@bjoernrochel/sevro 0.1.0-rc.2 → 0.1.0-rc.21

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 (243) hide show
  1. package/CONTRIBUTING.md +170 -0
  2. package/LICENSE +63 -136
  3. package/README.md +83 -66
  4. package/docs/README.md +44 -0
  5. package/docs/acknowledgements.md +14 -0
  6. package/docs/architecture.md +146 -0
  7. package/docs/assets/README.md +27 -0
  8. package/docs/assets/sevro-logo.png +0 -0
  9. package/docs/assets/sevro-symbol.svg +1 -0
  10. package/docs/assets/sevro-wordmark.svg +1 -0
  11. package/docs/builtin-graders-v1.md +41 -13
  12. package/docs/codex-host.md +184 -52
  13. package/docs/contracts.md +19 -0
  14. package/docs/creating-cases.md +139 -0
  15. package/docs/development-cli.md +190 -30
  16. package/docs/documentation-quality.md +61 -0
  17. package/docs/evaluation-workflow.md +110 -0
  18. package/docs/evidence/code-mode-controls-71.md +64 -0
  19. package/docs/evidence/code-mode-programs-85.md +64 -0
  20. package/docs/evidence/codex-state-73/README.md +204 -0
  21. package/docs/evidence/codex-state-73/darrow-migration.patch +182 -0
  22. package/docs/evidence/codex-state-73/results.json +661 -0
  23. package/docs/evidence/local-http-74.md +104 -0
  24. package/docs/evidence/runtime-startup-30/README.md +185 -0
  25. package/docs/evidence/runtime-startup-30/after.json +319 -0
  26. package/docs/evidence/runtime-startup-30/baseline-timeout.jsonl +2 -0
  27. package/docs/evidence/runtime-startup-30/before.json +403 -0
  28. package/docs/evidence/runtime-startup-30/consumer-smoke.json +63 -0
  29. package/docs/evidence/runtime-startup-30/full-policy-diagnostic.jsonl +3 -0
  30. package/docs/evidence/runtime-startup-30/minimal-policy-diagnostic.jsonl +3 -0
  31. package/docs/evidence/runtime-startup-30/plugin-template-assessment.json +20 -0
  32. package/docs/evidence/runtime-startup-30/seedless-timeout.jsonl +2 -0
  33. package/docs/evidence/skill-reads-70/results.json +823 -0
  34. package/docs/evidence.md +21 -0
  35. package/docs/extension-protocol-v1.md +50 -16
  36. package/docs/fixture-cleanup-validation.md +3 -0
  37. package/docs/getting-started.md +102 -0
  38. package/docs/guide-evaluation.md +89 -0
  39. package/docs/guide-verification.json +239 -0
  40. package/docs/identity-v1.md +35 -1
  41. package/docs/installing.md +88 -0
  42. package/docs/licenses/BUSL-1.1-0.1.0-rc.2.txt +159 -0
  43. package/docs/licensing.md +42 -0
  44. package/docs/native-controls.md +76 -5
  45. package/docs/native-hosts.md +228 -0
  46. package/docs/reading-results.md +141 -0
  47. package/docs/recipes.md +489 -0
  48. package/docs/releases.md +81 -40
  49. package/docs/report-v1.md +29 -1
  50. package/docs/results-v1.md +180 -0
  51. package/docs/run-settings-v1.md +147 -0
  52. package/docs/running-evaluations.md +31 -0
  53. package/docs/runtime-v1.md +614 -0
  54. package/docs/specs/repository-guide.md +58 -0
  55. package/docs/troubleshooting.md +30 -0
  56. package/docs/typescript-quality.md +234 -0
  57. package/docs/usability-study.md +114 -0
  58. package/examples/basic/host.ts +2 -2
  59. package/examples/recipes/candidate.ts +26 -0
  60. package/examples/recipes/cases/coding-change.json +21 -0
  61. package/examples/recipes/cases/host-tools.json +21 -0
  62. package/examples/recipes/cases/mise.json +21 -0
  63. package/examples/recipes/cases/node.json +21 -0
  64. package/examples/recipes/cases/python-uv.json +21 -0
  65. package/examples/recipes/cases/semantic.json +17 -0
  66. package/examples/recipes/coding-change/check.ts +23 -0
  67. package/examples/recipes/extension-artifacts.ts +65 -0
  68. package/examples/recipes/extension-methods.ts +174 -0
  69. package/examples/recipes/extension.ts +37 -0
  70. package/examples/recipes/repository-skill/.agents/skills/startup-guide/SKILL.md +7 -0
  71. package/examples/recipes/repository-skill/.claude/skills/startup-guide/SKILL.md +7 -0
  72. package/examples/recipes/repository-skill/README.md +3 -0
  73. package/examples/recipes/run-inputs.ts +188 -0
  74. package/examples/recipes/run-options.ts +131 -0
  75. package/examples/recipes/run.ts +63 -0
  76. package/examples/recipes/runtime/empty.json +1 -0
  77. package/examples/recipes/runtime/goals.json +1 -0
  78. package/examples/recipes/runtime/mise.json +6 -0
  79. package/examples/recipes/runtime/path.json +1 -0
  80. package/examples/recipes/runtime/plugins.json +4 -0
  81. package/examples/recipes/runtime/python-uv.json +10 -0
  82. package/examples/recipes/semantic-judge.ts +29 -0
  83. package/examples/templates/coding/adapter.ts +17 -0
  84. package/examples/templates/coding/case.json +15 -0
  85. package/examples/templates/repository-skill/.agents/skills/startup-guide/SKILL.md +6 -0
  86. package/examples/templates/repository-skill/extension-data.ts +64 -0
  87. package/examples/templates/repository-skill/extension.ts +31 -0
  88. package/examples/templates/response/adapter.ts +12 -0
  89. package/examples/templates/response/case.json +13 -0
  90. package/package.json +39 -6
  91. package/schemas/report-v1.schema.json +9 -0
  92. package/schemas/run-evidence-v1.schema.json +2 -1
  93. package/schemas/runtime-v1.schema.json +193 -0
  94. package/src/advisory.ts +10 -5
  95. package/src/candidate-artifacts.ts +208 -0
  96. package/src/candidate-codex-state.ts +185 -0
  97. package/src/case-validation.ts +105 -0
  98. package/src/cli/cancellation.ts +20 -0
  99. package/src/cli/command.ts +81 -0
  100. package/src/cli/completion.ts +15 -0
  101. package/src/cli/demo.ts +59 -0
  102. package/src/cli/evaluation.ts +55 -0
  103. package/src/cli/extension.ts +79 -0
  104. package/src/cli/help.ts +222 -0
  105. package/src/cli/hosts.ts +59 -0
  106. package/src/cli/init-settings.ts +30 -0
  107. package/src/cli/init.ts +111 -0
  108. package/src/cli/invocation.ts +627 -0
  109. package/src/cli/loading.ts +151 -0
  110. package/src/cli/preparation.ts +163 -0
  111. package/src/cli/presentation.ts +91 -0
  112. package/src/cli/progress.ts +6 -0
  113. package/src/cli/run-manifest.ts +133 -0
  114. package/src/cli/run-options.ts +101 -0
  115. package/src/cli/run-overrides.ts +83 -0
  116. package/src/cli/settings.ts +160 -0
  117. package/src/cli/summary.ts +43 -0
  118. package/src/cli.ts +13 -843
  119. package/src/contract-types.ts +94 -0
  120. package/src/engine.ts +19 -2373
  121. package/src/evaluation/advisory.ts +150 -0
  122. package/src/evaluation/check-plan.ts +510 -0
  123. package/src/evaluation/collection.ts +88 -0
  124. package/src/evaluation/context.ts +10 -0
  125. package/src/evaluation/grading-host.ts +39 -0
  126. package/src/evaluation/grading.ts +258 -0
  127. package/src/evaluation/invocation-plan.ts +548 -0
  128. package/src/evaluation/plan.ts +736 -0
  129. package/src/evaluation/progress.ts +19 -0
  130. package/src/evaluation/resources.ts +241 -0
  131. package/src/evaluation/run.ts +206 -0
  132. package/src/{trial-scheduler.ts → evaluation/scheduler.ts} +3 -1
  133. package/src/evaluation/semantic.ts +312 -0
  134. package/src/evaluation/trial.ts +866 -0
  135. package/src/evaluation/types.ts +207 -0
  136. package/src/evidence/host-results.ts +92 -0
  137. package/src/evidence/retained-artifacts.ts +18 -0
  138. package/src/evidence/trial-writer.ts +203 -0
  139. package/src/evidence/usage.ts +17 -0
  140. package/src/extension-client.ts +228 -148
  141. package/src/extension-session.ts +326 -195
  142. package/src/fixtures/advisory.ts +338 -0
  143. package/src/fixtures/bin.ts +57 -0
  144. package/src/{fixture-cleanup.ts → fixtures/cleanup.ts} +13 -7
  145. package/src/fixtures/create.ts +187 -0
  146. package/src/fixtures/generated.ts +300 -0
  147. package/src/fixtures/git-hooks.ts +24 -0
  148. package/src/fixtures/preparation.ts +270 -0
  149. package/src/{repository-fixture.ts → fixtures/repository.ts} +207 -85
  150. package/src/fixtures/setup.ts +224 -0
  151. package/src/fixtures/tools.ts +88 -0
  152. package/src/fixtures/workspace-state.ts +212 -0
  153. package/src/generated/cli-result.cjs +2 -0
  154. package/src/generated/cli-result.d.cts +3 -0
  155. package/src/generated/extension.cjs +2 -0
  156. package/src/generated/extension.d.cts +3 -0
  157. package/src/generated/report.cjs +2 -0
  158. package/src/generated/report.d.cts +3 -0
  159. package/src/generated/run-evidence.cjs +2 -0
  160. package/src/generated/run-evidence.d.cts +3 -0
  161. package/src/generated/runtime.cjs +2 -0
  162. package/src/generated/runtime.d.cts +3 -0
  163. package/src/graders/artifact.ts +56 -23
  164. package/src/graders/git-head.ts +79 -75
  165. package/src/graders/output.ts +222 -140
  166. package/src/graders/semantic.ts +102 -43
  167. package/src/graders/shell.ts +379 -207
  168. package/src/host-diagnostic-redaction.ts +52 -0
  169. package/src/host-diagnostics.ts +319 -0
  170. package/src/hosts/claude-continuation.ts +63 -18
  171. package/src/hosts/claude-credential.ts +43 -62
  172. package/src/hosts/claude-events.ts +86 -51
  173. package/src/hosts/claude-keychain.ts +59 -0
  174. package/src/hosts/claude-native-goal.ts +57 -33
  175. package/src/hosts/claude-process.ts +111 -0
  176. package/src/hosts/claude-repository-invocation.ts +208 -117
  177. package/src/hosts/claude-settings.ts +240 -21
  178. package/src/hosts/claude-skill-activations.ts +166 -0
  179. package/src/hosts/claude-state.ts +403 -0
  180. package/src/hosts/claude-tool-calls.ts +171 -106
  181. package/src/hosts/claude.ts +561 -451
  182. package/src/hosts/codex-app-server.ts +326 -178
  183. package/src/hosts/codex-configuration.ts +44 -18
  184. package/src/hosts/codex-events.ts +97 -60
  185. package/src/hosts/codex-exec-controls.ts +516 -0
  186. package/src/hosts/codex-invocation.ts +70 -0
  187. package/src/hosts/codex-marketplace.ts +283 -0
  188. package/src/hosts/codex-native-calls.ts +546 -500
  189. package/src/hosts/codex-process.ts +202 -0
  190. package/src/hosts/codex-profile.ts +257 -28
  191. package/src/hosts/codex-session-files.ts +106 -0
  192. package/src/hosts/codex-state.ts +492 -0
  193. package/src/hosts/codex.ts +783 -832
  194. package/src/hosts/isolation-common.ts +94 -0
  195. package/src/hosts/isolation-roots.ts +54 -27
  196. package/src/hosts/isolation.ts +10 -0
  197. package/src/hosts/linux-sandbox.ts +159 -0
  198. package/src/hosts/mac-sandbox.ts +187 -61
  199. package/src/hosts/native-controls.ts +152 -62
  200. package/src/hosts/process-lifecycle.ts +37 -0
  201. package/src/hosts/protected-worktrees.ts +51 -19
  202. package/src/hosts/skill-hook-catalog.ts +131 -0
  203. package/src/hosts/skill-hook-handler.ts +374 -0
  204. package/src/hosts/skill-hooks.ts +580 -0
  205. package/src/hosts/workspace-fingerprint.ts +55 -35
  206. package/src/identity.ts +102 -48
  207. package/src/instrumentation.ts +109 -48
  208. package/src/native-transcript-state.ts +131 -0
  209. package/src/native-transcript-view.ts +90 -0
  210. package/src/native-transcripts.ts +286 -0
  211. package/src/process-lifecycle.ts +150 -0
  212. package/src/provenance.ts +161 -81
  213. package/src/prune.ts +215 -0
  214. package/src/report.ts +172 -140
  215. package/src/resolved-case.ts +33 -0
  216. package/src/results-lock.ts +29 -0
  217. package/src/results.ts +133 -86
  218. package/src/run-owner.ts +154 -63
  219. package/src/runtime/config.ts +456 -0
  220. package/src/runtime/discovery-apple.ts +90 -0
  221. package/src/runtime/discovery-providers.ts +63 -0
  222. package/src/runtime/discovery.ts +70 -0
  223. package/src/runtime/hook-processes.ts +97 -0
  224. package/src/runtime/hooks.ts +476 -0
  225. package/src/runtime/paths.ts +313 -0
  226. package/src/runtime/query.ts +85 -0
  227. package/src/runtime/seeds.ts +134 -0
  228. package/src/runtime/shared-cache.ts +143 -0
  229. package/src/runtime/state.ts +143 -0
  230. package/src/schema-types.ts +198 -0
  231. package/src/schema.ts +15 -19
  232. package/src/transcript-retention.ts +146 -0
  233. package/src/value-guards.ts +20 -0
  234. package/src/advisory-fixture.ts +0 -282
  235. package/src/fixture-bin.ts +0 -87
  236. package/src/fixture-setup.ts +0 -163
  237. package/src/generated-fixture.ts +0 -220
  238. package/src/git-hooks.ts +0 -63
  239. package/src/hosts/claude-nested-skills.ts +0 -218
  240. package/src/hosts/codex-command-output.ts +0 -399
  241. package/src/hosts/codex-executor-command.ts +0 -195
  242. package/src/hosts/codex-skill-reads.ts +0 -391
  243. package/src/preparation.ts +0 -211
@@ -0,0 +1,94 @@
1
+ /** Shared structures; public schemas and runtime guards still validate external data. */
2
+ export type ExecutionStatus = "completed" | "failed" | "cancelled" | "not_run";
3
+ export type GradingStatus =
4
+ "completed" | "error" | "unavailable" | "not_requested";
5
+ export type TaskVerdict = "passed" | "failed" | "not_assessed";
6
+ export type OutcomeStatus = "passed" | "failed" | "unavailable";
7
+ export type ExitCode = 0 | 1 | 2 | 3 | 4 | 64 | 70 | 130 | 143;
8
+
9
+ export interface Assessment {
10
+ execution: { status: ExecutionStatus };
11
+ grading: { status: GradingStatus };
12
+ task: { verdict: TaskVerdict };
13
+ }
14
+
15
+ export type CheckDeclaration = {
16
+ id: string;
17
+ grader: string;
18
+ configuration: Record<string, unknown>;
19
+ };
20
+
21
+ /** Runner check outcomes include their grader and may lack evidence references. */
22
+ export interface CheckOutcome {
23
+ id: string;
24
+ grader: string;
25
+ status: OutcomeStatus;
26
+ detail?: string;
27
+ evidenceRefs?: string[];
28
+ }
29
+
30
+ /** Extension checks require references; their grader comes from the declaration. */
31
+ export type ExtensionCheckOutcome = {
32
+ id: string;
33
+ status: OutcomeStatus;
34
+ detail?: string;
35
+ evidenceRefs: string[];
36
+ };
37
+
38
+ /** Domain outcomes are retained independently of task checks and verdicts. */
39
+ export type DomainOutcome = {
40
+ id: string;
41
+ status: OutcomeStatus;
42
+ detail?: string;
43
+ evidenceRefs: string[];
44
+ data?: Record<string, unknown>;
45
+ };
46
+
47
+ export type ObservationCompleteness = "complete" | "partial" | "unavailable";
48
+
49
+ export type ObservationEvidence = {
50
+ id: string;
51
+ source: string;
52
+ completeness: ObservationCompleteness;
53
+ data: Record<string, unknown>;
54
+ };
55
+
56
+ export type MetricEvidence = {
57
+ id: string;
58
+ value: number | null;
59
+ unit: string;
60
+ };
61
+
62
+ /** Retained measurements keep unknown values and completeness separate. */
63
+ export type UsageEvidence = {
64
+ inputTokens: number | null;
65
+ outputTokens: number | null;
66
+ costUsd: number | null;
67
+ complete: boolean;
68
+ };
69
+
70
+ export type RawResultReference = {
71
+ source: string;
72
+ path: string | null;
73
+ sha256: string | null;
74
+ };
75
+
76
+ export type ArtifactReference = {
77
+ id: string;
78
+ path: string;
79
+ sha256: string;
80
+ gitExclude?: boolean;
81
+ executable?: boolean;
82
+ };
83
+
84
+ export type EvidenceRoute = {
85
+ role: "candidate" | "semantic" | "advisory";
86
+ host: string;
87
+ model: string;
88
+ effort: string;
89
+ };
90
+
91
+ export type Diagnostic = {
92
+ code: string;
93
+ message: string;
94
+ };