inspect-ai 0.3.70__py3-none-any.whl → 0.3.72__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.
Files changed (219) hide show
  1. inspect_ai/_cli/eval.py +14 -8
  2. inspect_ai/_display/core/display.py +2 -0
  3. inspect_ai/_display/core/footer.py +13 -3
  4. inspect_ai/_display/plain/display.py +6 -2
  5. inspect_ai/_display/rich/display.py +19 -6
  6. inspect_ai/_display/textual/app.py +6 -1
  7. inspect_ai/_display/textual/display.py +4 -0
  8. inspect_ai/_display/textual/widgets/transcript.py +10 -6
  9. inspect_ai/_eval/task/run.py +5 -8
  10. inspect_ai/_util/content.py +20 -1
  11. inspect_ai/_util/transcript.py +10 -4
  12. inspect_ai/_util/working.py +4 -0
  13. inspect_ai/_view/www/App.css +6 -0
  14. inspect_ai/_view/www/dist/assets/index.css +115 -87
  15. inspect_ai/_view/www/dist/assets/index.js +5324 -2276
  16. inspect_ai/_view/www/eslint.config.mjs +24 -1
  17. inspect_ai/_view/www/log-schema.json +283 -20
  18. inspect_ai/_view/www/package.json +8 -3
  19. inspect_ai/_view/www/src/App.tsx +2 -2
  20. inspect_ai/_view/www/src/components/AnsiDisplay.tsx +4 -3
  21. inspect_ai/_view/www/src/components/Card.tsx +9 -8
  22. inspect_ai/_view/www/src/components/DownloadButton.tsx +2 -1
  23. inspect_ai/_view/www/src/components/EmptyPanel.tsx +2 -2
  24. inspect_ai/_view/www/src/components/ErrorPanel.tsx +4 -3
  25. inspect_ai/_view/www/src/components/ExpandablePanel.tsx +13 -5
  26. inspect_ai/_view/www/src/components/FindBand.tsx +3 -3
  27. inspect_ai/_view/www/src/components/HumanBaselineView.tsx +3 -3
  28. inspect_ai/_view/www/src/components/LabeledValue.tsx +5 -4
  29. inspect_ai/_view/www/src/components/LargeModal.tsx +18 -13
  30. inspect_ai/_view/www/src/components/{LightboxCarousel.css → LightboxCarousel.module.css} +22 -18
  31. inspect_ai/_view/www/src/components/LightboxCarousel.tsx +36 -27
  32. inspect_ai/_view/www/src/components/MessageBand.tsx +2 -1
  33. inspect_ai/_view/www/src/components/NavPills.tsx +9 -8
  34. inspect_ai/_view/www/src/components/ProgressBar.tsx +2 -1
  35. inspect_ai/_view/www/src/components/TabSet.tsx +21 -15
  36. inspect_ai/_view/www/src/index.tsx +2 -2
  37. inspect_ai/_view/www/src/metadata/MetaDataGrid.tsx +11 -9
  38. inspect_ai/_view/www/src/metadata/MetaDataView.tsx +3 -2
  39. inspect_ai/_view/www/src/metadata/MetadataGrid.module.css +1 -0
  40. inspect_ai/_view/www/src/metadata/RenderedContent.tsx +16 -0
  41. inspect_ai/_view/www/src/plan/DatasetDetailView.tsx +3 -2
  42. inspect_ai/_view/www/src/plan/DetailStep.tsx +2 -1
  43. inspect_ai/_view/www/src/plan/PlanCard.tsx +2 -5
  44. inspect_ai/_view/www/src/plan/PlanDetailView.tsx +6 -9
  45. inspect_ai/_view/www/src/plan/ScorerDetailView.tsx +2 -1
  46. inspect_ai/_view/www/src/plan/SolverDetailView.tsx +3 -3
  47. inspect_ai/_view/www/src/samples/InlineSampleDisplay.tsx +2 -2
  48. inspect_ai/_view/www/src/samples/SampleDialog.tsx +3 -3
  49. inspect_ai/_view/www/src/samples/SampleDisplay.tsx +2 -2
  50. inspect_ai/_view/www/src/samples/SampleSummaryView.tsx +2 -2
  51. inspect_ai/_view/www/src/samples/SamplesTools.tsx +2 -1
  52. inspect_ai/_view/www/src/samples/chat/ChatMessage.tsx +3 -19
  53. inspect_ai/_view/www/src/samples/chat/ChatMessageRenderer.tsx +2 -1
  54. inspect_ai/_view/www/src/samples/chat/ChatMessageRow.tsx +2 -1
  55. inspect_ai/_view/www/src/samples/chat/ChatView.tsx +2 -1
  56. inspect_ai/_view/www/src/samples/chat/ChatViewVirtualList.tsx +22 -7
  57. inspect_ai/_view/www/src/samples/chat/MessageContent.tsx +35 -6
  58. inspect_ai/_view/www/src/samples/chat/MessageContents.tsx +2 -2
  59. inspect_ai/_view/www/src/samples/chat/messages.ts +15 -2
  60. inspect_ai/_view/www/src/samples/chat/tools/ToolCallView.tsx +13 -4
  61. inspect_ai/_view/www/src/samples/chat/tools/ToolInput.module.css +2 -2
  62. inspect_ai/_view/www/src/samples/chat/tools/ToolInput.tsx +18 -19
  63. inspect_ai/_view/www/src/samples/chat/tools/ToolOutput.module.css +1 -1
  64. inspect_ai/_view/www/src/samples/chat/tools/ToolOutput.tsx +4 -3
  65. inspect_ai/_view/www/src/samples/chat/tools/ToolTitle.tsx +2 -2
  66. inspect_ai/_view/www/src/samples/error/FlatSampleErrorView.tsx +2 -3
  67. inspect_ai/_view/www/src/samples/error/SampleErrorView.tsx +3 -2
  68. inspect_ai/_view/www/src/samples/list/SampleFooter.tsx +2 -1
  69. inspect_ai/_view/www/src/samples/list/SampleHeader.tsx +2 -1
  70. inspect_ai/_view/www/src/samples/list/SampleList.tsx +57 -45
  71. inspect_ai/_view/www/src/samples/list/SampleRow.tsx +2 -1
  72. inspect_ai/_view/www/src/samples/list/SampleSeparator.tsx +2 -1
  73. inspect_ai/_view/www/src/samples/sample-tools/EpochFilter.tsx +2 -2
  74. inspect_ai/_view/www/src/samples/sample-tools/SelectScorer.tsx +4 -3
  75. inspect_ai/_view/www/src/samples/sample-tools/SortFilter.tsx +2 -5
  76. inspect_ai/_view/www/src/samples/sample-tools/sample-filter/SampleFilter.tsx +2 -2
  77. inspect_ai/_view/www/src/samples/scores/SampleScoreView.tsx +2 -1
  78. inspect_ai/_view/www/src/samples/scores/SampleScores.tsx +2 -2
  79. inspect_ai/_view/www/src/samples/transcript/ApprovalEventView.tsx +2 -1
  80. inspect_ai/_view/www/src/samples/transcript/ErrorEventView.tsx +2 -1
  81. inspect_ai/_view/www/src/samples/transcript/InfoEventView.tsx +2 -1
  82. inspect_ai/_view/www/src/samples/transcript/InputEventView.tsx +2 -1
  83. inspect_ai/_view/www/src/samples/transcript/LoggerEventView.module.css +4 -0
  84. inspect_ai/_view/www/src/samples/transcript/LoggerEventView.tsx +12 -2
  85. inspect_ai/_view/www/src/samples/transcript/ModelEventView.module.css +1 -1
  86. inspect_ai/_view/www/src/samples/transcript/ModelEventView.tsx +25 -28
  87. inspect_ai/_view/www/src/samples/transcript/SampleInitEventView.tsx +2 -1
  88. inspect_ai/_view/www/src/samples/transcript/SampleLimitEventView.tsx +5 -4
  89. inspect_ai/_view/www/src/samples/transcript/SampleTranscript.tsx +2 -2
  90. inspect_ai/_view/www/src/samples/transcript/SandboxEventView.tsx +8 -7
  91. inspect_ai/_view/www/src/samples/transcript/ScoreEventView.tsx +2 -2
  92. inspect_ai/_view/www/src/samples/transcript/StepEventView.tsx +3 -3
  93. inspect_ai/_view/www/src/samples/transcript/SubtaskEventView.tsx +18 -14
  94. inspect_ai/_view/www/src/samples/transcript/ToolEventView.tsx +5 -5
  95. inspect_ai/_view/www/src/samples/transcript/TranscriptView.tsx +34 -15
  96. inspect_ai/_view/www/src/samples/transcript/event/EventNav.tsx +2 -1
  97. inspect_ai/_view/www/src/samples/transcript/event/EventNavs.tsx +2 -1
  98. inspect_ai/_view/www/src/samples/transcript/event/EventRow.tsx +3 -2
  99. inspect_ai/_view/www/src/samples/transcript/event/EventSection.tsx +2 -2
  100. inspect_ai/_view/www/src/samples/transcript/event/EventTimingPanel.module.css +28 -0
  101. inspect_ai/_view/www/src/samples/transcript/event/EventTimingPanel.tsx +115 -0
  102. inspect_ai/_view/www/src/samples/transcript/event/utils.ts +29 -0
  103. inspect_ai/_view/www/src/samples/transcript/state/StateDiffView.tsx +2 -1
  104. inspect_ai/_view/www/src/samples/transcript/state/StateEventRenderers.tsx +3 -3
  105. inspect_ai/_view/www/src/samples/transcript/state/StateEventView.tsx +11 -8
  106. inspect_ai/_view/www/src/types/log.d.ts +129 -34
  107. inspect_ai/_view/www/src/usage/ModelTokenTable.tsx +6 -10
  108. inspect_ai/_view/www/src/usage/ModelUsagePanel.module.css +4 -0
  109. inspect_ai/_view/www/src/usage/ModelUsagePanel.tsx +32 -9
  110. inspect_ai/_view/www/src/usage/TokenTable.tsx +4 -6
  111. inspect_ai/_view/www/src/usage/UsageCard.tsx +2 -1
  112. inspect_ai/_view/www/src/utils/format.ts +1 -1
  113. inspect_ai/_view/www/src/utils/json.ts +24 -0
  114. inspect_ai/_view/www/src/workspace/WorkSpace.tsx +6 -5
  115. inspect_ai/_view/www/src/workspace/WorkSpaceView.tsx +9 -2
  116. inspect_ai/_view/www/src/workspace/error/TaskErrorPanel.tsx +2 -1
  117. inspect_ai/_view/www/src/workspace/navbar/Navbar.tsx +2 -1
  118. inspect_ai/_view/www/src/workspace/navbar/PrimaryBar.tsx +3 -3
  119. inspect_ai/_view/www/src/workspace/navbar/ResultsPanel.tsx +4 -3
  120. inspect_ai/_view/www/src/workspace/navbar/SecondaryBar.tsx +5 -4
  121. inspect_ai/_view/www/src/workspace/navbar/StatusPanel.tsx +5 -8
  122. inspect_ai/_view/www/src/workspace/sidebar/EvalStatus.tsx +5 -4
  123. inspect_ai/_view/www/src/workspace/sidebar/LogDirectoryTitleView.tsx +2 -1
  124. inspect_ai/_view/www/src/workspace/sidebar/Sidebar.tsx +2 -1
  125. inspect_ai/_view/www/src/workspace/sidebar/SidebarLogEntry.tsx +2 -2
  126. inspect_ai/_view/www/src/workspace/sidebar/SidebarScoreView.tsx +2 -1
  127. inspect_ai/_view/www/src/workspace/sidebar/SidebarScoresView.tsx +2 -2
  128. inspect_ai/_view/www/src/workspace/tabs/InfoTab.tsx +2 -2
  129. inspect_ai/_view/www/src/workspace/tabs/JsonTab.tsx +2 -5
  130. inspect_ai/_view/www/src/workspace/tabs/SamplesTab.tsx +12 -11
  131. inspect_ai/_view/www/yarn.lock +241 -5
  132. inspect_ai/log/_condense.py +3 -0
  133. inspect_ai/log/_recorders/eval.py +6 -1
  134. inspect_ai/log/_transcript.py +58 -1
  135. inspect_ai/model/__init__.py +2 -0
  136. inspect_ai/model/_call_tools.py +7 -0
  137. inspect_ai/model/_chat_message.py +22 -7
  138. inspect_ai/model/_conversation.py +10 -8
  139. inspect_ai/model/_generate_config.py +25 -4
  140. inspect_ai/model/_model.py +133 -57
  141. inspect_ai/model/_model_output.py +3 -0
  142. inspect_ai/model/_openai.py +106 -40
  143. inspect_ai/model/_providers/anthropic.py +281 -153
  144. inspect_ai/model/_providers/google.py +27 -8
  145. inspect_ai/model/_providers/groq.py +9 -4
  146. inspect_ai/model/_providers/openai.py +57 -4
  147. inspect_ai/model/_providers/openai_o1.py +10 -0
  148. inspect_ai/model/_providers/providers.py +1 -1
  149. inspect_ai/model/_reasoning.py +15 -2
  150. inspect_ai/scorer/_model.py +23 -19
  151. inspect_ai/solver/_human_agent/agent.py +14 -10
  152. inspect_ai/solver/_human_agent/commands/__init__.py +7 -3
  153. inspect_ai/solver/_human_agent/commands/submit.py +76 -30
  154. inspect_ai/tool/__init__.py +2 -0
  155. inspect_ai/tool/_tool.py +3 -1
  156. inspect_ai/tool/_tools/_computer/_common.py +117 -58
  157. inspect_ai/tool/_tools/_computer/_computer.py +80 -57
  158. inspect_ai/tool/_tools/_computer/_resources/image_home_dir/.config/Code/User/settings.json +7 -1
  159. inspect_ai/tool/_tools/_computer/_resources/image_home_dir/.config/xfce4/xfconf/xfce-perchannel-xml/xfwm4.xml +91 -0
  160. inspect_ai/tool/_tools/_computer/_resources/tool/.pylintrc +8 -0
  161. inspect_ai/tool/_tools/_computer/_resources/tool/.vscode/settings.json +12 -0
  162. inspect_ai/tool/_tools/_computer/_resources/tool/_args.py +78 -0
  163. inspect_ai/tool/_tools/_computer/_resources/tool/_constants.py +20 -0
  164. inspect_ai/tool/_tools/_computer/_resources/tool/_run.py +1 -1
  165. inspect_ai/tool/_tools/_computer/_resources/tool/_x11_client.py +175 -113
  166. inspect_ai/tool/_tools/_computer/_resources/tool/computer_tool.py +76 -20
  167. inspect_ai/tool/_tools/_computer/_resources/tool/pyproject.toml +65 -0
  168. inspect_ai/tool/_tools/_computer/test_args.py +151 -0
  169. inspect_ai/tool/_tools/_web_browser/_resources/.pylintrc +8 -0
  170. inspect_ai/tool/_tools/_web_browser/_resources/.vscode/launch.json +24 -0
  171. inspect_ai/tool/_tools/_web_browser/_resources/.vscode/settings.json +25 -0
  172. inspect_ai/tool/_tools/_web_browser/_resources/Dockerfile +5 -6
  173. inspect_ai/tool/_tools/_web_browser/_resources/README.md +10 -11
  174. inspect_ai/tool/_tools/_web_browser/_resources/accessibility_tree.py +71 -0
  175. inspect_ai/tool/_tools/_web_browser/_resources/accessibility_tree_node.py +323 -0
  176. inspect_ai/tool/_tools/_web_browser/_resources/cdp/__init__.py +5 -0
  177. inspect_ai/tool/_tools/_web_browser/_resources/cdp/a11y.py +279 -0
  178. inspect_ai/tool/_tools/_web_browser/_resources/cdp/dom.py +9 -0
  179. inspect_ai/tool/_tools/_web_browser/_resources/cdp/dom_snapshot.py +293 -0
  180. inspect_ai/tool/_tools/_web_browser/_resources/cdp/page.py +94 -0
  181. inspect_ai/tool/_tools/_web_browser/_resources/constants.py +2 -0
  182. inspect_ai/tool/_tools/_web_browser/_resources/images/usage_diagram.svg +2 -0
  183. inspect_ai/tool/_tools/_web_browser/_resources/playwright_browser.py +50 -0
  184. inspect_ai/tool/_tools/_web_browser/_resources/playwright_crawler.py +31 -359
  185. inspect_ai/tool/_tools/_web_browser/_resources/playwright_page_crawler.py +280 -0
  186. inspect_ai/tool/_tools/_web_browser/_resources/pyproject.toml +65 -0
  187. inspect_ai/tool/_tools/_web_browser/_resources/rectangle.py +64 -0
  188. inspect_ai/tool/_tools/_web_browser/_resources/rpc_client_helpers.py +146 -0
  189. inspect_ai/tool/_tools/_web_browser/_resources/scale_factor.py +64 -0
  190. inspect_ai/tool/_tools/_web_browser/_resources/test_accessibility_tree_node.py +180 -0
  191. inspect_ai/tool/_tools/_web_browser/_resources/test_playwright_crawler.py +15 -9
  192. inspect_ai/tool/_tools/_web_browser/_resources/test_rectangle.py +15 -0
  193. inspect_ai/tool/_tools/_web_browser/_resources/test_web_client.py +44 -0
  194. inspect_ai/tool/_tools/_web_browser/_resources/web_browser_rpc_types.py +39 -0
  195. inspect_ai/tool/_tools/_web_browser/_resources/web_client.py +198 -48
  196. inspect_ai/tool/_tools/_web_browser/_resources/web_client_new_session.py +26 -25
  197. inspect_ai/tool/_tools/_web_browser/_resources/web_server.py +178 -39
  198. inspect_ai/tool/_tools/_web_browser/_web_browser.py +38 -19
  199. inspect_ai/util/__init__.py +2 -1
  200. inspect_ai/util/_display.py +12 -0
  201. inspect_ai/util/_sandbox/events.py +55 -21
  202. inspect_ai/util/_sandbox/self_check.py +131 -43
  203. inspect_ai/util/_subtask.py +11 -0
  204. {inspect_ai-0.3.70.dist-info → inspect_ai-0.3.72.dist-info}/METADATA +1 -1
  205. {inspect_ai-0.3.70.dist-info → inspect_ai-0.3.72.dist-info}/RECORD +209 -186
  206. {inspect_ai-0.3.70.dist-info → inspect_ai-0.3.72.dist-info}/WHEEL +1 -1
  207. inspect_ai/_view/www/src/components/VirtualList.module.css +0 -19
  208. inspect_ai/_view/www/src/components/VirtualList.tsx +0 -292
  209. inspect_ai/tool/_tools/_computer/_computer_split.py +0 -198
  210. inspect_ai/tool/_tools/_web_browser/_resources/accessibility_node.py +0 -312
  211. inspect_ai/tool/_tools/_web_browser/_resources/dm_env_servicer.py +0 -275
  212. inspect_ai/tool/_tools/_web_browser/_resources/images/usage_diagram.png +0 -0
  213. inspect_ai/tool/_tools/_web_browser/_resources/test_accessibility_node.py +0 -176
  214. inspect_ai/tool/_tools/_web_browser/_resources/test_dm_env_servicer.py +0 -135
  215. inspect_ai/tool/_tools/_web_browser/_resources/test_web_environment.py +0 -71
  216. inspect_ai/tool/_tools/_web_browser/_resources/web_environment.py +0 -184
  217. {inspect_ai-0.3.70.dist-info → inspect_ai-0.3.72.dist-info}/LICENSE +0 -0
  218. {inspect_ai-0.3.70.dist-info → inspect_ai-0.3.72.dist-info}/entry_points.txt +0 -0
  219. {inspect_ai-0.3.70.dist-info → inspect_ai-0.3.72.dist-info}/top_level.txt +0 -0
@@ -1,5 +1,7 @@
1
- import globals from "globals";
2
1
  import pluginJs from "@eslint/js";
2
+ import reactHooks from "eslint-plugin-react-hooks";
3
+ import globals from "globals";
4
+ import tseslint from "typescript-eslint";
3
5
 
4
6
  export default [
5
7
  {
@@ -14,4 +16,25 @@ export default [
14
16
  {
15
17
  ignores: ["libs/**", "preact/**", "dist/**"],
16
18
  },
19
+ // Add TypeScript support with customized rules for all files
20
+ {
21
+ files: ["**/*.{ts,tsx}"],
22
+ languageOptions: {
23
+ parser: tseslint.parser,
24
+ parserOptions: {
25
+ project: "./tsconfig.json",
26
+ },
27
+ },
28
+ plugins: {
29
+ "react-hooks": reactHooks,
30
+ },
31
+ rules: {
32
+ // React Hooks rules
33
+ "react-hooks/rules-of-hooks": "warn",
34
+ // "react-hooks/exhaustive-deps": "warn",
35
+
36
+ // These are disabled because we didn't have time to fix them, not because they are bad rules
37
+ "no-unused-vars": "off",
38
+ },
39
+ },
17
40
  ];
@@ -8,6 +8,10 @@
8
8
  "title": "Timestamp",
9
9
  "type": "string"
10
10
  },
11
+ "working_start": {
12
+ "title": "Working Start",
13
+ "type": "number"
14
+ },
11
15
  "pending": {
12
16
  "anyOf": [
13
17
  {
@@ -85,6 +89,7 @@
85
89
  },
86
90
  "required": [
87
91
  "timestamp",
92
+ "working_start",
88
93
  "pending",
89
94
  "event",
90
95
  "message",
@@ -216,6 +221,9 @@
216
221
  {
217
222
  "$ref": "#/$defs/ContentText"
218
223
  },
224
+ {
225
+ "$ref": "#/$defs/ContentReasoning"
226
+ },
219
227
  {
220
228
  "$ref": "#/$defs/ContentImage"
221
229
  },
@@ -262,26 +270,13 @@
262
270
  ],
263
271
  "default": null,
264
272
  "title": "Tool Calls"
265
- },
266
- "reasoning": {
267
- "anyOf": [
268
- {
269
- "type": "string"
270
- },
271
- {
272
- "type": "null"
273
- }
274
- ],
275
- "default": null,
276
- "title": "Reasoning"
277
273
  }
278
274
  },
279
275
  "required": [
280
276
  "role",
281
277
  "content",
282
278
  "source",
283
- "tool_calls",
284
- "reasoning"
279
+ "tool_calls"
285
280
  ],
286
281
  "title": "ChatMessageAssistant",
287
282
  "type": "object",
@@ -307,6 +302,9 @@
307
302
  {
308
303
  "$ref": "#/$defs/ContentText"
309
304
  },
305
+ {
306
+ "$ref": "#/$defs/ContentReasoning"
307
+ },
310
308
  {
311
309
  "$ref": "#/$defs/ContentImage"
312
310
  },
@@ -369,6 +367,9 @@
369
367
  {
370
368
  "$ref": "#/$defs/ContentText"
371
369
  },
370
+ {
371
+ "$ref": "#/$defs/ContentReasoning"
372
+ },
372
373
  {
373
374
  "$ref": "#/$defs/ContentImage"
374
375
  },
@@ -469,6 +470,9 @@
469
470
  {
470
471
  "$ref": "#/$defs/ContentText"
471
472
  },
473
+ {
474
+ "$ref": "#/$defs/ContentReasoning"
475
+ },
472
476
  {
473
477
  "$ref": "#/$defs/ContentImage"
474
478
  },
@@ -591,6 +595,47 @@
591
595
  "type": "object",
592
596
  "additionalProperties": false
593
597
  },
598
+ "ContentReasoning": {
599
+ "description": "Reasoning content.\n\nSee the specification for [thinking blocks](https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking#understanding-thinking-blocks) for Claude models.",
600
+ "properties": {
601
+ "type": {
602
+ "const": "reasoning",
603
+ "default": "reasoning",
604
+ "title": "Type",
605
+ "type": "string"
606
+ },
607
+ "reasoning": {
608
+ "title": "Reasoning",
609
+ "type": "string"
610
+ },
611
+ "signature": {
612
+ "anyOf": [
613
+ {
614
+ "type": "string"
615
+ },
616
+ {
617
+ "type": "null"
618
+ }
619
+ ],
620
+ "default": null,
621
+ "title": "Signature"
622
+ },
623
+ "redacted": {
624
+ "default": false,
625
+ "title": "Redacted",
626
+ "type": "boolean"
627
+ }
628
+ },
629
+ "required": [
630
+ "type",
631
+ "reasoning",
632
+ "signature",
633
+ "redacted"
634
+ ],
635
+ "title": "ContentReasoning",
636
+ "type": "object",
637
+ "additionalProperties": false
638
+ },
594
639
  "ContentText": {
595
640
  "description": "Text content.",
596
641
  "properties": {
@@ -653,6 +698,10 @@
653
698
  "title": "Timestamp",
654
699
  "type": "string"
655
700
  },
701
+ "working_start": {
702
+ "title": "Working Start",
703
+ "type": "number"
704
+ },
656
705
  "pending": {
657
706
  "anyOf": [
658
707
  {
@@ -677,6 +726,7 @@
677
726
  },
678
727
  "required": [
679
728
  "timestamp",
729
+ "working_start",
680
730
  "pending",
681
731
  "event",
682
732
  "error"
@@ -1210,6 +1260,7 @@
1210
1260
  "max_tool_output": null,
1211
1261
  "cache_prompt": null,
1212
1262
  "reasoning_effort": null,
1263
+ "reasoning_tokens": null,
1213
1264
  "reasoning_history": null
1214
1265
  }
1215
1266
  }
@@ -1556,6 +1607,18 @@
1556
1607
  "default": null,
1557
1608
  "title": "Working Time"
1558
1609
  },
1610
+ "uuid": {
1611
+ "anyOf": [
1612
+ {
1613
+ "type": "string"
1614
+ },
1615
+ {
1616
+ "type": "null"
1617
+ }
1618
+ ],
1619
+ "default": null,
1620
+ "title": "Uuid"
1621
+ },
1559
1622
  "error": {
1560
1623
  "anyOf": [
1561
1624
  {
@@ -1604,6 +1667,7 @@
1604
1667
  "model_usage",
1605
1668
  "total_time",
1606
1669
  "working_time",
1670
+ "uuid",
1607
1671
  "error",
1608
1672
  "attachments",
1609
1673
  "limit"
@@ -2466,10 +2530,28 @@
2466
2530
  "default": null,
2467
2531
  "title": "Reasoning Effort"
2468
2532
  },
2533
+ "reasoning_tokens": {
2534
+ "anyOf": [
2535
+ {
2536
+ "type": "integer"
2537
+ },
2538
+ {
2539
+ "type": "null"
2540
+ }
2541
+ ],
2542
+ "default": null,
2543
+ "title": "Reasoning Tokens"
2544
+ },
2469
2545
  "reasoning_history": {
2470
2546
  "anyOf": [
2471
2547
  {
2472
- "type": "boolean"
2548
+ "enum": [
2549
+ "none",
2550
+ "all",
2551
+ "last",
2552
+ "auto"
2553
+ ],
2554
+ "type": "string"
2473
2555
  },
2474
2556
  {
2475
2557
  "type": "null"
@@ -2504,6 +2586,7 @@
2504
2586
  "max_tool_output",
2505
2587
  "cache_prompt",
2506
2588
  "reasoning_effort",
2589
+ "reasoning_tokens",
2507
2590
  "reasoning_history"
2508
2591
  ],
2509
2592
  "additionalProperties": false
@@ -2516,6 +2599,10 @@
2516
2599
  "title": "Timestamp",
2517
2600
  "type": "string"
2518
2601
  },
2602
+ "working_start": {
2603
+ "title": "Working Start",
2604
+ "type": "number"
2605
+ },
2519
2606
  "pending": {
2520
2607
  "anyOf": [
2521
2608
  {
@@ -2552,6 +2639,7 @@
2552
2639
  },
2553
2640
  "required": [
2554
2641
  "timestamp",
2642
+ "working_start",
2555
2643
  "pending",
2556
2644
  "event",
2557
2645
  "source",
@@ -2569,6 +2657,10 @@
2569
2657
  "title": "Timestamp",
2570
2658
  "type": "string"
2571
2659
  },
2660
+ "working_start": {
2661
+ "title": "Working Start",
2662
+ "type": "number"
2663
+ },
2572
2664
  "pending": {
2573
2665
  "anyOf": [
2574
2666
  {
@@ -2598,6 +2690,7 @@
2598
2690
  },
2599
2691
  "required": [
2600
2692
  "timestamp",
2693
+ "working_start",
2601
2694
  "pending",
2602
2695
  "event",
2603
2696
  "input",
@@ -2667,6 +2760,10 @@
2667
2760
  "title": "Timestamp",
2668
2761
  "type": "string"
2669
2762
  },
2763
+ "working_start": {
2764
+ "title": "Working Start",
2765
+ "type": "number"
2766
+ },
2670
2767
  "pending": {
2671
2768
  "anyOf": [
2672
2769
  {
@@ -2691,6 +2788,7 @@
2691
2788
  },
2692
2789
  "required": [
2693
2790
  "timestamp",
2791
+ "working_start",
2694
2792
  "pending",
2695
2793
  "event",
2696
2794
  "message"
@@ -2861,6 +2959,7 @@
2861
2959
  "type": "null"
2862
2960
  }
2863
2961
  ],
2962
+ "default": null,
2864
2963
  "title": "Time"
2865
2964
  }
2866
2965
  },
@@ -2881,6 +2980,10 @@
2881
2980
  "title": "Timestamp",
2882
2981
  "type": "string"
2883
2982
  },
2983
+ "working_start": {
2984
+ "title": "Working Start",
2985
+ "type": "number"
2986
+ },
2884
2987
  "pending": {
2885
2988
  "anyOf": [
2886
2989
  {
@@ -2990,10 +3093,36 @@
2990
3093
  }
2991
3094
  ],
2992
3095
  "default": null
3096
+ },
3097
+ "completed": {
3098
+ "anyOf": [
3099
+ {
3100
+ "format": "date-time",
3101
+ "type": "string"
3102
+ },
3103
+ {
3104
+ "type": "null"
3105
+ }
3106
+ ],
3107
+ "default": null,
3108
+ "title": "Completed"
3109
+ },
3110
+ "working_time": {
3111
+ "anyOf": [
3112
+ {
3113
+ "type": "number"
3114
+ },
3115
+ {
3116
+ "type": "null"
3117
+ }
3118
+ ],
3119
+ "default": null,
3120
+ "title": "Working Time"
2993
3121
  }
2994
3122
  },
2995
3123
  "required": [
2996
3124
  "timestamp",
3125
+ "working_start",
2997
3126
  "pending",
2998
3127
  "event",
2999
3128
  "model",
@@ -3004,7 +3133,9 @@
3004
3133
  "output",
3005
3134
  "error",
3006
3135
  "cache",
3007
- "call"
3136
+ "call",
3137
+ "completed",
3138
+ "working_time"
3008
3139
  ],
3009
3140
  "title": "ModelEvent",
3010
3141
  "type": "object",
@@ -3126,6 +3257,18 @@
3126
3257
  ],
3127
3258
  "default": null,
3128
3259
  "title": "Input Tokens Cache Read"
3260
+ },
3261
+ "reasoning_tokens": {
3262
+ "anyOf": [
3263
+ {
3264
+ "type": "integer"
3265
+ },
3266
+ {
3267
+ "type": "null"
3268
+ }
3269
+ ],
3270
+ "default": null,
3271
+ "title": "Reasoning Tokens"
3129
3272
  }
3130
3273
  },
3131
3274
  "title": "ModelUsage",
@@ -3135,7 +3278,8 @@
3135
3278
  "output_tokens",
3136
3279
  "total_tokens",
3137
3280
  "input_tokens_cache_write",
3138
- "input_tokens_cache_read"
3281
+ "input_tokens_cache_read",
3282
+ "reasoning_tokens"
3139
3283
  ],
3140
3284
  "additionalProperties": false
3141
3285
  },
@@ -3286,6 +3430,10 @@
3286
3430
  "title": "Timestamp",
3287
3431
  "type": "string"
3288
3432
  },
3433
+ "working_start": {
3434
+ "title": "Working Start",
3435
+ "type": "number"
3436
+ },
3289
3437
  "pending": {
3290
3438
  "anyOf": [
3291
3439
  {
@@ -3313,6 +3461,7 @@
3313
3461
  },
3314
3462
  "required": [
3315
3463
  "timestamp",
3464
+ "working_start",
3316
3465
  "pending",
3317
3466
  "event",
3318
3467
  "sample",
@@ -3330,6 +3479,10 @@
3330
3479
  "title": "Timestamp",
3331
3480
  "type": "string"
3332
3481
  },
3482
+ "working_start": {
3483
+ "title": "Working Start",
3484
+ "type": "number"
3485
+ },
3333
3486
  "pending": {
3334
3487
  "anyOf": [
3335
3488
  {
@@ -3379,6 +3532,7 @@
3379
3532
  },
3380
3533
  "required": [
3381
3534
  "timestamp",
3535
+ "working_start",
3382
3536
  "pending",
3383
3537
  "event",
3384
3538
  "type",
@@ -3423,6 +3577,10 @@
3423
3577
  "title": "Timestamp",
3424
3578
  "type": "string"
3425
3579
  },
3580
+ "working_start": {
3581
+ "title": "Working Start",
3582
+ "type": "number"
3583
+ },
3426
3584
  "pending": {
3427
3585
  "anyOf": [
3428
3586
  {
@@ -3524,10 +3682,24 @@
3524
3682
  ],
3525
3683
  "default": null,
3526
3684
  "title": "Output"
3685
+ },
3686
+ "completed": {
3687
+ "anyOf": [
3688
+ {
3689
+ "format": "date-time",
3690
+ "type": "string"
3691
+ },
3692
+ {
3693
+ "type": "null"
3694
+ }
3695
+ ],
3696
+ "default": null,
3697
+ "title": "Completed"
3527
3698
  }
3528
3699
  },
3529
3700
  "required": [
3530
3701
  "timestamp",
3702
+ "working_start",
3531
3703
  "pending",
3532
3704
  "event",
3533
3705
  "action",
@@ -3536,7 +3708,8 @@
3536
3708
  "file",
3537
3709
  "input",
3538
3710
  "result",
3539
- "output"
3711
+ "output",
3712
+ "completed"
3540
3713
  ],
3541
3714
  "title": "SandboxEvent",
3542
3715
  "type": "object",
@@ -3658,6 +3831,10 @@
3658
3831
  "title": "Timestamp",
3659
3832
  "type": "string"
3660
3833
  },
3834
+ "working_start": {
3835
+ "title": "Working Start",
3836
+ "type": "number"
3837
+ },
3661
3838
  "pending": {
3662
3839
  "anyOf": [
3663
3840
  {
@@ -3705,6 +3882,7 @@
3705
3882
  },
3706
3883
  "required": [
3707
3884
  "timestamp",
3885
+ "working_start",
3708
3886
  "pending",
3709
3887
  "event",
3710
3888
  "score",
@@ -3723,6 +3901,10 @@
3723
3901
  "title": "Timestamp",
3724
3902
  "type": "string"
3725
3903
  },
3904
+ "working_start": {
3905
+ "title": "Working Start",
3906
+ "type": "number"
3907
+ },
3726
3908
  "pending": {
3727
3909
  "anyOf": [
3728
3910
  {
@@ -3751,6 +3933,7 @@
3751
3933
  },
3752
3934
  "required": [
3753
3935
  "timestamp",
3936
+ "working_start",
3754
3937
  "pending",
3755
3938
  "event",
3756
3939
  "changes"
@@ -3767,6 +3950,10 @@
3767
3950
  "title": "Timestamp",
3768
3951
  "type": "string"
3769
3952
  },
3953
+ "working_start": {
3954
+ "title": "Working Start",
3955
+ "type": "number"
3956
+ },
3770
3957
  "pending": {
3771
3958
  "anyOf": [
3772
3959
  {
@@ -3812,6 +3999,7 @@
3812
3999
  },
3813
4000
  "required": [
3814
4001
  "timestamp",
4002
+ "working_start",
3815
4003
  "pending",
3816
4004
  "event",
3817
4005
  "action",
@@ -3830,6 +4018,10 @@
3830
4018
  "title": "Timestamp",
3831
4019
  "type": "string"
3832
4020
  },
4021
+ "working_start": {
4022
+ "title": "Working Start",
4023
+ "type": "number"
4024
+ },
3833
4025
  "pending": {
3834
4026
  "anyOf": [
3835
4027
  {
@@ -3858,6 +4050,7 @@
3858
4050
  },
3859
4051
  "required": [
3860
4052
  "timestamp",
4053
+ "working_start",
3861
4054
  "pending",
3862
4055
  "event",
3863
4056
  "changes"
@@ -3874,6 +4067,10 @@
3874
4067
  "title": "Timestamp",
3875
4068
  "type": "string"
3876
4069
  },
4070
+ "working_start": {
4071
+ "title": "Working Start",
4072
+ "type": "number"
4073
+ },
3877
4074
  "pending": {
3878
4075
  "anyOf": [
3879
4076
  {
@@ -3968,17 +4165,45 @@
3968
4165
  },
3969
4166
  "title": "Events",
3970
4167
  "type": "array"
4168
+ },
4169
+ "completed": {
4170
+ "anyOf": [
4171
+ {
4172
+ "format": "date-time",
4173
+ "type": "string"
4174
+ },
4175
+ {
4176
+ "type": "null"
4177
+ }
4178
+ ],
4179
+ "default": null,
4180
+ "title": "Completed"
4181
+ },
4182
+ "working_time": {
4183
+ "anyOf": [
4184
+ {
4185
+ "type": "number"
4186
+ },
4187
+ {
4188
+ "type": "null"
4189
+ }
4190
+ ],
4191
+ "default": null,
4192
+ "title": "Working Time"
3971
4193
  }
3972
4194
  },
3973
4195
  "required": [
3974
4196
  "timestamp",
4197
+ "working_start",
3975
4198
  "pending",
3976
4199
  "event",
3977
4200
  "name",
3978
4201
  "type",
3979
4202
  "input",
3980
4203
  "result",
3981
- "events"
4204
+ "events",
4205
+ "completed",
4206
+ "working_time"
3982
4207
  ],
3983
4208
  "title": "SubtaskEvent",
3984
4209
  "type": "object",
@@ -4148,6 +4373,10 @@
4148
4373
  "title": "Timestamp",
4149
4374
  "type": "string"
4150
4375
  },
4376
+ "working_start": {
4377
+ "title": "Working Start",
4378
+ "type": "number"
4379
+ },
4151
4380
  "pending": {
4152
4381
  "anyOf": [
4153
4382
  {
@@ -4215,6 +4444,9 @@
4215
4444
  {
4216
4445
  "$ref": "#/$defs/ContentText"
4217
4446
  },
4447
+ {
4448
+ "$ref": "#/$defs/ContentReasoning"
4449
+ },
4218
4450
  {
4219
4451
  "$ref": "#/$defs/ContentImage"
4220
4452
  },
@@ -4230,6 +4462,9 @@
4230
4462
  {
4231
4463
  "$ref": "#/$defs/ContentText"
4232
4464
  },
4465
+ {
4466
+ "$ref": "#/$defs/ContentReasoning"
4467
+ },
4233
4468
  {
4234
4469
  "$ref": "#/$defs/ContentImage"
4235
4470
  },
@@ -4331,10 +4566,36 @@
4331
4566
  },
4332
4567
  "title": "Events",
4333
4568
  "type": "array"
4569
+ },
4570
+ "completed": {
4571
+ "anyOf": [
4572
+ {
4573
+ "format": "date-time",
4574
+ "type": "string"
4575
+ },
4576
+ {
4577
+ "type": "null"
4578
+ }
4579
+ ],
4580
+ "default": null,
4581
+ "title": "Completed"
4582
+ },
4583
+ "working_time": {
4584
+ "anyOf": [
4585
+ {
4586
+ "type": "number"
4587
+ },
4588
+ {
4589
+ "type": "null"
4590
+ }
4591
+ ],
4592
+ "default": null,
4593
+ "title": "Working Time"
4334
4594
  }
4335
4595
  },
4336
4596
  "required": [
4337
4597
  "timestamp",
4598
+ "working_start",
4338
4599
  "pending",
4339
4600
  "event",
4340
4601
  "type",
@@ -4345,7 +4606,9 @@
4345
4606
  "result",
4346
4607
  "truncated",
4347
4608
  "error",
4348
- "events"
4609
+ "events",
4610
+ "completed",
4611
+ "working_time"
4349
4612
  ],
4350
4613
  "title": "ToolEvent",
4351
4614
  "type": "object",