synth-ai 0.2.9.dev5__py3-none-any.whl → 0.2.10__py3-none-any.whl

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.

Potentially problematic release.


This version of synth-ai might be problematic. Click here for more details.

Files changed (349) hide show
  1. examples/__init__.py +16 -0
  2. examples/crafter_debug_render.py +23 -17
  3. examples/dev/qwen3_32b_qlora_4xh100.toml +40 -0
  4. examples/multi_step/crafter_rl_lora.md +29 -0
  5. examples/qwen_coder/README.md +102 -0
  6. examples/qwen_coder/_shared.py +113 -0
  7. examples/qwen_coder/configs/coder_lora_30b.toml +61 -0
  8. examples/qwen_coder/configs/coder_lora_4b.toml +57 -0
  9. examples/qwen_coder/configs/coder_lora_small.toml +58 -0
  10. examples/qwen_coder/generate_dataset.py +98 -0
  11. examples/qwen_coder/infer_ft_smoke.py +65 -0
  12. examples/qwen_coder/infer_prod_proxy.py +73 -0
  13. examples/qwen_coder/infer_via_synth.py +87 -0
  14. examples/qwen_coder/scripts/infer_coder.sh +19 -0
  15. examples/qwen_coder/scripts/train_coder_30b.sh +22 -0
  16. examples/qwen_coder/sft_full_17b.py +103 -0
  17. examples/qwen_coder/sft_lora_30b.py +110 -0
  18. examples/qwen_coder/subset_jsonl.py +39 -0
  19. examples/qwen_coder/todos.md +38 -0
  20. examples/qwen_coder/validate_jsonl.py +60 -0
  21. examples/rl/configs/eval_base_qwen.toml +1 -1
  22. examples/rl/configs/rl_from_base_qwen17.toml +1 -1
  23. examples/rl/download_dataset.py +26 -10
  24. examples/rl/run_eval.py +53 -52
  25. examples/rl/run_rl_and_save.py +29 -12
  26. examples/rl/task_app/math_single_step.py +180 -41
  27. examples/rl/task_app/math_task_app.py +14 -6
  28. examples/sft/README.md +139 -0
  29. examples/sft/configs/crafter_fft_qwen0p6b.toml +44 -0
  30. examples/sft/configs/crafter_lora_qwen0p6b.toml +45 -0
  31. examples/sft/evaluate.py +117 -0
  32. examples/sft/export_dataset.py +117 -0
  33. examples/sft/generate_traces.py +162 -0
  34. examples/swe/__init__.py +12 -0
  35. examples/swe/task_app/README.md +105 -0
  36. examples/swe/task_app/__init__.py +2 -0
  37. examples/swe/task_app/grpo_swe_mini.py +571 -0
  38. examples/swe/task_app/grpo_swe_mini_task_app.py +136 -0
  39. examples/swe/task_app/hosted/README.md +173 -0
  40. examples/swe/task_app/hosted/__init__.py +5 -0
  41. examples/swe/task_app/hosted/branching.py +143 -0
  42. examples/swe/task_app/hosted/environment_routes.py +1289 -0
  43. examples/swe/task_app/hosted/envs/__init__.py +1 -0
  44. examples/swe/task_app/hosted/envs/crafter/__init__.py +6 -0
  45. examples/swe/task_app/hosted/envs/crafter/app.py +1 -0
  46. examples/swe/task_app/hosted/envs/crafter/environment.py +522 -0
  47. examples/swe/task_app/hosted/envs/crafter/policy.py +478 -0
  48. examples/swe/task_app/hosted/envs/crafter/react_agent.py +108 -0
  49. examples/swe/task_app/hosted/envs/crafter/shared.py +305 -0
  50. examples/swe/task_app/hosted/envs/crafter/tools.py +47 -0
  51. examples/swe/task_app/hosted/envs/mini_swe/__init__.py +8 -0
  52. examples/swe/task_app/hosted/envs/mini_swe/environment.py +1164 -0
  53. examples/swe/task_app/hosted/envs/mini_swe/policy.py +355 -0
  54. examples/swe/task_app/hosted/envs/mini_swe/shared.py +83 -0
  55. examples/swe/task_app/hosted/envs/mini_swe/tools.py +96 -0
  56. examples/swe/task_app/hosted/hosted_app.py +204 -0
  57. examples/swe/task_app/hosted/inference/__init__.py +5 -0
  58. examples/swe/task_app/hosted/inference/openai_client.py +618 -0
  59. examples/swe/task_app/hosted/main.py +100 -0
  60. examples/swe/task_app/hosted/policy_routes.py +1079 -0
  61. examples/swe/task_app/hosted/registry.py +195 -0
  62. examples/swe/task_app/hosted/rollout.py +1869 -0
  63. examples/swe/task_app/hosted/storage/__init__.py +5 -0
  64. examples/swe/task_app/hosted/storage/volume.py +211 -0
  65. examples/swe/task_app/hosted/test_agents.py +161 -0
  66. examples/swe/task_app/hosted/test_service.py +137 -0
  67. examples/swe/task_app/hosted/utils.py +62 -0
  68. examples/vlm/PROPOSAL.md +53 -0
  69. examples/vlm/README.md +68 -0
  70. examples/vlm/configs/crafter_vlm_gpt4o.toml +44 -0
  71. examples/vlm/crafter_image_only_agent.py +207 -0
  72. examples/vlm/crafter_openai_vlm_agent.py +277 -0
  73. examples/vlm/filter_image_rows.py +63 -0
  74. examples/vlm/run_crafter_vlm_benchmark.py +316 -0
  75. examples/warming_up_to_rl/analyze_trace_db.py +12 -10
  76. examples/warming_up_to_rl/configs/rl_from_base_qwen4b.toml +11 -1
  77. examples/warming_up_to_rl/export_trace_sft.py +218 -36
  78. examples/warming_up_to_rl/groq_test.py +15 -8
  79. examples/warming_up_to_rl/manage_secrets.py +29 -25
  80. examples/warming_up_to_rl/readme.md +9 -2
  81. examples/warming_up_to_rl/run_eval.py +137 -61
  82. examples/warming_up_to_rl/run_fft_and_save.py +131 -60
  83. examples/warming_up_to_rl/run_local_rollout.py +88 -39
  84. examples/warming_up_to_rl/run_local_rollout_modal.py +114 -28
  85. examples/warming_up_to_rl/run_local_rollout_parallel.py +81 -20
  86. examples/warming_up_to_rl/run_local_rollout_traced.py +126 -23
  87. examples/warming_up_to_rl/run_rl_and_save.py +35 -12
  88. examples/warming_up_to_rl/run_rollout_remote.py +44 -19
  89. examples/warming_up_to_rl/task_app/README.md +6 -2
  90. examples/warming_up_to_rl/task_app/grpo_crafter.py +319 -57
  91. examples/warming_up_to_rl/task_app/grpo_crafter_task_app.py +11 -30
  92. examples/warming_up_to_rl/task_app/synth_envs_hosted/__init__.py +1 -1
  93. examples/warming_up_to_rl/task_app/synth_envs_hosted/branching.py +9 -11
  94. examples/warming_up_to_rl/task_app/synth_envs_hosted/environment_routes.py +137 -182
  95. examples/warming_up_to_rl/task_app/synth_envs_hosted/envs/__init__.py +1 -1
  96. examples/warming_up_to_rl/task_app/synth_envs_hosted/envs/crafter/__init__.py +1 -1
  97. examples/warming_up_to_rl/task_app/synth_envs_hosted/envs/crafter/app.py +1 -1
  98. examples/warming_up_to_rl/task_app/synth_envs_hosted/envs/crafter/environment.py +150 -57
  99. examples/warming_up_to_rl/task_app/synth_envs_hosted/envs/crafter/policy.py +105 -69
  100. examples/warming_up_to_rl/task_app/synth_envs_hosted/envs/crafter/react_agent.py +19 -7
  101. examples/warming_up_to_rl/task_app/synth_envs_hosted/envs/crafter/shared.py +45 -42
  102. examples/warming_up_to_rl/task_app/synth_envs_hosted/envs/crafter/tools.py +1 -1
  103. examples/warming_up_to_rl/task_app/synth_envs_hosted/hosted_app.py +47 -45
  104. examples/warming_up_to_rl/task_app/synth_envs_hosted/inference/__init__.py +1 -1
  105. examples/warming_up_to_rl/task_app/synth_envs_hosted/inference/openai_client.py +198 -92
  106. examples/warming_up_to_rl/task_app/synth_envs_hosted/main.py +0 -2
  107. examples/warming_up_to_rl/task_app/synth_envs_hosted/policy_routes.py +361 -263
  108. examples/warming_up_to_rl/task_app/synth_envs_hosted/registry.py +21 -23
  109. examples/warming_up_to_rl/task_app/synth_envs_hosted/rollout.py +394 -274
  110. examples/warming_up_to_rl/task_app/synth_envs_hosted/storage/__init__.py +1 -1
  111. examples/warming_up_to_rl/task_app/synth_envs_hosted/storage/volume.py +56 -62
  112. examples/warming_up_to_rl/task_app/synth_envs_hosted/test_agents.py +1 -0
  113. examples/warming_up_to_rl/task_app/synth_envs_hosted/test_service.py +6 -15
  114. examples/warming_up_to_rl/task_app/synth_envs_hosted/utils.py +4 -3
  115. synth_ai/__init__.py +1 -0
  116. synth_ai/api/models/supported.py +376 -0
  117. synth_ai/api/train/builders.py +157 -26
  118. synth_ai/api/train/cli.py +213 -57
  119. synth_ai/api/train/config_finder.py +65 -5
  120. synth_ai/api/train/env_resolver.py +33 -15
  121. synth_ai/api/train/pollers.py +13 -4
  122. synth_ai/api/train/supported_algos.py +139 -0
  123. synth_ai/api/train/task_app.py +5 -3
  124. synth_ai/api/train/utils.py +33 -48
  125. synth_ai/cli/__init__.py +19 -4
  126. synth_ai/cli/_modal_wrapper.py +28 -0
  127. synth_ai/cli/_typer_patch.py +49 -0
  128. synth_ai/cli/balance.py +2 -3
  129. synth_ai/cli/calc.py +1 -1
  130. synth_ai/cli/demo.py +21 -6
  131. synth_ai/cli/recent.py +2 -2
  132. synth_ai/cli/rl_demo.py +77 -17
  133. synth_ai/cli/root.py +116 -39
  134. synth_ai/cli/status.py +2 -2
  135. synth_ai/cli/task_apps.py +1699 -259
  136. synth_ai/cli/traces.py +7 -4
  137. synth_ai/cli/turso.py +73 -0
  138. synth_ai/cli/watch.py +12 -18
  139. synth_ai/core/experiment.py +0 -2
  140. synth_ai/demo_registry.py +68 -31
  141. synth_ai/demos/core/cli.py +516 -194
  142. synth_ai/demos/demo_task_apps/__init__.py +3 -3
  143. synth_ai/demos/demo_task_apps/core.py +64 -28
  144. synth_ai/demos/demo_task_apps/crafter/configs/crafter_fft_4b.toml +2 -3
  145. synth_ai/demos/demo_task_apps/crafter/grpo_crafter_task_app.py +37 -30
  146. synth_ai/demos/demo_task_apps/math/_common.py +1 -2
  147. synth_ai/demos/demo_task_apps/math/app.py +2 -1
  148. synth_ai/demos/demo_task_apps/math/deploy_modal.py +3 -6
  149. synth_ai/demos/demo_task_apps/math/modal_task_app.py +183 -82
  150. synth_ai/demos/demo_task_apps/math/task_app_entry.py +0 -2
  151. synth_ai/environments/examples/bandit/engine.py +12 -4
  152. synth_ai/environments/examples/bandit/taskset.py +4 -4
  153. synth_ai/environments/examples/crafter_classic/environment.py +76 -1
  154. synth_ai/environments/reproducibility/tree.py +5 -6
  155. synth_ai/environments/service/app.py +11 -12
  156. synth_ai/environments/service/core_routes.py +10 -9
  157. synth_ai/environments/stateful/engine.py +1 -1
  158. synth_ai/environments/tasks/core.py +1 -0
  159. synth_ai/environments/tasks/filters.py +5 -6
  160. synth_ai/environments/tasks/utils.py +4 -5
  161. synth_ai/evals/base.py +0 -2
  162. synth_ai/handshake.py +11 -9
  163. synth_ai/http.py +1 -1
  164. synth_ai/http_client.py +43 -11
  165. synth_ai/inference/__init__.py +0 -2
  166. synth_ai/inference/client.py +20 -6
  167. synth_ai/jobs/client.py +103 -78
  168. synth_ai/learning/__init__.py +41 -6
  169. synth_ai/learning/algorithms.py +14 -0
  170. synth_ai/learning/client.py +121 -29
  171. synth_ai/learning/config.py +2 -40
  172. synth_ai/learning/constants.py +0 -2
  173. synth_ai/learning/ft_client.py +4 -56
  174. synth_ai/learning/health.py +13 -7
  175. synth_ai/learning/jobs.py +43 -47
  176. synth_ai/{rl → learning/rl}/__init__.py +14 -5
  177. synth_ai/learning/rl/client.py +267 -0
  178. synth_ai/learning/rl/config.py +31 -0
  179. synth_ai/{rl → learning/rl}/contracts.py +5 -10
  180. synth_ai/{rl → learning/rl}/env_keys.py +45 -16
  181. synth_ai/learning/rl/secrets.py +13 -0
  182. synth_ai/learning/rl_client.py +2 -253
  183. synth_ai/learning/sft/__init__.py +29 -0
  184. synth_ai/learning/sft/client.py +68 -0
  185. synth_ai/learning/sft/config.py +270 -0
  186. synth_ai/learning/sft/data.py +295 -0
  187. synth_ai/learning/sse.py +25 -26
  188. synth_ai/learning/validators.py +25 -24
  189. synth_ai/lm/__init__.py +21 -47
  190. synth_ai/task/__init__.py +26 -27
  191. synth_ai/task/apps/__init__.py +18 -19
  192. synth_ai/task/auth.py +35 -23
  193. synth_ai/task/client.py +15 -13
  194. synth_ai/task/contracts.py +37 -35
  195. synth_ai/task/datasets.py +9 -6
  196. synth_ai/task/errors.py +11 -10
  197. synth_ai/task/health.py +17 -11
  198. synth_ai/task/json.py +58 -24
  199. synth_ai/task/proxy.py +15 -14
  200. synth_ai/task/rubrics.py +22 -15
  201. synth_ai/task/server.py +43 -17
  202. synth_ai/task/tracing_utils.py +12 -7
  203. synth_ai/task/validators.py +0 -1
  204. synth_ai/task/vendors.py +5 -7
  205. synth_ai/tracing_v3/__init__.py +2 -0
  206. synth_ai/tracing_v3/abstractions.py +21 -4
  207. synth_ai/tracing_v3/db_config.py +26 -1
  208. synth_ai/tracing_v3/decorators.py +18 -15
  209. synth_ai/tracing_v3/examples/basic_usage.py +3 -2
  210. synth_ai/tracing_v3/hooks.py +6 -4
  211. synth_ai/tracing_v3/llm_call_record_helpers.py +6 -6
  212. synth_ai/tracing_v3/replica_sync.py +1 -0
  213. synth_ai/tracing_v3/session_tracer.py +63 -16
  214. synth_ai/tracing_v3/storage/base.py +89 -1
  215. synth_ai/tracing_v3/storage/config.py +21 -8
  216. synth_ai/tracing_v3/storage/factory.py +10 -8
  217. synth_ai/tracing_v3/storage/utils.py +4 -2
  218. synth_ai/tracing_v3/turso/daemon.py +7 -2
  219. synth_ai/tracing_v3/turso/models.py +5 -2
  220. synth_ai/tracing_v3/turso/native_manager.py +1173 -0
  221. synth_ai/tracing_v3/utils.py +4 -3
  222. synth_ai/v0/api/__init__.py +8 -0
  223. synth_ai/v0/api/models/__init__.py +8 -0
  224. synth_ai/v0/api/models/supported.py +8 -0
  225. synth_ai/v0/config/__init__.py +15 -0
  226. synth_ai/v0/config/base_url.py +12 -0
  227. synth_ai/v0/lm/__init__.py +51 -0
  228. synth_ai/{lm → v0/lm}/caching/ephemeral.py +3 -5
  229. synth_ai/{lm → v0/lm}/caching/handler.py +4 -4
  230. synth_ai/{lm → v0/lm}/caching/initialize.py +1 -1
  231. synth_ai/{lm → v0/lm}/caching/persistent.py +1 -1
  232. synth_ai/{lm → v0/lm}/config.py +6 -1
  233. synth_ai/{lm → v0/lm}/core/all.py +9 -9
  234. synth_ai/{lm → v0/lm}/core/exceptions.py +0 -2
  235. synth_ai/{lm → v0/lm}/core/main.py +19 -7
  236. synth_ai/{lm → v0/lm}/core/main_v3.py +10 -10
  237. synth_ai/{lm → v0/lm}/core/synth_models.py +2 -15
  238. synth_ai/{lm → v0/lm}/core/vendor_clients.py +6 -4
  239. synth_ai/{lm → v0/lm}/overrides.py +4 -4
  240. synth_ai/{lm → v0/lm}/provider_support/anthropic.py +4 -4
  241. synth_ai/{lm → v0/lm}/provider_support/openai.py +5 -5
  242. synth_ai/{lm → v0/lm}/structured_outputs/handler.py +5 -5
  243. synth_ai/{lm → v0/lm}/structured_outputs/rehabilitate.py +1 -1
  244. synth_ai/{lm → v0/lm}/vendors/core/anthropic_api.py +16 -16
  245. synth_ai/{lm → v0/lm}/vendors/core/gemini_api.py +5 -5
  246. synth_ai/{lm → v0/lm}/vendors/core/mistral_api.py +5 -5
  247. synth_ai/{lm → v0/lm}/vendors/core/openai_api.py +12 -10
  248. synth_ai/{lm → v0/lm}/vendors/openai_standard.py +11 -9
  249. synth_ai/{lm → v0/lm}/vendors/openai_standard_responses.py +8 -5
  250. synth_ai/{lm → v0/lm}/vendors/supported/custom_endpoint.py +4 -6
  251. synth_ai/{lm → v0/lm}/vendors/supported/deepseek.py +2 -2
  252. synth_ai/{lm → v0/lm}/vendors/supported/grok.py +2 -2
  253. synth_ai/{lm → v0/lm}/vendors/supported/groq.py +1 -1
  254. synth_ai/{lm → v0/lm}/vendors/supported/ollama.py +1 -1
  255. synth_ai/{lm → v0/lm}/vendors/supported/openrouter.py +3 -3
  256. synth_ai/{lm → v0/lm}/vendors/supported/together.py +1 -1
  257. synth_ai/{lm → v0/lm}/vendors/synth_client.py +38 -11
  258. synth_ai/v0/tracing/upload.py +32 -135
  259. synth_ai/v0/tracing_v3/__init__.py +10 -0
  260. synth_ai/v0/tracing_v3/abstractions.py +3 -0
  261. synth_ai/v0/tracing_v3/decorators.py +3 -0
  262. synth_ai/v0/tracing_v3/llm_call_record_helpers.py +3 -0
  263. synth_ai/v0/tracing_v3/session_tracer.py +3 -0
  264. {synth_ai-0.2.9.dev5.dist-info → synth_ai-0.2.10.dist-info}/METADATA +10 -7
  265. {synth_ai-0.2.9.dev5.dist-info → synth_ai-0.2.10.dist-info}/RECORD +294 -258
  266. examples/common_old/backend.py +0 -21
  267. examples/evals_old/README.md +0 -98
  268. examples/evals_old/__init__.py +0 -6
  269. examples/evals_old/compare_models.py +0 -1037
  270. examples/evals_old/example_log.md +0 -145
  271. examples/evals_old/run_demo.sh +0 -126
  272. examples/evals_old/trace_analysis.py +0 -270
  273. examples/finetuning_old/_backup_synth_qwen/config.toml +0 -29
  274. examples/finetuning_old/_backup_synth_qwen/example_log.md +0 -324
  275. examples/finetuning_old/_backup_synth_qwen/filter_traces.py +0 -60
  276. examples/finetuning_old/_backup_synth_qwen/filter_traces_achievements.py +0 -239
  277. examples/finetuning_old/_backup_synth_qwen/purge_v3_traces.py +0 -109
  278. examples/finetuning_old/_backup_synth_qwen/react_agent_lm.py +0 -1924
  279. examples/finetuning_old/_backup_synth_qwen/readme.md +0 -49
  280. examples/finetuning_old/_backup_synth_qwen/run_crafter_qwen4b.py +0 -114
  281. examples/finetuning_old/_backup_synth_qwen/run_demo.sh +0 -195
  282. examples/finetuning_old/_backup_synth_qwen/sft_kickoff.py +0 -118
  283. examples/finetuning_old/synth_qwen_v1/README.md +0 -68
  284. examples/finetuning_old/synth_qwen_v1/filter_traces.py +0 -60
  285. examples/finetuning_old/synth_qwen_v1/filter_traces_achievements.py +0 -239
  286. examples/finetuning_old/synth_qwen_v1/finetune.py +0 -46
  287. examples/finetuning_old/synth_qwen_v1/hello_ft_model.py +0 -71
  288. examples/finetuning_old/synth_qwen_v1/infer.py +0 -37
  289. examples/finetuning_old/synth_qwen_v1/poll.py +0 -44
  290. examples/finetuning_old/synth_qwen_v1/prepare_data.py +0 -35
  291. examples/finetuning_old/synth_qwen_v1/purge_v3_traces.py +0 -109
  292. examples/finetuning_old/synth_qwen_v1/react_agent_lm.py +0 -1932
  293. examples/finetuning_old/synth_qwen_v1/run_crafter_sft_job.py +0 -207
  294. examples/finetuning_old/synth_qwen_v1/run_ft_job.py +0 -232
  295. examples/finetuning_old/synth_qwen_v1/upload_data.py +0 -34
  296. examples/finetuning_old/synth_qwen_v1/util.py +0 -147
  297. examples/rl_old/task_app.py +0 -962
  298. synth_ai/experimental/synth_oss.py +0 -446
  299. synth_ai/install_sqld.sh +0 -40
  300. synth_ai/learning/filtering.py +0 -0
  301. synth_ai/learning/offline/dpo.py +0 -0
  302. synth_ai/learning/offline/providers.py +0 -7
  303. synth_ai/learning/offline/sft.py +0 -0
  304. synth_ai/learning/offline/shared.py +0 -0
  305. synth_ai/learning/online/grpo.py +0 -0
  306. synth_ai/learning/online/irft.py +0 -0
  307. synth_ai/learning/prompts/banking77_injection_eval.py +0 -168
  308. synth_ai/learning/prompts/gepa.py +0 -0
  309. synth_ai/learning/prompts/hello_world_in_context_injection_ex.py +0 -213
  310. synth_ai/learning/prompts/mipro.py +0 -289
  311. synth_ai/learning/prompts/random_search.py +0 -246
  312. synth_ai/learning/prompts/run_mipro_banking77.py +0 -172
  313. synth_ai/learning/prompts/run_random_search_banking77.py +0 -324
  314. synth_ai/rl/secrets.py +0 -19
  315. synth_ai/scripts/verify_rewards.py +0 -100
  316. synth_ai/tracing/__init__.py +0 -30
  317. synth_ai/tracing_v1/__init__.py +0 -33
  318. synth_ai/tracing_v3/turso/__init__.py +0 -25
  319. synth_ai/tracing_v3/turso/manager.py +0 -774
  320. synth_ai/zyk/__init__.py +0 -30
  321. /synth_ai/{lm → v0/lm}/caching/__init__.py +0 -0
  322. /synth_ai/{lm → v0/lm}/caching/constants.py +0 -0
  323. /synth_ai/{lm → v0/lm}/caching/dbs.py +0 -0
  324. /synth_ai/{lm → v0/lm}/constants.py +0 -0
  325. /synth_ai/{lm → v0/lm}/core/__init__.py +0 -0
  326. /synth_ai/{lm → v0/lm}/cost/__init__.py +0 -0
  327. /synth_ai/{lm → v0/lm}/cost/monitor.py +0 -0
  328. /synth_ai/{lm → v0/lm}/cost/statefulness.py +0 -0
  329. /synth_ai/{lm → v0/lm}/injection.py +0 -0
  330. /synth_ai/{lm → v0/lm}/provider_support/__init__.py +0 -0
  331. /synth_ai/{lm → v0/lm}/provider_support/suppress_logging.py +0 -0
  332. /synth_ai/{lm → v0/lm}/structured_outputs/__init__.py +0 -0
  333. /synth_ai/{lm → v0/lm}/structured_outputs/inject.py +0 -0
  334. /synth_ai/{lm → v0/lm}/tools/__init__.py +0 -0
  335. /synth_ai/{lm → v0/lm}/tools/base.py +0 -0
  336. /synth_ai/{lm → v0/lm}/unified_interface.py +0 -0
  337. /synth_ai/{lm → v0/lm}/vendors/__init__.py +0 -0
  338. /synth_ai/{lm → v0/lm}/vendors/base.py +0 -0
  339. /synth_ai/{lm → v0/lm}/vendors/core/__init__.py +0 -0
  340. /synth_ai/{lm → v0/lm}/vendors/core/synth_dev_api.py +0 -0
  341. /synth_ai/{lm → v0/lm}/vendors/local/__init__.py +0 -0
  342. /synth_ai/{lm → v0/lm}/vendors/local/ollama.py +0 -0
  343. /synth_ai/{lm → v0/lm}/vendors/retries.py +0 -0
  344. /synth_ai/{lm → v0/lm}/vendors/supported/__init__.py +0 -0
  345. /synth_ai/{lm → v0/lm}/warmup.py +0 -0
  346. {synth_ai-0.2.9.dev5.dist-info → synth_ai-0.2.10.dist-info}/WHEEL +0 -0
  347. {synth_ai-0.2.9.dev5.dist-info → synth_ai-0.2.10.dist-info}/entry_points.txt +0 -0
  348. {synth_ai-0.2.9.dev5.dist-info → synth_ai-0.2.10.dist-info}/licenses/LICENSE +0 -0
  349. {synth_ai-0.2.9.dev5.dist-info → synth_ai-0.2.10.dist-info}/top_level.txt +0 -0
@@ -0,0 +1,1289 @@
1
+ from __future__ import annotations
2
+
3
+ import contextlib
4
+ import json
5
+ import logging
6
+ from typing import Any
7
+ from uuid import uuid4
8
+
9
+ from fastapi import APIRouter, HTTPException
10
+ from pydantic import BaseModel
11
+
12
+ # Import the actual classes from synth-ai
13
+ from synth_ai.environments.examples.crafter_classic.environment import (
14
+ CrafterClassicEnvironment,
15
+ )
16
+ from synth_ai.environments.examples.crafter_classic.taskset import (
17
+ CrafterTaskInstance,
18
+ CrafterTaskInstanceMetadata,
19
+ )
20
+ from synth_ai.environments.tasks.core import Impetus, Intent
21
+
22
+ from .envs.crafter.environment import CrafterEnvironmentWrapper
23
+ from .envs.mini_swe.environment import MiniSweEnvironmentWrapper
24
+ from .registry import registry
25
+ from .storage.volume import storage
26
+
27
+ logger = logging.getLogger(__name__)
28
+
29
+ router = APIRouter()
30
+
31
+
32
+ async def validate_environment_observation(observation: Any, context: str) -> None:
33
+ """
34
+ Validate that an environment observation has the correct structure.
35
+ This ensures the environment wrapper is producing valid observations.
36
+
37
+ Args:
38
+ observation: The observation to validate
39
+ context: Context string for logging (e.g., "initialize", "step")
40
+ """
41
+ if observation is None:
42
+ raise ValueError(f"Environment observation cannot be None in {context}")
43
+
44
+ if not isinstance(observation, dict):
45
+ raise ValueError(
46
+ f"Environment observation must be dict in {context}, got {type(observation)}"
47
+ )
48
+
49
+ # For Wordle environments, validate specific structure
50
+ # Check if this looks like a Wordle observation by looking for Wordle-specific keys
51
+ wordle_keys = {
52
+ "text",
53
+ "status",
54
+ "remaining_guesses",
55
+ "guesses",
56
+ "feedback",
57
+ "reward_last",
58
+ "total_reward",
59
+ "terminated",
60
+ }
61
+ if wordle_keys.issubset(set(observation.keys())):
62
+ logger.info(f"🔍 ENV_ROUTES: Validating Wordle observation structure in {context}")
63
+ logger.info(f"🔍 ENV_ROUTES: Observation keys: {list(observation.keys())}")
64
+
65
+ missing_keys = wordle_keys - set(observation.keys())
66
+ if missing_keys:
67
+ logger.error(
68
+ f"❌ ENV_ROUTES: Wordle observation missing required keys in {context}: {missing_keys}"
69
+ )
70
+ logger.error(f"❌ ENV_ROUTES: Full observation: {observation}")
71
+ raise ValueError(
72
+ f"Wordle observation missing required keys in {context}: {missing_keys}"
73
+ )
74
+
75
+ # Validate data types
76
+ if not isinstance(observation.get("text"), str):
77
+ raise ValueError(
78
+ f"Wordle observation 'text' must be string in {context}, got {type(observation.get('text'))}"
79
+ )
80
+
81
+ if not isinstance(observation.get("guesses"), list):
82
+ raise ValueError(
83
+ f"Wordle observation 'guesses' must be list in {context}, got {type(observation.get('guesses'))}"
84
+ )
85
+
86
+ if not isinstance(observation.get("feedback"), list):
87
+ raise ValueError(
88
+ f"Wordle observation 'feedback' must be list in {context}, got {type(observation.get('feedback'))}"
89
+ )
90
+
91
+ logger.info(
92
+ f"✅ ENV_ROUTES: Wordle observation structure validated successfully in {context}"
93
+ )
94
+ else:
95
+ logger.debug(
96
+ f"🔍 ENV_ROUTES: Observation doesn't appear to be Wordle in {context}, skipping validation"
97
+ )
98
+
99
+
100
+ class EnvCreateRequest(BaseModel):
101
+ env_name: str
102
+ config: dict[str, Any] = {}
103
+ seed: int | None = None
104
+ parent_env_id: str | None = None
105
+ rl_run_id: str
106
+
107
+
108
+ class EnvCreateResponse(BaseModel):
109
+ env_id: str
110
+ observation: dict[str, Any]
111
+ info: dict[str, Any] | None = None
112
+ step_idx: int
113
+
114
+
115
+ class EnvResetRequest(BaseModel):
116
+ env_id: str
117
+ seed: int | None = None
118
+
119
+
120
+ class EnvResetResponse(BaseModel):
121
+ observation: dict[str, Any]
122
+ info: dict[str, Any] | None = None
123
+ step_idx: int
124
+
125
+
126
+ class EnvStepRequest(BaseModel):
127
+ env_id: str
128
+ tool_calls: list[dict[str, Any]]
129
+
130
+
131
+ class EnvStepResponse(BaseModel):
132
+ observation: dict[str, Any]
133
+ done: bool
134
+ info: dict[str, Any] | None = None
135
+ reward: float | None = None
136
+ truncated: bool | None = None
137
+ step_idx: int
138
+
139
+
140
+ class EnvSnapshotRequest(BaseModel):
141
+ env_id: str
142
+
143
+
144
+ class EnvSnapshotResponse(BaseModel):
145
+ snapshot_id: str
146
+ path: str
147
+ rl_run_id: str
148
+ size: int
149
+
150
+
151
+ class EnvRestoreRequest(BaseModel):
152
+ snapshot_id: str
153
+
154
+
155
+ class EnvRestoreResponse(BaseModel):
156
+ env_id: str
157
+ observation: dict[str, Any]
158
+ info: dict[str, Any] | None = None
159
+ step_idx: int
160
+
161
+
162
+ class EnvTerminateRequest(BaseModel):
163
+ env_id: str
164
+
165
+
166
+ class EnvTerminateResponse(BaseModel):
167
+ ok: bool
168
+
169
+
170
+ @router.post("/create", response_model=EnvCreateResponse)
171
+ async def create_environment(request: EnvCreateRequest) -> EnvCreateResponse:
172
+ """Create a new environment instance."""
173
+ try:
174
+ # Create the underlying synth-ai environment
175
+ env_name_lower = request.env_name.lower()
176
+ if env_name_lower == "crafter":
177
+ # Build a minimal Crafter task instance
178
+ difficulty = (request.config or {}).get("difficulty", "normal")
179
+ seed_value = request.seed if request.seed is not None else 0
180
+ # Task object is part of the ecosystem; not required for instantiation here
181
+ impetus = Impetus(instructions="Survive and unlock achievements.")
182
+ intent = Intent(
183
+ rubric={"goal": "Unlock achievements"},
184
+ gold_trajectories=None,
185
+ gold_state_diff={},
186
+ )
187
+ metadata = CrafterTaskInstanceMetadata(
188
+ difficulty=difficulty,
189
+ seed=seed_value,
190
+ num_trees_radius=0,
191
+ num_cows_radius=0,
192
+ num_hostiles_radius=0,
193
+ )
194
+ instance = CrafterTaskInstance(
195
+ id=uuid4(),
196
+ impetus=impetus,
197
+ intent=intent,
198
+ metadata=metadata,
199
+ is_reproducible=True,
200
+ initial_engine_snapshot=None,
201
+ )
202
+ # Create CrafterClassicEnvironment from task instance
203
+ base_env = CrafterClassicEnvironment(task_instance=instance)
204
+
205
+ # Wrap it for our API
206
+ wrapper = CrafterEnvironmentWrapper(
207
+ env=base_env,
208
+ seed=request.seed,
209
+ )
210
+
211
+ # Initialize the environment
212
+ result = await wrapper.initialize()
213
+
214
+ # Log a world signature for sanity: seed + starting public state hash
215
+ try:
216
+ pub_state = base_env.engine._get_public_state_from_env() # type: ignore[attr-defined]
217
+ import hashlib
218
+ import json as _json
219
+
220
+ sig_src = {
221
+ "player_position": list(pub_state.player_position),
222
+ "player_direction": pub_state.player_direction,
223
+ "semantic_map": pub_state.semantic_map,
224
+ "inventory": {k: v for k, v in pub_state.inventory.items() if v},
225
+ }
226
+ sig_str = _json.dumps(sig_src, sort_keys=True)
227
+ sig = hashlib.md5(sig_str.encode("utf-8")).hexdigest()[:12]
228
+ logger.info(
229
+ "Crafter init signature: seed=%s sig=%s pos=%s inv=%s",
230
+ str(seed_value),
231
+ sig,
232
+ list(pub_state.player_position),
233
+ {k: v for k, v in pub_state.inventory.items() if v},
234
+ )
235
+ except Exception as _e:
236
+ pass
237
+
238
+ # Handle the observation structure consistently
239
+ # For Crafter, the result might still have the old nested structure, so we need to handle both
240
+ if isinstance(result, dict) and "observation" in result:
241
+ # Old nested structure - extract the inner observation
242
+ observation_for_registry = result["observation"].copy()
243
+ else:
244
+ # New flat structure - remove non-observation fields
245
+ observation_for_registry = result.copy()
246
+ for key in ["step_idx", "info"]:
247
+ if key in observation_for_registry:
248
+ del observation_for_registry[key]
249
+
250
+ # Register in memory
251
+ env_id = registry.register_env(
252
+ env=wrapper,
253
+ seed=request.seed,
254
+ rl_run_id=request.rl_run_id,
255
+ last_observation=observation_for_registry,
256
+ last_info=result.get("info"),
257
+ )
258
+
259
+ # Update step index in registry
260
+ handle = registry.get_env(env_id)
261
+ if handle:
262
+ handle.step_idx = result["step_idx"]
263
+
264
+ return EnvCreateResponse(
265
+ env_id=env_id,
266
+ observation=observation_for_registry,
267
+ info=result.get("info"),
268
+ step_idx=result["step_idx"],
269
+ )
270
+ elif env_name_lower in {"swe-mini", "swe_mini", "swe-mini-task"}:
271
+ cfg = request.config or {}
272
+ task_payload = cfg.get("task") or {}
273
+ env_cfg = cfg.get("environment") or {}
274
+ submit_command = cfg.get("submit_command")
275
+
276
+ if not task_payload:
277
+ raise HTTPException(
278
+ status_code=422,
279
+ detail="swe-mini environment requires config.task to be provided",
280
+ )
281
+
282
+ wrapper = MiniSweEnvironmentWrapper(
283
+ task=task_payload,
284
+ env_config=env_cfg,
285
+ submit_command=submit_command,
286
+ )
287
+ result = await wrapper.initialize()
288
+ observation = result.get("observation") or {}
289
+
290
+ env_id = registry.register_env(
291
+ env=wrapper,
292
+ seed=request.seed,
293
+ rl_run_id=request.rl_run_id,
294
+ last_observation=observation,
295
+ last_info=result.get("info"),
296
+ )
297
+ handle = registry.get_env(env_id)
298
+ if handle:
299
+ handle.step_idx = result.get("step_idx", 0)
300
+
301
+ return EnvCreateResponse(
302
+ env_id=env_id,
303
+ observation=observation,
304
+ info=result.get("info"),
305
+ step_idx=result.get("step_idx", 0),
306
+ )
307
+ elif env_name_lower == "wordle":
308
+ # Defer imports to avoid hard dependency when not used
309
+ try:
310
+ from synth_ai.environments.examples.wordle.environment import (
311
+ WordleEnvironment,
312
+ )
313
+ from synth_ai.environments.examples.wordle.taskset import (
314
+ WordleTaskInstance,
315
+ WordleTaskInstanceMetadata,
316
+ )
317
+ except Exception as e:
318
+ raise HTTPException(
319
+ status_code=500, detail=f"Wordle modules unavailable: {e}"
320
+ ) from e
321
+
322
+ # Lazy import of wrapper within branch
323
+ try:
324
+ from .envs.wordle.environment import WordleEnvironmentWrapper
325
+ except Exception as e:
326
+ raise HTTPException(
327
+ status_code=500, detail=f"Wordle wrapper unavailable: {e}"
328
+ ) from e
329
+ else:
330
+ wordle_wrapper_cls = WordleEnvironmentWrapper
331
+
332
+ cfg = request.config or {}
333
+ word_length = int(cfg.get("word_length", 5))
334
+ max_guesses = int(cfg.get("max_guesses", 6))
335
+
336
+ # Build a single Wordle task instance with proper seed usage
337
+ md = WordleTaskInstanceMetadata(
338
+ word_length=word_length,
339
+ max_guesses=max_guesses,
340
+ target_word=None, # Let seed determine the word
341
+ enforce_wordlist=True,
342
+ seed=request.seed,
343
+ consume_invalid_attempts=True,
344
+ )
345
+ instance = WordleTaskInstance(
346
+ id=uuid4(),
347
+ impetus=Impetus(instructions="Play Wordle. Submit one 5-letter word per turn."),
348
+ intent=Intent(rubric="guess the word", gold_trajectories=None, gold_state_diff={}),
349
+ metadata=md,
350
+ is_reproducible=True,
351
+ initial_engine_snapshot=None,
352
+ )
353
+ base_env = WordleEnvironment(task_instance=instance)
354
+
355
+ # Try to preserve the exact puzzle snapshot for reproducibility
356
+ init_snap = getattr(instance, "initial_engine_snapshot", None)
357
+
358
+ wrapper = wordle_wrapper_cls(
359
+ env=base_env,
360
+ seed=request.seed,
361
+ word_length=word_length,
362
+ max_guesses=max_guesses,
363
+ initial_engine_snapshot=init_snap,
364
+ )
365
+
366
+ result = await wrapper.initialize()
367
+
368
+ # Validate Wordle observation structure
369
+ # After our fix, the result is now flat, so we need to extract the observation fields
370
+ # that should be passed to the registry and response
371
+ if isinstance(result, dict) and "observation" in result:
372
+ # Old nested structure - extract the inner observation
373
+ observation_for_registry = result["observation"].copy()
374
+ else:
375
+ # New flat structure - remove non-observation fields
376
+ observation_for_registry = result.copy()
377
+ for key in ["step_idx", "info"]:
378
+ if key in observation_for_registry:
379
+ del observation_for_registry[key]
380
+
381
+ await validate_environment_observation(observation_for_registry, "initialize")
382
+
383
+ env_id = registry.register_env(
384
+ env=wrapper,
385
+ seed=request.seed,
386
+ rl_run_id=request.rl_run_id,
387
+ last_observation=observation_for_registry,
388
+ last_info=result.get("info"),
389
+ )
390
+ handle = registry.get_env(env_id)
391
+ if handle:
392
+ handle.step_idx = result["step_idx"]
393
+ return EnvCreateResponse(
394
+ env_id=env_id,
395
+ observation=observation_for_registry,
396
+ info=result.get("info"),
397
+ step_idx=result["step_idx"],
398
+ )
399
+
400
+ elif env_name_lower == "sokoban":
401
+ try:
402
+ from synth_ai.environments.examples.sokoban.environment import (
403
+ SokobanEnvironment,
404
+ )
405
+ from synth_ai.environments.examples.sokoban.taskset import (
406
+ SokobanTaskInstance,
407
+ SokobanTaskInstanceMetadata,
408
+ )
409
+ except Exception as e:
410
+ raise HTTPException(
411
+ status_code=500, detail=f"Sokoban modules unavailable: {e}"
412
+ ) from e
413
+
414
+ # Lazy import of wrapper within branch
415
+ try:
416
+ from .envs.sokoban.environment import SokobanEnvironmentWrapper
417
+ except Exception as e:
418
+ raise HTTPException(
419
+ status_code=500, detail=f"Sokoban wrapper unavailable: {e}"
420
+ ) from e
421
+
422
+ cfg = request.config or {}
423
+ difficulty = cfg.get("difficulty", "easy")
424
+ initial_state = cfg.get("initial_state") # Optional engine snapshot
425
+
426
+ metadata = SokobanTaskInstanceMetadata(
427
+ difficulty=difficulty,
428
+ )
429
+ instance = SokobanTaskInstance(
430
+ id=uuid4(),
431
+ impetus=Impetus(instructions="Push boxes to targets."),
432
+ intent=Intent(
433
+ rubric={"goal": "Solve the Sokoban puzzle"},
434
+ gold_trajectories=None,
435
+ gold_state_diff={},
436
+ ),
437
+ metadata=metadata,
438
+ is_reproducible=True,
439
+ initial_engine_snapshot=initial_state,
440
+ )
441
+ base_env = SokobanEnvironment(task_instance=instance)
442
+
443
+ wrapper = SokobanEnvironmentWrapper(env=base_env, seed=request.seed, config=cfg)
444
+ result = await wrapper.initialize()
445
+
446
+ # Handle the observation structure consistently for Sokoban
447
+ if isinstance(result, dict) and "observation" in result:
448
+ # Old nested structure - extract the inner observation
449
+ observation_for_registry = result["observation"].copy()
450
+ else:
451
+ # New flat structure - remove non-observation fields
452
+ observation_for_registry = result.copy()
453
+ for key in ["step_idx", "info"]:
454
+ if key in observation_for_registry:
455
+ del observation_for_registry[key]
456
+
457
+ env_id = registry.register_env(
458
+ env=wrapper,
459
+ seed=request.seed,
460
+ rl_run_id=request.rl_run_id,
461
+ last_observation=observation_for_registry,
462
+ last_info=result.get("info"),
463
+ )
464
+ handle = registry.get_env(env_id)
465
+ if handle:
466
+ handle.step_idx = result["step_idx"]
467
+ return EnvCreateResponse(
468
+ env_id=env_id,
469
+ observation=observation_for_registry,
470
+ info=result.get("info"),
471
+ step_idx=result["step_idx"],
472
+ )
473
+
474
+ elif env_name_lower == "math":
475
+ # Single-step math env (GSM8K-style)
476
+ cfg = request.config or {}
477
+ # Lazy import of wrapper within branch
478
+ try:
479
+ from .envs.math.environment import MathEnvironmentWrapper
480
+ except Exception as e:
481
+ raise HTTPException(status_code=500, detail=f"Math wrapper unavailable: {e}") from e
482
+
483
+ wrapper = MathEnvironmentWrapper(
484
+ seed=request.seed,
485
+ problem_id=cfg.get("problem_id"),
486
+ problem_text=cfg.get("problem_text"),
487
+ )
488
+ result = await wrapper.initialize()
489
+
490
+ observation_for_registry = (
491
+ result["observation"].copy()
492
+ if isinstance(result, dict) and "observation" in result
493
+ else result.copy()
494
+ )
495
+ for key in ["step_idx", "info"]:
496
+ if key in observation_for_registry:
497
+ del observation_for_registry[key]
498
+
499
+ env_id = registry.register_env(
500
+ env=wrapper,
501
+ seed=request.seed,
502
+ rl_run_id=request.rl_run_id,
503
+ last_observation=observation_for_registry,
504
+ last_info=result.get("info"),
505
+ )
506
+ handle = registry.get_env(env_id)
507
+ if handle:
508
+ handle.step_idx = result["step_idx"]
509
+ return EnvCreateResponse(
510
+ env_id=env_id,
511
+ observation=observation_for_registry,
512
+ info=result.get("info"),
513
+ step_idx=result["step_idx"],
514
+ )
515
+ else:
516
+ raise HTTPException(
517
+ status_code=422,
518
+ detail=f"Unknown environment name: {request.env_name}",
519
+ )
520
+
521
+ except Exception as e:
522
+ logger.error(f"Failed to create environment: {e}")
523
+ raise HTTPException(status_code=500, detail=str(e)) from e
524
+
525
+
526
+ # --- Compatibility routes for existing eval scripts that expect CrafterClassic paths ---
527
+ @router.post("/CrafterClassic/initialize", response_model=EnvCreateResponse)
528
+ async def compat_initialize(payload: dict) -> EnvCreateResponse:
529
+ seed = payload.get("seed")
530
+ wc = payload.get("world_config")
531
+ cfg = payload.get("config")
532
+ difficulty: str = "normal"
533
+ if isinstance(wc, str) and wc:
534
+ difficulty = wc
535
+ elif isinstance(wc, dict) and wc.get("difficulty"):
536
+ difficulty = str(wc.get("difficulty"))
537
+ elif isinstance(cfg, dict) and cfg.get("difficulty"):
538
+ difficulty = str(cfg.get("difficulty"))
539
+ req = EnvCreateRequest(
540
+ env_name="crafter", config={"difficulty": difficulty}, seed=seed, rl_run_id="eval"
541
+ )
542
+ return await create_environment(req)
543
+
544
+
545
+ @router.post("/CrafterClassic/step", response_model=EnvStepResponse)
546
+ async def compat_step(payload: dict) -> EnvStepResponse:
547
+ env_id = payload.get("env_id")
548
+ # eval script wraps action as {"tool_calls":[{"tool":"interact","args":{"action": <id>}}]}
549
+ action = payload.get("action") or {}
550
+ tool_calls = action.get("tool_calls") if isinstance(action, dict) else None
551
+ if not isinstance(tool_calls, list):
552
+ tool_calls = []
553
+ # Fallback: support {action: {actions: [..]}} by expanding into tool_calls
554
+ actions_list = action.get("actions") if isinstance(action, dict) else None
555
+ if isinstance(actions_list, list) and actions_list:
556
+ for a in actions_list:
557
+ tool_calls.append(
558
+ {
559
+ "tool": "interact",
560
+ "args": {"action": a},
561
+ }
562
+ )
563
+ req = EnvStepRequest(env_id=env_id, tool_calls=tool_calls)
564
+ return await step_environment(req)
565
+
566
+
567
+ @router.post("/CrafterClassic/terminate", response_model=EnvTerminateResponse)
568
+ async def compat_terminate(payload: dict) -> EnvTerminateResponse:
569
+ env_id = payload.get("env_id")
570
+ req = EnvTerminateRequest(env_id=env_id)
571
+ return await terminate_environment(req)
572
+
573
+
574
+ @router.post("/reset", response_model=EnvResetResponse)
575
+ async def reset_environment(request: EnvResetRequest) -> EnvResetResponse:
576
+ """Reset an environment to its initial state."""
577
+ handle = registry.get_env(request.env_id)
578
+ if not handle:
579
+ raise HTTPException(status_code=404, detail=f"Environment {request.env_id} not found")
580
+
581
+ try:
582
+ # Determine wrapper type and rebuild base env if a new seed is provided
583
+ wrapper = handle.env
584
+ if isinstance(wrapper, CrafterEnvironmentWrapper):
585
+ if request.seed is not None:
586
+ try:
587
+ difficulty = "normal"
588
+ seed_value = int(request.seed)
589
+ metadata = CrafterTaskInstanceMetadata(
590
+ difficulty=difficulty,
591
+ seed=seed_value,
592
+ num_trees_radius=0,
593
+ num_cows_radius=0,
594
+ num_hostiles_radius=0,
595
+ )
596
+ instance = CrafterTaskInstance(
597
+ id=uuid4(),
598
+ impetus=Impetus(instructions="Reset"),
599
+ intent=Intent(
600
+ rubric={"goal": "Reset"},
601
+ gold_trajectories=None,
602
+ gold_state_diff={},
603
+ ),
604
+ metadata=metadata,
605
+ is_reproducible=True,
606
+ initial_engine_snapshot=None,
607
+ )
608
+ new_base_env = CrafterClassicEnvironment(task_instance=instance)
609
+ wrapper.env = new_base_env
610
+ wrapper.seed = seed_value
611
+ handle.seed = seed_value
612
+ except Exception:
613
+ wrapper.seed = request.seed
614
+ handle.seed = request.seed
615
+
616
+ elif True:
617
+ # Try to dynamically import Wordle wrapper and check instance safely
618
+ wordle_wrapper_cls = None
619
+ with contextlib.suppress(Exception):
620
+ from .envs.wordle.environment import WordleEnvironmentWrapper
621
+
622
+ wordle_wrapper_cls = WordleEnvironmentWrapper # type: ignore[assignment]
623
+
624
+ if wordle_wrapper_cls is not None and isinstance(wrapper, wordle_wrapper_cls):
625
+ # Rebuild Wordle env with the same configuration; if we have a preserved
626
+ # initial_engine_snapshot, prefer constructing the instance directly.
627
+ try:
628
+ from synth_ai.environments.examples.wordle.environment import (
629
+ WordleEnvironment,
630
+ )
631
+ from synth_ai.environments.examples.wordle.taskset import (
632
+ WordleTaskInstance,
633
+ WordleTaskInstanceMetadata,
634
+ create_wordle_taskset,
635
+ )
636
+ except Exception as e:
637
+ raise HTTPException(
638
+ status_code=500, detail=f"Wordle modules unavailable: {e}"
639
+ ) from e
640
+
641
+ init_snap = getattr(wrapper, "initial_engine_snapshot", None)
642
+ if init_snap is not None:
643
+ metadata = WordleTaskInstanceMetadata(
644
+ word_length=int(wrapper.word_length),
645
+ max_guesses=int(wrapper.max_guesses),
646
+ )
647
+ instance = WordleTaskInstance(
648
+ id=uuid4(),
649
+ impetus=Impetus(instructions="Reset"),
650
+ intent=Intent(
651
+ rubric={"goal": "Reset"},
652
+ gold_trajectories=None,
653
+ gold_state_diff={},
654
+ ),
655
+ metadata=metadata,
656
+ is_reproducible=True,
657
+ initial_engine_snapshot=init_snap,
658
+ )
659
+ new_base_env = WordleEnvironment(task_instance=instance)
660
+ else:
661
+ ts = await create_wordle_taskset(
662
+ sample_size=1,
663
+ word_length=int(wrapper.word_length),
664
+ max_guesses=int(wrapper.max_guesses),
665
+ )
666
+ instance = ts.instances[0]
667
+ new_base_env = WordleEnvironment(task_instance=instance)
668
+ wrapper.env = new_base_env
669
+ if request.seed is not None:
670
+ wrapper.seed = int(request.seed)
671
+ handle.seed = int(request.seed)
672
+ else:
673
+ pass
674
+ # Rebuild Wordle env with the same configuration; if we have a preserved
675
+ # initial_engine_snapshot, prefer constructing the instance directly.
676
+ try:
677
+ from synth_ai.environments.examples.wordle.environment import (
678
+ WordleEnvironment,
679
+ )
680
+ from synth_ai.environments.examples.wordle.taskset import (
681
+ WordleTaskInstance,
682
+ WordleTaskInstanceMetadata,
683
+ create_wordle_taskset,
684
+ )
685
+ except Exception as e:
686
+ raise HTTPException(
687
+ status_code=500, detail=f"Wordle modules unavailable: {e}"
688
+ ) from e
689
+
690
+ init_snap = getattr(wrapper, "initial_engine_snapshot", None)
691
+ if init_snap is not None:
692
+ metadata = WordleTaskInstanceMetadata(
693
+ word_length=int(wrapper.word_length),
694
+ max_guesses=int(wrapper.max_guesses),
695
+ )
696
+ instance = WordleTaskInstance(
697
+ id=uuid4(),
698
+ impetus=Impetus(instructions="Reset"),
699
+ intent=Intent(
700
+ rubric={"goal": "Reset"},
701
+ gold_trajectories=None,
702
+ gold_state_diff={},
703
+ ),
704
+ metadata=metadata,
705
+ is_reproducible=True,
706
+ initial_engine_snapshot=init_snap,
707
+ )
708
+ new_base_env = WordleEnvironment(task_instance=instance)
709
+ else:
710
+ ts = await create_wordle_taskset(
711
+ sample_size=1,
712
+ word_length=int(wrapper.word_length),
713
+ max_guesses=int(wrapper.max_guesses),
714
+ )
715
+ instance = ts.instances[0]
716
+ new_base_env = WordleEnvironment(task_instance=instance)
717
+ wrapper.env = new_base_env
718
+ if request.seed is not None:
719
+ wrapper.seed = int(request.seed)
720
+ handle.seed = int(request.seed)
721
+
722
+ elif True:
723
+ # Try to dynamically import Sokoban wrapper and check instance safely
724
+ sokoban_wrapper_cls = None
725
+ with contextlib.suppress(Exception):
726
+ from .envs.sokoban.environment import SokobanEnvironmentWrapper
727
+
728
+ sokoban_wrapper_cls = SokobanEnvironmentWrapper # type: ignore[assignment]
729
+
730
+ if sokoban_wrapper_cls is not None and isinstance(wrapper, sokoban_wrapper_cls):
731
+ # Rebuild Sokoban env using stored config snapshot
732
+ try:
733
+ from synth_ai.environments.examples.sokoban.environment import (
734
+ SokobanEnvironment,
735
+ )
736
+ from synth_ai.environments.examples.sokoban.taskset import (
737
+ SokobanTaskInstance,
738
+ SokobanTaskInstanceMetadata,
739
+ )
740
+ except Exception as e:
741
+ raise HTTPException(
742
+ status_code=500, detail=f"Sokoban modules unavailable: {e}"
743
+ ) from e
744
+
745
+ cfg = dict(wrapper.config or {})
746
+ metadata = SokobanTaskInstanceMetadata(
747
+ difficulty=cfg.get("difficulty", "easy"),
748
+ )
749
+ instance = SokobanTaskInstance(
750
+ id=uuid4(),
751
+ impetus=Impetus(instructions="Reset"),
752
+ intent=Intent(
753
+ rubric={"goal": "Reset"}, gold_trajectories=None, gold_state_diff={}
754
+ ),
755
+ metadata=metadata,
756
+ is_reproducible=True,
757
+ initial_engine_snapshot=cfg.get("initial_state"),
758
+ )
759
+ new_base_env = SokobanEnvironment(task_instance=instance)
760
+ wrapper.env = new_base_env
761
+ if request.seed is not None:
762
+ wrapper.seed = int(request.seed)
763
+ handle.seed = int(request.seed)
764
+ else:
765
+ pass
766
+ # Rebuild Sokoban env using stored config snapshot
767
+ try:
768
+ from synth_ai.environments.examples.sokoban.environment import (
769
+ SokobanEnvironment,
770
+ )
771
+ from synth_ai.environments.examples.sokoban.taskset import (
772
+ SokobanTaskInstance,
773
+ SokobanTaskInstanceMetadata,
774
+ )
775
+ except Exception as e:
776
+ raise HTTPException(
777
+ status_code=500, detail=f"Sokoban modules unavailable: {e}"
778
+ ) from e
779
+
780
+ cfg = dict(wrapper.config or {})
781
+ metadata = SokobanTaskInstanceMetadata(
782
+ difficulty=cfg.get("difficulty", "easy"),
783
+ )
784
+ instance = SokobanTaskInstance(
785
+ id=uuid4(),
786
+ impetus=Impetus(instructions="Reset"),
787
+ intent=Intent(rubric={"goal": "Reset"}, gold_trajectories=None, gold_state_diff={}),
788
+ metadata=metadata,
789
+ is_reproducible=True,
790
+ initial_engine_snapshot=cfg.get("initial_state"),
791
+ )
792
+ new_base_env = SokobanEnvironment(task_instance=instance)
793
+ wrapper.env = new_base_env
794
+ if request.seed is not None:
795
+ wrapper.seed = int(request.seed)
796
+ handle.seed = int(request.seed)
797
+
798
+ # Reset the environment regardless of type
799
+ result = await wrapper.initialize()
800
+
801
+ # Log a world signature after reset for sanity
802
+ try:
803
+ base_env = handle.env.env # type: ignore[attr-defined]
804
+ pub_state = base_env.engine._get_public_state_from_env() # type: ignore[attr-defined]
805
+ import hashlib
806
+ import json as _json
807
+
808
+ sig_src = {
809
+ "player_position": list(pub_state.player_position),
810
+ "player_direction": pub_state.player_direction,
811
+ "semantic_map": pub_state.semantic_map,
812
+ "inventory": {k: v for k, v in pub_state.inventory.items() if v},
813
+ }
814
+ sig_str = _json.dumps(sig_src, sort_keys=True)
815
+ sig = hashlib.md5(sig_str.encode("utf-8")).hexdigest()[:12]
816
+ logger.info(
817
+ "Crafter reset signature: seed=%s sig=%s pos=%s inv=%s",
818
+ str(handle.seed),
819
+ sig,
820
+ list(pub_state.player_position),
821
+ {k: v for k, v in pub_state.inventory.items() if v},
822
+ )
823
+ except Exception as _:
824
+ pass
825
+
826
+ # Update registry
827
+ handle.step_idx = result["step_idx"]
828
+ handle.last_observation = result["observation"]
829
+ handle.last_info = result.get("info")
830
+
831
+ return EnvResetResponse(
832
+ observation=result["observation"],
833
+ info=result.get("info"),
834
+ step_idx=result["step_idx"],
835
+ )
836
+
837
+ except Exception as e:
838
+ logger.error(f"Failed to reset environment {request.env_id}: {e}")
839
+ raise HTTPException(status_code=500, detail=str(e)) from e
840
+
841
+
842
+ @router.post("/step", response_model=EnvStepResponse)
843
+ async def step_environment(request: EnvStepRequest) -> EnvStepResponse:
844
+ """Execute a step in the environment."""
845
+ handle = registry.get_env(request.env_id)
846
+ if not handle:
847
+ raise HTTPException(status_code=404, detail=f"Environment {request.env_id} not found")
848
+
849
+ try:
850
+ # Execute the step, pre-normalizing invalid Wordle guesses to avoid hard failures
851
+ wrapper = handle.env
852
+ wordle_wrapper_cls = None
853
+ with contextlib.suppress(Exception):
854
+ from .envs.wordle.environment import WordleEnvironmentWrapper
855
+
856
+ wordle_wrapper_cls = WordleEnvironmentWrapper # type: ignore[assignment]
857
+
858
+ if wordle_wrapper_cls is not None and isinstance(wrapper, wordle_wrapper_cls):
859
+ expected_len = int(getattr(wrapper, "word_length", 5))
860
+ normalized: list[dict[str, Any]] = []
861
+ for tc in request.tool_calls or []:
862
+ tool = tc.get("tool") or tc.get("tool_name") or tc.get("name") or "interact"
863
+ args = tc.get("arguments") or tc.get("args") or {}
864
+ if isinstance(args, str):
865
+ try:
866
+ args = json.loads(args)
867
+ except Exception:
868
+ args = {}
869
+ guess = None
870
+ if isinstance(args, dict):
871
+ guess = args.get("guess") or args.get("word")
872
+ if isinstance(guess, str):
873
+ g = guess.strip().lower()
874
+ if (not g.isalpha()) or (len(g) != expected_len):
875
+ normalized.append(
876
+ {"tool": "invalid_guess", "args": {"original_guess": guess}}
877
+ )
878
+ else:
879
+ # Preserve the original tool name (interact or submit) for the environment to handle
880
+ normalized.append({"tool": tool, "args": {"guess": g}})
881
+ else:
882
+ normalized.append({"tool": "invalid_guess", "args": {"original_guess": guess}})
883
+ result = await wrapper.step(normalized)
884
+ else:
885
+ result = await handle.env.step(request.tool_calls)
886
+
887
+ # Validate observation structure for Wordle environments
888
+ env_name = getattr(handle.env, "env", None)
889
+ if (
890
+ env_name
891
+ and hasattr(env_name, "__class__")
892
+ and "wordle" in env_name.__class__.__name__.lower()
893
+ ):
894
+ # Extract observation fields from the flat result structure for validation
895
+ observation_for_validation = result.copy()
896
+ # Remove step_idx, done, info, reward, truncated from the observation since they're separate fields
897
+ for key in ["step_idx", "done", "info", "reward", "truncated"]:
898
+ if key in observation_for_validation:
899
+ del observation_for_validation[key]
900
+ await validate_environment_observation(observation_for_validation, "step")
901
+
902
+ # Update registry
903
+ handle.step_idx = result["step_idx"]
904
+
905
+ # Extract the observation fields from the result structure (handle both old nested and new flat)
906
+ if isinstance(result, dict) and "observation" in result:
907
+ # Old nested structure - extract the inner observation
908
+ observation_for_registry = result["observation"].copy()
909
+ else:
910
+ # New flat structure - remove non-observation fields
911
+ observation_for_registry = result.copy()
912
+ for key in ["step_idx", "done", "info", "reward", "truncated"]:
913
+ if key in observation_for_registry:
914
+ del observation_for_registry[key]
915
+
916
+ handle.last_observation = observation_for_registry
917
+ handle.last_info = result.get("info")
918
+
919
+ return EnvStepResponse(
920
+ observation=observation_for_registry,
921
+ done=result["done"],
922
+ info=result.get("info"),
923
+ reward=result.get("reward"),
924
+ truncated=result.get("truncated"),
925
+ step_idx=result["step_idx"],
926
+ )
927
+
928
+ except Exception as e:
929
+ logger.error(f"Failed to step environment {request.env_id}: {e}")
930
+ # Fallback for Wordle: convert invalid guesses into 'invalid_guess' tool calls and retry once
931
+ try:
932
+ wordle_wrapper_cls = None
933
+ with contextlib.suppress(Exception):
934
+ from .envs.wordle.environment import WordleEnvironmentWrapper
935
+
936
+ wordle_wrapper_cls = WordleEnvironmentWrapper # type: ignore[assignment]
937
+
938
+ wrapper = handle.env
939
+ if wordle_wrapper_cls is not None and isinstance(wrapper, wordle_wrapper_cls):
940
+ expected_len = int(getattr(wrapper, "word_length", 5))
941
+ normalized: list[dict[str, Any]] = []
942
+ for tc in request.tool_calls or []:
943
+ tool = tc.get("tool") or tc.get("tool_name") or tc.get("name") or "interact"
944
+ args = tc.get("arguments") or tc.get("args") or {}
945
+ if isinstance(args, str):
946
+ try:
947
+ args = json.loads(args)
948
+ except Exception:
949
+ args = {}
950
+ guess = None
951
+ if isinstance(args, dict):
952
+ guess = args.get("guess") or args.get("word")
953
+ if isinstance(guess, str):
954
+ g = guess.strip().lower()
955
+ if (not g.isalpha()) or (len(g) != expected_len):
956
+ normalized.append(
957
+ {
958
+ "tool": "invalid_guess",
959
+ "args": {"original_guess": guess},
960
+ }
961
+ )
962
+ else:
963
+ normalized.append({"tool": "interact", "args": {"guess": g}})
964
+ else:
965
+ normalized.append(
966
+ {"tool": "invalid_guess", "args": {"original_guess": guess}}
967
+ )
968
+
969
+ # Retry with normalized calls, allowing the wrapper to synthesize an observation
970
+ result = await wrapper.step(normalized)
971
+
972
+ # Update registry and return as usual
973
+ handle.step_idx = result["step_idx"]
974
+ if isinstance(result, dict) and "observation" in result:
975
+ observation_for_registry = result["observation"].copy()
976
+ else:
977
+ observation_for_registry = result.copy()
978
+ for key in ["step_idx", "done", "info", "reward", "truncated"]:
979
+ if key in observation_for_registry:
980
+ del observation_for_registry[key]
981
+ handle.last_observation = observation_for_registry
982
+ handle.last_info = result.get("info")
983
+ return EnvStepResponse(
984
+ observation=observation_for_registry,
985
+ done=result["done"],
986
+ info=result.get("info"),
987
+ reward=result.get("reward"),
988
+ truncated=result.get("truncated"),
989
+ step_idx=result["step_idx"],
990
+ )
991
+ except Exception:
992
+ # Ignore fallback errors; fall through to generic error
993
+ pass
994
+
995
+ raise HTTPException(status_code=500, detail=f"{type(e).__name__}: {e}") from e
996
+
997
+
998
+ @router.post("/snapshot", response_model=EnvSnapshotResponse)
999
+ async def snapshot_environment(request: EnvSnapshotRequest) -> EnvSnapshotResponse:
1000
+ """Create a snapshot of the environment state."""
1001
+ handle = registry.get_env(request.env_id)
1002
+ if not handle:
1003
+ raise HTTPException(status_code=404, detail=f"Environment {request.env_id} not found")
1004
+
1005
+ try:
1006
+ # Serialize environment state
1007
+ state_dict = await handle.env.serialize()
1008
+
1009
+ # Save to volume
1010
+ snapshot_id, path, size = storage.save_snapshot(
1011
+ rl_run_id=handle.rl_run_id,
1012
+ kind="env",
1013
+ state_dict=state_dict,
1014
+ config={"seed": handle.seed},
1015
+ )
1016
+
1017
+ # Register snapshot
1018
+ registry.register_snapshot(
1019
+ kind="env",
1020
+ rl_run_id=handle.rl_run_id,
1021
+ size=size,
1022
+ path=path,
1023
+ )
1024
+
1025
+ return EnvSnapshotResponse(
1026
+ snapshot_id=snapshot_id,
1027
+ path=path,
1028
+ rl_run_id=handle.rl_run_id,
1029
+ size=size,
1030
+ )
1031
+
1032
+ except Exception as e:
1033
+ logger.error(f"Failed to snapshot environment {request.env_id}: {e}")
1034
+ raise HTTPException(status_code=500, detail=str(e)) from e
1035
+
1036
+
1037
+ @router.post("/restore", response_model=EnvRestoreResponse)
1038
+ async def restore_environment(request: EnvRestoreRequest) -> EnvRestoreResponse:
1039
+ """Restore an environment from a snapshot."""
1040
+ snapshot = registry.get_snapshot(request.snapshot_id)
1041
+ if not snapshot:
1042
+ raise HTTPException(status_code=404, detail=f"Snapshot {request.snapshot_id} not found")
1043
+
1044
+ if snapshot.kind != "env":
1045
+ raise HTTPException(
1046
+ status_code=422,
1047
+ detail=f"Snapshot {request.snapshot_id} is not an environment snapshot",
1048
+ )
1049
+
1050
+ try:
1051
+ # Load snapshot from volume
1052
+ state_dict, meta = storage.load_snapshot(
1053
+ rl_run_id=snapshot.rl_run_id,
1054
+ kind="env",
1055
+ snapshot_id=request.snapshot_id,
1056
+ )
1057
+
1058
+ # Recreate environment
1059
+ env_name = state_dict.get("name", "crafter")
1060
+ name_lower = str(env_name).lower()
1061
+ if name_lower == "crafter":
1062
+ # Create base environment
1063
+ # Recreate classic env from snapshot metadata
1064
+ seed_value = state_dict["config"]["seed"]
1065
+ metadata = CrafterTaskInstanceMetadata(
1066
+ difficulty="normal",
1067
+ seed=seed_value,
1068
+ num_trees_radius=0,
1069
+ num_cows_radius=0,
1070
+ num_hostiles_radius=0,
1071
+ )
1072
+ instance = CrafterTaskInstance(
1073
+ id=uuid4(),
1074
+ impetus=Impetus(instructions="Restore"),
1075
+ intent=Intent(
1076
+ rubric={"goal": "Restore"},
1077
+ gold_trajectories=None,
1078
+ gold_state_diff={},
1079
+ ),
1080
+ metadata=metadata,
1081
+ is_reproducible=True,
1082
+ initial_engine_snapshot=None,
1083
+ )
1084
+ base_env = CrafterClassicEnvironment(task_instance=instance)
1085
+
1086
+ # Deserialize into wrapper
1087
+ wrapper = await CrafterEnvironmentWrapper.deserialize(
1088
+ payload=state_dict,
1089
+ env=base_env,
1090
+ )
1091
+
1092
+ # Register new instance
1093
+ env_id = registry.register_env(
1094
+ env=wrapper,
1095
+ seed=wrapper.seed,
1096
+ rl_run_id=snapshot.rl_run_id,
1097
+ last_observation=wrapper.last_observation,
1098
+ last_info=wrapper.last_info,
1099
+ )
1100
+
1101
+ # Update step index
1102
+ handle = registry.get_env(env_id)
1103
+ if handle:
1104
+ handle.step_idx = wrapper.step_idx
1105
+
1106
+ return EnvRestoreResponse(
1107
+ env_id=env_id,
1108
+ observation=wrapper.last_observation or {},
1109
+ info=wrapper.last_info,
1110
+ step_idx=wrapper.step_idx,
1111
+ )
1112
+ elif name_lower == "wordle":
1113
+ try:
1114
+ from synth_ai.environments.examples.wordle.environment import (
1115
+ WordleEnvironment,
1116
+ )
1117
+ from synth_ai.environments.examples.wordle.taskset import (
1118
+ WordleTaskInstance,
1119
+ WordleTaskInstanceMetadata,
1120
+ create_wordle_taskset,
1121
+ )
1122
+ except Exception as e:
1123
+ raise HTTPException(
1124
+ status_code=500, detail=f"Wordle modules unavailable: {e}"
1125
+ ) from e
1126
+
1127
+ cfg = state_dict.get("config", {}) or {}
1128
+ word_length = int(cfg.get("word_length", 5))
1129
+ max_guesses = int(cfg.get("max_guesses", 6))
1130
+ init_snap = cfg.get("initial_engine_snapshot")
1131
+ if init_snap is not None:
1132
+ metadata = WordleTaskInstanceMetadata(
1133
+ word_length=word_length, max_guesses=max_guesses
1134
+ )
1135
+ instance = WordleTaskInstance(
1136
+ id=uuid4(),
1137
+ impetus=Impetus(instructions="Restore"),
1138
+ intent=Intent(
1139
+ rubric={"goal": "Restore"},
1140
+ gold_trajectories=None,
1141
+ gold_state_diff={},
1142
+ ),
1143
+ metadata=metadata,
1144
+ is_reproducible=True,
1145
+ initial_engine_snapshot=init_snap,
1146
+ )
1147
+ base_env = WordleEnvironment(task_instance=instance)
1148
+ else:
1149
+ ts = await create_wordle_taskset(
1150
+ sample_size=1, word_length=word_length, max_guesses=max_guesses
1151
+ )
1152
+ instance = ts.instances[0]
1153
+ base_env = WordleEnvironment(task_instance=instance)
1154
+ # Lazy import of wrapper only when needed
1155
+ try:
1156
+ from .envs.wordle.environment import WordleEnvironmentWrapper
1157
+ except Exception as e:
1158
+ raise HTTPException(
1159
+ status_code=500, detail=f"Wordle wrapper unavailable: {e}"
1160
+ ) from e
1161
+ wrapper = await WordleEnvironmentWrapper.deserialize(payload=state_dict, env=base_env)
1162
+
1163
+ env_id = registry.register_env(
1164
+ env=wrapper,
1165
+ seed=wrapper.seed,
1166
+ rl_run_id=snapshot.rl_run_id,
1167
+ last_observation=wrapper.last_observation,
1168
+ last_info=wrapper.last_info,
1169
+ )
1170
+ handle = registry.get_env(env_id)
1171
+ if handle:
1172
+ handle.step_idx = wrapper.step_idx
1173
+ return EnvRestoreResponse(
1174
+ env_id=env_id,
1175
+ observation=wrapper.last_observation or {},
1176
+ info=wrapper.last_info,
1177
+ step_idx=wrapper.step_idx,
1178
+ )
1179
+
1180
+ elif name_lower in {"swe-mini", "swe_mini", "swe-mini-task"}:
1181
+ cfg = state_dict.get("config", {}) or {}
1182
+ payload = {
1183
+ "config": cfg,
1184
+ "state": state_dict.get("state", {}),
1185
+ }
1186
+ wrapper = await MiniSweEnvironmentWrapper.deserialize(payload)
1187
+
1188
+ env_id = registry.register_env(
1189
+ env=wrapper,
1190
+ seed=None,
1191
+ rl_run_id=snapshot.rl_run_id,
1192
+ last_observation=wrapper._build_observation(wrapper.last_result),
1193
+ last_info=None,
1194
+ )
1195
+ handle = registry.get_env(env_id)
1196
+ if handle:
1197
+ handle.step_idx = wrapper.state.step_idx
1198
+ return EnvRestoreResponse(
1199
+ env_id=env_id,
1200
+ observation=wrapper._build_observation(wrapper.last_result),
1201
+ info=None,
1202
+ step_idx=wrapper.state.step_idx,
1203
+ )
1204
+
1205
+ elif name_lower == "sokoban":
1206
+ try:
1207
+ from synth_ai.environments.examples.sokoban.environment import (
1208
+ SokobanEnvironment,
1209
+ )
1210
+ from synth_ai.environments.examples.sokoban.taskset import (
1211
+ SokobanTaskInstance,
1212
+ SokobanTaskInstanceMetadata,
1213
+ )
1214
+ except Exception as e:
1215
+ raise HTTPException(
1216
+ status_code=500, detail=f"Sokoban modules unavailable: {e}"
1217
+ ) from e
1218
+
1219
+ cfg = state_dict.get("config", {}) or {}
1220
+ metadata = SokobanTaskInstanceMetadata(difficulty=cfg.get("difficulty", "easy"))
1221
+ instance = SokobanTaskInstance(
1222
+ id=uuid4(),
1223
+ impetus=Impetus(instructions="Restore"),
1224
+ intent=Intent(
1225
+ rubric={"goal": "Restore"},
1226
+ gold_trajectories=None,
1227
+ gold_state_diff={},
1228
+ ),
1229
+ metadata=metadata,
1230
+ is_reproducible=True,
1231
+ initial_engine_snapshot=cfg.get("initial_state"),
1232
+ )
1233
+ base_env = SokobanEnvironment(task_instance=instance)
1234
+ # Lazy import of wrapper only when needed
1235
+ try:
1236
+ from .envs.sokoban.environment import SokobanEnvironmentWrapper
1237
+ except Exception as e:
1238
+ raise HTTPException(
1239
+ status_code=500, detail=f"Sokoban wrapper unavailable: {e}"
1240
+ ) from e
1241
+ wrapper = await SokobanEnvironmentWrapper.deserialize(payload=state_dict, env=base_env)
1242
+
1243
+ env_id = registry.register_env(
1244
+ env=wrapper,
1245
+ seed=wrapper.seed,
1246
+ rl_run_id=snapshot.rl_run_id,
1247
+ last_observation=wrapper.last_observation,
1248
+ last_info=wrapper.last_info,
1249
+ )
1250
+ handle = registry.get_env(env_id)
1251
+ if handle:
1252
+ handle.step_idx = wrapper.step_idx
1253
+ return EnvRestoreResponse(
1254
+ env_id=env_id,
1255
+ observation=wrapper.last_observation or {},
1256
+ info=wrapper.last_info,
1257
+ step_idx=wrapper.step_idx,
1258
+ )
1259
+
1260
+ else:
1261
+ raise HTTPException(
1262
+ status_code=422,
1263
+ detail=f"Unknown environment name in snapshot: {env_name}",
1264
+ )
1265
+
1266
+ except Exception as e:
1267
+ logger.error(f"Failed to restore environment from snapshot {request.snapshot_id}: {e}")
1268
+ raise HTTPException(status_code=500, detail=str(e)) from e
1269
+
1270
+
1271
+ @router.post("/terminate", response_model=EnvTerminateResponse)
1272
+ async def terminate_environment(request: EnvTerminateRequest) -> EnvTerminateResponse:
1273
+ """Terminate an environment and clean up resources."""
1274
+ handle = registry.get_env(request.env_id)
1275
+ if not handle:
1276
+ raise HTTPException(status_code=404, detail=f"Environment {request.env_id} not found")
1277
+
1278
+ try:
1279
+ # Call terminate on the environment
1280
+ await handle.env.terminate()
1281
+
1282
+ # Remove from registry
1283
+ registry.remove_env(request.env_id)
1284
+
1285
+ return EnvTerminateResponse(ok=True)
1286
+
1287
+ except Exception as e:
1288
+ logger.error(f"Failed to terminate environment {request.env_id}: {e}")
1289
+ raise HTTPException(status_code=500, detail=str(e)) from e