engineering-behavior-observatory 0.2.1

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 (178) hide show
  1. package/LICENSE +202 -0
  2. package/README.md +151 -0
  3. package/contracts/codex-app-server-0.150.1/manifest.json +25 -0
  4. package/contracts/codex-app-server-0.150.1/schema/InitializeParams.json +84 -0
  5. package/contracts/codex-app-server-0.150.1/schema/JSONRPCMessage.json +137 -0
  6. package/contracts/codex-app-server-0.150.1/schema/v2/ThreadReadParams.json +17 -0
  7. package/contracts/codex-app-server-0.150.1/schema/v2/ThreadTokenUsageUpdatedNotification.json +82 -0
  8. package/contracts/codex-app-server-0.150.1/schema/v2/TurnInterruptParams.json +17 -0
  9. package/contracts/codex-app-server-0.150.1/types/AskForApproval.ts +5 -0
  10. package/contracts/codex-app-server-0.150.1/types/ClientNotification.ts +5 -0
  11. package/contracts/codex-app-server-0.150.1/types/SandboxMode.ts +5 -0
  12. package/contracts/codex-app-server-0.150.1/types/ThreadReadParams.ts +9 -0
  13. package/contracts/codex-app-server-0.150.1/types/TokenUsageBreakdown.ts +5 -0
  14. package/contracts/codex-app-server-0.150.1/types/TurnInterruptParams.ts +5 -0
  15. package/contracts/codex-app-server-0.153.4/manifest.json +27 -0
  16. package/contracts/codex-app-server-0.153.4/schema/JSONRPCMessage.json +137 -0
  17. package/contracts/codex-app-server-0.153.4/schema/v1/InitializeParams.json +84 -0
  18. package/contracts/codex-app-server-0.153.4/schema/v2/ThreadReadParams.json +17 -0
  19. package/contracts/codex-app-server-0.153.4/schema/v2/ThreadStartParams.json +515 -0
  20. package/contracts/codex-app-server-0.153.4/schema/v2/ThreadTokenUsageUpdatedNotification.json +82 -0
  21. package/contracts/codex-app-server-0.153.4/schema/v2/TurnInterruptParams.json +17 -0
  22. package/contracts/codex-app-server-0.153.4/schema/v2/TurnStartParams.json +921 -0
  23. package/contracts/codex-app-server-0.153.4/types/AskForApproval.ts +5 -0
  24. package/contracts/codex-app-server-0.153.4/types/ClientNotification.ts +5 -0
  25. package/contracts/codex-app-server-0.153.4/types/SandboxMode.ts +5 -0
  26. package/contracts/codex-app-server-0.153.4/types/ThreadReadParams.ts +12 -0
  27. package/contracts/codex-app-server-0.153.4/types/TokenUsageBreakdown.ts +5 -0
  28. package/contracts/codex-app-server-0.153.4/types/TurnInterruptParams.ts +5 -0
  29. package/contracts/openhands-agent-server-v1.44.1.json +65 -0
  30. package/contracts/openhands-agent-server-v1.46.0.json +46 -0
  31. package/dist/contracts/codex-app-server-0.153.4/types/AskForApproval.d.ts +9 -0
  32. package/dist/contracts/codex-app-server-0.153.4/types/AskForApproval.js +2 -0
  33. package/dist/contracts/codex-app-server-0.153.4/types/ClientNotification.d.ts +3 -0
  34. package/dist/contracts/codex-app-server-0.153.4/types/ClientNotification.js +2 -0
  35. package/dist/contracts/codex-app-server-0.153.4/types/SandboxMode.d.ts +1 -0
  36. package/dist/contracts/codex-app-server-0.153.4/types/SandboxMode.js +2 -0
  37. package/dist/contracts/codex-app-server-0.153.4/types/ThreadReadParams.d.ts +10 -0
  38. package/dist/contracts/codex-app-server-0.153.4/types/ThreadReadParams.js +2 -0
  39. package/dist/contracts/codex-app-server-0.153.4/types/TokenUsageBreakdown.d.ts +8 -0
  40. package/dist/contracts/codex-app-server-0.153.4/types/TokenUsageBreakdown.js +2 -0
  41. package/dist/contracts/codex-app-server-0.153.4/types/TurnInterruptParams.d.ts +4 -0
  42. package/dist/contracts/codex-app-server-0.153.4/types/TurnInterruptParams.js +2 -0
  43. package/dist/src/agent-sdk-normalizer.d.ts +22 -0
  44. package/dist/src/agent-sdk-normalizer.js +995 -0
  45. package/dist/src/agent-sdk-run.d.ts +31 -0
  46. package/dist/src/agent-sdk-run.js +242 -0
  47. package/dist/src/agent-sdk-runner.d.ts +96 -0
  48. package/dist/src/agent-sdk-runner.js +407 -0
  49. package/dist/src/agent-sdk.d.ts +294 -0
  50. package/dist/src/agent-sdk.js +644 -0
  51. package/dist/src/aggregation.d.ts +183 -0
  52. package/dist/src/aggregation.js +513 -0
  53. package/dist/src/artifacts.d.ts +36 -0
  54. package/dist/src/artifacts.js +1372 -0
  55. package/dist/src/atlas-grafana.d.ts +377 -0
  56. package/dist/src/atlas-grafana.js +47 -0
  57. package/dist/src/atlas-html.d.ts +2 -0
  58. package/dist/src/atlas-html.js +34 -0
  59. package/dist/src/atlas.d.ts +130 -0
  60. package/dist/src/atlas.js +235 -0
  61. package/dist/src/behavior-assertions.d.ts +95 -0
  62. package/dist/src/behavior-assertions.js +154 -0
  63. package/dist/src/cli.d.ts +5 -0
  64. package/dist/src/cli.js +957 -0
  65. package/dist/src/codex-judge.d.ts +44 -0
  66. package/dist/src/codex-judge.js +214 -0
  67. package/dist/src/codex-run.d.ts +114 -0
  68. package/dist/src/codex-run.js +519 -0
  69. package/dist/src/codex.d.ts +187 -0
  70. package/dist/src/codex.js +1028 -0
  71. package/dist/src/contracts.d.ts +155 -0
  72. package/dist/src/contracts.js +866 -0
  73. package/dist/src/corpus.d.ts +67 -0
  74. package/dist/src/corpus.js +540 -0
  75. package/dist/src/cursor-sdk-runner.d.ts +81 -0
  76. package/dist/src/cursor-sdk-runner.js +362 -0
  77. package/dist/src/cursor-sdk.d.ts +81 -0
  78. package/dist/src/cursor-sdk.js +1078 -0
  79. package/dist/src/deepseek-adapter.d.ts +152 -0
  80. package/dist/src/deepseek-adapter.js +777 -0
  81. package/dist/src/exports.d.ts +69 -0
  82. package/dist/src/exports.js +966 -0
  83. package/dist/src/human-calibration.d.ts +167 -0
  84. package/dist/src/human-calibration.js +618 -0
  85. package/dist/src/index.d.ts +71 -0
  86. package/dist/src/index.js +38 -0
  87. package/dist/src/lifecycle.d.ts +196 -0
  88. package/dist/src/lifecycle.js +1889 -0
  89. package/dist/src/normalization-integrity.d.ts +128 -0
  90. package/dist/src/normalization-integrity.js +429 -0
  91. package/dist/src/openhands-run.d.ts +36 -0
  92. package/dist/src/openhands-run.js +415 -0
  93. package/dist/src/openhands.d.ts +130 -0
  94. package/dist/src/openhands.js +859 -0
  95. package/dist/src/pi.d.ts +203 -0
  96. package/dist/src/pi.js +1345 -0
  97. package/dist/src/process-protocol.d.ts +245 -0
  98. package/dist/src/process-protocol.js +1261 -0
  99. package/dist/src/retained-evidence.d.ts +12 -0
  100. package/dist/src/retained-evidence.js +201 -0
  101. package/dist/src/run-bundles.d.ts +193 -0
  102. package/dist/src/run-bundles.js +993 -0
  103. package/dist/src/scheduler.d.ts +109 -0
  104. package/dist/src/scheduler.js +1080 -0
  105. package/dist/src/semantic-judge.d.ts +194 -0
  106. package/dist/src/semantic-judge.js +875 -0
  107. package/dist/src/structural-observations.d.ts +124 -0
  108. package/dist/src/structural-observations.js +671 -0
  109. package/dist/src/task-packets.d.ts +115 -0
  110. package/dist/src/task-packets.js +683 -0
  111. package/dist/src/uniform-events.d.ts +120 -0
  112. package/dist/src/uniform-events.js +158 -0
  113. package/dist/src/verifiers.d.ts +116 -0
  114. package/dist/src/verifiers.js +819 -0
  115. package/dist/src/workspaces.d.ts +56 -0
  116. package/dist/src/workspaces.js +1314 -0
  117. package/docs/README.md +60 -0
  118. package/docs/development/README.md +71 -0
  119. package/docs/development/documentation-sync.md +58 -0
  120. package/docs/development/extension-contracts.md +307 -0
  121. package/docs/evaluation/README.md +48 -0
  122. package/docs/evaluation/aggregation.md +112 -0
  123. package/docs/evaluation/behavior-assertions.md +45 -0
  124. package/docs/evaluation/human-calibration.md +170 -0
  125. package/docs/evaluation/normalization-integrity.md +80 -0
  126. package/docs/evaluation/semantic-judge.md +188 -0
  127. package/docs/evaluation/structural-observations.md +84 -0
  128. package/docs/evaluation/uniform-events.md +131 -0
  129. package/docs/guides/atlas.md +144 -0
  130. package/docs/guides/evidence-and-sharing.md +100 -0
  131. package/docs/guides/operator-guide.md +424 -0
  132. package/docs/guides/quickstart.md +152 -0
  133. package/docs/guides/telemetry.md +66 -0
  134. package/docs/harnesses/README.md +31 -0
  135. package/docs/harnesses/claude-agent-sdk.md +61 -0
  136. package/docs/harnesses/codex-harness.md +174 -0
  137. package/docs/harnesses/cursor-sdk.md +149 -0
  138. package/docs/harnesses/deepseek-harness.md +134 -0
  139. package/docs/harnesses/openhands-agent-server.md +153 -0
  140. package/docs/harnesses/pi-sdk.md +66 -0
  141. package/docs/reference/README.md +19 -0
  142. package/docs/reference/agent-sdk-operational-runner.md +118 -0
  143. package/docs/reference/cli.md +114 -0
  144. package/docs/reference/contracts.md +222 -0
  145. package/docs/reference/run-bundle-contract.md +354 -0
  146. package/docs/reference/run-lifecycle.md +54 -0
  147. package/examples/cursor-sdk/README.md +11 -0
  148. package/examples/cursor-sdk/capture-profile.json +1 -0
  149. package/examples/cursor-sdk/harness.json +1 -0
  150. package/examples/cursor-sdk/model.json +1 -0
  151. package/examples/cursor-sdk/native-limits.json +1 -0
  152. package/examples/cursor-sdk/native-tool-policy.json +1 -0
  153. package/ontology/behavior-categories.v1.json +46 -0
  154. package/package.json +66 -0
  155. package/release/0.1.0/KNOWN_LIMITATIONS.md +35 -0
  156. package/release/0.1.0/README.md +46 -0
  157. package/release/0.1.0/reproducibility.json +61 -0
  158. package/release/0.2.0/KNOWN_LIMITATIONS.md +40 -0
  159. package/release/0.2.0/README.md +42 -0
  160. package/release/0.2.0/reproducibility.json +73 -0
  161. package/release/0.2.1/KNOWN_LIMITATIONS.md +18 -0
  162. package/release/0.2.1/README.md +63 -0
  163. package/release/0.2.1/reproducibility.json +73 -0
  164. package/release/README.md +13 -0
  165. package/schemas/aggregation.v1.json +293 -0
  166. package/schemas/behavior-assertions.v1.json +220 -0
  167. package/schemas/deepseek-runtime-composition.v1.schema.json +122 -0
  168. package/schemas/experiment.v1.schema.json +188 -0
  169. package/schemas/human-calibration.v1.json +336 -0
  170. package/schemas/normalization-integrity.v1.json +302 -0
  171. package/schemas/run-bundles/v1.json +1040 -0
  172. package/schemas/run-queue.v1.schema.json +206 -0
  173. package/schemas/semantic-judge.v1.json +380 -0
  174. package/schemas/structural-observations.v1.json +131 -0
  175. package/schemas/task-packet-freeze.v1.schema.json +106 -0
  176. package/schemas/task-packet.v1.schema.json +234 -0
  177. package/schemas/uniform-events/v1.json +289 -0
  178. package/scripts/atlas-grafana.sh +18 -0
@@ -0,0 +1,293 @@
1
+ {
2
+ "$schema": "https://json-schema.org/draft/2020-12/schema",
3
+ "$id": "urn:ebo:schema:aggregation:v1",
4
+ "$defs": {
5
+ "identifier": { "type": "string", "minLength": 1, "maxLength": 1024, "pattern": ".*\\S.*" },
6
+ "path": { "type": "string", "minLength": 1, "maxLength": 4096, "pattern": ".*\\S.*" },
7
+ "derivedSource": {
8
+ "type": "object", "additionalProperties": false, "required": ["bundleRoot", "path"],
9
+ "properties": { "bundleRoot": { "$ref": "#/$defs/path" }, "path": { "$ref": "#/$defs/path" } }
10
+ },
11
+ "eligibilityGate": {
12
+ "type": "object", "additionalProperties": false, "required": ["request", "report"],
13
+ "properties": { "request": { "$ref": "#/$defs/path" }, "report": { "$ref": "#/$defs/path" } }
14
+ },
15
+ "dimension": { "enum": ["task", "model", "harness", "trial", "capture-qualification"] },
16
+ "filters": {
17
+ "type": "object",
18
+ "additionalProperties": false,
19
+ "properties": {
20
+ "task": { "$ref": "#/$defs/identifier" },
21
+ "model": { "$ref": "#/$defs/identifier" },
22
+ "harness": { "$ref": "#/$defs/identifier" },
23
+ "trial": { "$ref": "#/$defs/identifier" },
24
+ "capture-qualification": { "$ref": "#/$defs/identifier" }
25
+ }
26
+ },
27
+ "request": {
28
+ "type": "object",
29
+ "additionalProperties": false,
30
+ "required": ["schemaVersion", "sources", "groupBy", "selectedAttemptPolicy", "recurrence", "comparisons"],
31
+ "properties": {
32
+ "schemaVersion": { "const": "ebo.aggregation-request/v1" },
33
+ "sources": {
34
+ "type": "object",
35
+ "additionalProperties": false,
36
+ "required": ["corpusRoot", "corpusIndex", "observationSets", "assertions", "calibrations"],
37
+ "properties": {
38
+ "corpusRoot": { "$ref": "#/$defs/path" },
39
+ "corpusIndex": { "$ref": "#/$defs/path" },
40
+ "observationSets": { "type": "array", "maxItems": 100000, "items": { "$ref": "#/$defs/derivedSource" } },
41
+ "assertions": { "type": "array", "maxItems": 100000, "items": { "$ref": "#/$defs/derivedSource" } },
42
+ "calibrations": {
43
+ "type": "array",
44
+ "maxItems": 100000,
45
+ "items": {
46
+ "type": "object",
47
+ "additionalProperties": false,
48
+ "required": ["selection", "history"],
49
+ "properties": { "selection": { "$ref": "#/$defs/path" }, "history": { "$ref": "#/$defs/path" } }
50
+ }
51
+ }
52
+ }
53
+ },
54
+ "groupBy": { "type": "array", "maxItems": 5, "uniqueItems": true, "items": { "$ref": "#/$defs/dimension" } },
55
+ "selectedAttemptPolicy": { "enum": ["all-attempts", "latest-attempt-per-run"] },
56
+ "recurrence": {
57
+ "type": "object", "additionalProperties": false, "required": ["minimumOccurrences"],
58
+ "properties": { "minimumOccurrences": { "type": "integer", "minimum": 2 } }
59
+ },
60
+ "comparisons": {
61
+ "type": "array",
62
+ "maxItems": 100000,
63
+ "items": {
64
+ "type": "object",
65
+ "additionalProperties": false,
66
+ "required": ["id", "measure", "left", "right", "matchBy", "eligibilityGates"],
67
+ "properties": {
68
+ "id": { "$ref": "#/$defs/identifier" },
69
+ "measure": { "$ref": "#/$defs/identifier" },
70
+ "left": { "$ref": "#/$defs/filters" },
71
+ "right": { "$ref": "#/$defs/filters" },
72
+ "matchBy": { "type": "array", "minItems": 1, "maxItems": 5, "uniqueItems": true, "items": { "$ref": "#/$defs/dimension" } },
73
+ "eligibilityGates": { "type": "array", "minItems": 1, "maxItems": 100000, "items": { "$ref": "#/$defs/eligibilityGate" } }
74
+ }
75
+ }
76
+ }
77
+ }
78
+ },
79
+ "quantity": {
80
+ "type": "object", "additionalProperties": false, "required": ["value", "unit"],
81
+ "properties": { "value": { "type": "number" }, "unit": { "$ref": "#/$defs/identifier" } }
82
+ },
83
+ "exclusion": {
84
+ "type": "object", "additionalProperties": false, "required": ["reason", "count", "unit"],
85
+ "properties": { "reason": { "$ref": "#/$defs/identifier" }, "count": { "type": "integer", "minimum": 0 }, "unit": { "$ref": "#/$defs/identifier" } }
86
+ },
87
+ "measurement": {
88
+ "type": "object",
89
+ "additionalProperties": false,
90
+ "required": ["status", "numerator", "denominator", "exclusions"],
91
+ "properties": {
92
+ "status": { "enum": ["available", "unavailable"] },
93
+ "numerator": { "$ref": "#/$defs/quantity" },
94
+ "denominator": { "$ref": "#/$defs/quantity" },
95
+ "rate": { "type": "number" },
96
+ "exclusions": { "type": "array", "maxItems": 100000, "items": { "$ref": "#/$defs/exclusion" } },
97
+ "reason": { "$ref": "#/$defs/identifier" }
98
+ },
99
+ "allOf": [
100
+ { "if": { "properties": { "status": { "const": "available" } } }, "then": { "required": ["rate"] } },
101
+ { "if": { "properties": { "status": { "const": "unavailable" } } }, "then": { "required": ["reason"], "not": { "required": ["rate"] } } }
102
+ ]
103
+ },
104
+ "metric": {
105
+ "type": "object", "additionalProperties": false, "required": ["id", "population", "claimStatus", "measurement"],
106
+ "properties": {
107
+ "id": { "$ref": "#/$defs/identifier" },
108
+ "population": { "enum": ["run", "attempt", "operation", "assertion", "reviewed-assertion"] },
109
+ "claimStatus": { "enum": ["descriptive", "unavailable"] },
110
+ "measurement": { "$ref": "#/$defs/measurement" }
111
+ }
112
+ },
113
+ "sourceLineage": {
114
+ "type": "object", "additionalProperties": false,
115
+ "required": ["requestDigest", "corpusIndexDigest", "manifests", "observationSets", "assertions", "calibrations", "comparisonGates"],
116
+ "properties": {
117
+ "requestDigest": { "$ref": "urn:ebo:schema:normalization-integrity:v1#/$defs/digest" },
118
+ "corpusIndexDigest": { "$ref": "urn:ebo:schema:normalization-integrity:v1#/$defs/digest" },
119
+ "manifests": {
120
+ "type": "array", "maxItems": 100000,
121
+ "items": { "type": "object", "additionalProperties": false, "required": ["runId", "attemptId", "digest"], "properties": {
122
+ "runId": { "$ref": "#/$defs/identifier" }, "attemptId": { "$ref": "#/$defs/identifier" }, "digest": { "$ref": "urn:ebo:schema:normalization-integrity:v1#/$defs/digest" }
123
+ } }
124
+ },
125
+ "observationSets": {
126
+ "type": "array", "maxItems": 100000,
127
+ "items": { "type": "object", "additionalProperties": false, "required": ["runId", "attemptId", "digest"], "properties": {
128
+ "runId": { "$ref": "#/$defs/identifier" }, "attemptId": { "$ref": "#/$defs/identifier" }, "digest": { "$ref": "urn:ebo:schema:normalization-integrity:v1#/$defs/digest" }
129
+ } }
130
+ },
131
+ "assertions": {
132
+ "type": "array", "maxItems": 100000,
133
+ "items": { "type": "object", "additionalProperties": false, "required": ["runId", "attemptId", "id", "digest"], "properties": {
134
+ "runId": { "$ref": "#/$defs/identifier" }, "attemptId": { "$ref": "#/$defs/identifier" }, "id": { "$ref": "#/$defs/identifier" }, "digest": { "$ref": "urn:ebo:schema:normalization-integrity:v1#/$defs/digest" }
135
+ } }
136
+ },
137
+ "calibrations": {
138
+ "type": "array", "maxItems": 100000,
139
+ "items": { "type": "object", "additionalProperties": false, "required": ["selectionDigest", "historyDigest"], "properties": {
140
+ "selectionDigest": { "$ref": "urn:ebo:schema:normalization-integrity:v1#/$defs/digest" }, "historyDigest": { "$ref": "urn:ebo:schema:normalization-integrity:v1#/$defs/digest" }
141
+ } }
142
+ },
143
+ "comparisonGates": {
144
+ "type": "array", "maxItems": 100000,
145
+ "items": { "type": "object", "additionalProperties": false, "required": ["requestDigest", "reportDigest"], "properties": {
146
+ "requestDigest": { "$ref": "urn:ebo:schema:normalization-integrity:v1#/$defs/digest" }, "reportDigest": { "$ref": "urn:ebo:schema:normalization-integrity:v1#/$defs/digest" }
147
+ } }
148
+ }
149
+ }
150
+ },
151
+ "behaviorAggregate": {
152
+ "type": "object",
153
+ "additionalProperties": false,
154
+ "required": ["behavior","rubric","evaluator","population","assessments","assertions"],
155
+ "properties": {
156
+ "behavior": {
157
+ "type": "object",
158
+ "additionalProperties": false,
159
+ "required": ["vocabularyVersion","categoryId","dimensionId"],
160
+ "properties": {
161
+ "vocabularyVersion": {"$ref":"#/$defs/identifier"},
162
+ "categoryId": {"$ref":"#/$defs/identifier"},
163
+ "dimensionId": {"$ref":"#/$defs/identifier"}
164
+ }
165
+ },
166
+ "rubric": {
167
+ "type": "object",
168
+ "additionalProperties": false,
169
+ "required": ["id","version"],
170
+ "properties": {"id":{"$ref":"#/$defs/identifier"},"version":{"$ref":"#/$defs/identifier"}}
171
+ },
172
+ "evaluator": {
173
+ "type": "object",
174
+ "additionalProperties": false,
175
+ "required": ["id","version"],
176
+ "properties": {
177
+ "id": {"$ref":"#/$defs/identifier"},
178
+ "version": {"$ref":"#/$defs/identifier"},
179
+ "configurationDigest": {"type":"string","pattern":"^sha256:[a-f0-9]{64}$"}
180
+ }
181
+ },
182
+ "population": {"const":"confirmed-attempt-dimension"},
183
+ "assessments": {
184
+ "type": "array",
185
+ "minItems": 4,
186
+ "maxItems": 4,
187
+ "allOf": [
188
+ {"contains":{"properties":{"assessment":{"const":"constructive"}},"required":["assessment"]},"minContains":1,"maxContains":1},
189
+ {"contains":{"properties":{"assessment":{"const":"adverse"}},"required":["assessment"]},"minContains":1,"maxContains":1},
190
+ {"contains":{"properties":{"assessment":{"const":"mixed"}},"required":["assessment"]},"minContains":1,"maxContains":1},
191
+ {"contains":{"properties":{"assessment":{"const":"context-dependent"}},"required":["assessment"]},"minContains":1,"maxContains":1}
192
+ ],
193
+ "items": {
194
+ "type": "object",
195
+ "additionalProperties": false,
196
+ "required": ["assessment","measurement"],
197
+ "properties": {
198
+ "assessment": {"enum":["constructive","adverse","mixed","context-dependent"]},
199
+ "measurement": {"$ref":"#/$defs/measurement"}
200
+ }
201
+ }
202
+ },
203
+ "assertions": {
204
+ "type": "array",
205
+ "items": {
206
+ "type": "object",
207
+ "additionalProperties": false,
208
+ "required": ["runId","attemptId","id","digest","reviewOutcome","disputed","included"],
209
+ "properties": {
210
+ "runId": {"$ref":"#/$defs/identifier"},
211
+ "attemptId": {"$ref":"#/$defs/identifier"},
212
+ "id": {"$ref":"#/$defs/identifier"},
213
+ "digest": {"$ref":"urn:ebo:schema:normalization-integrity:v1#/$defs/digest"},
214
+ "reviewOutcome": {"enum":["confirmed","rejected","unresolved","unreviewed","judge-abstained"]},
215
+ "disputed": {"type":"boolean"},
216
+ "included": {"type":"boolean"}
217
+ }
218
+ }
219
+ }
220
+ }
221
+ },
222
+ "report": {
223
+ "type": "object",
224
+ "additionalProperties": false,
225
+ "required": ["schemaVersion", "policy", "sourcePopulation", "sourceLineage", "groups", "comparisons", "limitations"],
226
+ "properties": {
227
+ "schemaVersion": { "const": "ebo.aggregation-report/v1" },
228
+ "policy": {
229
+ "type": "object", "additionalProperties": false, "required": ["groupBy", "selectedAttemptPolicy", "recurrence"],
230
+ "properties": {
231
+ "groupBy": { "type": "array", "maxItems": 5, "uniqueItems": true, "items": { "$ref": "#/$defs/dimension" } },
232
+ "selectedAttemptPolicy": { "enum": ["all-attempts", "latest-attempt-per-run"] },
233
+ "recurrence": {
234
+ "type": "object", "additionalProperties": false, "required": ["minimumOccurrences", "establishesCausality", "establishesStatisticalSignificance"],
235
+ "properties": {
236
+ "minimumOccurrences": { "type": "integer", "minimum": 2 },
237
+ "establishesCausality": { "const": false },
238
+ "establishesStatisticalSignificance": { "const": false }
239
+ }
240
+ }
241
+ }
242
+ },
243
+ "sourcePopulation": {
244
+ "type": "object", "additionalProperties": false, "required": ["uniqueRuns", "uniqueAttempts", "selectedAttempts", "duplicateInputsIgnored"],
245
+ "properties": {
246
+ "uniqueRuns": { "type": "integer", "minimum": 0 }, "uniqueAttempts": { "type": "integer", "minimum": 0 },
247
+ "selectedAttempts": { "type": "integer", "minimum": 0 }, "duplicateInputsIgnored": { "type": "integer", "minimum": 0 }
248
+ }
249
+ },
250
+ "sourceLineage": { "$ref": "#/$defs/sourceLineage" },
251
+ "groups": {
252
+ "type": "array", "maxItems": 100000,
253
+ "items": {
254
+ "type": "object", "additionalProperties": false, "required": ["dimensions", "metrics", "variations"],
255
+ "properties": {
256
+ "dimensions": { "$ref": "#/$defs/filters" },
257
+ "behaviors": { "type": "array", "items": { "$ref": "#/$defs/behaviorAggregate" } },
258
+ "metrics": { "type": "array", "maxItems": 100000, "items": { "$ref": "#/$defs/metric" } },
259
+ "variations": {
260
+ "type": "array", "maxItems": 2,
261
+ "items": {
262
+ "type": "object", "additionalProperties": false, "required": ["measure", "distribution", "claimStatus"],
263
+ "properties": {
264
+ "measure": { "enum": ["terminal-state", "verifier-status"] },
265
+ "distribution": { "type": "array", "items": { "$ref": "#/$defs/metric" } },
266
+ "claimStatus": { "enum": ["no-variation", "case-study", "recurring-description", "unavailable"] }
267
+ }
268
+ }
269
+ }
270
+ }
271
+ }
272
+ },
273
+ "comparisons": {
274
+ "type": "array", "maxItems": 100000,
275
+ "items": {
276
+ "type": "object", "additionalProperties": false,
277
+ "required": ["id", "measure", "eligibility", "matchedDifference", "differingMatchedUnits", "claimStatus", "limitations"],
278
+ "properties": {
279
+ "id": { "$ref": "#/$defs/identifier" }, "measure": { "$ref": "#/$defs/identifier" },
280
+ "eligibility": { "type": "array", "minItems": 1, "maxItems": 100000, "items": { "$ref": "urn:ebo:schema:normalization-integrity:v1#/$defs/comparisonReportV2" } },
281
+ "matchedDifference": { "$ref": "#/$defs/measurement" },
282
+ "differingMatchedUnits": { "type": "integer", "minimum": 0 },
283
+ "claimStatus": { "enum": ["no-difference", "case-study", "recurring-description", "unavailable"] },
284
+ "limitations": { "type": "array", "items": { "$ref": "#/$defs/identifier" } }
285
+ }
286
+ }
287
+ },
288
+ "limitations": { "type": "array", "minItems": 1, "items": { "$ref": "#/$defs/identifier" } }
289
+ }
290
+ }
291
+ },
292
+ "oneOf": [{ "$ref": "#/$defs/request" }, { "$ref": "#/$defs/report" }]
293
+ }
@@ -0,0 +1,220 @@
1
+ {
2
+ "$schema": "https://json-schema.org/draft/2020-12/schema",
3
+ "$id": "urn:ebo:schema:behavior-assertions:v1",
4
+ "$defs": {
5
+ "identifier": {
6
+ "type": "string",
7
+ "minLength": 1,
8
+ "maxLength": 256,
9
+ "pattern": ".*\\S.*"
10
+ },
11
+ "text": {
12
+ "type": "string",
13
+ "minLength": 1,
14
+ "maxLength": 8192,
15
+ "pattern": ".*\\S.*"
16
+ },
17
+ "digest": {
18
+ "type": "string",
19
+ "pattern": "^sha256:[a-f0-9]{64}$"
20
+ },
21
+ "versionedIdentity": {
22
+ "type": "object",
23
+ "additionalProperties": false,
24
+ "required": ["id", "version"],
25
+ "properties": {
26
+ "id": { "$ref": "#/$defs/identifier" },
27
+ "version": { "$ref": "#/$defs/identifier" }
28
+ }
29
+ },
30
+ "evaluatorIdentity": {
31
+ "type": "object",
32
+ "additionalProperties": false,
33
+ "required": ["id", "version"],
34
+ "properties": {
35
+ "id": { "$ref": "#/$defs/identifier" },
36
+ "version": { "$ref": "#/$defs/identifier" },
37
+ "configurationDigest": { "type": "string", "pattern": "^sha256:[a-f0-9]{64}$" }
38
+ }
39
+ },
40
+ "dimension": {
41
+ "type": "object",
42
+ "additionalProperties": false,
43
+ "required": ["id", "name"],
44
+ "properties": {
45
+ "id": { "$ref": "#/$defs/identifier" },
46
+ "name": { "$ref": "#/$defs/identifier" }
47
+ }
48
+ },
49
+ "category": {
50
+ "type": "object",
51
+ "additionalProperties": false,
52
+ "required": ["id", "name", "dimensions"],
53
+ "properties": {
54
+ "id": { "$ref": "#/$defs/identifier" },
55
+ "name": { "$ref": "#/$defs/identifier" },
56
+ "dimensions": {
57
+ "type": "array",
58
+ "minItems": 1,
59
+ "maxItems": 64,
60
+ "items": { "$ref": "#/$defs/dimension" }
61
+ }
62
+ }
63
+ },
64
+ "vocabulary": {
65
+ "type": "object",
66
+ "additionalProperties": false,
67
+ "required": ["schemaVersion", "version", "categories"],
68
+ "properties": {
69
+ "schemaVersion": { "const": "ebo.behavior-vocabulary/v1" },
70
+ "version": { "$ref": "#/$defs/identifier" },
71
+ "categories": {
72
+ "type": "array",
73
+ "minItems": 1,
74
+ "maxItems": 64,
75
+ "items": { "$ref": "#/$defs/category" }
76
+ }
77
+ }
78
+ },
79
+ "citation": {
80
+ "type": "object",
81
+ "additionalProperties": false,
82
+ "required": ["eventId", "nativeReference"],
83
+ "properties": {
84
+ "eventId": { "$ref": "#/$defs/identifier" },
85
+ "nativeReference": { "$ref": "urn:ebo:schema:uniform-events:v1#/$defs/nativeReference" }
86
+ }
87
+ },
88
+ "assessedJudgment": {
89
+ "type": "object",
90
+ "additionalProperties": false,
91
+ "required": ["disposition", "assessment", "confidence", "rationale", "alternativeExplanation", "citations"],
92
+ "properties": {
93
+ "disposition": { "const": "assessed" },
94
+ "assessment": { "enum": ["constructive", "adverse", "mixed", "context-dependent"] },
95
+ "confidence": {
96
+ "type": "object",
97
+ "additionalProperties": false,
98
+ "required": ["value", "scale"],
99
+ "properties": {
100
+ "value": { "type": "number", "minimum": 0, "maximum": 1 },
101
+ "scale": { "const": "evaluator-reported-0-to-1" }
102
+ }
103
+ },
104
+ "rationale": { "$ref": "#/$defs/text" },
105
+ "alternativeExplanation": { "$ref": "#/$defs/text" },
106
+ "citations": {
107
+ "type": "array",
108
+ "minItems": 1,
109
+ "maxItems": 256,
110
+ "items": { "$ref": "#/$defs/citation" }
111
+ }
112
+ }
113
+ },
114
+ "abstainedJudgment": {
115
+ "type": "object",
116
+ "additionalProperties": false,
117
+ "required": ["disposition", "reason", "rationale", "alternativeExplanation", "citations"],
118
+ "properties": {
119
+ "disposition": { "const": "abstained" },
120
+ "reason": { "$ref": "#/$defs/text" },
121
+ "missingEvidenceCapability": {
122
+ "type": "string",
123
+ "pattern": "^(?:family:(?:message|model-request|tool|context|permission|delegation|artifact|validation|runtime|outcome)|evidence:(?:nativeOrder|nativeTime|parentage|content))$"
124
+ },
125
+ "rationale": { "$ref": "#/$defs/text" },
126
+ "alternativeExplanation": { "$ref": "#/$defs/text" },
127
+ "citations": {
128
+ "type": "array",
129
+ "maxItems": 256,
130
+ "items": { "$ref": "#/$defs/citation" }
131
+ }
132
+ }
133
+ },
134
+ "assertion": {
135
+ "type": "object",
136
+ "additionalProperties": false,
137
+ "required": ["schemaVersion", "id", "runId", "attemptId", "dataset", "behavior", "rubric", "evaluator", "judgment"],
138
+ "properties": {
139
+ "schemaVersion": { "const": "ebo.behavior-assertion/v1" },
140
+ "id": { "$ref": "#/$defs/identifier" },
141
+ "runId": { "$ref": "#/$defs/identifier" },
142
+ "attemptId": { "$ref": "#/$defs/identifier" },
143
+ "dataset": {
144
+ "type": "object",
145
+ "additionalProperties": false,
146
+ "required": ["schemaVersion", "digest"],
147
+ "properties": {
148
+ "schemaVersion": { "const": "ebo.normalized-dataset/v1" },
149
+ "digest": { "$ref": "#/$defs/digest" }
150
+ }
151
+ },
152
+ "behavior": {
153
+ "type": "object",
154
+ "additionalProperties": false,
155
+ "required": ["vocabularyVersion", "categoryId", "dimensionId"],
156
+ "properties": {
157
+ "vocabularyVersion": { "$ref": "#/$defs/identifier" },
158
+ "categoryId": { "$ref": "#/$defs/identifier" },
159
+ "dimensionId": { "$ref": "#/$defs/identifier" }
160
+ }
161
+ },
162
+ "rubric": { "$ref": "#/$defs/versionedIdentity" },
163
+ "evaluator": { "$ref": "#/$defs/evaluatorIdentity" },
164
+ "judgment": {
165
+ "oneOf": [
166
+ { "$ref": "#/$defs/assessedJudgment" },
167
+ { "$ref": "#/$defs/abstainedJudgment" }
168
+ ]
169
+ }
170
+ }
171
+ },
172
+ "assertionBinding": {
173
+ "type": "object",
174
+ "additionalProperties": false,
175
+ "required": ["id", "schemaVersion", "digest"],
176
+ "properties": {
177
+ "id": { "$ref": "#/$defs/identifier" },
178
+ "schemaVersion": { "const": "ebo.behavior-assertion/v1" },
179
+ "digest": { "$ref": "#/$defs/digest" }
180
+ }
181
+ },
182
+ "reviewer": {
183
+ "type": "object",
184
+ "additionalProperties": false,
185
+ "required": ["kind", "id"],
186
+ "properties": {
187
+ "kind": { "const": "human" },
188
+ "id": { "$ref": "#/$defs/identifier" }
189
+ }
190
+ },
191
+ "review": {
192
+ "type": "object",
193
+ "additionalProperties": false,
194
+ "required": ["schemaVersion", "id", "assertion", "state"],
195
+ "properties": {
196
+ "schemaVersion": { "const": "ebo.behavior-review/v1" },
197
+ "id": { "$ref": "#/$defs/identifier" },
198
+ "assertion": { "$ref": "#/$defs/assertionBinding" },
199
+ "state": { "enum": ["proposed", "confirmed", "disputed", "rejected", "insufficient-evidence"] },
200
+ "reviewer": { "$ref": "#/$defs/reviewer" },
201
+ "rationale": { "$ref": "#/$defs/text" }
202
+ },
203
+ "allOf": [
204
+ {
205
+ "if": {
206
+ "required": ["state"],
207
+ "properties": { "state": { "const": "proposed" } }
208
+ },
209
+ "then": { "not": { "required": ["reviewer"] } },
210
+ "else": { "required": ["reviewer", "rationale"] }
211
+ }
212
+ ]
213
+ }
214
+ },
215
+ "oneOf": [
216
+ { "$ref": "#/$defs/vocabulary" },
217
+ { "$ref": "#/$defs/assertion" },
218
+ { "$ref": "#/$defs/review" }
219
+ ]
220
+ }
@@ -0,0 +1,122 @@
1
+ {
2
+ "$schema": "https://json-schema.org/draft/2020-12/schema",
3
+ "$id": "https://ebo.dev/schemas/deepseek-runtime-composition.v1.schema.json",
4
+ "type": "object",
5
+ "additionalProperties": false,
6
+ "required": ["schemaVersion", "id", "launch", "workspaceCwd", "runtime", "route", "cordis", "patches", "plugins", "environment", "telemetry"],
7
+ "properties": {
8
+ "schemaVersion": { "const": "ebo.deepseek-runtime-composition/v1" },
9
+ "id": { "$ref": "#/$defs/nonempty" },
10
+ "launch": {
11
+ "type": "object",
12
+ "additionalProperties": false,
13
+ "required": ["command", "args", "processCwd", "profile", "runtimeArtifact"],
14
+ "properties": {
15
+ "command": { "$ref": "#/$defs/nonempty" },
16
+ "args": {
17
+ "type": "array",
18
+ "minItems": 3,
19
+ "maxItems": 64,
20
+ "items": { "$ref": "#/$defs/nonempty" }
21
+ },
22
+ "processCwd": { "$ref": "#/$defs/nonempty" },
23
+ "profile": { "$ref": "#/$defs/nonempty" },
24
+ "dshHome": { "$ref": "#/$defs/nonempty" },
25
+ "runtimeArtifact": { "$ref": "#/$defs/fileReference" }
26
+ }
27
+ },
28
+ "workspaceCwd": { "$ref": "#/$defs/nonempty" },
29
+ "runtime": {
30
+ "type": "object",
31
+ "additionalProperties": false,
32
+ "required": ["clientPackage", "clientVersion", "protocolPackage", "protocolVersion"],
33
+ "properties": {
34
+ "clientPackage": { "const": "@deepseek-ai/dsh-sdk-client" },
35
+ "clientVersion": { "$ref": "#/$defs/nonempty" },
36
+ "protocolPackage": { "const": "@deepseek-ai/dsh-sdk-protocol" },
37
+ "protocolVersion": { "$ref": "#/$defs/nonempty" }
38
+ }
39
+ },
40
+ "route": {
41
+ "type": "object",
42
+ "additionalProperties": false,
43
+ "required": ["provider", "model"],
44
+ "properties": {
45
+ "provider": { "$ref": "#/$defs/nonempty" },
46
+ "model": { "$ref": "#/$defs/nonempty" },
47
+ "maxTokens": { "type": "integer", "minimum": 1, "maximum": 9007199254740991 }
48
+ }
49
+ },
50
+ "cordis": { "$ref": "#/$defs/fileReference" },
51
+ "patches": {
52
+ "type": "array",
53
+ "minItems": 1,
54
+ "maxItems": 32,
55
+ "items": { "$ref": "#/$defs/fileReference" }
56
+ },
57
+ "plugins": {
58
+ "type": "array",
59
+ "minItems": 1,
60
+ "maxItems": 128,
61
+ "items": {
62
+ "type": "object",
63
+ "additionalProperties": false,
64
+ "required": ["name"],
65
+ "properties": {
66
+ "name": { "$ref": "#/$defs/nonempty" },
67
+ "configuration": { "$ref": "#/$defs/fileReference" }
68
+ }
69
+ }
70
+ },
71
+ "environment": {
72
+ "type": "object",
73
+ "additionalProperties": false,
74
+ "required": ["mode", "allowedKeys", "secretKeys"],
75
+ "properties": {
76
+ "mode": { "const": "replace" },
77
+ "allowedKeys": {
78
+ "type": "array",
79
+ "maxItems": 128,
80
+ "uniqueItems": true,
81
+ "items": { "$ref": "#/$defs/envKey" }
82
+ },
83
+ "secretKeys": {
84
+ "type": "array",
85
+ "maxItems": 32,
86
+ "uniqueItems": true,
87
+ "items": { "$ref": "#/$defs/envKey" }
88
+ }
89
+ }
90
+ },
91
+ "telemetry": {
92
+ "type": "object",
93
+ "additionalProperties": false,
94
+ "required": ["status", "nativeSpans"],
95
+ "properties": {
96
+ "status": { "enum": ["enabled", "disabled"] },
97
+ "nativeSpans": { "enum": ["available", "unsupported", "not-checked"] },
98
+ "artifact": { "$ref": "#/$defs/fileReference" }
99
+ },
100
+ "allOf": [
101
+ {
102
+ "if": { "properties": { "status": { "const": "disabled" } }, "required": ["status"] },
103
+ "then": { "properties": { "nativeSpans": { "const": "unsupported" } }, "not": { "required": ["artifact"] } }
104
+ }
105
+ ]
106
+ }
107
+ },
108
+ "$defs": {
109
+ "nonempty": { "type": "string", "minLength": 1, "maxLength": 4096, "pattern": ".*\\S.*" },
110
+ "envKey": { "type": "string", "pattern": "^[A-Za-z_][A-Za-z0-9_]*$", "maxLength": 255 },
111
+ "digest": { "type": "string", "pattern": "^sha256:[a-f0-9]{64}$" },
112
+ "fileReference": {
113
+ "type": "object",
114
+ "additionalProperties": false,
115
+ "required": ["locator", "digest"],
116
+ "properties": {
117
+ "locator": { "$ref": "#/$defs/nonempty" },
118
+ "digest": { "$ref": "#/$defs/digest" }
119
+ }
120
+ }
121
+ }
122
+ }