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 +1,289 @@
1
+ {
2
+ "$schema": "https://json-schema.org/draft/2020-12/schema",
3
+ "$id": "urn:ebo:schema:uniform-events:v1",
4
+ "$defs": {
5
+ "identifier": {
6
+ "type": "string",
7
+ "minLength": 1,
8
+ "maxLength": 256,
9
+ "pattern": ".*\\S.*"
10
+ },
11
+ "unavailable": {
12
+ "oneOf": [
13
+ {
14
+ "type": "object",
15
+ "additionalProperties": false,
16
+ "required": ["status", "reason"],
17
+ "properties": {
18
+ "status": { "const": "unknown" },
19
+ "reason": { "$ref": "#/$defs/identifier" }
20
+ }
21
+ },
22
+ {
23
+ "type": "object",
24
+ "additionalProperties": false,
25
+ "required": ["status", "reason"],
26
+ "properties": {
27
+ "status": { "const": "unsupported" },
28
+ "reason": { "$ref": "#/$defs/identifier" }
29
+ }
30
+ }
31
+ ]
32
+ },
33
+ "nativeReference": {
34
+ "type": "object",
35
+ "additionalProperties": false,
36
+ "required": ["artifactId", "recordLocator"],
37
+ "properties": {
38
+ "artifactId": { "$ref": "#/$defs/identifier" },
39
+ "recordLocator": { "$ref": "#/$defs/identifier" }
40
+ }
41
+ },
42
+ "nativeOrder": {
43
+ "oneOf": [
44
+ {
45
+ "type": "object",
46
+ "additionalProperties": false,
47
+ "required": ["status", "value", "domain"],
48
+ "properties": {
49
+ "status": { "const": "known" },
50
+ "value": { "type": "integer", "minimum": 0, "maximum": 9007199254740991 },
51
+ "domain": { "$ref": "#/$defs/identifier" }
52
+ }
53
+ },
54
+ { "$ref": "#/$defs/unavailable" }
55
+ ]
56
+ },
57
+ "nativeTime": {
58
+ "oneOf": [
59
+ {
60
+ "type": "object",
61
+ "additionalProperties": false,
62
+ "required": ["status", "value"],
63
+ "properties": {
64
+ "status": { "const": "known" },
65
+ "value": { "type": "string", "format": "date-time" }
66
+ }
67
+ },
68
+ { "$ref": "#/$defs/unavailable" }
69
+ ]
70
+ },
71
+ "eventIdentity": {
72
+ "oneOf": [
73
+ {
74
+ "type": "object",
75
+ "additionalProperties": false,
76
+ "required": ["status", "value"],
77
+ "properties": {
78
+ "status": { "const": "known" },
79
+ "value": { "$ref": "#/$defs/identifier" }
80
+ }
81
+ },
82
+ { "$ref": "#/$defs/unavailable" }
83
+ ]
84
+ },
85
+ "contentReference": {
86
+ "type": "object",
87
+ "additionalProperties": false,
88
+ "required": ["nativeReference"],
89
+ "properties": {
90
+ "nativeReference": { "$ref": "#/$defs/nativeReference" },
91
+ "mediaType": { "$ref": "#/$defs/identifier" },
92
+ "role": { "$ref": "#/$defs/identifier" }
93
+ }
94
+ },
95
+ "content": {
96
+ "oneOf": [
97
+ {
98
+ "type": "object",
99
+ "additionalProperties": false,
100
+ "required": ["status", "value"],
101
+ "properties": {
102
+ "status": { "const": "known" },
103
+ "value": {
104
+ "type": "array",
105
+ "maxItems": 16,
106
+ "items": { "$ref": "#/$defs/contentReference" }
107
+ }
108
+ }
109
+ },
110
+ { "$ref": "#/$defs/unavailable" }
111
+ ]
112
+ },
113
+ "attributeScalar": {
114
+ "oneOf": [
115
+ { "type": "string", "maxLength": 512 },
116
+ { "type": "number" },
117
+ { "type": "boolean" },
118
+ { "type": "null" }
119
+ ]
120
+ },
121
+ "attributeValue": {
122
+ "oneOf": [
123
+ { "$ref": "#/$defs/attributeScalar" },
124
+ {
125
+ "type": "array",
126
+ "maxItems": 16,
127
+ "items": { "$ref": "#/$defs/attributeScalar" }
128
+ }
129
+ ]
130
+ },
131
+ "uniformEvent": {
132
+ "type": "object",
133
+ "additionalProperties": false,
134
+ "required": [
135
+ "schemaVersion",
136
+ "id",
137
+ "runId",
138
+ "attemptId",
139
+ "source",
140
+ "nativeOrder",
141
+ "nativeTime",
142
+ "actor",
143
+ "family",
144
+ "phase",
145
+ "scope",
146
+ "relations",
147
+ "attributes",
148
+ "content"
149
+ ],
150
+ "properties": {
151
+ "schemaVersion": { "const": "ebo.uniform-event/v1" },
152
+ "id": { "$ref": "#/$defs/identifier" },
153
+ "runId": { "$ref": "#/$defs/identifier" },
154
+ "attemptId": { "$ref": "#/$defs/identifier" },
155
+ "source": {
156
+ "type": "object",
157
+ "additionalProperties": false,
158
+ "required": ["harness", "nativeType", "nativeReference"],
159
+ "properties": {
160
+ "harness": { "$ref": "#/$defs/identifier" },
161
+ "nativeType": { "$ref": "#/$defs/identifier" },
162
+ "nativeReference": { "$ref": "#/$defs/nativeReference" }
163
+ }
164
+ },
165
+ "nativeOrder": { "$ref": "#/$defs/nativeOrder" },
166
+ "nativeTime": { "$ref": "#/$defs/nativeTime" },
167
+ "actor": {
168
+ "type": "object",
169
+ "additionalProperties": false,
170
+ "required": ["kind"],
171
+ "properties": {
172
+ "kind": { "enum": ["user", "agent", "model", "tool", "harness", "operator", "system"] },
173
+ "id": { "$ref": "#/$defs/identifier" }
174
+ }
175
+ },
176
+ "family": {
177
+ "enum": [
178
+ "message",
179
+ "model-request",
180
+ "tool",
181
+ "context",
182
+ "permission",
183
+ "delegation",
184
+ "artifact",
185
+ "validation",
186
+ "runtime",
187
+ "outcome"
188
+ ]
189
+ },
190
+ "phase": { "enum": ["before", "during", "after", "instant"] },
191
+ "scope": {
192
+ "type": "object",
193
+ "additionalProperties": false,
194
+ "required": ["kind"],
195
+ "properties": {
196
+ "kind": { "enum": ["run", "attempt", "session", "turn", "operation", "workspace", "artifact"] },
197
+ "id": { "$ref": "#/$defs/identifier" }
198
+ }
199
+ },
200
+ "relations": {
201
+ "type": "object",
202
+ "additionalProperties": false,
203
+ "required": ["parent", "known"],
204
+ "properties": {
205
+ "parent": { "$ref": "#/$defs/eventIdentity" },
206
+ "known": {
207
+ "type": "array",
208
+ "maxItems": 4096,
209
+ "items": {
210
+ "type": "object",
211
+ "additionalProperties": false,
212
+ "required": ["kind", "eventId"],
213
+ "properties": {
214
+ "kind": { "enum": ["child", "caused-by", "correlates-with"] },
215
+ "eventId": { "$ref": "#/$defs/identifier" }
216
+ }
217
+ }
218
+ }
219
+ }
220
+ },
221
+ "attributes": {
222
+ "type": "object",
223
+ "maxProperties": 32,
224
+ "propertyNames": { "type": "string", "minLength": 1, "maxLength": 64, "pattern": "^[A-Za-z][A-Za-z0-9._-]*$" },
225
+ "additionalProperties": { "$ref": "#/$defs/attributeValue" }
226
+ },
227
+ "content": { "$ref": "#/$defs/content" }
228
+ }
229
+ },
230
+ "capability": {
231
+ "type": "object",
232
+ "additionalProperties": false,
233
+ "required": ["status"],
234
+ "properties": {
235
+ "status": { "enum": ["available", "partial", "unsupported"] },
236
+ "detail": { "type": "string", "minLength": 1, "maxLength": 1024 }
237
+ }
238
+ },
239
+ "capabilityProfile": {
240
+ "type": "object",
241
+ "additionalProperties": false,
242
+ "required": ["schemaVersion", "adapterId", "harness", "nativeTypes", "families", "evidence"],
243
+ "properties": {
244
+ "schemaVersion": { "const": "ebo.adapter-capability-profile/v1" },
245
+ "adapterId": { "$ref": "#/$defs/identifier" },
246
+ "harness": { "$ref": "#/$defs/identifier" },
247
+ "nativeTypes": {
248
+ "type": "array",
249
+ "minItems": 1,
250
+ "maxItems": 128,
251
+ "uniqueItems": true,
252
+ "items": { "$ref": "#/$defs/identifier" }
253
+ },
254
+ "families": {
255
+ "type": "object",
256
+ "additionalProperties": false,
257
+ "required": ["message", "model-request", "tool", "context", "permission", "delegation", "artifact", "validation", "runtime", "outcome"],
258
+ "properties": {
259
+ "message": { "$ref": "#/$defs/capability" },
260
+ "model-request": { "$ref": "#/$defs/capability" },
261
+ "tool": { "$ref": "#/$defs/capability" },
262
+ "context": { "$ref": "#/$defs/capability" },
263
+ "permission": { "$ref": "#/$defs/capability" },
264
+ "delegation": { "$ref": "#/$defs/capability" },
265
+ "artifact": { "$ref": "#/$defs/capability" },
266
+ "validation": { "$ref": "#/$defs/capability" },
267
+ "runtime": { "$ref": "#/$defs/capability" },
268
+ "outcome": { "$ref": "#/$defs/capability" }
269
+ }
270
+ },
271
+ "evidence": {
272
+ "type": "object",
273
+ "additionalProperties": false,
274
+ "required": ["nativeOrder", "nativeTime", "parentage", "content"],
275
+ "properties": {
276
+ "nativeOrder": { "$ref": "#/$defs/capability" },
277
+ "nativeTime": { "$ref": "#/$defs/capability" },
278
+ "parentage": { "$ref": "#/$defs/capability" },
279
+ "content": { "$ref": "#/$defs/capability" }
280
+ }
281
+ }
282
+ }
283
+ }
284
+ },
285
+ "oneOf": [
286
+ { "$ref": "#/$defs/uniformEvent" },
287
+ { "$ref": "#/$defs/capabilityProfile" }
288
+ ]
289
+ }
@@ -0,0 +1,18 @@
1
+ #!/bin/sh
2
+ set -eu
3
+ if [ "$#" -ne 2 ]; then
4
+ echo 'Usage: sh scripts/atlas-grafana.sh <Grafana 13.2.0 home> <Atlas output root>' >&2
5
+ exit 1
6
+ fi
7
+ atlas_grafana_home=$(cd "$1" && pwd -P)
8
+ atlas_output=$(cd "$2/grafana" && pwd -P)
9
+ export GF_PATHS_DATA="$atlas_output/data"
10
+ export GF_PATHS_LOGS="$atlas_output/logs"
11
+ export GF_PATHS_PLUGINS="$atlas_output/plugins"
12
+ export GF_PATHS_PROVISIONING="$atlas_output/provisioning"
13
+ export EBO_ATLAS_DASHBOARDS="$atlas_output/dashboards"
14
+ mkdir -p "$GF_PATHS_DATA" "$GF_PATHS_LOGS" "$GF_PATHS_PLUGINS"
15
+ if [ ! -d "$GF_PATHS_PLUGINS/yesoreyeram-infinity-datasource" ]; then
16
+ "$atlas_grafana_home/bin/grafana" cli --homepath "$atlas_grafana_home" --pluginsDir "$GF_PATHS_PLUGINS" plugins install yesoreyeram-infinity-datasource 4.0.0
17
+ fi
18
+ exec "$atlas_grafana_home/bin/grafana" server --homepath "$atlas_grafana_home" --config "$atlas_output/grafana.ini"