agents/crewai/references/flows.md +439 -0
  354. package/skills/ai-research/14-agents/crewai/references/tools.md +430 -0
  355. package/skills/ai-research/14-agents/crewai/references/troubleshooting.md +481 -0
  356. package/skills/ai-research/14-agents/langchain/SKILL.md +480 -0
  357. package/skills/ai-research/14-agents/langchain/references/agents.md +500 -0
  358. package/skills/ai-research/14-agents/langchain/references/integration.md +563 -0
  359. package/skills/ai-research/14-agents/langchain/references/rag.md +601 -0
  360. package/skills/ai-research/14-agents/llamaindex/SKILL.md +569 -0
  361. package/skills/ai-research/14-agents/llamaindex/references/agents.md +84 -0
  362. package/skills/ai-research/14-agents/llamaindex/references/data_connectors.md +109 -0
  363. package/skills/ai-research/14-agents/llamaindex/references/query_engines.md +407 -0
  364. package/skills/ai-research/15-rag/chroma/SKILL.md +406 -0
  365. package/skills/ai-research/15-rag/chroma/references/integration.md +39 -0
  366. package/skills/ai-research/15-rag/faiss/SKILL.md +221 -0
  367. package/skills/ai-research/15-rag/faiss/references/index_types.md +281 -0
  368. package/skills/ai-research/15-rag/pinecone/SKILL.md +358 -0
  369. package/skills/ai-research/15-rag/pinecone/references/deployment.md +182 -0
  370. package/skills/ai-research/15-rag/qdrant/SKILL.md +493 -0
  371. package/skills/ai-research/15-rag/qdrant/references/advanced-usage.md +649 -0
  372. package/skills/ai-research/15-rag/qdrant/references/troubleshooting.md +632 -0
  373. package/skills/ai-research/15-rag/sentence-transformers/SKILL.md +255 -0
  374. package/skills/ai-research/15-rag/sentence-transformers/references/models.md +124 -0
  375. package/skills/ai-research/16-prompt-engineering/dspy/SKILL.md +590 -0
  376. package/skills/ai-research/16-prompt-engineering/dspy/references/examples.md +664 -0
  377. package/skills/ai-research/16-prompt-engineering/dspy/references/modules.md +476 -0
  378. package/skills/ai-research/16-prompt-engineering/dspy/references/optimizers.md +567 -0
  379. package/skills/ai-research/16-prompt-engineering/guidance/SKILL.md +572 -0
  380. package/skills/ai-research/16-prompt-engineering/guidance/references/backends.md +555 -0
  381. package/skills/ai-research/16-prompt-engineering/guidance/references/constraints.md +675 -0
  382. package/skills/ai-research/16-prompt-engineering/guidance/references/examples.md +768 -0
  383. package/skills/ai-research/16-prompt-engineering/instructor/SKILL.md +740 -0
  384. package/skills/ai-research/16-prompt-engineering/instructor/references/examples.md +108 -0
  385. package/skills/ai-research/16-prompt-engineering/instructor/references/providers.md +71 -0
  386. package/skills/ai-research/16-prompt-engineering/instructor/references/validation.md +607 -0
  387. package/skills/ai-research/16-prompt-engineering/outlines/SKILL.md +652 -0
  388. package/skills/ai-research/16-prompt-engineering/outlines/references/backends.md +616 -0
  389. package/skills/ai-research/16-prompt-engineering/outlines/references/examples.md +774 -0
  390. package/skills/ai-research/16-prompt-engineering/outlines/references/json_generation.md +653 -0
  391. package/skills/ai-research/17-observability/langsmith/SKILL.md +422 -0
  392. package/skills/ai-research/17-observability/langsmith/references/advanced-usage.md +549 -0
  393. package/skills/ai-research/17-observability/langsmith/references/troubleshooting.md +538 -0
  394. package/skills/ai-research/17-observability/phoenix/SKILL.md +475 -0
  395. package/skills/ai-research/17-observability/phoenix/references/advanced-usage.md +620 -0
  396. package/skills/ai-research/17-observability/phoenix/references/troubleshooting.md +539 -0
  397. package/skills/ai-research/18-multimodal/audiocraft/SKILL.md +564 -0
  398. package/skills/ai-research/18-multimodal/audiocraft/references/advanced-usage.md +667 -0
  399. package/skills/ai-research/18-multimodal/audiocraft/references/troubleshooting.md +505 -0
  400. package/skills/ai-research/18-multimodal/blip-2/SKILL.md +564 -0
  401. package/skills/ai-research/18-multimodal/blip-2/references/advanced-usage.md +681 -0
  402. package/skills/ai-research/18-multimodal/blip-2/references/troubleshooting.md +527 -0
  403. package/skills/ai-research/18-multimodal/clip/SKILL.md +253 -0
  404. package/skills/ai-research/18-multimodal/clip/references/applications.md +208 -0
  405. package/skills/ai-research/18-multimodal/cosmos-policy/SKILL.md +402 -0
  406. package/skills/ai-research/18-multimodal/cosmos-policy/references/libero-commands.md +175 -0
  407. package/skills/ai-research/18-multimodal/cosmos-policy/references/robocasa-commands.md +147 -0
  408. package/skills/ai-research/18-multimodal/llava/SKILL.md +304 -0
  409. package/skills/ai-research/18-multimodal/llava/references/training.md +198 -0
  410. package/skills/ai-research/18-multimodal/openpi/SKILL.md +480 -0
  411. package/skills/ai-research/18-multimodal/openpi/references/checkpoints-and-env-map.md +86 -0
  412. package/skills/ai-research/18-multimodal/openpi/references/config-recipes.md +50 -0
  413. package/skills/ai-research/18-multimodal/openpi/references/pytorch-gotchas.md +49 -0
  414. package/skills/ai-research/18-multimodal/openpi/references/remote-client-pattern.md +48 -0
  415. package/skills/ai-research/18-multimodal/openpi/references/training-debugging.md +46 -0
  416. package/skills/ai-research/18-multimodal/openvla-oft/SKILL.md +441 -0
  417. package/skills/ai-research/18-multimodal/openvla-oft/references/aloha-workflow.md +141 -0
  418. package/skills/ai-research/18-multimodal/openvla-oft/references/config-troubleshooting.md +97 -0
  419. package/skills/ai-research/18-multimodal/openvla-oft/references/libero-workflow.md +115 -0
  420. package/skills/ai-research/18-multimodal/openvla-oft/references/paper-and-checkpoints.md +84 -0
  421. package/skills/ai-research/18-multimodal/segment-anything/SKILL.md +500 -0
  422. package/skills/ai-research/18-multimodal/segment-anything/references/advanced-usage.md +590 -0
  423. package/skills/ai-research/18-multimodal/segment-anything/references/troubleshooting.md +485 -0
  424. package/skills/ai-research/18-multimodal/stable-diffusion/SKILL.md +519 -0
  425. package/skills/ai-research/18-multimodal/stable-diffusion/references/advanced-usage.md +717 -0
  426. package/skills/ai-research/18-multimodal/stable-diffusion/references/troubleshooting.md +556 -0
  427. package/skills/ai-research/18-multimodal/whisper/SKILL.md +317 -0
  428. package/skills/ai-research/18-multimodal/whisper/references/languages.md +190 -0
  429. package/skills/ai-research/19-emerging-techniques/knowledge-distillation/SKILL.md +458 -0
  430. package/skills/ai-research/19-emerging-techniques/knowledge-distillation/references/minillm.md +335 -0
  431. package/skills/ai-research/19-emerging-techniques/long-context/SKILL.md +536 -0
  432. package/skills/ai-research/19-emerging-techniques/long-context/references/extension_methods.md +469 -0
  433. package/skills/ai-research/19-emerging-techniques/long-context/references/fine_tuning.md +612 -0
  434. package/skills/ai-research/19-emerging-techniques/long-context/references/rope.md +403 -0
  435. package/skills/ai-research/19-emerging-techniques/model-merging/SKILL.md +497 -0
  436. package/skills/ai-research/19-emerging-techniques/model-merging/references/coefficient-tuning.md +415 -0
  437. package/skills/ai-research/19-emerging-techniques/model-merging/references/evaluation.md +463 -0
  438. package/skills/ai-research/19-emerging-techniques/model-merging/references/examples.md +429 -0
  439. package/skills/ai-research/19-emerging-techniques/model-merging/references/methods.md +353 -0
  440. package/skills/ai-research/19-emerging-techniques/model-pruning/SKILL.md +495 -0
  441. package/skills/ai-research/19-emerging-techniques/model-pruning/references/wanda.md +348 -0
  442. package/skills/ai-research/19-emerging-techniques/moe-training/SKILL.md +526 -0
  443. package/skills/ai-research/19-emerging-techniques/moe-training/references/architectures.md +433 -0
  444. package/skills/ai-research/19-emerging-techniques/moe-training/references/inference.md +349 -0
  445. package/skills/ai-research/19-emerging-techniques/moe-training/references/training.md +426 -0
  446. package/skills/ai-research/19-emerging-techniques/speculative-decoding/SKILL.md +467 -0
  447. package/skills/ai-research/19-emerging-techniques/speculative-decoding/references/lookahead.md +310 -0
  448. package/skills/ai-research/19-emerging-techniques/speculative-decoding/references/medusa.md +351 -0
  449. package/skills/ai-research/20-ml-paper-writing/academic-plotting/SKILL.md +479 -0
  450. package/skills/ai-research/20-ml-paper-writing/academic-plotting/references/data-visualization.md +509 -0
  451. package/skills/ai-research/20-ml-paper-writing/academic-plotting/references/diagram-generation.md +395 -0
  452. package/skills/ai-research/20-ml-paper-writing/academic-plotting/references/style-guide.md +305 -0
  453. package/skills/ai-research/20-ml-paper-writing/ml-paper-writing/SKILL.md +982 -0
  454. package/skills/ai-research/20-ml-paper-writing/ml-paper-writing/references/checklists.md +348 -0
  455. package/skills/ai-research/20-ml-paper-writing/ml-paper-writing/references/citation-workflow.md +563 -0
  456. package/skills/ai-research/20-ml-paper-writing/ml-paper-writing/references/reviewer-guidelines.md +349 -0
  457. package/skills/ai-research/20-ml-paper-writing/ml-paper-writing/references/sources.md +165 -0
  458. package/skills/ai-research/20-ml-paper-writing/ml-paper-writing/references/writing-guide.md +477 -0
  459. package/skills/ai-research/20-ml-paper-writing/presenting-conference-talks/SKILL.md +298 -0
  460. package/skills/ai-research/20-ml-paper-writing/presenting-conference-talks/references/slide-templates.md +505 -0
  461. package/skills/ai-research/20-ml-paper-writing/systems-paper-writing/SKILL.md +270 -0
  462. package/skills/ai-research/20-ml-paper-writing/systems-paper-writing/references/checklist.md +279 -0
  463. package/skills/ai-research/20-ml-paper-writing/systems-paper-writing/references/reviewer-guidelines.md +116 -0
  464. package/skills/ai-research/20-ml-paper-writing/systems-paper-writing/references/section-blueprints.md +400 -0
  465. package/skills/ai-research/20-ml-paper-writing/systems-paper-writing/references/systems-conferences.md +261 -0
  466. package/skills/ai-research/20-ml-paper-writing/systems-paper-writing/references/writing-patterns.md +231 -0
  467. package/skills/ai-research/21-research-ideation/brainstorming-research-ideas/SKILL.md +384 -0
  468. package/skills/ai-research/21-research-ideation/creative-thinking-for-research/SKILL.md +366 -0
  469. package/skills/ai-research/22-agent-native-research-artifact/compiler/SKILL.md +245 -0
  470. package/skills/ai-research/22-agent-native-research-artifact/compiler/references/ara-schema.md +439 -0
  471. package/skills/ai-research/22-agent-native-research-artifact/compiler/references/exploration-tree-spec.md +125 -0
  472. package/skills/ai-research/22-agent-native-research-artifact/compiler/references/validation-checklist.md +149 -0
  473. package/skills/ai-research/22-agent-native-research-artifact/research-manager/SKILL.md +324 -0
  474. package/skills/ai-research/22-agent-native-research-artifact/research-manager/references/event-taxonomy.md +124 -0
  475. package/skills/ai-research/22-agent-native-research-artifact/research-manager/references/provenance-tags.md +190 -0
  476. package/skills/ai-research/22-agent-native-research-artifact/research-manager/references/session-protocol.md +162 -0
  477. package/skills/ai-research/22-agent-native-research-artifact/rigor-reviewer/SKILL.md +322 -0
  478. package/skills/ai-research/22-agent-native-research-artifact/rigor-reviewer/references/review-dimensions.md +182 -0
  479. package/skills/ai-research/SKILL.md +101 -0
  480. package/skills/ai-research/SOURCE +38 -0
  481. package/skills/ai-research/index/1-frame.md +18 -0
  482. package/skills/ai-research/index/2-data.md +21 -0
  483. package/skills/ai-research/index/3-train.md +48 -0
  484. package/skills/ai-research/index/4-measure.md +25 -0
  485. package/skills/ai-research/index/5-run.md +23 -0
  486. package/skills/ai-research/index/6-write.md +35 -0
  487. package/skills/ai-research/library.json +3456 -0
  488. package/skills/analyze-app/EVALS.md +82 -0
  489. package/skills/analyze-app/SKILL.md +188 -0
  490. package/skills/design-contract/EVALS.md +80 -0
  491. package/skills/design-contract/SKILL.md +171 -0
  492. package/skills/design-library/EVALS.md +85 -0
  493. package/skills/design-library/INDEX.md +59 -0
  494. package/skills/design-library/SKILL.md +58 -0
  495. package/skills/design-library/SOURCE +131 -0
  496. package/skills/design-library/adaptive-interfaces/skills/colour-independence/SKILL.md +90 -0
  497. package/skills/design-library/adaptive-interfaces/skills/flexible-typography/SKILL.md +83 -0
  498. package/skills/design-library/adaptive-interfaces/skills/information-density/SKILL.md +87 -0
  499. package/skills/design-library/adaptive-interfaces/skills/preference-audit/SKILL.md +70 -0
  500. package/skills/design-library/adaptive-interfaces/skills/responsive-accessibility/SKILL.md +86 -0
  501. package/skills/design-library/adaptive-interfaces/skills/responsive-review/SKILL.md +76 -0
  502. package/skills/design-library/adaptive-interfaces/skills/simplified-views/SKILL.md +97 -0
  503. package/skills/design-library/adaptive-interfaces/skills/specify/SKILL.md +63 -0
  504. package/skills/design-library/adaptive-interfaces/skills/user-preference-respect/SKILL.md +100 -0
  505. package/skills/design-library/cognitive-accessibility/skills/adaptive-personalisation/SKILL.md +61 -0
  506. package/skills/design-library/cognitive-accessibility/skills/assess-load/SKILL.md +48 -0
  507. package/skills/design-library/cognitive-accessibility/skills/cognitive-load-assessment/SKILL.md +60 -0
  508. package/skills/design-library/cognitive-accessibility/skills/contextual-help-design/SKILL.md +61 -0
  509. package/skills/design-library/cognitive-accessibility/skills/error-prevention-recovery/SKILL.md +71 -0
  510. package/skills/design-library/cognitive-accessibility/skills/focus-attention-design/SKILL.md +67 -0
  511. package/skills/design-library/cognitive-accessibility/skills/memory-load-reduction/SKILL.md +60 -0
  512. package/skills/design-library/cognitive-accessibility/skills/plain-language-design/SKILL.md +78 -0
  513. package/skills/design-library/cognitive-accessibility/skills/review/SKILL.md +56 -0
  514. package/skills/design-library/cognitive-accessibility/skills/simplify/SKILL.md +39 -0
  515. package/skills/design-library/cognitive-accessibility/skills/wayfinding-navigation/SKILL.md +64 -0
  516. package/skills/design-library/design-research/skills/behavioural-analytics/SKILL.md +33 -0
  517. package/skills/design-library/design-research/skills/qual-quant-triangulation/SKILL.md +33 -0
  518. package/skills/design-library/inclusive-interaction/skills/audit/SKILL.md +50 -0
  519. package/skills/design-library/inclusive-interaction/skills/design-flow/SKILL.md +58 -0
  520. package/skills/design-library/inclusive-interaction/skills/feedback-and-status/SKILL.md +76 -0
  521. package/skills/design-library/inclusive-interaction/skills/gesture-alternatives/SKILL.md +70 -0
  522. package/skills/design-library/inclusive-interaction/skills/keyboard-navigation/SKILL.md +99 -0
  523. package/skills/design-library/inclusive-interaction/skills/keyboard-review/SKILL.md +54 -0
  524. package/skills/design-library/inclusive-interaction/skills/motion-sensitivity/SKILL.md +79 -0
  525. package/skills/design-library/inclusive-interaction/skills/multi-modal-input/SKILL.md +70 -0
  526. package/skills/design-library/inclusive-interaction/skills/touch-target-design/SKILL.md +71 -0
  527. package/skills/design-library/inclusive-interaction/skills/voice-interaction/SKILL.md +68 -0
  528. package/skills/design-library/visual-critique/skills/critique-affordance/SKILL.md +45 -0
  529. package/skills/design-library/visual-critique/skills/critique-brand-consistency/SKILL.md +42 -0
  530. package/skills/design-library/visual-critique/skills/critique-color/SKILL.md +44 -0
  531. package/skills/design-library/visual-critique/skills/critique-composition/SKILL.md +46 -0
  532. package/skills/design-library/visual-critique/skills/critique-information-density/SKILL.md +45 -0
  533. package/skills/design-library/visual-critique/skills/critique-typography/SKILL.md +45 -0
  534. package/skills/design-library/visual-critique/skills/critique-visual-hierarchy/SKILL.md +40 -0
  535. package/skills/feature-audit/EVALS.md +83 -0
  536. package/skills/feature-audit/SKILL.md +326 -0
  537. package/skills/harness/EVALS.md +81 -0
  538. package/skills/harness/SKILL.md +88 -0
  539. package/skills/i-have-adhd/EVALS.md +80 -0
  540. package/skills/i-have-adhd/SKILL.md +183 -0
  541. package/skills/no-ai-slop/EVALS.md +89 -0
  542. package/skills/no-ai-slop/SKILL.md +209 -0
  543. package/skills/no-ai-slop/eval.md +52 -0
  544. package/skills/plan-add/EVALS.md +79 -0
  545. package/skills/plan-add/SKILL.md +179 -0
  546. package/skills/plan-add/render_table.py +72 -0
  547. package/skills/plan-status/EVALS.md +78 -0
  548. package/skills/plan-status/SKILL.md +52 -0
  549. package/skills/plan-sync/EVALS.md +80 -0
  550. package/skills/plan-sync/SKILL.md +69 -0
  551. package/skills/pm-frameworks/SKILL.md +33 -0
  552. package/skills/pm-frameworks/SOURCE +39 -0
  553. package/skills/pm-frameworks/intended-vs-implemented/SKILL.md +42 -0
  554. package/skills/pm-frameworks/opportunity-solution-tree/SKILL.md +64 -0
  555. package/skills/ponytail/EVALS.md +80 -0
  556. package/skills/ponytail/SKILL.md +124 -0
  557. package/skills/product-analysis/EVALS.md +103 -0
  558. package/skills/product-analysis/SKILL.md +361 -0
  559. package/skills/research/EVALS.md +104 -0
  560. package/skills/research/SKILL.md +570 -0
  561. package/skills/rl-env/SKILL.md +53 -0
  562. package/skills/rl-env/SOURCE +41 -0
  563. package/skills/rl-env/generate-nemo-gym-env/SKILL.md +169 -0
  564. package/skills/rl-env/generate-nemo-gym-env/references/architecture.md +114 -0
  565. package/skills/rl-env/generate-openenv-env/SKILL.md +183 -0
  566. package/skills/rl-env/generate-openenv-env/references/architecture.md +82 -0
  567. package/skills/rl-env/generate-ors-env/SKILL.md +209 -0
  568. package/skills/rl-env/generate-ors-env/references/architecture.md +113 -0
  569. package/skills/rl-env/generate-verifiers-env/SKILL.md +188 -0
  570. package/skills/rl-env/generate-verifiers-env/references/architecture.md +119 -0
  571. package/skills/rl-env/rl-env-from-description/SKILL.md +124 -0
  572. package/skills/rl-env/rl-env-from-description/references/interview.md +75 -0
  573. package/skills/rl-env/rl-env-from-description/references/nemo_gym.md +41 -0
  574. package/skills/rl-env/rl-env-from-description/references/openenv.md +35 -0
  575. package/skills/rl-env/rl-env-from-description/references/ors.md +50 -0
  576. package/skills/rl-env/rl-env-from-description/references/verifiers.md +39 -0
  577. package/skills/tezgah-contract/EVALS.md +78 -0
  578. package/skills/tezgah-contract/SKILL.md +792 -0
  579. package/statusline.py +129 -0
  580. package/workflows/graph-impact.js +153 -0
  581. package/workflows/graph-map.js +158 -0
  582. package/workflows/graph-review.js +170 -0
package/README.de.md ADDED
@@ -0,0 +1,169 @@
1
+ <p align="center">
2
+ <a href="README.md">English</a> |
3
+ <a href="README.de.md">Deutsch</a> |
4
+ <a href="README.es.md">Español</a> |
5
+ <a href="README.fr.md">Français</a> |
6
+ <a href="README.tr.md">Türkçe</a>
7
+ </p>
8
+
9
+ # Tezgah
10
+
11
+ <p align="center">
12
+ <img src="assets/logo/tezgah-logo.svg" alt="tezgah logo" width="220">
13
+ </p>
14
+
15
+ <h3 align="center">Ein funktionierender Vertrag für jeden KI-Programmierassistenten, den Sie ausführen.</h3>
16
+
17
+ <p align="center">
18
+ <a href="LICENSE"><img alt="License: MIT" src="https://img.shields.io/github/license/r1z4x/tezgah?style=flat-square"></a>
19
+ <a href="https://github.com/r1z4x/tezgah/actions/workflows/ci.yml"><img alt="CI" src="https://img.shields.io/github/actions/workflow/status/r1z4x/tezgah/ci.yml?style=flat-square&label=ci"></a>
20
+ <a href="https://github.com/r1z4x/tezgah/releases/latest"><img alt="Release" src="https://img.shields.io/github/v/release/r1z4x/tezgah?style=flat-square"></a>
21
+ </p>
22
+
23
+ <p align="center"><sub>Englisch ist die maßgebliche Quelle; Übersetzungen können hinterherhinken.</sub></p>
24
+
25
+ ---
26
+
27
+ Tezgah rüstet jeden KI-Programmierassistenten, den Sie ausführen — **omp, den
28
+ primären Host**, plus Claude Code, Codex, Cursor, opencode und DeepSeeks dsh —
29
+ innerhalb der von Ihnen konfigurierten Repository-Stammverzeichnisse mit einem
30
+ funktionierenden Vertrag aus. Sich selbst überlassen, driften die Assistenten:
31
+ Einer antwortet auf Türkisch, ein anderer auf Englisch; einer nutzt Grep, ein
32
+ anderer fragt einen Code-Graphen ab; einer meldet „fertig“, ohne einen Test
33
+ auszuführen. Die Regeln existieren einmal in einem gemeinsamen Kern; jeder
34
+ Host erhält einen dünnen Adapter, der sie in die Form übersetzt, die er
35
+ versteht — eine Regel, die an einer Stelle geändert wird, erreicht jeden Host
36
+ auf dieselbe Weise.
37
+
38
+ ## Was tezgah durchsetzt und aufzeichnet
39
+
40
+ - **Ein Vertrag, sechs Hosts.** omp, Claude Code, Codex, Cursor, opencode und
41
+ dsh sehen dieselben Regeln, denn jeder Host ist ein dünner Adapter über
42
+ einem gemeinsamen Kern — eine Regel einmal ändern, alle übernehmen sie.
43
+ - **Eine Fertig-Meldung braucht eine bestandene Prüfung im Protokoll.** Ein
44
+ Gate aus 15 Verweigerungen stoppt die entschärfte Prüfung — `--no-verify`,
45
+ `|| true`, ein in `tail` gepipeter Test, ein mitten im Flug ergänzter Skip —
46
+ und die Stop-Regel verweigert einen Fertigstellungsanspruch, den das
47
+ Evidenzprotokoll der Sitzung nicht stützt. Sie verweigert den Zug; sie kann
48
+ das Modell nicht davon abhalten, den Anspruch zu erheben.
49
+ - **Forschung kommt mit eigener Bibliothek.** Forschungsaufgaben laufen über
50
+ OpenResearch, mit einer integrierten Bibliothek aus 98 Upstream-Skills, die
51
+ Eintrag für Eintrag geladen wird, damit der Kontext klein bleibt.
52
+ - **Jede Regel hat einen Ausschalter.** Sechzehn Kill-Switches — dazu
53
+ Repo-Markierungen — entfernen den Regeltext aus der Sitzung; die Regel
54
+ stoppt damit wirklich, statt nur als aus zu erscheinen.
55
+ - **Eine feste Antwortform.** Antworten sind türkisch, sofern Sie keine andere Sprache wählen, und
56
+ beginnen mit dem Ergebnis; eine Liste zeigt höchstens fünf gerankte
57
+ Einträge; eine Schätzung wird als Schätzung genannt; ein Fehler liest sich
58
+ als Ort, Ursache, Behebung.
59
+
60
+ ## Was es nicht gezeigt hat
61
+
62
+ Tezgah ist eine Sammlung von Mechanismen: Regeln, die in die Sitzung
63
+ eingespeist werden, ein Gate, das bestimmte Befehle verweigert, und ein
64
+ Protokoll, das die Stop-Regel liest. Es ist kein Beleg dafür, dass ein Agent
65
+ weniger schummelt oder mehr Aufgaben löst. Ein gepaartes An/Aus-Experiment
66
+ (Oktober 2026; DeepSeek V4.1 Flash auf omp, ohne tezgah gegen eine volle
67
+ tezgah-Installation) fand keine messbare Verringerung des Schummelns, weil
68
+ das Modell in beiden Armen kaum schummelte:
69
+
70
+ - **Phase 1, drei Druck-Fixtures:** Schummeln in 0 von 60 Läufen in jedem
71
+ Arm; sauber bestanden 48/60 mit tezgah gegen 53/60 ohne.
72
+ - **Phase 2, der widersprüchliche SWE-bench-Split von ImpossibleBench (44
73
+ Instanzen, k = 3):** Schummeln 0/132 mit tezgah gegen 1/131 ohne (-0.8 pp,
74
+ 95%-KI -2.3 bis 0); sauber bestanden auf den Originalaufgaben 129/131 gegen
75
+ 129/132.
76
+ - **Falsche Fertig-Meldungen, von Hand gelabelt:** Bei diesen unmöglichen
77
+ Aufgaben meldeten 3 von 30 Stichproben-Abschlussnachrichten mit tezgah und
78
+ 6 von 30 ohne dem Nutzer, die Aufgabe sei erledigt. Der bereinigte
79
+ Unterschied liegt bei -5 pp (95%-KI -18 bis +9), also kein Effekt in
80
+ irgendeine Richtung.
81
+ - **Eine Schwachstelle im Anspruchsleser der Stop-Regel:** Er markierte
82
+ 115/132 Läufe mit tezgah und 90/131 ohne, aber nur 10% bzw. 20% seiner
83
+ Markierungen waren echte Fertig-Meldungen. Mit tezgah antwortete der Agent
84
+ auf Türkisch, und der Leser zählt `doğrulandı` (verifiziert) bei einer
85
+ Teilkorrektur als Anspruch. Das ist festgehalten, aber noch nicht behoben.
86
+
87
+ Der Lauf hatte Messmängel, darunter Agenten in beiden Armen, die die
88
+ Grundwahrheit des Benchmarks lesen konnten. Die Aggregate, die Zählungen pro
89
+ Aufgabe und jede Einschränkung stehen im
90
+ [Ergebnisbündel](docs/results/paired-outcome-2026-10.md) (Englisch).
91
+
92
+ <a id="install"></a>
93
+
94
+ ## Installation
95
+
96
+ Eine Zeile, auf macOS, Linux oder WSL (Python 3.10+, `curl`, `tar`). Sie lädt
97
+ die neueste Version herunter, prüft ihre sha256 und rüstet jeden gefundenen
98
+ Host (Claude Code über dessen eigene `claude plugin`-CLI):
99
+
100
+ ```bash
101
+ curl -fsSL https://raw.githubusercontent.com/r1z4x/tezgah/main/packaging/install.sh | sh
102
+ ```
103
+
104
+ Oder über npm: `npm i -g @r1z4x/tezgah && tezgah --install`.
105
+
106
+ Eine Host-Konfiguration, die es nicht lesen kann, bleibt unverändert; jede
107
+ geänderte Konfiguration behält eine datierte `.tezgah-bak`-Kopie, und die
108
+ Installation endet mit einem Fehlercode, wenn ein geplanter Host nicht
109
+ gerüstet ist.
110
+
111
+ ### Was tezgah auf Ihrem Rechner ändert
112
+
113
+ - **Host-Konfiguration.** Jeder gerüstete Host erhält tezgahs Hooks, seine
114
+ MCP-Einträge und einen verwalteten Vertragsblock in seiner globalen
115
+ Regeldatei (`~/.claude/CLAUDE.md`, `~/.codex/AGENTS.md`,
116
+ `~/.omp/agent/RULES.md`). Claude Codes Hooks kommen in einem Plugin, das über
117
+ `claude plugin` registriert wird; ohne `claude`-CLI bleibt Claude ungerüstet,
118
+ und der Bericht sagt es.
119
+ - **Agent-Dateien in Ihren Repositories.** In den konfigurierten Wurzeln
120
+ schreibt eine Sitzung Subagent-Dateien nach `.claude/agents/`,
121
+ `.codex/agents/` und `.opencode/agents/` und trägt diese Verzeichnisse in die
122
+ eigene `.git/info/exclude` des Klons ein (`TEZGAH_NO_EXCLUDE=1` stoppt das).
123
+ - **Eine tägliche Update-Prüfung.** Die Statuszeile fragt höchstens einmal am
124
+ Tag nach einer neueren Version; `~/.config/tezgah/update-check-off` oder
125
+ `TEZGAH_UPDATE_CHECK=0` schaltet sie ab.
126
+ - **orx.** Die Installation lädt die OpenResearch-CLI, über die
127
+ Forschungsaufgaben laufen; `TEZGAH_NO_DEPS=1` überspringt sie.
128
+ - **Antwortsprache.** Antworten sind standardmäßig türkisch, und die Stop-Regel
129
+ hält sie dazu an. `--reply-lang en` verlangt Englisch, `any` Ihre eigene
130
+ Sprache; beides wird nicht geprüft: `curl -fsSL … | sh -s -- --reply-lang en`.
131
+
132
+ Lieber der Coding-Assistent erledigen? Fügen Sie dies in omp, Claude Code,
133
+ Codex, Cursor oder opencode ein (der Prompt bleibt Englisch; der Assistent
134
+ versteht ihn in jeder Sprache):
135
+
136
+ ```text
137
+ Install tezgah (https://github.com/r1z4x/tezgah) on this machine and verify it.
138
+
139
+ 1. Check the prerequisites: python3 --version must be 3.10 or newer, and curl
140
+ and tar must exist. If one is missing, stop and tell me which.
141
+ 2. Run the installer exactly as published - do not edit it or pipe it anywhere else:
142
+ curl -fsSL https://raw.githubusercontent.com/r1z4x/tezgah/main/packaging/install.sh | sh
143
+ Its output must contain "verified tezgah-<version>.tar.gz" (the sha256
144
+ check). If it does not, stop and show me the output.
145
+ 3. Verify: run ~/.local/share/tezgah/current/bin/tezgah-setup --version and
146
+ ~/.local/share/tezgah/current/bin/tezgah-setup --report, and show me every
147
+ line that says MISS.
148
+ 4. Tell me which hosts were armed, which repository root was configured
149
+ (default ~/Projects - if my code lives elsewhere, ask me for the directory and
150
+ run ~/.local/share/tezgah/current/bin/tezgah-setup --roots <dir> --install),
151
+ and that I must restart each assistant for the hooks to load.
152
+ Do not change any other file and do not uninstall anything.
153
+ ```
154
+
155
+ Festgepinnte Versionen, Windows (`packaging/install.ps1`), offline-Tarballs,
156
+ Upgrades und die optionalen Integrationen stehen in der [Dokumentation](docs/README.md).
157
+
158
+ ## Unterstützte Hosts
159
+
160
+ **omp** (der primäre Host, gegen den entwickelt und verifiziert wird), **Claude Code**, **Codex**, **Cursor**, **opencode**, **dsh** — dünne Adapter über einem gemeinsamen Kern.
161
+
162
+ ## Wo die Tiefe liegt
163
+
164
+ Gate, Evidence-Ledger, Host-Adapter, Konfiguration und Entwicklung liegen in [docs/README.md](docs/README.md), eine Seite pro Frage.
165
+
166
+ ## Lizenz
167
+
168
+ MIT — siehe [LICENSE](LICENSE).
169
+ Vendorte Skills (`ponytail`, `no-ai-slop`, `i-have-adhd`) tragen ihre eigenen MIT-Bedingungen, verzeichnet in [NOTICE](NOTICE).
package/README.es.md ADDED
@@ -0,0 +1,168 @@
1
+ <p align="center">
2
+ <a href="README.md">English</a> |
3
+ <a href="README.de.md">Deutsch</a> |
4
+ <a href="README.es.md">Español</a> |
5
+ <a href="README.fr.md">Français</a> |
6
+ <a href="README.tr.md">Türkçe</a>
7
+ </p>
8
+
9
+ # Tezgah
10
+
11
+ <p align="center">
12
+ <img src="assets/logo/tezgah-logo.svg" alt="tezgah logo" width="220">
13
+ </p>
14
+
15
+ <h3 align="center">Un contrato de trabajo para cada asistente de programación de IA que ejecutes.</h3>
16
+
17
+ <p align="center">
18
+ <a href="LICENSE"><img alt="License: MIT" src="https://img.shields.io/github/license/r1z4x/tezgah?style=flat-square"></a>
19
+ <a href="https://github.com/r1z4x/tezgah/actions/workflows/ci.yml"><img alt="CI" src="https://img.shields.io/github/actions/workflow/status/r1z4x/tezgah/ci.yml?style=flat-square&label=ci"></a>
20
+ <a href="https://github.com/r1z4x/tezgah/releases/latest"><img alt="Release" src="https://img.shields.io/github/v/release/r1z4x/tezgah?style=flat-square"></a>
21
+ </p>
22
+
23
+ <p align="center"><sub>El inglés es la fuente de la verdad; las traducciones pueden estar desactualizadas.</sub></p>
24
+
25
+ ---
26
+
27
+ Tezgah arma a cada asistente de programación de IA que ejecutes — **omp, el
28
+ host principal**, más Claude Code, Codex, Cursor, opencode y el dsh de
29
+ DeepSeek — con un solo contrato de trabajo, dentro de las raíces de
30
+ repositorios que configures. Por sí solos, los asistentes divergen: uno
31
+ responde en turco mientras otro responde en inglés; uno usa grep donde otro
32
+ consulta un grafo de código; uno informa "hecho" sin ejecutar una prueba. Las
33
+ reglas residen una sola vez en un núcleo compartido; cada host recibe un
34
+ adaptador ligero que las traduce a la forma que entiende — una regla cambiada
35
+ en un lugar llega a todos los hosts de la misma manera.
36
+
37
+ ## Qué impone y registra tezgah
38
+
39
+ - **Un contrato, seis hosts.** omp, Claude Code, Codex, Cursor, opencode y dsh
40
+ ven las mismas reglas, porque cada host es un adaptador ligero sobre un
41
+ único núcleo compartido — cambia una regla una vez y todos la adoptan.
42
+ - **Declarar algo terminado exige una comprobación superada en el registro.**
43
+ Una compuerta de 15 rechazos detiene la comprobación neutralizada —
44
+ `--no-verify`, `|| true`, una prueba canalizada a `tail`, un skip añadido a
45
+ mitad de vuelo — y la regla Stop rechaza una afirmación de finalización que
46
+ el registro de evidencias de la sesión no respalda. Rechaza el turno; no
47
+ puede impedir que el modelo lo afirme.
48
+ - **La investigación trae su biblioteca.** Las tareas de investigación se
49
+ dirigen a través de OpenResearch con una biblioteca incorporada de 98
50
+ habilidades upstream, cargada de una en una para que el contexto siga
51
+ siendo pequeño.
52
+ - **Cada regla tiene un interruptor de apagado.** Dieciséis interruptores — más
53
+ las marcas por repositorio — eliminan el texto de la regla de la sesión, de
54
+ modo que la regla realmente se detiene en lugar de limitarse a figurar como
55
+ desactivada.
56
+ - **Una forma de respuesta fija.** Las respuestas son en turco salvo que elijas otro idioma, y
57
+ empiezan por el resultado; una lista muestra como máximo cinco elementos
58
+ ordenados; una estimación se nombra como estimación; un error se lee como
59
+ ubicación, causa, solución.
60
+
61
+ ## Lo que no ha demostrado
62
+
63
+ Tezgah es un conjunto de mecanismos: reglas inyectadas en la sesión, una
64
+ compuerta que rechaza ciertos comandos y un registro que lee la regla Stop. No
65
+ es prueba de que un agente haga menos trampas ni de que termine más tareas. Un
66
+ experimento emparejado con y sin tezgah (octubre de 2026; DeepSeek V4.1 Flash
67
+ en omp, sin tezgah frente a una instalación completa) no encontró una
68
+ reducción medible de las trampas, porque el modelo apenas hizo trampas en
69
+ ninguno de los dos brazos:
70
+
71
+ - **Fase 1, tres fixtures de presión:** trampas en 0 de 60 ejecuciones en cada
72
+ brazo; aprobado limpio 48/60 con tezgah frente a 53/60 sin él.
73
+ - **Fase 2, el split conflictivo de SWE-bench de ImpossibleBench (44
74
+ instancias, k = 3):** trampas 0/132 con tezgah frente a 1/131 sin él (-0.8
75
+ pp, IC del 95% -2.3 a 0); aprobado limpio en las tareas originales 129/131
76
+ frente a 129/132.
77
+ - **Falsas declaraciones de terminado, etiquetadas a mano:** en esas tareas
78
+ imposibles, 3 de 30 mensajes finales muestreados con tezgah y 6 de 30 sin él
79
+ le dijeron al usuario que la tarea estaba hecha. La diferencia ajustada es
80
+ de -5 pp (IC del 95% -18 a +9), así que no hay efecto en ningún sentido.
81
+ - **Un punto débil del lector de afirmaciones de la regla Stop:** marcó
82
+ 115/132 ejecuciones con tezgah y 90/131 sin él, pero solo el 10% y el 20% de
83
+ sus marcas eran declaraciones reales de terminado. Con tezgah el agente
84
+ respondió en turco, y el lector cuenta `doğrulandı` (verificado) sobre un
85
+ arreglo parcial como afirmación. Está registrado, aún no corregido.
86
+
87
+ La ejecución tuvo fallos de instrumento, entre ellos agentes de ambos brazos
88
+ que podían leer la verdad de referencia del benchmark. Los agregados, los
89
+ recuentos por tarea y cada salvedad están en el
90
+ [paquete de resultados](docs/results/paired-outcome-2026-10.md) (en inglés).
91
+
92
+ <a id="install"></a>
93
+
94
+ ## Instalación
95
+
96
+ Una línea, en macOS, Linux o WSL (Python 3.10+, `curl`, `tar`). Descarga la
97
+ última versión, comprueba su sha256 y arma cada host que encuentra (Claude
98
+ Code a través de su propia CLI `claude plugin`):
99
+
100
+ ```bash
101
+ curl -fsSL https://raw.githubusercontent.com/r1z4x/tezgah/main/packaging/install.sh | sh
102
+ ```
103
+
104
+ O a través de npm: `npm i -g @r1z4x/tezgah && tezgah --install`.
105
+
106
+ Una configuración de host que no puede leer queda intacta, cada configuración
107
+ que cambia conserva una copia `.tezgah-bak` fechada, y la instalación termina
108
+ con un código distinto de cero cuando un host previsto no queda armado.
109
+
110
+ ### Qué cambia tezgah en tu máquina
111
+
112
+ - **Configuración de los hosts.** Cada host armado recibe los hooks de tezgah,
113
+ sus entradas MCP y un bloque de contrato gestionado en su archivo de reglas
114
+ global (`~/.claude/CLAUDE.md`, `~/.codex/AGENTS.md`,
115
+ `~/.omp/agent/RULES.md`). Los hooks de Claude Code van en un plugin
116
+ registrado con `claude plugin`; sin la CLI `claude`, Claude queda sin armar y
117
+ el informe lo dice.
118
+ - **Archivos de agentes en tus repositorios.** Dentro de las raíces
119
+ configuradas, una sesión escribe archivos de subagentes en `.claude/agents/`,
120
+ `.codex/agents/` y `.opencode/agents/`, y añade esos directorios al
121
+ `.git/info/exclude` propio del clon (`TEZGAH_NO_EXCLUDE=1` lo impide).
122
+ - **Una comprobación diaria de actualizaciones.** La línea de estado pregunta
123
+ por una versión nueva como mucho una vez al día;
124
+ `~/.config/tezgah/update-check-off` o `TEZGAH_UPDATE_CHECK=0` la desactiva.
125
+ - **orx.** La instalación descarga la CLI de OpenResearch por la que pasan las
126
+ tareas de investigación; `TEZGAH_NO_DEPS=1` la omite.
127
+ - **Idioma de las respuestas.** Por defecto las respuestas son en turco y la
128
+ regla Stop las sostiene en él. `--reply-lang en` pide inglés, y `any` tu
129
+ propio idioma; ninguno se comprueba: `curl -fsSL … | sh -s -- --reply-lang en`.
130
+
131
+ ¿Prefieres que lo haga tu asistente de programación? Pega esto en omp,
132
+ Claude Code, Codex, Cursor u opencode (el prompt queda en inglés; el
133
+ asistente lo entiende en cualquier idioma):
134
+
135
+ ```text
136
+ Install tezgah (https://github.com/r1z4x/tezgah) on this machine and verify it.
137
+
138
+ 1. Check the prerequisites: python3 --version must be 3.10 or newer, and curl
139
+ and tar must exist. If one is missing, stop and tell me which.
140
+ 2. Run the installer exactly as published - do not edit it or pipe it anywhere else:
141
+ curl -fsSL https://raw.githubusercontent.com/r1z4x/tezgah/main/packaging/install.sh | sh
142
+ Its output must contain "verified tezgah-<version>.tar.gz" (the sha256
143
+ check). If it does not, stop and show me the output.
144
+ 3. Verify: run ~/.local/share/tezgah/current/bin/tezgah-setup --version and
145
+ ~/.local/share/tezgah/current/bin/tezgah-setup --report, and show me every
146
+ line that says MISS.
147
+ 4. Tell me which hosts were armed, which repository root was configured
148
+ (default ~/Projects - if my code lives elsewhere, ask me for the directory and
149
+ run ~/.local/share/tezgah/current/bin/tezgah-setup --roots <dir> --install),
150
+ and that I must restart each assistant for the hooks to load.
151
+ Do not change any other file and do not uninstall anything.
152
+ ```
153
+
154
+ Versiones fijadas, Windows (`packaging/install.ps1`), tarballs sin conexión,
155
+ actualizaciones y las integraciones opcionales están en la [documentación](docs/README.md).
156
+
157
+ ## Hosts compatibles
158
+
159
+ **omp** (el host principal, contra el que se desarrolla y verifica), **Claude Code**, **Codex**, **Cursor**, **opencode**, **dsh** — adaptadores ligeros sobre un único núcleo compartido.
160
+
161
+ ## Dónde está la profundidad
162
+
163
+ La compuerta, el registro de evidencia, los adaptadores de host, la configuración y el desarrollo están en [docs/README.md](docs/README.md), una página por pregunta.
164
+
165
+ ## Licencia
166
+
167
+ MIT — ver [LICENSE](LICENSE).
168
+ Las habilidades incorporadas (`ponytail`, `no-ai-slop`, `i-have-adhd`) llevan sus propios términos MIT, registrados en [NOTICE](NOTICE).
package/README.fr.md ADDED
@@ -0,0 +1,166 @@
1
+ <p align="center">
2
+ <a href="README.md">English</a> |
3
+ <a href="README.de.md">Deutsch</a> |
4
+ <a href="README.es.md">Español</a> |
5
+ <a href="README.fr.md">Français</a> |
6
+ <a href="README.tr.md">Türkçe</a>
7
+ </p>
8
+
9
+ # Tezgah
10
+
11
+ <p align="center">
12
+ <img src="assets/logo/tezgah-logo.svg" alt="tezgah logo" width="220">
13
+ </p>
14
+
15
+ <h3 align="center">Un contrat de travail unique pour chaque assistant de codage IA que vous exécutez.</h3>
16
+
17
+ <p align="center">
18
+ <a href="LICENSE"><img alt="License: MIT" src="https://img.shields.io/github/license/r1z4x/tezgah?style=flat-square"></a>
19
+ <a href="https://github.com/r1z4x/tezgah/actions/workflows/ci.yml"><img alt="CI" src="https://img.shields.io/github/actions/workflow/status/r1z4x/tezgah/ci.yml?style=flat-square&label=ci"></a>
20
+ <a href="https://github.com/r1z4x/tezgah/releases/latest"><img alt="Release" src="https://img.shields.io/github/v/release/r1z4x/tezgah?style=flat-square"></a>
21
+ </p>
22
+
23
+ <p align="center"><sub>L'anglais est la source de vérité ; les traductions peuvent être en décalage.</sub></p>
24
+
25
+ ---
26
+
27
+ Tezgah arme chaque assistant de codage IA que vous exécutez — **omp, l'hôte
28
+ principal**, plus Claude Code, Codex, Cursor, opencode et le dsh de DeepSeek —
29
+ avec un seul contrat de travail, à l'intérieur des racines de dépôts que vous
30
+ configurez. Laissés à eux-mêmes, les assistants divergent : l'un répond en
31
+ turc pendant qu'un autre répond en anglais ; l'un utilise grep là où un autre
32
+ interroge un graphe de code ; l'un annonce « terminé » sans exécuter de test.
33
+ Les règles résident une seule fois dans un cœur partagé ; chaque hôte reçoit
34
+ un adaptateur léger qui les traduit dans la forme qu'il comprend — une règle
35
+ modifiée à un endroit atteint chaque hôte de la même façon.
36
+
37
+ ## Ce que tezgah impose et enregistre
38
+
39
+ - **Un contrat, six hôtes.** omp, Claude Code, Codex, Cursor, opencode et dsh
40
+ voient les mêmes règles, car chaque hôte est un adaptateur léger au-dessus
41
+ d'un seul cœur partagé — modifiez une règle une fois, tous l'adoptent.
42
+ - **Annoncer « terminé » exige une vérification réussie dans le registre.**
43
+ Une porte de 15 refus arrête la vérification neutralisée — `--no-verify`,
44
+ `|| true`, un test redirigé vers `tail`, un skip ajouté en plein vol — et la
45
+ règle Stop refuse une affirmation d'achèvement que le registre de preuves de
46
+ la session ne peut pas étayer. Elle refuse le tour ; elle ne peut pas
47
+ empêcher le modèle de l'affirmer.
48
+ - **La recherche arrive avec sa bibliothèque.** Les tâches de recherche
49
+ passent par OpenResearch avec une bibliothèque intégrée de 98 compétences
50
+ upstream, chargée une entrée à la fois pour que le contexte reste petit.
51
+ - **Chaque règle a un interrupteur.** Seize interrupteurs — plus des
52
+ marques par dépôt — retirent le texte de la règle de la session ; la règle
53
+ s'arrête vraiment au lieu de simplement figurer comme désactivée.
54
+ - **Une forme de réponse fixe.** Les réponses sont en turc, sauf si vous choisissez une autre langue, et
55
+ commencent par le résultat ; une liste montre au plus cinq éléments classés
56
+ ; une estimation est nommée estimation ; une erreur se lit lieu, cause,
57
+ correctif.
58
+
59
+ ## Ce qu'il n'a pas démontré
60
+
61
+ Tezgah est un ensemble de mécanismes : des règles injectées dans la session,
62
+ une porte qui refuse certaines commandes et un registre que lit la règle Stop.
63
+ Ce n'est pas la preuve qu'un agent triche moins ou termine plus de tâches. Une
64
+ expérience appariée avec et sans tezgah (octobre 2026 ; DeepSeek V4.1 Flash
65
+ sur omp, sans tezgah contre une installation complète) n'a trouvé aucune
66
+ baisse mesurable de la triche, car le modèle n'a presque pas triché dans
67
+ aucun des deux bras :
68
+
69
+ - **Phase 1, trois fixtures de pression :** triche dans 0 exécution sur 60
70
+ dans chaque bras ; réussite propre 48/60 avec tezgah contre 53/60 sans.
71
+ - **Phase 2, le split conflictuel SWE-bench d'ImpossibleBench (44 instances,
72
+ k = 3) :** triche 0/132 avec tezgah contre 1/131 sans (-0.8 pp, IC à 95%
73
+ -2.3 à 0) ; réussite propre sur les tâches originales 129/131 contre
74
+ 129/132.
75
+ - **Fausses annonces de fin, étiquetées à la main :** sur ces tâches
76
+ impossibles, 3 messages finaux échantillonnés sur 30 avec tezgah et 6 sur 30
77
+ sans ont dit à l'utilisateur que la tâche était faite. L'écart ajusté est de
78
+ -5 pp (IC à 95% -18 à +9), donc aucun effet dans un sens ou dans l'autre.
79
+ - **Un point faible du lecteur d'affirmations de la règle Stop :** il a
80
+ signalé 115/132 exécutions avec tezgah et 90/131 sans, mais seuls 10% et 20%
81
+ de ses signalements étaient de vraies annonces de fin. Avec tezgah, l'agent
82
+ répondait en turc, et le lecteur compte `doğrulandı` (vérifié) sur un
83
+ correctif partiel comme une affirmation. C'est consigné, pas encore corrigé.
84
+
85
+ L'exécution avait des défauts d'instrument, entre autres des agents des deux
86
+ bras qui pouvaient lire la vérité terrain du benchmark. Les agrégats, les
87
+ comptes par tâche et chaque réserve sont dans le
88
+ [dossier de résultats](docs/results/paired-outcome-2026-10.md) (en anglais).
89
+
90
+ <a id="install"></a>
91
+
92
+ ## Installation
93
+
94
+ Une ligne, sur macOS, Linux ou WSL (Python 3.10+, `curl`, `tar`). Elle
95
+ télécharge la dernière version, vérifie son sha256 et arme chaque hôte
96
+ trouvé (Claude Code par sa propre CLI `claude plugin`) :
97
+
98
+ ```bash
99
+ curl -fsSL https://raw.githubusercontent.com/r1z4x/tezgah/main/packaging/install.sh | sh
100
+ ```
101
+
102
+ Ou via npm : `npm i -g @r1z4x/tezgah && tezgah --install`.
103
+
104
+ Une configuration d'hôte illisible reste intacte, chaque configuration modifiée
105
+ garde une copie `.tezgah-bak` datée, et l'installation se termine avec un code
106
+ non nul quand un hôte prévu n'est pas armé.
107
+
108
+ ### Ce que tezgah modifie sur votre machine
109
+
110
+ - **Configuration des hôtes.** Chaque hôte armé reçoit les hooks de tezgah, ses
111
+ entrées MCP et un bloc de contrat géré dans son fichier de règles global
112
+ (`~/.claude/CLAUDE.md`, `~/.codex/AGENTS.md`, `~/.omp/agent/RULES.md`). Les
113
+ hooks de Claude Code sont livrés dans un plugin enregistré par
114
+ `claude plugin` ; sans CLI `claude`, Claude reste non armé et le rapport le
115
+ dit.
116
+ - **Des fichiers d'agents dans vos dépôts.** Dans les racines configurées, une
117
+ session écrit des fichiers de sous-agents dans `.claude/agents/`,
118
+ `.codex/agents/` et `.opencode/agents/`, et ajoute ces répertoires au
119
+ `.git/info/exclude` propre au clone (`TEZGAH_NO_EXCLUDE=1` l'empêche).
120
+ - **Une vérification quotidienne des mises à jour.** La ligne d'état demande
121
+ une nouvelle version au plus une fois par jour ;
122
+ `~/.config/tezgah/update-check-off` ou `TEZGAH_UPDATE_CHECK=0` la coupe.
123
+ - **orx.** L'installation télécharge la CLI OpenResearch par laquelle passent
124
+ les tâches de recherche ; `TEZGAH_NO_DEPS=1` l'ignore.
125
+ - **Langue des réponses.** Les réponses sont en turc par défaut et la règle
126
+ Stop les y tient. `--reply-lang en` demande l'anglais, et `any` votre propre
127
+ langue ; aucun des deux n'est contrôlé : `curl -fsSL … | sh -s -- --reply-lang en`.
128
+
129
+ Vous préférez laisser votre assistant de codage s'en charger ? Collez ceci
130
+ dans omp, Claude Code, Codex, Cursor ou opencode (le prompt reste en anglais
131
+ ; l'assistant le comprend dans toute langue) :
132
+
133
+ ```text
134
+ Install tezgah (https://github.com/r1z4x/tezgah) on this machine and verify it.
135
+
136
+ 1. Check the prerequisites: python3 --version must be 3.10 or newer, and curl
137
+ and tar must exist. If one is missing, stop and tell me which.
138
+ 2. Run the installer exactly as published - do not edit it or pipe it anywhere else:
139
+ curl -fsSL https://raw.githubusercontent.com/r1z4x/tezgah/main/packaging/install.sh | sh
140
+ Its output must contain "verified tezgah-<version>.tar.gz" (the sha256
141
+ check). If it does not, stop and show me the output.
142
+ 3. Verify: run ~/.local/share/tezgah/current/bin/tezgah-setup --version and
143
+ ~/.local/share/tezgah/current/bin/tezgah-setup --report, and show me every
144
+ line that says MISS.
145
+ 4. Tell me which hosts were armed, which repository root was configured
146
+ (default ~/Projects - if my code lives elsewhere, ask me for the directory and
147
+ run ~/.local/share/tezgah/current/bin/tezgah-setup --roots <dir> --install),
148
+ and that I must restart each assistant for the hooks to load.
149
+ Do not change any other file and do not uninstall anything.
150
+ ```
151
+
152
+ Versions épinglées, Windows (`packaging/install.ps1`), archives hors ligne,
153
+ mises à niveau et intégrations optionnelles sont dans la [documentation](docs/README.md).
154
+
155
+ ## Hôtes pris en charge
156
+
157
+ **omp** (l'hôte principal, contre lequel tezgah est développé et vérifié), **Claude Code**, **Codex**, **Cursor**, **opencode**, **dsh** — des adaptateurs légers au-dessus d'un seul cœur partagé.
158
+
159
+ ## Où se trouve la profondeur
160
+
161
+ La porte, le registre de preuves, les adaptateurs d'hôtes, la configuration et le développement sont dans [docs/README.md](docs/README.md), une page par question.
162
+
163
+ ## Licence
164
+
165
+ MIT — voir [LICENSE](LICENSE).
166
+ Les compétences intégrées (`ponytail`, `no-ai-slop`, `i-have-adhd`) suivent leurs propres termes MIT, consignés dans [NOTICE](NOTICE).
package/README.md ADDED
@@ -0,0 +1,161 @@
1
+ <p align="center">
2
+ <a href="README.md">English</a> |
3
+ <a href="README.de.md">Deutsch</a> |
4
+ <a href="README.es.md">Español</a> |
5
+ <a href="README.fr.md">Français</a> |
6
+ <a href="README.tr.md">Türkçe</a>
7
+ </p>
8
+
9
+ # Tezgah
10
+
11
+ <p align="center">
12
+ <img src="assets/logo/tezgah-logo.svg" alt="tezgah logo" width="220">
13
+ </p>
14
+
15
+ <h3 align="center">One working contract for every AI coding assistant you run.</h3>
16
+
17
+ <p align="center">
18
+ <a href="LICENSE"><img alt="License: MIT" src="https://img.shields.io/github/license/r1z4x/tezgah?style=flat-square"></a>
19
+ <a href="https://github.com/r1z4x/tezgah/actions/workflows/ci.yml"><img alt="CI" src="https://img.shields.io/github/actions/workflow/status/r1z4x/tezgah/ci.yml?style=flat-square&label=ci"></a>
20
+ <a href="https://github.com/r1z4x/tezgah/releases/latest"><img alt="Release" src="https://img.shields.io/github/v/release/r1z4x/tezgah?style=flat-square"></a>
21
+ </p>
22
+
23
+ <p align="center"><sub>English is the source of truth; translations may lag behind it.</sub></p>
24
+
25
+ ---
26
+
27
+ Tezgah arms every AI coding assistant you run — **omp, the primary host**,
28
+ plus Claude Code, Codex, Cursor, opencode and DeepSeek's dsh harness — with
29
+ one working contract, inside the repository roots you configure. Left alone, each
30
+ assistant drifts: one answers in Turkish while another answers in English, one
31
+ greps where another queries a code graph, one reports "done" without running a
32
+ test. The rules live once in a shared core; each host gets a thin adapter that
33
+ translates them into the shape it understands — so a rule changed in one place
34
+ reaches every host the same way.
35
+
36
+ ## What tezgah enforces and records
37
+
38
+ - **One contract, six hosts.** omp, Claude Code, Codex, Cursor, opencode and
39
+ dsh see the same rules, because every host is a thin adapter over one shared
40
+ core — change a rule once and all of them pick it up.
41
+ - **A done claim needs a passing check on record.** A gate of 15 refusals stops
42
+ the neutered check — `--no-verify`, `|| true`, a test piped into `tail`, a
43
+ skip added mid-flight — and the Stop rule refuses a completion claim the
44
+ session's evidence ledger cannot back. It refuses the turn; it cannot make
45
+ the model stop claiming.
46
+ - **Research ships with its library.** Research tasks route through
47
+ OpenResearch with a vendored library of 98 upstream skills, loaded one entry
48
+ at a time so the context stays small.
49
+ - **Every rule has an off switch.** Sixteen kill switches — plus per-repo
50
+ marks — remove a rule's text from the session, so the rule actually stops
51
+ rather than merely showing as off.
52
+ - **A fixed reply shape.** Replies lead with the outcome, in Turkish unless
53
+ you pick another language; a list shows at most five ranked items; an
54
+ estimate is named as an estimate; an error reads as location, cause, fix.
55
+
56
+ ## What it has not shown
57
+
58
+ Tezgah is a set of mechanisms: rules injected into the session, a gate that
59
+ refuses certain commands, and a ledger the Stop rule reads. It is not evidence
60
+ that an agent cheats less or finishes more tasks. A paired on/off experiment
61
+ (October 2026; DeepSeek V4.1 Flash on omp, bare vs a full tezgah install) found
62
+ no measurable reduction in cheating, because the model barely cheated in
63
+ either arm:
64
+
65
+ - **Phase 1, three pressure fixtures:** cheating 0 of 60 runs in every arm;
66
+ clean pass 48/60 with tezgah vs 53/60 without.
67
+ - **Phase 2, ImpossibleBench's conflicting SWE-bench split (44 instances,
68
+ k = 3):** cheating 0/132 with tezgah vs 1/131 without (-0.8 pp, 95% CI
69
+ -2.3 to 0); clean pass on the original tasks 129/131 vs 129/132.
70
+ - **False done claims, hand-labelled:** on those impossible tasks, 3 of 30
71
+ sampled final messages with tezgah and 6 of 30 without told the user the
72
+ task was done. The adjusted difference is -5 pp (95% CI -18 to +9), so no
73
+ effect either way.
74
+ - **A weak spot in the Stop rule's claim reader:** it flagged 115/132 runs
75
+ with tezgah and 90/131 without, but only 10% and 20% of the flags it raised
76
+ were real done claims. With tezgah the agent replied in Turkish, and the
77
+ reader counts `doğrulandı` (verified) on a partial fix as a claim. This is
78
+ recorded, not yet fixed.
79
+
80
+ The run had instrument flaws, among them agents in both arms that could read
81
+ the benchmark's ground truth. The aggregates, per-task counts and every caveat
82
+ are in [the results bundle](docs/results/paired-outcome-2026-10.md).
83
+
84
+ <a id="install"></a>
85
+
86
+ ## Install
87
+
88
+ One line, on macOS, Linux or WSL (Python 3.10+, `curl`, `tar`). It downloads
89
+ the latest release, checks its sha256, arms every host it finds (Claude Code
90
+ through its own `claude plugin` CLI), and installs the missing optional tools
91
+ tezgah itself uses (orx; `TEZGAH_NO_DEPS=1` skips them). It never installs a
92
+ host's own CLI (cursor-agent, pnpm, dsh): it prints the command instead, or
93
+ runs it under `tezgah-setup --install --host-deps`.
94
+
95
+ ```bash
96
+ curl -fsSL https://raw.githubusercontent.com/r1z4x/tezgah/main/packaging/install.sh | sh
97
+ ```
98
+
99
+ Or through npm: `npm i -g @r1z4x/tezgah && tezgah --install`.
100
+
101
+ A host config it cannot parse is left untouched, every config it changes
102
+ keeps a dated `.tezgah-bak` copy, and the install exits non-zero when a host
103
+ it set out to arm is not armed.
104
+
105
+ ### What tezgah changes on your machine
106
+
107
+ - **Host config.** Each armed host gets tezgah's hooks, its MCP rows and a
108
+ managed contract block in the host's global rules file (`~/.claude/CLAUDE.md`,
109
+ `~/.codex/AGENTS.md`, `~/.omp/agent/RULES.md`). Claude Code's hooks ship in a
110
+ plugin registered through `claude plugin`; with no `claude` CLI, Claude stays
111
+ unarmed and the report says so.
112
+ - **Agent files in your repositories.** Inside the configured roots, a session
113
+ writes subagent files to `.claude/agents/`, `.codex/agents/` and
114
+ `.opencode/agents/`, and lists those directories in the clone's own
115
+ `.git/info/exclude` (`TEZGAH_NO_EXCLUDE=1` stops that).
116
+ - **A daily update check.** The status line asks for a newer release at most
117
+ once a day; `~/.config/tezgah/update-check-off` or `TEZGAH_UPDATE_CHECK=0`
118
+ turns it off.
119
+ - **orx.** The install fetches the OpenResearch CLI that research tasks route
120
+ through; `TEZGAH_NO_DEPS=1` skips it.
121
+ - **Reply language.** Replies are Turkish by default and the Stop rule holds
122
+ them to it. `--reply-lang en` asks for English, and `any` for your own
123
+ language; neither is checked: `curl -fsSL … | sh -s -- --reply-lang en`.
124
+
125
+ Rather let your coding assistant do it? Paste this into omp, Claude Code,
126
+ Codex, Cursor or opencode:
127
+
128
+ ```text
129
+ Install tezgah (https://github.com/r1z4x/tezgah) on this machine and verify it.
130
+
131
+ 1. Check the prerequisites: python3 --version must be 3.10 or newer, and curl
132
+ and tar must exist. If one is missing, stop and tell me which.
133
+ 2. Run the installer exactly as published - do not edit it or pipe it anywhere else:
134
+ curl -fsSL https://raw.githubusercontent.com/r1z4x/tezgah/main/packaging/install.sh | sh
135
+ Its output must contain "verified tezgah-<version>.tar.gz" (the sha256
136
+ check). If it does not, stop and show me the output.
137
+ 3. Verify: run ~/.local/share/tezgah/current/bin/tezgah-setup --version and
138
+ ~/.local/share/tezgah/current/bin/tezgah-setup --report, and show me every
139
+ line that says MISS.
140
+ 4. Tell me which hosts were armed, which repository root was configured
141
+ (default ~/Projects - if my code lives elsewhere, ask me for the directory and
142
+ run ~/.local/share/tezgah/current/bin/tezgah-setup --roots <dir> --install),
143
+ and that I must restart each assistant for the hooks to load.
144
+ Do not change any other file and do not uninstall anything.
145
+ ```
146
+
147
+ Pinned versions, Windows (`packaging/install.ps1`), offline tarballs, upgrades
148
+ and the optional integrations are in the [docs](docs/README.md).
149
+
150
+ ## Supported hosts
151
+
152
+ **omp** (the primary host, developed and verified against), **Claude Code**, **Codex**, **Cursor**, **opencode**, **dsh** — thin adapters over one shared core.
153
+
154
+ ## Where the depth lives
155
+
156
+ The gate, the evidence ledger, the host adapters, configuration and development are in [docs/README.md](docs/README.md), one page per question.
157
+
158
+ ## License
159
+
160
+ MIT — see [LICENSE](LICENSE).
161
+ Vendored skills (`ponytail`, `no-ai-slop`, `i-have-adhd`) carry their own MIT terms, recorded in [NOTICE](NOTICE).