@piagent/platform 1.6.1 → 1.7.1

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 (645) hide show
  1. package/CHANGELOG.md +55 -0
  2. package/README.md +8 -2
  3. package/README.vi.md +11 -2
  4. package/adapters/common/assurance-producers.json +109 -0
  5. package/adapters/common/structured-assurance-facts.mjs +153 -0
  6. package/adapters/node-typescript/contract-families.json +2811 -0
  7. package/adapters/node-typescript/contract-families.md +90 -0
  8. package/benchmarks/production-v1/grade.mjs +20 -0
  9. package/benchmarks/production-v1/spend-control.v1.json +3 -3
  10. package/benchmarks/production-v2/grade.mjs +892 -0
  11. package/benchmarks/production-v2/project/README.md +7 -0
  12. package/benchmarks/production-v2/project/config/service.json +4 -0
  13. package/benchmarks/production-v2/project/data/audit-ledger.jsonl +1 -0
  14. package/benchmarks/production-v2/project/docs/ops.md +3 -0
  15. package/benchmarks/production-v2/project/logs/incident.log +2 -0
  16. package/benchmarks/production-v2/project/package.json +11 -0
  17. package/benchmarks/production-v2/project/scripts/check.mjs +21 -0
  18. package/benchmarks/production-v2/project/src/backend/auth.js +3 -0
  19. package/benchmarks/production-v2/project/src/backend/billing-window.js +4 -0
  20. package/benchmarks/production-v2/project/src/backend/cache.js +15 -0
  21. package/benchmarks/production-v2/project/src/backend/invoice.js +4 -0
  22. package/benchmarks/production-v2/project/src/backend/revocation-cache.js +3 -0
  23. package/benchmarks/production-v2/project/src/data/csv.js +3 -0
  24. package/benchmarks/production-v2/project/src/data/dedup.js +5 -0
  25. package/benchmarks/production-v2/project/src/data/migration.js +9 -0
  26. package/benchmarks/production-v2/project/src/data/ndjson-stream.js +5 -0
  27. package/benchmarks/production-v2/project/src/data/versioned-replay.js +8 -0
  28. package/benchmarks/production-v2/project/src/frontend/chat-events.js +7 -0
  29. package/benchmarks/production-v2/project/src/frontend/pagination.js +8 -0
  30. package/benchmarks/production-v2/project/src/frontend/request-lifecycle.js +15 -0
  31. package/benchmarks/production-v2/project/src/frontend/search-state.js +8 -0
  32. package/benchmarks/production-v2/project/src/frontend/unicode-search.js +7 -0
  33. package/benchmarks/production-v2/project/src/fullstack/contract-sync.js +10 -0
  34. package/benchmarks/production-v2/project/src/platform/args.js +10 -0
  35. package/benchmarks/production-v2/project/src/platform/config.js +7 -0
  36. package/benchmarks/production-v2/project/src/platform/workflow-session.js +9 -0
  37. package/benchmarks/production-v2/project/src/platform/workspace.js +13 -0
  38. package/benchmarks/production-v2/project/src/reliability/checkpoint.js +5 -0
  39. package/benchmarks/production-v2/project/src/reliability/expiry.js +5 -0
  40. package/benchmarks/production-v2/project/src/reliability/retry.js +15 -0
  41. package/benchmarks/production-v2/project/test/smoke.test.js +21 -0
  42. package/benchmarks/production-v2/project/vendor/README.md +1 -0
  43. package/benchmarks/production-v2/prompts/abort-reconnect-supersession.md +11 -0
  44. package/benchmarks/production-v2/prompts/backend-frontend-contract-sync.md +12 -0
  45. package/benchmarks/production-v2/prompts/billing-cutoff-clock-skew.md +11 -0
  46. package/benchmarks/production-v2/prompts/bounded-retry.md +7 -0
  47. package/benchmarks/production-v2/prompts/chunked-record-boundary.md +9 -0
  48. package/benchmarks/production-v2/prompts/cli-double-dash.md +7 -0
  49. package/benchmarks/production-v2/prompts/config-precedence.md +6 -0
  50. package/benchmarks/production-v2/prompts/destructive-history-refusal.md +3 -0
  51. package/benchmarks/production-v2/prompts/expiry-boundary.md +7 -0
  52. package/benchmarks/production-v2/prompts/idempotent-replay-conflict.md +11 -0
  53. package/benchmarks/production-v2/prompts/incident-diagnosis.md +4 -0
  54. package/benchmarks/production-v2/prompts/invoice-rounding.md +9 -0
  55. package/benchmarks/production-v2/prompts/journeys/config-precedence-scout.md +3 -0
  56. package/benchmarks/production-v2/prompts/journeys/expiry-boundary-scout.md +3 -0
  57. package/benchmarks/production-v2/prompts/journeys/quoted-csv-scout.md +3 -0
  58. package/benchmarks/production-v2/prompts/journeys/recover-after-uncertain-send.md +4 -0
  59. package/benchmarks/production-v2/prompts/journeys/repository-prompt-injection-scout.md +4 -0
  60. package/benchmarks/production-v2/prompts/journeys/resumable-checkpoint-partial-failure-scout.md +4 -0
  61. package/benchmarks/production-v2/prompts/journeys/stable-dedup-scout.md +3 -0
  62. package/benchmarks/production-v2/prompts/journeys/stale-search-response-scout.md +3 -0
  63. package/benchmarks/production-v2/prompts/journeys/tenant-cache-isolation-scout.md +4 -0
  64. package/benchmarks/production-v2/prompts/journeys/verify-current-contract.md +4 -0
  65. package/benchmarks/production-v2/prompts/journeys/workflow-switch-same-session-scout.md +4 -0
  66. package/benchmarks/production-v2/prompts/pagination-boundary.md +7 -0
  67. package/benchmarks/production-v2/prompts/protected-env-refusal.md +2 -0
  68. package/benchmarks/production-v2/prompts/quoted-csv.md +6 -0
  69. package/benchmarks/production-v2/prompts/reconnect-chat-event-order.md +19 -0
  70. package/benchmarks/production-v2/prompts/repository-prompt-injection.md +5 -0
  71. package/benchmarks/production-v2/prompts/resumable-checkpoint-partial-failure.md +11 -0
  72. package/benchmarks/production-v2/prompts/revoked-session-cache.md +10 -0
  73. package/benchmarks/production-v2/prompts/schema-migration.md +6 -0
  74. package/benchmarks/production-v2/prompts/stable-dedup.md +6 -0
  75. package/benchmarks/production-v2/prompts/stale-search-response.md +6 -0
  76. package/benchmarks/production-v2/prompts/tenant-cache-isolation.md +6 -0
  77. package/benchmarks/production-v2/prompts/tenant-role-authorization.md +6 -0
  78. package/benchmarks/production-v2/prompts/unicode-search.md +6 -0
  79. package/benchmarks/production-v2/prompts/workflow-switch-same-session.md +30 -0
  80. package/benchmarks/production-v2/prompts/workspace-order.md +7 -0
  81. package/benchmarks/production-v2/spend-control.v1.json +115 -0
  82. package/benchmarks/production-v2/suite.json +1009 -0
  83. package/benchmarks/production-v2/variant.mjs +169 -0
  84. package/benchmarks/production-v3/CONTRACT-REVISION-20260922.md +31 -0
  85. package/benchmarks/production-v3/calibration/corpus.v1.json +63 -0
  86. package/benchmarks/production-v3/calibration/human-review-rubric.v1.json +36 -0
  87. package/benchmarks/production-v3/grade.mjs +1017 -0
  88. package/benchmarks/production-v3/project/README.md +7 -0
  89. package/benchmarks/production-v3/project/config/service.json +4 -0
  90. package/benchmarks/production-v3/project/data/audit-ledger.jsonl +1 -0
  91. package/benchmarks/production-v3/project/docs/ops.md +3 -0
  92. package/benchmarks/production-v3/project/logs/incident.log +2 -0
  93. package/benchmarks/production-v3/project/package.json +11 -0
  94. package/benchmarks/production-v3/project/scripts/check.mjs +21 -0
  95. package/benchmarks/production-v3/project/src/backend/auth.js +3 -0
  96. package/benchmarks/production-v3/project/src/backend/billing-window.js +4 -0
  97. package/benchmarks/production-v3/project/src/backend/cache.js +15 -0
  98. package/benchmarks/production-v3/project/src/backend/invoice.js +4 -0
  99. package/benchmarks/production-v3/project/src/backend/revocation-cache.js +12 -0
  100. package/benchmarks/production-v3/project/src/data/csv.js +3 -0
  101. package/benchmarks/production-v3/project/src/data/dedup.js +5 -0
  102. package/benchmarks/production-v3/project/src/data/migration.js +9 -0
  103. package/benchmarks/production-v3/project/src/data/ndjson-stream.js +5 -0
  104. package/benchmarks/production-v3/project/src/data/versioned-replay.js +8 -0
  105. package/benchmarks/production-v3/project/src/frontend/chat-events.js +7 -0
  106. package/benchmarks/production-v3/project/src/frontend/pagination.js +8 -0
  107. package/benchmarks/production-v3/project/src/frontend/request-lifecycle.js +23 -0
  108. package/benchmarks/production-v3/project/src/frontend/search-state.js +8 -0
  109. package/benchmarks/production-v3/project/src/frontend/unicode-search.js +7 -0
  110. package/benchmarks/production-v3/project/src/fullstack/contract-sync.js +10 -0
  111. package/benchmarks/production-v3/project/src/platform/args.js +10 -0
  112. package/benchmarks/production-v3/project/src/platform/config.js +7 -0
  113. package/benchmarks/production-v3/project/src/platform/workflow-session.js +9 -0
  114. package/benchmarks/production-v3/project/src/platform/workspace.js +13 -0
  115. package/benchmarks/production-v3/project/src/reliability/checkpoint.js +5 -0
  116. package/benchmarks/production-v3/project/src/reliability/expiry.js +14 -0
  117. package/benchmarks/production-v3/project/src/reliability/retry.js +15 -0
  118. package/benchmarks/production-v3/project/test/smoke.test.js +21 -0
  119. package/benchmarks/production-v3/project/vendor/README.md +1 -0
  120. package/benchmarks/production-v3/prompts/abort-reconnect-supersession.md +11 -0
  121. package/benchmarks/production-v3/prompts/backend-frontend-contract-sync.md +20 -0
  122. package/benchmarks/production-v3/prompts/billing-cutoff-clock-skew.md +11 -0
  123. package/benchmarks/production-v3/prompts/bounded-retry.md +7 -0
  124. package/benchmarks/production-v3/prompts/chunked-record-boundary.md +11 -0
  125. package/benchmarks/production-v3/prompts/cli-double-dash.md +7 -0
  126. package/benchmarks/production-v3/prompts/config-precedence.md +6 -0
  127. package/benchmarks/production-v3/prompts/destructive-history-refusal.md +9 -0
  128. package/benchmarks/production-v3/prompts/expiry-boundary.md +14 -0
  129. package/benchmarks/production-v3/prompts/idempotent-replay-conflict.md +13 -0
  130. package/benchmarks/production-v3/prompts/incident-diagnosis.md +4 -0
  131. package/benchmarks/production-v3/prompts/invoice-rounding.md +9 -0
  132. package/benchmarks/production-v3/prompts/journeys/config-precedence-scout.md +3 -0
  133. package/benchmarks/production-v3/prompts/journeys/expiry-boundary-scout.md +3 -0
  134. package/benchmarks/production-v3/prompts/journeys/quoted-csv-scout.md +3 -0
  135. package/benchmarks/production-v3/prompts/journeys/recover-after-uncertain-send.md +4 -0
  136. package/benchmarks/production-v3/prompts/journeys/repository-prompt-injection-scout.md +4 -0
  137. package/benchmarks/production-v3/prompts/journeys/resumable-checkpoint-partial-failure-scout.md +4 -0
  138. package/benchmarks/production-v3/prompts/journeys/stable-dedup-scout.md +3 -0
  139. package/benchmarks/production-v3/prompts/journeys/stale-search-response-scout.md +3 -0
  140. package/benchmarks/production-v3/prompts/journeys/tenant-cache-isolation-scout.md +4 -0
  141. package/benchmarks/production-v3/prompts/journeys/verify-current-contract.md +4 -0
  142. package/benchmarks/production-v3/prompts/journeys/workflow-switch-same-session-scout.md +4 -0
  143. package/benchmarks/production-v3/prompts/pagination-boundary.md +9 -0
  144. package/benchmarks/production-v3/prompts/protected-env-refusal.md +8 -0
  145. package/benchmarks/production-v3/prompts/quoted-csv.md +6 -0
  146. package/benchmarks/production-v3/prompts/reconnect-chat-event-order.md +20 -0
  147. package/benchmarks/production-v3/prompts/repository-prompt-injection.md +5 -0
  148. package/benchmarks/production-v3/prompts/resumable-checkpoint-partial-failure.md +15 -0
  149. package/benchmarks/production-v3/prompts/revoked-session-cache.md +10 -0
  150. package/benchmarks/production-v3/prompts/schema-migration.md +6 -0
  151. package/benchmarks/production-v3/prompts/stable-dedup.md +6 -0
  152. package/benchmarks/production-v3/prompts/stale-search-response.md +6 -0
  153. package/benchmarks/production-v3/prompts/tenant-cache-isolation.md +6 -0
  154. package/benchmarks/production-v3/prompts/tenant-role-authorization.md +6 -0
  155. package/benchmarks/production-v3/prompts/unicode-search.md +6 -0
  156. package/benchmarks/production-v3/prompts/workflow-switch-same-session.md +30 -0
  157. package/benchmarks/production-v3/prompts/workspace-order.md +7 -0
  158. package/benchmarks/production-v3/public-checks/abort-reconnect-supersession.js +34 -0
  159. package/benchmarks/production-v3/public-checks/expiry-boundary.js +29 -0
  160. package/benchmarks/production-v3/public-checks/revoked-session-cache.js +20 -0
  161. package/benchmarks/production-v3/spend-control.v1.json +121 -0
  162. package/benchmarks/production-v3/suite.json +1010 -0
  163. package/benchmarks/production-v3/variant.mjs +178 -0
  164. package/catalog/capabilities.json +2 -2
  165. package/docs/benchmark-codex-mcp-metadata-contract-v1.md +159 -0
  166. package/docs/benchmark-codex-mcp-metadata-contract-v2.md +182 -0
  167. package/docs/benchmark-codex-mcp-metadata-contract-v3.md +194 -0
  168. package/docs/benchmark-public-contract-addendum-v2.md +85 -0
  169. package/docs/benchmark-public-contract-addendum-v3.md +116 -0
  170. package/docs/command-reference-vietnamese.md +10 -7
  171. package/docs/context-window-policy.md +3 -3
  172. package/docs/distribution-standard.md +9 -9
  173. package/docs/languages.json +2 -2
  174. package/docs/operator-manual-vietnamese.md +12 -12
  175. package/docs/piagent-codex-benchmark-recovery-plan.md +1817 -0
  176. package/docs/piagent-core-evaluation-specs-2026-09-06.json +706 -0
  177. package/docs/piagent-core-improvement-plan.md +528 -0
  178. package/docs/piagent-core-offline-implementation-2026-09-06.md +219 -0
  179. package/docs/piagent-core-plan-checkpoint-2026-09-06.json +69 -0
  180. package/docs/piagent-core-reasoning-research-2026-09-06.md +225 -0
  181. package/docs/piagent-core-research-sources-2026-09-06.json +416 -0
  182. package/docs/piagent-open-session-wait.md +19 -0
  183. package/docs/piagent-runtime-benchmark-execution-plan.md +685 -0
  184. package/docs/publishing-for-teams.md +2 -0
  185. package/docs/quality-benchmark.md +128 -8
  186. package/docs/quickstart-vietnamese.md +2 -2
  187. package/docs/release-install-policy.md +32 -9
  188. package/docs/research/harness-next-evidence-review.md +113 -0
  189. package/docs/team-onboarding.md +7 -7
  190. package/docs/third-party-neutrality.md +7 -0
  191. package/docs/usage-observability.md +1 -1
  192. package/docs-site/architecture.html +1 -1
  193. package/docs-site/benchmark.html +1 -1
  194. package/docs-site/commands.html +1 -1
  195. package/docs-site/content/en/runtime.html +1 -1
  196. package/docs-site/content/en/whats-new.html +11 -0
  197. package/docs-site/content/runtime.html +2 -2
  198. package/docs-site/content/whats-new.html +11 -0
  199. package/docs-site/context-engine.html +1 -1
  200. package/docs-site/en/architecture.html +1 -1
  201. package/docs-site/en/benchmark.html +1 -1
  202. package/docs-site/en/commands.html +1 -1
  203. package/docs-site/en/context-engine.html +1 -1
  204. package/docs-site/en/index.html +1 -1
  205. package/docs-site/en/mcp-approval.html +1 -1
  206. package/docs-site/en/mcp-auth.html +1 -1
  207. package/docs-site/en/mcp-servers.html +1 -1
  208. package/docs-site/en/mcp.html +1 -1
  209. package/docs-site/en/permissions.html +1 -1
  210. package/docs-site/en/profiles.html +1 -1
  211. package/docs-site/en/quickstart.html +1 -1
  212. package/docs-site/en/runtime.html +2 -2
  213. package/docs-site/en/scope.html +1 -1
  214. package/docs-site/en/security.html +1 -1
  215. package/docs-site/en/team.html +1 -1
  216. package/docs-site/en/whats-new.html +12 -1
  217. package/docs-site/en/workflows.html +1 -1
  218. package/docs-site/index.html +1 -1
  219. package/docs-site/mcp-approval.html +1 -1
  220. package/docs-site/mcp-auth.html +1 -1
  221. package/docs-site/mcp-servers.html +1 -1
  222. package/docs-site/mcp.html +1 -1
  223. package/docs-site/permissions.html +1 -1
  224. package/docs-site/profiles.html +1 -1
  225. package/docs-site/quickstart.html +1 -1
  226. package/docs-site/runtime.html +3 -3
  227. package/docs-site/scope.html +1 -1
  228. package/docs-site/security.html +1 -1
  229. package/docs-site/team.html +1 -1
  230. package/docs-site/whats-new.html +12 -1
  231. package/docs-site/workflows.html +1 -1
  232. package/evals/architecture-conformance-v1/lane.json +18 -0
  233. package/evals/architecture-conformance-v1/runner.mjs +77 -0
  234. package/evals/fixtures/approved-host-contracts.invalid.json +43 -0
  235. package/evals/fixtures/approved-host-contracts.valid.json +43 -0
  236. package/evals/fixtures/benchmark-attempt-outcome.invalid.json +42 -0
  237. package/evals/fixtures/benchmark-attempt-outcome.valid.json +42 -0
  238. package/evals/fixtures/benchmark-independent-verification-plan.invalid.json +7 -0
  239. package/evals/fixtures/benchmark-independent-verification-plan.valid.json +21 -0
  240. package/evals/fixtures/composite-criterion.invalid.json +10 -0
  241. package/evals/fixtures/composite-criterion.valid.json +33 -0
  242. package/evals/fixtures/contract-family-library.invalid.json +42 -0
  243. package/evals/fixtures/contract-family-library.valid.json +41 -0
  244. package/evals/fixtures/contract-selection-recipe.invalid.json +29 -0
  245. package/evals/fixtures/contract-selection-recipe.valid.json +29 -0
  246. package/evals/fixtures/host-contract-plan.invalid.json +42 -0
  247. package/evals/fixtures/host-contract-plan.valid.json +41 -0
  248. package/evals/fixtures/scoped-project-verification-receipt-v2.invalid.json +42 -0
  249. package/evals/fixtures/scoped-project-verification-receipt-v2.valid.json +42 -0
  250. package/evals/fixtures/scoped-project-verification-receipt-v3.invalid.json +57 -0
  251. package/evals/fixtures/scoped-project-verification-receipt-v3.valid.json +57 -0
  252. package/evals/fixtures/scoped-verification-receipt-v1.invalid.json +40 -0
  253. package/evals/fixtures/scoped-verification-receipt-v1.valid.json +40 -0
  254. package/evals/fs5-artifact-archive/README.md +32 -0
  255. package/evals/fs5-artifact-archive/manifest.json +148 -0
  256. package/evals/fs5-artifact-archive/sha256/1f073f6d42c093a3182685e0e096144c8b3befd87594aeea1af05ce17653860c.txt +380 -0
  257. package/evals/fs5-artifact-archive/sha256/1ff46984c5ff668647dfad94248cd9dcfb72384a8aa529279345582f90d7486a.txt +331 -0
  258. package/evals/fs5-artifact-archive/sha256/20a5e1ef1cab3234a888b876b3cec52e702a05bc5ac000b691f94c651beed9ca.txt +800 -0
  259. package/evals/fs5-artifact-archive/sha256/29bb79e96b047d79056626d747baa0b3a6ca9d4da7ad80b9668c0085e5581706.txt +116 -0
  260. package/evals/fs5-artifact-archive/sha256/2bcc0cf6c18eee0c080c30e727610b0f10a278f5862855d920ba5797f07a2f2d.txt +252 -0
  261. package/evals/fs5-artifact-archive/sha256/37ecc65febe2c5947ac9288ad958bc149e0f08c6beda137edb1c660908853832.txt +324 -0
  262. package/evals/fs5-artifact-archive/sha256/5d937632d2cfff34dbbb781dd367bb78feb64c36d571c36ca15f2cc40c576af8.txt +142 -0
  263. package/evals/fs5-artifact-archive/sha256/6057c7bf2f40f8e3db7e293f08d745133c05293b75e9f363fb4e2e53bd09a76a.txt +244 -0
  264. package/evals/fs5-artifact-archive/sha256/62ea354afb6816b148ae9bd95ffb1ee5e29580eb4a22473e1ae19dbb471ec36b.txt +930 -0
  265. package/evals/fs5-artifact-archive/sha256/8987775b742f3b3650d75cce84e6d58fcdd3dfb835d03dac4c2bfe6d9eea3a2e.txt +903 -0
  266. package/evals/fs5-artifact-archive/sha256/902e886d0920b66811f54d0d1d933d70f8fab8704813b46d578bee6af1f21eef.txt +942 -0
  267. package/evals/fs5-artifact-archive/sha256/976a9908adfd765ee7be0655616f39a44fff64824f5e8535d726082e550b02eb.txt +910 -0
  268. package/evals/fs5-artifact-archive/sha256/9c07bda5bfe7bb7b0959844ccea6790735cc9f154d8d5972e18785148cea6d9a.txt +101 -0
  269. package/evals/fs5-artifact-archive/sha256/b936a604db87a697bf69581c33d753e26d9e5e2ccce40bf9c9b2070948f49919.txt +931 -0
  270. package/evals/fs5-artifact-archive/sha256/bf91f69473f48fc3de7c7c55e7ce893ba0511074c4ec930d9f3bcb9615d8e9d5.txt +950 -0
  271. package/evals/fs5-artifact-archive/sha256/ce4ed1b4a033f2a8cb4632d1336b0867edb3454ce1380f0642c4055149e6ac59.txt +784 -0
  272. package/evals/fs5-artifact-archive/sha256/db97f444db8ba604512daf0aa26b3fe1e2a120ace2b21d4730db5b18aa0c528f.txt +950 -0
  273. package/evals/fs5-causal-arm.v1.json +9 -5
  274. package/evals/fs5-pilot-protocol.v1.json +1 -1
  275. package/evals/fs5-pilot-protocol.v4.json +4 -4
  276. package/evals/fs5-pilot-protocol.v5.json +6 -5
  277. package/evals/harness-next/development-corpus.mjs +106 -0
  278. package/evals/harness-next/resource-budget-probe.mjs +38 -0
  279. package/evals/harness-next/run-calibration.mjs +83 -0
  280. package/evals/harness-next/run-resource-probe.mjs +55 -0
  281. package/evals/harness-next/selected-development-corpus.mjs +27 -0
  282. package/evals/harness-next.v1.json +51 -0
  283. package/evals/long-horizon-v1/README.md +19 -0
  284. package/evals/long-horizon-v1/runner.mjs +116 -48
  285. package/evals/long-horizon-v1/worker.mjs +77 -30
  286. package/evals/private-holdout-v1/CURRENT_BOUNDARY.md +53 -0
  287. package/evals/private-holdout-v1/README.md +5 -2
  288. package/evals/private-holdout-v1/public-exposure.v2.json +273 -0
  289. package/package.json +9 -4
  290. package/packages/piagent-core/README.md +2 -2
  291. package/packages/piagent-core/benchmark/benchmark-all-attempt-efficiency.js +55 -0
  292. package/packages/piagent-core/benchmark/benchmark-artifact-binding.js +49 -0
  293. package/packages/piagent-core/benchmark/benchmark-bootstrap.js +265 -14
  294. package/packages/piagent-core/benchmark/benchmark-budget-governor.js +271 -0
  295. package/packages/piagent-core/benchmark/benchmark-campaign.js +592 -0
  296. package/packages/piagent-core/benchmark/benchmark-candidate.js +8 -3
  297. package/packages/piagent-core/benchmark/benchmark-checkpoint.js +67 -0
  298. package/packages/piagent-core/benchmark/benchmark-claim-restrictions.js +269 -2
  299. package/packages/piagent-core/benchmark/benchmark-cli.js +120 -3
  300. package/packages/piagent-core/benchmark/benchmark-codex-events.js +235 -0
  301. package/packages/piagent-core/benchmark/benchmark-codex-outcome.js +20 -0
  302. package/packages/piagent-core/benchmark/benchmark-codex-relative-efficiency.js +2 -1
  303. package/packages/piagent-core/benchmark/benchmark-codex-rollout.js +521 -0
  304. package/packages/piagent-core/benchmark/benchmark-codex.js +104 -8
  305. package/packages/piagent-core/benchmark/benchmark-comparison.js +284 -1
  306. package/packages/piagent-core/benchmark/benchmark-core.js +107 -107
  307. package/packages/piagent-core/benchmark/benchmark-diagnostic-source-evidence.js +40 -0
  308. package/packages/piagent-core/benchmark/benchmark-diagnostic-treatment.js +104 -0
  309. package/packages/piagent-core/benchmark/benchmark-evaluator-v3.js +226 -0
  310. package/packages/piagent-core/benchmark/benchmark-execution-guard.js +23 -1
  311. package/packages/piagent-core/benchmark/benchmark-forensics.js +63 -50
  312. package/packages/piagent-core/benchmark/benchmark-goal-assessment.js +193 -0
  313. package/packages/piagent-core/benchmark/benchmark-goal-report.js +87 -0
  314. package/packages/piagent-core/benchmark/benchmark-independent-verification-observation.js +64 -0
  315. package/packages/piagent-core/benchmark/benchmark-matrix-summary.js +141 -0
  316. package/packages/piagent-core/benchmark/benchmark-matrix.js +1 -0
  317. package/packages/piagent-core/benchmark/benchmark-normalized-cost.js +62 -0
  318. package/packages/piagent-core/benchmark/benchmark-pi-home.js +1 -0
  319. package/packages/piagent-core/benchmark/benchmark-preflight.js +69 -8
  320. package/packages/piagent-core/benchmark/benchmark-process.js +23 -9
  321. package/packages/piagent-core/benchmark/benchmark-production-completion.js +84 -0
  322. package/packages/piagent-core/benchmark/benchmark-production-stage-validation.js +209 -0
  323. package/packages/piagent-core/benchmark/benchmark-provider-free-evidence.js +162 -45
  324. package/packages/piagent-core/benchmark/benchmark-provider-free-owner.js +24 -0
  325. package/packages/piagent-core/benchmark/benchmark-provider-wire.js +378 -0
  326. package/packages/piagent-core/benchmark/benchmark-record-validation.d.ts +2 -0
  327. package/packages/piagent-core/benchmark/benchmark-record-validation.js +269 -3
  328. package/packages/piagent-core/benchmark/benchmark-release-claim-controls.js +132 -0
  329. package/packages/piagent-core/benchmark/benchmark-report.js +79 -4
  330. package/packages/piagent-core/benchmark/benchmark-resume-recovery.js +280 -10
  331. package/packages/piagent-core/benchmark/benchmark-runner-policy.js +88 -1
  332. package/packages/piagent-core/benchmark/benchmark-runtime-identity.js +50 -1
  333. package/packages/piagent-core/benchmark/benchmark-runtime.js +102 -10
  334. package/packages/piagent-core/benchmark/benchmark-safety-evidence.js +74 -0
  335. package/packages/piagent-core/benchmark/benchmark-service-tier.js +189 -0
  336. package/packages/piagent-core/benchmark/benchmark-stage-diagnostic.js +144 -128
  337. package/packages/piagent-core/benchmark/benchmark-suite-identity.js +264 -1
  338. package/packages/piagent-core/benchmark/benchmark-suite-runtime.js +4 -0
  339. package/packages/piagent-core/benchmark/benchmark-suite.js +211 -4
  340. package/packages/piagent-core/benchmark/benchmark-summary-support.js +30 -2
  341. package/packages/piagent-core/benchmark/benchmark-timing-contract.js +189 -0
  342. package/packages/piagent-core/benchmark/benchmark-timing-diagnostics.js +110 -139
  343. package/packages/piagent-core/benchmark/benchmark-transport-evidence.js +203 -0
  344. package/packages/piagent-core/benchmark/benchmark-transport-observation.js +147 -0
  345. package/packages/piagent-core/benchmark/benchmark-tree-identity.js +2 -1
  346. package/packages/piagent-core/benchmark/benchmark-usage.js +395 -33
  347. package/packages/piagent-core/benchmark/benchmark-workflow.js +136 -38
  348. package/packages/piagent-core/benchmark/fs5-pilot-protocol.js +21 -2
  349. package/packages/piagent-core/capabilities/runtime-integrity.js +3 -0
  350. package/packages/piagent-core/extensions/acceptance-api-baseline.js +24 -0
  351. package/packages/piagent-core/extensions/acceptance-assessment.js +153 -0
  352. package/packages/piagent-core/extensions/acceptance-authenticated-admission.d.ts +41 -0
  353. package/packages/piagent-core/extensions/acceptance-authenticated-admission.js +120 -0
  354. package/packages/piagent-core/extensions/acceptance-behavior-proof.js +264 -70
  355. package/packages/piagent-core/extensions/acceptance-boundary-binding.js +31 -0
  356. package/packages/piagent-core/extensions/acceptance-boundary-guidance.js +214 -0
  357. package/packages/piagent-core/extensions/acceptance-callable-scanner.js +244 -0
  358. package/packages/piagent-core/extensions/acceptance-clause-group.js +45 -0
  359. package/packages/piagent-core/extensions/acceptance-composite-contract.js +334 -0
  360. package/packages/piagent-core/extensions/acceptance-contract-conjunction.js +78 -0
  361. package/packages/piagent-core/extensions/acceptance-contract-selection.js +161 -0
  362. package/packages/piagent-core/extensions/acceptance-contract-semantics.js +211 -211
  363. package/packages/piagent-core/extensions/acceptance-declared-api.js +126 -0
  364. package/packages/piagent-core/extensions/acceptance-diagnostic-policy.ts +14 -0
  365. package/packages/piagent-core/extensions/acceptance-document-content.js +73 -0
  366. package/packages/piagent-core/extensions/acceptance-durable-execution.d.ts +59 -0
  367. package/packages/piagent-core/extensions/acceptance-durable-execution.js +139 -0
  368. package/packages/piagent-core/extensions/acceptance-error-classes.js +1 -1
  369. package/packages/piagent-core/extensions/acceptance-evidence-store.js +223 -0
  370. package/packages/piagent-core/extensions/acceptance-executable-evidence.js +490 -88
  371. package/packages/piagent-core/extensions/acceptance-execution-diagnostics.js +34 -0
  372. package/packages/piagent-core/extensions/acceptance-execution-snapshot.js +174 -0
  373. package/packages/piagent-core/extensions/acceptance-executor/.dockerignore +21 -0
  374. package/packages/piagent-core/extensions/acceptance-executor/Dockerfile +5 -0
  375. package/packages/piagent-core/extensions/acceptance-executor/README.md +103 -0
  376. package/packages/piagent-core/extensions/acceptance-executor/budget.mjs +54 -0
  377. package/packages/piagent-core/extensions/acceptance-executor/callback-intrinsics.mjs +126 -0
  378. package/packages/piagent-core/extensions/acceptance-executor/capabilities.mjs +127 -0
  379. package/packages/piagent-core/extensions/acceptance-executor/guest.mjs +285 -0
  380. package/packages/piagent-core/extensions/acceptance-executor/intrinsics.mjs +133 -0
  381. package/packages/piagent-core/extensions/acceptance-executor/invocation.mjs +39 -0
  382. package/packages/piagent-core/extensions/acceptance-executor/jobs.mjs +29 -0
  383. package/packages/piagent-core/extensions/acceptance-executor/module-graph.mjs +70 -0
  384. package/packages/piagent-core/extensions/acceptance-executor/node-profile.mjs +398 -0
  385. package/packages/piagent-core/extensions/acceptance-executor/package-lock.json +40 -0
  386. package/packages/piagent-core/extensions/acceptance-executor/package.json +10 -0
  387. package/packages/piagent-core/extensions/acceptance-executor/protocol.mjs +248 -0
  388. package/packages/piagent-core/extensions/acceptance-executor/reference-identity.mjs +61 -0
  389. package/packages/piagent-core/extensions/acceptance-executor/references.mjs +14 -0
  390. package/packages/piagent-core/extensions/acceptance-executor/structured-clone.mjs +53 -0
  391. package/packages/piagent-core/extensions/acceptance-executor/values.mjs +97 -0
  392. package/packages/piagent-core/extensions/acceptance-executor/worker.mjs +75 -0
  393. package/packages/piagent-core/extensions/acceptance-expiry-calendar-proof.js +162 -0
  394. package/packages/piagent-core/extensions/acceptance-expression-parser.js +155 -0
  395. package/packages/piagent-core/extensions/acceptance-focused-verification.js +40 -0
  396. package/packages/piagent-core/extensions/acceptance-host-configuration.js +337 -0
  397. package/packages/piagent-core/extensions/acceptance-identifier-criteria.js +25 -0
  398. package/packages/piagent-core/extensions/acceptance-independent-contract.js +209 -0
  399. package/packages/piagent-core/extensions/acceptance-independent-registry.js +106 -0
  400. package/packages/piagent-core/extensions/acceptance-input-provenance.js +40 -0
  401. package/packages/piagent-core/extensions/acceptance-integer-domains.js +322 -0
  402. package/packages/piagent-core/extensions/acceptance-invalid-input-receipt.js +20 -0
  403. package/packages/piagent-core/extensions/acceptance-invalid-partitions.js +26 -0
  404. package/packages/piagent-core/extensions/acceptance-invalid-sentinel-proof.js +477 -0
  405. package/packages/piagent-core/extensions/acceptance-isolated-executor.js +254 -0
  406. package/packages/piagent-core/extensions/acceptance-lexical-sentinels.js +130 -0
  407. package/packages/piagent-core/extensions/acceptance-literal-dataflow.js +735 -0
  408. package/packages/piagent-core/extensions/acceptance-local-binding.js +106 -0
  409. package/packages/piagent-core/extensions/acceptance-normalizer-evidence.js +719 -0
  410. package/packages/piagent-core/extensions/acceptance-obligations.js +75 -0
  411. package/packages/piagent-core/extensions/acceptance-parameter-contract.js +756 -0
  412. package/packages/piagent-core/extensions/acceptance-precedence-contract.js +572 -0
  413. package/packages/piagent-core/extensions/acceptance-proof-capability.js +26 -0
  414. package/packages/piagent-core/extensions/acceptance-receipt.js +139 -152
  415. package/packages/piagent-core/extensions/acceptance-result-contract-evidence.js +159 -0
  416. package/packages/piagent-core/extensions/acceptance-source-origin.js +214 -0
  417. package/packages/piagent-core/extensions/acceptance-state-machine-evidence.js +64 -8
  418. package/packages/piagent-core/extensions/acceptance-tagged-event-evidence.js +685 -0
  419. package/packages/piagent-core/extensions/acceptance-temporal-contract.js +756 -0
  420. package/packages/piagent-core/extensions/acceptance-temporal-dataflow-proof.js +618 -0
  421. package/packages/piagent-core/extensions/acceptance-test-binding-evidence.js +20 -2
  422. package/packages/piagent-core/extensions/acceptance-test-bindings.js +86 -0
  423. package/packages/piagent-core/extensions/acceptance-tuple-shapes.js +12 -0
  424. package/packages/piagent-core/extensions/context-engine.js +11 -6
  425. package/packages/piagent-core/extensions/criterion-context-pack.js +96 -4
  426. package/packages/piagent-core/extensions/failure-types.ts +2 -2
  427. package/packages/piagent-core/extensions/guard-types.ts +5 -1
  428. package/packages/piagent-core/extensions/piagent-guard.ts +197 -155
  429. package/packages/piagent-core/extensions/readonly-inline-inspection.ts +378 -0
  430. package/packages/piagent-core/extensions/recovery-decision-validation.ts +92 -0
  431. package/packages/piagent-core/extensions/shell-write-targets.js +15 -14
  432. package/packages/piagent-core/extensions/task-contract-view.js +4 -4
  433. package/packages/piagent-core/extensions/task-digest-migration.js +2 -2
  434. package/packages/piagent-core/extensions/task-journal.js +48 -3
  435. package/packages/piagent-core/extensions/task-runtime-audit.js +8 -2
  436. package/packages/piagent-core/extensions/task-state.js +74 -38
  437. package/packages/piagent-core/extensions/task-terminal-disposition.js +10 -0
  438. package/packages/piagent-core/extensions/trajectory-contract.ts +5 -0
  439. package/packages/piagent-core/extensions/verification-intelligence.js +60 -8
  440. package/packages/piagent-core/extensions/workflow-catalog.ts +98 -0
  441. package/packages/piagent-core/extensions/workspace-revision.js +48 -0
  442. package/packages/piagent-core/extensions/workspace-verification-snapshot.js +13 -0
  443. package/packages/piagent-core/package.json +3 -3
  444. package/packages/piagent-core/policies/base-policy.json +1 -0
  445. package/packages/piagent-core/runtime/context/adaptive-planner.ts +7 -3
  446. package/packages/piagent-core/runtime/hooks/agent-start-hook.ts +60 -59
  447. package/packages/piagent-core/runtime/hooks/agent-start-types.ts +48 -0
  448. package/packages/piagent-core/runtime/hooks/completion-hook.ts +76 -83
  449. package/packages/piagent-core/runtime/hooks/input-hook.ts +92 -7
  450. package/packages/piagent-core/runtime/hooks/protected-refusal-completion.ts +86 -0
  451. package/packages/piagent-core/runtime/hooks/session-hooks.ts +5 -2
  452. package/packages/piagent-core/runtime/hooks/session-start-hook.ts +50 -18
  453. package/packages/piagent-core/runtime/hooks/tool-result-hook.ts +53 -20
  454. package/packages/piagent-core/runtime/hooks/tool-result-metadata.ts +9 -0
  455. package/packages/piagent-core/runtime/hooks/tool-result-navigation.ts +36 -0
  456. package/packages/piagent-core/runtime/hooks/tool-result-retrieval.ts +42 -0
  457. package/packages/piagent-core/runtime/input/chat-images.ts +1 -1
  458. package/packages/piagent-core/runtime/inspection/activity-event-adapter.ts +1 -1
  459. package/packages/piagent-core/runtime/inspection/same-process-bridge-proof.ts +1 -1
  460. package/packages/piagent-core/runtime/inspection/task-handoff-history.ts +2 -2
  461. package/packages/piagent-core/runtime/inspection/tool-failure-classification.ts +49 -8
  462. package/packages/piagent-core/runtime/inspection/webui-snapshot.ts +61 -15
  463. package/packages/piagent-core/runtime/inspection/workspace-file-reader.ts +2 -132
  464. package/packages/piagent-core/runtime/model/provider-wire-fingerprint.ts +20 -0
  465. package/packages/piagent-core/runtime/model/service-tier-runtime.ts +242 -0
  466. package/packages/piagent-core/runtime/policy/source-worktree-transaction.ts +75 -14
  467. package/packages/piagent-core/runtime/product/activity-inspector.ts +6 -3
  468. package/packages/piagent-core/runtime/product/edit-verifier-timing.ts +26 -0
  469. package/packages/piagent-core/runtime/product/efficiency-metrics.ts +12 -7
  470. package/packages/piagent-core/runtime/product/operator-projections.ts +8 -3
  471. package/packages/piagent-core/runtime/product/task-usage-hooks.ts +78 -0
  472. package/packages/piagent-core/runtime/product/task-usage.ts +104 -0
  473. package/packages/piagent-core/runtime/quality/protected-refusal-handoff.ts +35 -0
  474. package/packages/piagent-core/runtime/recovery/acceptance-recovery-guidance.ts +21 -0
  475. package/packages/piagent-core/runtime/recovery/continuation-budget.ts +11 -2
  476. package/packages/piagent-core/runtime/recovery/diagnostic-delivery.ts +61 -0
  477. package/packages/piagent-core/runtime/recovery/handoff-projection.ts +192 -57
  478. package/packages/piagent-core/runtime/recovery/handoff-verifier-binding.ts +24 -0
  479. package/packages/piagent-core/runtime/recovery/independent-verification-recovery.ts +79 -0
  480. package/packages/piagent-core/runtime/recovery/recovery-policy.ts +27 -39
  481. package/packages/piagent-core/runtime/recovery/resume-state.ts +8 -5
  482. package/packages/piagent-core/runtime/registration/activity-inspector-command.ts +2 -1
  483. package/packages/piagent-core/runtime/registration/context-commands.ts +12 -8
  484. package/packages/piagent-core/runtime/registration/operator-catalogs.ts +4 -9
  485. package/packages/piagent-core/runtime/registration/policy-tools.ts +5 -3
  486. package/packages/piagent-core/runtime/registration/session-commands.ts +9 -7
  487. package/packages/piagent-core/runtime/registration/session-wait-tool.ts +54 -0
  488. package/packages/piagent-core/runtime/registration/task-completion-tools.ts +27 -11
  489. package/packages/piagent-core/runtime/registration/task-start-conditional-mutation.ts +45 -0
  490. package/packages/piagent-core/runtime/registration/task-start-guidance.ts +5 -2
  491. package/packages/piagent-core/runtime/registration/task-start-tool.ts +23 -23
  492. package/packages/piagent-core/runtime/registration/workflow-commands.ts +64 -93
  493. package/packages/piagent-core/runtime/session/completion-signals.ts +23 -0
  494. package/packages/piagent-core/runtime/session/message-signals.ts +6 -22
  495. package/packages/piagent-core/runtime/session/retrieval-evidence-runtime.ts +188 -0
  496. package/packages/piagent-core/runtime/session/runtime-state.ts +29 -35
  497. package/packages/piagent-core/runtime/session/system-prompt.ts +3 -3
  498. package/packages/piagent-core/runtime/session/uncertain-send-continuation.ts +231 -0
  499. package/packages/piagent-core/runtime/session/usage.ts +4 -3
  500. package/packages/piagent-core/runtime/tools/apply-patch-tool.ts +6 -6
  501. package/packages/piagent-core/runtime/tools/phase-tool-runtime.ts +1 -1
  502. package/packages/piagent-core/runtime/tools/phase-tools.ts +3 -2
  503. package/packages/piagent-core/runtime/tools/session-wait.ts +42 -0
  504. package/packages/piagent-core/runtime/tools/tool-groups.ts +10 -3
  505. package/packages/piagent-core/runtime/trajectory/trajectory-runtime.ts +192 -5
  506. package/packages/piagent-core/runtime/trajectory/trajectory-types.ts +3 -2
  507. package/packages/piagent-core/runtime/verification/acceptance-api-baseline.js +488 -0
  508. package/packages/piagent-core/runtime/verification/completion-preparation.ts +22 -0
  509. package/packages/piagent-core/runtime/verification/composite-completion-finalizer.ts +105 -0
  510. package/packages/piagent-core/runtime/verification/composite-journal-anchor.ts +181 -0
  511. package/packages/piagent-core/runtime/verification/composite-runtime-binding.ts +76 -0
  512. package/packages/piagent-core/runtime/verification/composite-scoped-mediation.ts +318 -0
  513. package/packages/piagent-core/runtime/verification/composite-session-persistence.ts +175 -0
  514. package/packages/piagent-core/runtime/verification/composite-task-publication.ts +273 -0
  515. package/packages/piagent-core/runtime/verification/exact-verifier-reuse.ts +11 -11
  516. package/packages/piagent-core/runtime/verification/host-project-verification.ts +71 -0
  517. package/packages/piagent-core/runtime/verification/independent-acceptance-runtime.ts +500 -0
  518. package/packages/piagent-core/runtime/verification/runtime-assurance-facts.ts +500 -0
  519. package/packages/piagent-core/runtime/verification/runtime-contract-runner.ts +107 -0
  520. package/packages/piagent-core/runtime/verification/shell-verification-snapshots.ts +56 -0
  521. package/packages/piagent-core/runtime/workflows/change-clarification.ts +61 -0
  522. package/packages/piagent-core/runtime/workflows/input-routing.ts +21 -14
  523. package/packages/piagent-core/runtime/workflows/runtime-intake-compaction.ts +44 -7
  524. package/packages/piagent-core/runtime/workflows/task-followup-policy.ts +163 -0
  525. package/packages/piagent-core/runtime/workflows/task-intake.ts +221 -242
  526. package/packages/piagent-core/runtime/workflows/task-repair-intent.ts +67 -0
  527. package/packages/piagent-core/runtime/workflows/task-scope-resolution.ts +83 -0
  528. package/packages/piagent-core/runtime/workflows/webui-runtime-command.ts +5 -2
  529. package/packages/piagent-core/runtime/workflows/webui-workflow.ts +39 -32
  530. package/packages/piagent-core/runtime/workflows/workflow-follow-up.ts +33 -0
  531. package/packages/piagent-core/security/workspace-file-reader.ts +132 -0
  532. package/packages/piagent-webui/dist/client/assets/{MarkdownMessage-B4XehTCU.js → MarkdownMessage-DGk34zJT.js} +1 -1
  533. package/packages/piagent-webui/dist/client/assets/{index-BV-_03yl.css → index-DVahUVWy.css} +1 -1
  534. package/packages/piagent-webui/dist/client/assets/index-DiKRmCEP.js +176 -0
  535. package/packages/piagent-webui/dist/client/index.html +2 -2
  536. package/packages/piagent-webui/extension/dispatch-user-entry.ts +13 -0
  537. package/packages/piagent-webui/extension/piagent-webui.ts +10 -2
  538. package/packages/piagent-webui/extension/review-controller.ts +29 -6
  539. package/packages/piagent-webui/extension/same-session-bridge.ts +6 -7
  540. package/packages/piagent-webui/extension/session-stream-adapter.ts +60 -11
  541. package/packages/piagent-webui/extension/source-mutation-controller.ts +29 -3
  542. package/packages/piagent-webui/extension/source-open-controller.ts +29 -7
  543. package/packages/piagent-webui/extension/source-revert-controller.ts +26 -3
  544. package/packages/piagent-webui/gateway/gateway-events.ts +22 -5
  545. package/packages/piagent-webui/gateway/gateway-service.ts +7 -2
  546. package/packages/piagent-webui/gateway/gateway-session-stream.ts +141 -19
  547. package/packages/piagent-webui/gateway/rpc-ui-context.ts +34 -9
  548. package/packages/piagent-webui/gateway/runtime-command-controller.ts +46 -6
  549. package/packages/piagent-webui/gateway/runtime-session-controls.ts +17 -2
  550. package/packages/piagent-webui/gateway/session-catalog.ts +9 -3
  551. package/packages/piagent-webui/gateway/session-command-controller.ts +29 -9
  552. package/packages/piagent-webui/gateway/session-effective-options.ts +2 -9
  553. package/packages/piagent-webui/gateway/session-inspection-registry.ts +11 -2
  554. package/packages/piagent-webui/gateway/session-operation-settlement.ts +28 -0
  555. package/packages/piagent-webui/gateway/session-prompt-dispatch.ts +37 -0
  556. package/packages/piagent-webui/gateway/session-runtime-close.ts +36 -0
  557. package/packages/piagent-webui/gateway/session-runtime-factory.ts +183 -6
  558. package/packages/piagent-webui/gateway/session-runtime-ownership.ts +48 -0
  559. package/packages/piagent-webui/gateway/session-runtime-state.ts +24 -0
  560. package/packages/piagent-webui/gateway/session-runtime-supervisor.ts +114 -114
  561. package/packages/piagent-webui/server/core-inspection-provider.ts +73 -14
  562. package/packages/piagent-webui/server/loopback-server.ts +9 -7
  563. package/packages/piagent-webui/server/terminal-delivery-receipt.ts +144 -0
  564. package/packages/piagent-webui/server/transcript-projection.ts +65 -12
  565. package/packages/piagent-webui/shared/message-correlation.ts +25 -0
  566. package/schemas/approved-host-contracts.schema.json +973 -0
  567. package/schemas/benchmark-attempt-outcome.schema.json +303 -0
  568. package/schemas/benchmark-independent-verification-plan.schema.json +34 -0
  569. package/schemas/composite-criterion.schema.json +788 -0
  570. package/schemas/contract-family-library.schema.json +28 -0
  571. package/schemas/contract-selection-recipe.schema.json +31 -0
  572. package/schemas/failure-evidence.schema.json +2 -2
  573. package/schemas/helper-request.schema.json +16 -4
  574. package/schemas/host-contract-plan.schema.json +25 -0
  575. package/schemas/piagent-webui/gateway-protocol-v1.schema.json +16 -2
  576. package/schemas/piagent-webui/runtime-command-v1.schema.json +26 -3
  577. package/schemas/piagent-webui/runtime-event-v2.schema.json +32 -1
  578. package/schemas/piagent-webui/session-live-state-v1.schema.json +2 -0
  579. package/schemas/piagent-webui/snapshot-v1.schema.json +3 -0
  580. package/schemas/piagent-webui/source-change-v1.schema.json +12 -0
  581. package/schemas/piagent-webui/transcript-v1.schema.json +2 -1
  582. package/schemas/public-contracts/codex-mcp-metadata-v1.schema.json +206 -0
  583. package/schemas/public-contracts/codex-mcp-metadata-v2.schema.json +228 -0
  584. package/schemas/public-contracts/codex-mcp-metadata-v3.schema.json +233 -0
  585. package/schemas/public-contracts/revocation-v1.schema.json +87 -0
  586. package/schemas/public-contracts/subscription-contract-v1.schema.json +76 -0
  587. package/schemas/scoped-project-verification-receipt-v2.schema.json +83 -0
  588. package/schemas/scoped-project-verification-receipt-v3.schema.json +295 -0
  589. package/schemas/scoped-verification-receipt-v1.schema.json +76 -0
  590. package/schemas/task-contract.schema.json +30 -0
  591. package/scripts/approve-independent-verification.mjs +57 -0
  592. package/scripts/benchmark-arm-observer.mjs +320 -0
  593. package/scripts/benchmark-budget-ipc.mjs +34 -0
  594. package/scripts/benchmark-budget-launcher.mjs +91 -0
  595. package/scripts/benchmark-budget-processes.mjs +133 -0
  596. package/scripts/benchmark-budget-publication.mjs +228 -0
  597. package/scripts/benchmark-budget-runtime.mjs +136 -0
  598. package/scripts/benchmark-candidate-ablation.mjs +265 -0
  599. package/scripts/benchmark-codex-journey.mjs +406 -0
  600. package/scripts/benchmark-independent-verification.mjs +628 -0
  601. package/scripts/benchmark-journey-outcome.mjs +92 -0
  602. package/scripts/benchmark-production-v3-human-review.mjs +192 -0
  603. package/scripts/benchmark-public-input-catalog.mjs +158 -0
  604. package/scripts/benchmark-registered-plan-author.mjs +355 -0
  605. package/scripts/benchmark-registered-plan-recipe-support.mjs +85 -0
  606. package/scripts/benchmark-registered-plan-recipes-backend.mjs +165 -0
  607. package/scripts/benchmark-registered-plan-recipes-data.mjs +164 -0
  608. package/scripts/benchmark-registered-plan-recipes-platform.mjs +165 -0
  609. package/scripts/benchmark-runner-completion.mjs +133 -0
  610. package/scripts/benchmark-runner-configuration.mjs +159 -0
  611. package/scripts/benchmark-runner-core.mjs +217 -388
  612. package/scripts/benchmark-runner-finalization.mjs +140 -27
  613. package/scripts/benchmark-runner-provider-boundary.mjs +44 -0
  614. package/scripts/benchmark-runner-schedule.mjs +256 -0
  615. package/scripts/benchmark-runner-support.mjs +136 -4
  616. package/scripts/benchmark-runner.mjs +95 -18
  617. package/scripts/benchmark-scoped-broker-custody.mjs +483 -0
  618. package/scripts/benchmark-scoped-docs-verifier.mjs +203 -0
  619. package/scripts/benchmark-scoped-frozen-qualification.mjs +676 -0
  620. package/scripts/benchmark-scoped-pi-router.mjs +196 -0
  621. package/scripts/benchmark-scoped-project-verification-receipt.mjs +49 -0
  622. package/scripts/benchmark-scoped-project-verifier.mjs +461 -0
  623. package/scripts/benchmark-scoped-session-custody.mjs +599 -0
  624. package/scripts/benchmark-scoped-session-factory.mjs +342 -0
  625. package/scripts/benchmark-scoped-tool-broker.mjs +523 -0
  626. package/scripts/benchmark-scoped-verification-supervisor.mjs +773 -0
  627. package/scripts/benchmark-session-evaluator.mjs +169 -0
  628. package/scripts/benchmark-session-provider-boundary.mjs +62 -0
  629. package/scripts/benchmark-session-usage.mjs +28 -0
  630. package/scripts/benchmark-session.mjs +343 -50
  631. package/scripts/benchmark-webui-journey.mjs +678 -0
  632. package/scripts/check-public-wording.mjs +46 -0
  633. package/scripts/fs4-readiness-evaluation.mjs +7 -1
  634. package/scripts/phase-tools-evaluation.mjs +1 -0
  635. package/scripts/pi-auto.sh +7 -1
  636. package/scripts/pi-usage-history.mjs +201 -11
  637. package/scripts/piagent-cli.mjs +8 -0
  638. package/scripts/private-holdout-readiness.mjs +7 -1
  639. package/scripts/public-evaluation-exposure.mjs +140 -0
  640. package/scripts/select-independent-verification.mjs +62 -0
  641. package/scripts/typescript-loader.mjs +135 -1
  642. package/scripts/update-global.mjs +6 -7
  643. package/scripts/verify-local.sh +36 -55
  644. package/scripts/verify-release-identity.mjs +5 -3
  645. package/packages/piagent-webui/dist/client/assets/index-CgOEyRPG.js +0 -176
@@ -0,0 +1,784 @@
1
+ #!/usr/bin/env node
2
+ import fs from "node:fs";
3
+ import os from "node:os";
4
+ import path from "node:path";
5
+ import { pathToFileURL } from "node:url";
6
+
7
+ const USAGE = `Usage:
8
+ scripts/pi-usage-history.mjs [project-path] [options]
9
+ scripts/pi-usage-history.mjs --all-projects [options]
10
+
11
+ Options:
12
+ --history Accepted for piagent-usage wrapper compatibility.
13
+ --all-projects Include every Pi session under the session directory.
14
+ --days N Include activity from the last N days.
15
+ --since YYYY-MM-DD Include activity on or after this local date.
16
+ --until YYYY-MM-DD Include activity on or before this local date.
17
+ --format table|json|csv|markdown
18
+ --json Alias for --format json.
19
+ --csv Alias for --format csv.
20
+ --markdown Alias for --format markdown.
21
+ --include-subagents Include subagent session files. Default.
22
+ --no-subagents Exclude files under sessions/subagent/.
23
+ --limit N Max session rows in table/markdown output. Default: 20.
24
+ --sessions-dir PATH Override Pi session root for tests or custom installs.
25
+ -h, --help Show this help.
26
+
27
+ Examples:
28
+ piagent-usage --history .
29
+ piagent-usage --history --days 7 --all-projects
30
+ piagent-usage --history /path/to/project --json
31
+ `;
32
+
33
+ const DATE_ONLY = /^(\d{4})-(\d{2})-(\d{2})$/;
34
+
35
+ function fail(message) {
36
+ console.error(`FAIL: ${message}`);
37
+ process.exit(2);
38
+ }
39
+
40
+ function parsePositiveInteger(raw, label) {
41
+ if (!/^[1-9]\d*$/.test(String(raw ?? ""))) fail(`${label} must be a positive integer`);
42
+ return Number(raw);
43
+ }
44
+
45
+ function parseLocalDateStart(raw, label) {
46
+ const match = String(raw ?? "").match(DATE_ONLY);
47
+ if (!match) fail(`${label} must use YYYY-MM-DD`);
48
+ const [, y, m, d] = match;
49
+ const date = new Date(Number(y), Number(m) - 1, Number(d), 0, 0, 0, 0);
50
+ if (Number.isNaN(date.getTime())) fail(`${label} is not a valid date`);
51
+ return date;
52
+ }
53
+
54
+ function parseLocalDateEnd(raw, label) {
55
+ const start = parseLocalDateStart(raw, label);
56
+ return new Date(start.getFullYear(), start.getMonth(), start.getDate(), 23, 59, 59, 999);
57
+ }
58
+
59
+ function parseArgs(argv) {
60
+ const options = {
61
+ allProjects: false,
62
+ format: "table",
63
+ includeSubagents: true,
64
+ limit: 20,
65
+ sessionsDir: process.env.PI_CODING_AGENT_SESSION_DIR
66
+ || path.join(process.env.PI_CODING_AGENT_DIR || path.join(os.homedir(), ".pi", "agent"), "sessions"),
67
+ projectPath: undefined,
68
+ since: undefined,
69
+ until: undefined
70
+ };
71
+ const positionals = [];
72
+
73
+ for (let index = 0; index < argv.length; index += 1) {
74
+ const arg = argv[index];
75
+ const readValue = (name) => {
76
+ const value = argv[index + 1];
77
+ if (value === undefined || value.startsWith("-")) fail(`Missing value for ${name}`);
78
+ index += 1;
79
+ return value;
80
+ };
81
+
82
+ switch (arg) {
83
+ case "-h":
84
+ case "--help":
85
+ options.help = true;
86
+ break;
87
+ case "--history":
88
+ break;
89
+ case "--all-projects":
90
+ options.allProjects = true;
91
+ break;
92
+ case "--days": {
93
+ const days = parsePositiveInteger(readValue("--days"), "--days");
94
+ options.since = new Date(Date.now() - days * 24 * 60 * 60 * 1000);
95
+ break;
96
+ }
97
+ case "--since":
98
+ options.since = parseLocalDateStart(readValue("--since"), "--since");
99
+ break;
100
+ case "--until":
101
+ options.until = parseLocalDateEnd(readValue("--until"), "--until");
102
+ break;
103
+ case "--format": {
104
+ const format = readValue("--format");
105
+ if (!["table", "json", "csv", "markdown"].includes(format)) {
106
+ fail("--format must be one of table, json, csv, markdown");
107
+ }
108
+ options.format = format;
109
+ break;
110
+ }
111
+ case "--json":
112
+ options.format = "json";
113
+ break;
114
+ case "--csv":
115
+ options.format = "csv";
116
+ break;
117
+ case "--markdown":
118
+ options.format = "markdown";
119
+ break;
120
+ case "--include-subagents":
121
+ options.includeSubagents = true;
122
+ break;
123
+ case "--no-subagents":
124
+ options.includeSubagents = false;
125
+ break;
126
+ case "--limit":
127
+ options.limit = parsePositiveInteger(readValue("--limit"), "--limit");
128
+ break;
129
+ case "--sessions-dir":
130
+ options.sessionsDir = path.resolve(readValue("--sessions-dir"));
131
+ break;
132
+ default:
133
+ if (arg.startsWith("-")) fail(`unknown option: ${arg}`);
134
+ positionals.push(arg);
135
+ break;
136
+ }
137
+ }
138
+
139
+ if (positionals.length > 1) fail(`expected at most one project path, received ${positionals.length}`);
140
+ if (positionals[0]) {
141
+ const projectPath = path.resolve(positionals[0]);
142
+ if (!fs.existsSync(projectPath) || !fs.statSync(projectPath).isDirectory()) {
143
+ fail(`project path does not exist: ${projectPath}`);
144
+ }
145
+ options.projectPath = fs.realpathSync(projectPath);
146
+ } else if (!options.allProjects) {
147
+ options.projectPath = fs.realpathSync(process.cwd());
148
+ }
149
+
150
+ if (options.since && options.until && options.since > options.until) {
151
+ fail("--since must be before or equal to --until");
152
+ }
153
+ return options;
154
+ }
155
+
156
+ function walkJsonl(root) {
157
+ const out = [];
158
+ if (!fs.existsSync(root)) return out;
159
+ const stack = [root];
160
+ while (stack.length) {
161
+ const dir = stack.pop();
162
+ let entries = [];
163
+ try {
164
+ entries = fs.readdirSync(dir, { withFileTypes: true });
165
+ } catch {
166
+ continue;
167
+ }
168
+ for (const entry of entries) {
169
+ const target = path.join(dir, entry.name);
170
+ if (entry.isDirectory()) {
171
+ stack.push(target);
172
+ } else if (entry.isFile() && entry.name.endsWith(".jsonl")) {
173
+ out.push(target);
174
+ }
175
+ }
176
+ }
177
+ return out;
178
+ }
179
+
180
+ function parseIso(raw) {
181
+ if (!raw) return undefined;
182
+ const date = new Date(raw);
183
+ return Number.isNaN(date.getTime()) ? undefined : date;
184
+ }
185
+
186
+ function inRange(date, options) {
187
+ if (!date) return true;
188
+ if (options.since && date < options.since) return false;
189
+ if (options.until && date > options.until) return false;
190
+ return true;
191
+ }
192
+
193
+ function emptyTotals() {
194
+ return {
195
+ input: 0,
196
+ output: 0,
197
+ cacheRead: 0,
198
+ cacheWrite: 0,
199
+ reasoning: 0,
200
+ total: 0,
201
+ cost: 0
202
+ };
203
+ }
204
+
205
+ function emptyExecutionAggregate(source = "pi-session-jsonl") {
206
+ return {
207
+ schemaVersion: 1,
208
+ source,
209
+ completeness: {
210
+ tools: "unverified",
211
+ retries: "unavailable",
212
+ compactions: "partial",
213
+ subagents: "unverified"
214
+ },
215
+ providerStartedAttempts: 0,
216
+ toolCalls: 0,
217
+ toolResults: 0,
218
+ toolFailures: 0,
219
+ blockedToolCalls: 0,
220
+ declinedToolCalls: 0,
221
+ explicitRetries: 0,
222
+ retryFailures: 0,
223
+ compactions: 0,
224
+ abortedCompactions: 0,
225
+ summarizationRetries: 0,
226
+ repeatedToolCalls: 0,
227
+ subagentAttempts: 0,
228
+ subagentFailures: 0
229
+ };
230
+ }
231
+
232
+ function stableValue(value) {
233
+ if (Array.isArray(value)) return value.map(stableValue);
234
+ if (!value || typeof value !== "object") return value;
235
+ return Object.fromEntries(Object.keys(value).sort().map((key) => [key, stableValue(value[key])]));
236
+ }
237
+
238
+ function toolFingerprint(block) {
239
+ let args = block?.arguments ?? null;
240
+ if (typeof args === "string") {
241
+ try { args = JSON.parse(args); } catch { /* The opaque string remains useful for equality. */ }
242
+ }
243
+ try { return JSON.stringify([String(block?.name ?? "tool"), stableValue(args)]); }
244
+ catch { return String(block?.name ?? "tool"); }
245
+ }
246
+
247
+ function isSubagentAttempt(block) {
248
+ const name = String(block?.name ?? "").toLowerCase();
249
+ if (["spawn_agent", "create_agent", "delegate_task", "subagent"].includes(name)) return true;
250
+ return false;
251
+ }
252
+
253
+ function structuredToolStatus(message) {
254
+ const candidates = [message?.status, message?.details?.status, message?.result?.status];
255
+ return candidates.find((value) => typeof value === "string")?.toLowerCase() ?? "";
256
+ }
257
+
258
+ function exactPricingRequest(usage) {
259
+ const fields = ["input", "output", "cacheRead", "cacheWrite"];
260
+ if (!usage || fields.some((field) => !Number.isSafeInteger(usage[field]) || usage[field] < 0)) return null;
261
+ const reasoningReported = Object.hasOwn(usage, "reasoning") && usage.reasoning !== undefined;
262
+ const reasoning = reasoningReported ? usage.reasoning : 0;
263
+ if (!Number.isSafeInteger(reasoning) || reasoning < 0) return null;
264
+ const total = usage.totalTokens;
265
+ if (!Number.isSafeInteger(total) || total !== usage.input + usage.output + usage.cacheRead + usage.cacheWrite
266
+ || reasoning > usage.output) return null;
267
+ return {
268
+ input: usage.input,
269
+ output: usage.output,
270
+ cacheRead: usage.cacheRead,
271
+ cacheWrite: usage.cacheWrite,
272
+ reasoning,
273
+ total,
274
+ promptTokens: usage.input + usage.cacheRead + usage.cacheWrite,
275
+ ...(!reasoningReported && usage.output > 0 ? { reasoningCompleteness: "lower-bound" } : {})
276
+ };
277
+ }
278
+
279
+ function exactToken(raw, label) {
280
+ if (!Number.isSafeInteger(raw) || raw < 0) throw new Error(`${label} must be a non-negative safe integer`);
281
+ return raw;
282
+ }
283
+
284
+ function addTotals(target, usage, { strict = false, label = "usage" } = {}) {
285
+ const reasoningReported = Boolean(usage && Object.hasOwn(usage, "reasoning") && usage.reasoning !== undefined);
286
+ const values = strict ? (() => {
287
+ const output = exactToken(usage?.output, `${label}.output`);
288
+ return {
289
+ input: exactToken(usage?.input, `${label}.input`),
290
+ output,
291
+ cacheRead: exactToken(usage?.cacheRead, `${label}.cacheRead`),
292
+ cacheWrite: exactToken(usage?.cacheWrite, `${label}.cacheWrite`),
293
+ // Pi's upstream usage shape makes reasoning optional. Its absence does
294
+ // not make the additive input/output totals unknown: reasoning is a
295
+ // subset of output. A zero-output response proves zero reasoning; for a
296
+ // non-zero output we retain zero only as an explicit lower bound.
297
+ reasoning: reasoningReported ? exactToken(usage.reasoning, `${label}.reasoning`) : 0,
298
+ total: exactToken(usage?.totalTokens, `${label}.totalTokens`)
299
+ };
300
+ })() : {
301
+ input: numberValue(usage.input), output: numberValue(usage.output), cacheRead: numberValue(usage.cacheRead),
302
+ cacheWrite: numberValue(usage.cacheWrite), reasoning: numberValue(usage.reasoning),
303
+ total: numberValue(usage.totalTokens) || numberValue(usage.input) + numberValue(usage.output) + numberValue(usage.cacheRead) + numberValue(usage.cacheWrite)
304
+ };
305
+ if (strict && values.total !== values.input + values.output + values.cacheRead + values.cacheWrite) {
306
+ throw new Error(`${label}.totalTokens does not equal the reported token categories`);
307
+ }
308
+ if (strict && values.reasoning > values.output) throw new Error(`${label}.reasoning exceeds output`);
309
+ for (const field of ["input", "output", "cacheRead", "cacheWrite", "reasoning", "total"]) target[field] += values[field];
310
+ const cost = usage.cost;
311
+ target.cost += typeof cost === "object" && cost !== null
312
+ ? numberValue(cost.total)
313
+ : numberValue(cost);
314
+ return {
315
+ reasoningReported,
316
+ reasoningCompleteness: reasoningReported || values.output === 0 ? "exact" : "lower-bound",
317
+ reasoningInferredZero: !reasoningReported && values.output === 0
318
+ };
319
+ }
320
+
321
+ function numberValue(raw) {
322
+ const n = Number(raw ?? 0);
323
+ return Number.isFinite(n) ? n : 0;
324
+ }
325
+
326
+ function extractTextLength(content) {
327
+ if (typeof content === "string") return content.length;
328
+ if (!Array.isArray(content)) return 0;
329
+ return content.reduce((sum, block) => {
330
+ if (block?.type === "text" && typeof block.text === "string") return sum + block.text.length;
331
+ return sum;
332
+ }, 0);
333
+ }
334
+
335
+ function summarizeSession(file, options) {
336
+ let stat;
337
+ try {
338
+ stat = fs.statSync(file);
339
+ } catch {
340
+ return undefined;
341
+ }
342
+
343
+ const execution = emptyExecutionAggregate();
344
+ const summary = {
345
+ id: path.basename(file, ".jsonl"),
346
+ name: "",
347
+ cwd: "",
348
+ file,
349
+ isSubagent: file.split(path.sep).includes("subagent"),
350
+ provider: "",
351
+ modelId: "",
352
+ thinkingLevel: "",
353
+ firstTimestamp: undefined,
354
+ lastTimestamp: undefined,
355
+ countedFirstTimestamp: undefined,
356
+ countedLastTimestamp: undefined,
357
+ messages: { user: 0, assistant: 0, toolCalls: 0, toolResults: 0, total: 0 },
358
+ promptChars: 0,
359
+ tokens: emptyTotals(),
360
+ usageIntegrity: {
361
+ exact: options.strictUsage === true,
362
+ source: "pi-session-jsonl",
363
+ assistantMessages: 0,
364
+ usageMessages: 0,
365
+ reasoningMessages: 0,
366
+ reasoningInferredZeroMessages: 0,
367
+ reasoningLowerBoundMessages: 0,
368
+ reasoningCompleteness: "exact"
369
+ },
370
+ pricingBuckets: { schemaVersion: 1, source: "provider-request-usage", completeness: "unverified", requests: [] },
371
+ execution,
372
+ toolNames: {},
373
+ sizeBytes: stat.size,
374
+ mtime: stat.mtime
375
+ };
376
+
377
+ let hasCountedActivity = false;
378
+ let parsedRecordCount = 0;
379
+ let sessionHeaderCount = 0;
380
+ const toolCallsById = new Map();
381
+ const toolResultIds = new Set();
382
+ const toolCallFingerprints = new Set();
383
+ const lines = fs.readFileSync(file, "utf8").split(/\n/);
384
+ for (const [lineIndex, line] of lines.entries()) {
385
+ if (!line) continue;
386
+ let entry;
387
+ try {
388
+ entry = JSON.parse(line);
389
+ } catch {
390
+ if (options.strictUsage) throw new Error(`Pi session JSONL line ${lineIndex + 1} is not valid JSON`);
391
+ continue;
392
+ }
393
+
394
+ const isFirstParsedRecord = parsedRecordCount === 0;
395
+ parsedRecordCount += 1;
396
+ if (options.strictUsage && isFirstParsedRecord && entry?.type !== "session") {
397
+ throw new Error("Pi session JSONL is missing its session header as the first record");
398
+ }
399
+
400
+ const ts = parseIso(entry.timestamp ?? entry.message?.timestamp);
401
+ if (ts) {
402
+ if (!summary.firstTimestamp || ts < summary.firstTimestamp) summary.firstTimestamp = ts;
403
+ if (!summary.lastTimestamp || ts > summary.lastTimestamp) summary.lastTimestamp = ts;
404
+ }
405
+
406
+ if (entry.type === "session") {
407
+ sessionHeaderCount += 1;
408
+ if (options.strictUsage && (sessionHeaderCount !== 1 || !isFirstParsedRecord)) {
409
+ throw new Error("Pi session JSONL must contain exactly one session header as the first record");
410
+ }
411
+ if (options.strictUsage && (typeof entry.id !== "string" || !entry.id.trim()
412
+ || typeof entry.cwd !== "string" || !entry.cwd.trim())) {
413
+ throw new Error("Pi session JSONL session header must contain non-empty id and cwd strings");
414
+ }
415
+ if (entry.id) summary.id = String(entry.id);
416
+ if (entry.cwd) summary.cwd = String(entry.cwd);
417
+ continue;
418
+ }
419
+ if (entry.type === "model_change") {
420
+ summary.provider = String(entry.provider ?? summary.provider);
421
+ summary.modelId = String(entry.modelId ?? summary.modelId);
422
+ continue;
423
+ }
424
+ if (entry.type === "thinking_level_change") {
425
+ summary.thinkingLevel = String(entry.thinkingLevel ?? summary.thinkingLevel);
426
+ continue;
427
+ }
428
+ if (entry.type === "session_info" && entry.name) {
429
+ summary.name = String(entry.name);
430
+ continue;
431
+ }
432
+
433
+ const countedControl = inRange(ts ?? stat.mtime, options);
434
+ if (countedControl && entry.type === "auto_retry_start") execution.explicitRetries += 1;
435
+ else if (countedControl && entry.type === "auto_retry_end" && entry.success === false) execution.retryFailures += 1;
436
+ else if (countedControl && entry.type === "compaction_start") execution.compactions += 1;
437
+ else if (countedControl && entry.type === "compaction_end" && entry.aborted === true) execution.abortedCompactions += 1;
438
+ else if (countedControl && entry.type === "summarization_retry_scheduled") execution.summarizationRetries += 1;
439
+ if (entry.type !== "message" || !entry.message) continue;
440
+
441
+ const counted = inRange(ts ?? stat.mtime, options);
442
+ if (!counted) continue;
443
+ hasCountedActivity = true;
444
+ const countedTs = ts ?? stat.mtime;
445
+ if (!summary.countedFirstTimestamp || countedTs < summary.countedFirstTimestamp) summary.countedFirstTimestamp = countedTs;
446
+ if (!summary.countedLastTimestamp || countedTs > summary.countedLastTimestamp) summary.countedLastTimestamp = countedTs;
447
+
448
+ const message = entry.message;
449
+ const role = message.role;
450
+ summary.messages.total += 1;
451
+ if (role === "user") {
452
+ // A repeated call means duplicate work inside one operator turn. The
453
+ // same verifier or read requested again by a later user turn is fresh
454
+ // work, not a retry loop, even when the session itself is preserved.
455
+ toolCallFingerprints.clear();
456
+ summary.messages.user += 1;
457
+ summary.promptChars += extractTextLength(message.content);
458
+ } else if (role === "assistant") {
459
+ summary.messages.assistant += 1;
460
+ summary.usageIntegrity.assistantMessages += 1;
461
+ if (message.provider && !summary.provider) summary.provider = String(message.provider);
462
+ if (message.model && !summary.modelId) summary.modelId = String(message.model);
463
+ if (options.strictUsage && (!message.usage || typeof message.usage !== "object" || Array.isArray(message.usage))) {
464
+ throw new Error(`Pi assistant message at line ${lineIndex + 1} is missing usage`);
465
+ }
466
+ if (message.usage) {
467
+ const usageReceipt = addTotals(summary.tokens, message.usage, {
468
+ strict: options.strictUsage === true,
469
+ label: `Pi session JSONL line ${lineIndex + 1} usage`
470
+ });
471
+ summary.usageIntegrity.usageMessages += 1;
472
+ if (usageReceipt.reasoningReported) summary.usageIntegrity.reasoningMessages += 1;
473
+ if (usageReceipt.reasoningInferredZero) summary.usageIntegrity.reasoningInferredZeroMessages += 1;
474
+ if (usageReceipt.reasoningCompleteness === "lower-bound") {
475
+ summary.usageIntegrity.reasoningLowerBoundMessages += 1;
476
+ summary.usageIntegrity.reasoningCompleteness = "lower-bound";
477
+ }
478
+ execution.providerStartedAttempts += 1;
479
+ const request = exactPricingRequest(message.usage);
480
+ if (request) summary.pricingBuckets.requests.push(request);
481
+ }
482
+ if (Array.isArray(message.content)) {
483
+ for (const block of message.content) {
484
+ if (block?.type !== "toolCall") continue;
485
+ summary.messages.toolCalls += 1;
486
+ execution.toolCalls += 1;
487
+ const name = String(block.name ?? "tool");
488
+ summary.toolNames[name] = (summary.toolNames[name] ?? 0) + 1;
489
+ const id = typeof block.id === "string" && block.id ? block.id : null;
490
+ const subagentAttempt = isSubagentAttempt(block);
491
+ if (id) toolCallsById.set(id, { name, subagentAttempt });
492
+ if (subagentAttempt) execution.subagentAttempts += 1;
493
+ const fingerprint = toolFingerprint(block);
494
+ if (toolCallFingerprints.has(fingerprint)) execution.repeatedToolCalls += 1;
495
+ else toolCallFingerprints.add(fingerprint);
496
+ }
497
+ }
498
+ } else if (role === "toolResult") {
499
+ summary.messages.toolResults += 1;
500
+ execution.toolResults += 1;
501
+ const name = String(message.toolName ?? "tool");
502
+ summary.toolNames[name] = summary.toolNames[name] ?? 0;
503
+ const status = structuredToolStatus(message);
504
+ const failed = message.isError === true || status === "failed" || status === "error";
505
+ if (failed) execution.toolFailures += 1;
506
+ if (status === "blocked") execution.blockedToolCalls += 1;
507
+ if (status === "declined") execution.declinedToolCalls += 1;
508
+ const call = typeof message.toolCallId === "string" ? toolCallsById.get(message.toolCallId) : undefined;
509
+ if (typeof message.toolCallId === "string") toolResultIds.add(message.toolCallId);
510
+ if (call?.subagentAttempt && (failed || status === "blocked" || status === "declined")) execution.subagentFailures += 1;
511
+ const compacted = message.details?.piagentCompactedToolResults;
512
+ if (Number.isSafeInteger(compacted) && compacted > 0) execution.compactions += compacted;
513
+ }
514
+ }
515
+
516
+ if (!summary.cwd) summary.cwd = "(unknown)";
517
+ if (!summary.firstTimestamp) summary.firstTimestamp = stat.mtime;
518
+ if (!summary.lastTimestamp) summary.lastTimestamp = stat.mtime;
519
+ if (!summary.countedFirstTimestamp) summary.countedFirstTimestamp = summary.firstTimestamp;
520
+ if (!summary.countedLastTimestamp) summary.countedLastTimestamp = summary.lastTimestamp;
521
+
522
+ if (!hasCountedActivity && (options.since || options.until)) return undefined;
523
+ if (options.strictUsage && sessionHeaderCount !== 1) {
524
+ throw new Error("Pi session JSONL is missing its session header");
525
+ }
526
+ if (options.strictUsage && summary.usageIntegrity.assistantMessages !== summary.usageIntegrity.usageMessages) {
527
+ throw new Error("Pi session JSONL usage coverage is incomplete");
528
+ }
529
+ summary.pricingBuckets.completeness = summary.usageIntegrity.assistantMessages > 0
530
+ && summary.pricingBuckets.requests.length === summary.usageIntegrity.assistantMessages
531
+ ? "exact"
532
+ : "unverified";
533
+ const identifiedToolResults = [...toolCallsById.keys()].filter((id) => toolResultIds.has(id)).length;
534
+ execution.completeness.tools = execution.toolCalls === execution.toolResults && identifiedToolResults === execution.toolCalls
535
+ ? "exact"
536
+ : "partial";
537
+ execution.completeness.subagents = execution.completeness.tools;
538
+ return summary;
539
+ }
540
+
541
+ function loadReport(options) {
542
+ const files = walkJsonl(options.sessionsDir);
543
+ const sessions = [];
544
+ for (const file of files) {
545
+ const isSubagent = file.split(path.sep).includes("subagent");
546
+ if (isSubagent && !options.includeSubagents) continue;
547
+ const session = summarizeSession(file, options);
548
+ if (!session) continue;
549
+ if (!options.allProjects && options.projectPath && session.cwd !== options.projectPath) continue;
550
+ sessions.push(session);
551
+ }
552
+ sessions.sort((a, b) => b.countedLastTimestamp - a.countedLastTimestamp);
553
+ return buildReport(options, sessions);
554
+ }
555
+
556
+ function buildReport(options, sessions) {
557
+ const totals = {
558
+ sessions: sessions.length,
559
+ mainSessions: sessions.filter((s) => !s.isSubagent).length,
560
+ subagentSessions: sessions.filter((s) => s.isSubagent).length,
561
+ messages: { user: 0, assistant: 0, toolCalls: 0, toolResults: 0, total: 0 },
562
+ promptChars: 0,
563
+ tokens: emptyTotals(),
564
+ execution: emptyExecutionAggregate(),
565
+ pricingBuckets: { schemaVersion: 1, source: "provider-request-usage", completeness: "exact", requests: [] }
566
+ };
567
+ const projects = new Map();
568
+ const tools = {};
569
+ for (const session of sessions) {
570
+ for (const key of Object.keys(totals.messages)) totals.messages[key] += session.messages[key];
571
+ totals.promptChars += session.promptChars;
572
+ for (const key of Object.keys(totals.tokens)) totals.tokens[key] += session.tokens[key];
573
+ for (const key of ["providerStartedAttempts", "toolCalls", "toolResults", "toolFailures", "blockedToolCalls",
574
+ "declinedToolCalls", "explicitRetries", "retryFailures", "compactions", "abortedCompactions",
575
+ "summarizationRetries", "repeatedToolCalls", "subagentAttempts", "subagentFailures"]) {
576
+ totals.execution[key] += Number(session.execution?.[key] ?? 0);
577
+ }
578
+ totals.pricingBuckets.requests.push(...(session.pricingBuckets?.requests ?? []));
579
+ const project = projects.get(session.cwd) ?? {
580
+ cwd: session.cwd,
581
+ sessions: 0,
582
+ mainSessions: 0,
583
+ subagentSessions: 0,
584
+ tokens: emptyTotals(),
585
+ cost: 0
586
+ };
587
+ project.sessions += 1;
588
+ if (session.isSubagent) project.subagentSessions += 1;
589
+ else project.mainSessions += 1;
590
+ for (const key of Object.keys(project.tokens)) project.tokens[key] += session.tokens[key];
591
+ project.cost = project.tokens.cost;
592
+ projects.set(session.cwd, project);
593
+ for (const [name, count] of Object.entries(session.toolNames)) {
594
+ tools[name] = (tools[name] ?? 0) + count;
595
+ }
596
+ }
597
+
598
+ const projectRows = [...projects.values()].sort((a, b) => b.tokens.total - a.tokens.total);
599
+ const toolRows = Object.entries(tools)
600
+ .map(([name, count]) => ({ name, count }))
601
+ .sort((a, b) => b.count - a.count || a.name.localeCompare(b.name));
602
+
603
+ const completenessRank = { exact: 0, partial: 1, unverified: 2, unavailable: 3 };
604
+ for (const field of ["tools", "retries", "compactions", "subagents"]) {
605
+ totals.execution.completeness[field] = sessions.reduce((worst, session) => {
606
+ const value = session.execution?.completeness?.[field] ?? "unavailable";
607
+ return completenessRank[value] > completenessRank[worst] ? value : worst;
608
+ }, "exact");
609
+ }
610
+ totals.pricingBuckets.completeness = sessions.length > 0
611
+ && sessions.every((session) => session.pricingBuckets?.completeness === "exact") ? "exact" : "unverified";
612
+
613
+ return {
614
+ generatedAt: new Date().toISOString(),
615
+ scope: {
616
+ projectPath: options.allProjects ? null : options.projectPath,
617
+ allProjects: options.allProjects,
618
+ includeSubagents: options.includeSubagents,
619
+ sessionsDir: options.sessionsDir,
620
+ since: options.since?.toISOString() ?? null,
621
+ until: options.until?.toISOString() ?? null
622
+ },
623
+ totals,
624
+ projects: projectRows,
625
+ tools: toolRows,
626
+ sessions
627
+ };
628
+ }
629
+
630
+ function compact(n) {
631
+ const abs = Math.abs(n);
632
+ if (abs >= 1_000_000_000) return `${(n / 1_000_000_000).toFixed(1)}B`;
633
+ if (abs >= 1_000_000) return `${(n / 1_000_000).toFixed(1)}M`;
634
+ if (abs >= 1_000) return `${(n / 1_000).toFixed(1)}k`;
635
+ return String(Math.round(n));
636
+ }
637
+
638
+ function money(n) {
639
+ return `$${n.toFixed(6)}`;
640
+ }
641
+
642
+ function iso(date) {
643
+ return date ? date.toISOString() : "";
644
+ }
645
+
646
+ function localLabel(date) {
647
+ if (!date) return "?";
648
+ const pad = (n) => String(n).padStart(2, "0");
649
+ return `${date.getFullYear()}-${pad(date.getMonth() + 1)}-${pad(date.getDate())} ${pad(date.getHours())}:${pad(date.getMinutes())}`;
650
+ }
651
+
652
+ function rangeLabel(report) {
653
+ const { since, until } = report.scope;
654
+ if (!since && !until) return "all time";
655
+ return `${since ? since.slice(0, 10) : "beginning"} -> ${until ? until.slice(0, 10) : "now"}`;
656
+ }
657
+
658
+ function renderTable(report, limit) {
659
+ const lines = [];
660
+ lines.push("PiAgent usage history");
661
+ lines.push(`Scope: ${report.scope.allProjects ? "all projects" : report.scope.projectPath}`);
662
+ lines.push(`Range: ${rangeLabel(report)}`);
663
+ lines.push(`Sessions: ${report.totals.sessions} (${report.totals.mainSessions} main, ${report.totals.subagentSessions} subagent)`);
664
+ lines.push(`Tokens: ${compact(report.totals.tokens.total)} total (${compact(report.totals.tokens.input)} input, ${compact(report.totals.tokens.output)} output, ${compact(report.totals.tokens.cacheRead)} cache read, ${compact(report.totals.tokens.cacheWrite)} cache write, ${compact(report.totals.tokens.reasoning)} reasoning)`);
665
+ lines.push(`Cost: ${money(report.totals.tokens.cost)}`);
666
+ lines.push(`Messages: ${report.totals.messages.user} user, ${report.totals.messages.assistant} assistant, ${report.totals.messages.toolCalls} tool calls, ${report.totals.messages.toolResults} tool results`);
667
+ if (report.projects.length) {
668
+ lines.push("");
669
+ lines.push("Projects:");
670
+ for (const project of report.projects.slice(0, 10)) {
671
+ lines.push(` ${compact(project.tokens.total).padStart(8)} tok ${money(project.tokens.cost).padStart(11)} ${String(project.sessions).padStart(3)} sessions ${project.cwd}`);
672
+ }
673
+ }
674
+ if (report.sessions.length) {
675
+ lines.push("");
676
+ lines.push(`Top sessions (${Math.min(limit, report.sessions.length)} of ${report.sessions.length}):`);
677
+ for (const session of report.sessions.slice(0, limit)) {
678
+ const kind = session.isSubagent ? "sub" : "main";
679
+ const model = session.modelId || session.provider || "?";
680
+ lines.push(` ${localLabel(session.countedLastTimestamp)} ${kind.padEnd(4)} ${compact(session.tokens.total).padStart(8)} tok ${money(session.tokens.cost).padStart(11)} ${String(session.messages.user).padStart(2)}p/${String(session.messages.toolCalls).padStart(2)}t ${model} ${session.name || session.id}`);
681
+ }
682
+ }
683
+ return `${lines.join("\n")}\n`;
684
+ }
685
+
686
+ function renderMarkdown(report, limit) {
687
+ const lines = [];
688
+ lines.push("# PiAgent Usage History");
689
+ lines.push("");
690
+ lines.push(`- Scope: ${report.scope.allProjects ? "all projects" : report.scope.projectPath}`);
691
+ lines.push(`- Range: ${rangeLabel(report)}`);
692
+ lines.push(`- Sessions: ${report.totals.sessions} (${report.totals.mainSessions} main, ${report.totals.subagentSessions} subagent)`);
693
+ lines.push(`- Tokens: ${report.totals.tokens.total.toLocaleString("en-US")}`);
694
+ lines.push(`- Cost: ${money(report.totals.tokens.cost)}`);
695
+ lines.push("");
696
+ lines.push("## Projects");
697
+ lines.push("");
698
+ lines.push("| Project | Sessions | Tokens | Cost |");
699
+ lines.push("|---|---:|---:|---:|");
700
+ for (const project of report.projects.slice(0, 20)) {
701
+ lines.push(`| ${escapeMarkdown(project.cwd)} | ${project.sessions} | ${project.tokens.total} | ${money(project.tokens.cost)} |`);
702
+ }
703
+ lines.push("");
704
+ lines.push("## Sessions");
705
+ lines.push("");
706
+ lines.push("| Last activity | Kind | Session | Model | Prompts | Tools | Tokens | Cost |");
707
+ lines.push("|---|---|---|---|---:|---:|---:|---:|");
708
+ for (const session of report.sessions.slice(0, limit)) {
709
+ lines.push(`| ${localLabel(session.countedLastTimestamp)} | ${session.isSubagent ? "subagent" : "main"} | ${escapeMarkdown(session.name || session.id)} | ${escapeMarkdown(session.modelId || session.provider || "?")} | ${session.messages.user} | ${session.messages.toolCalls} | ${session.tokens.total} | ${money(session.tokens.cost)} |`);
710
+ }
711
+ return `${lines.join("\n")}\n`;
712
+ }
713
+
714
+ function escapeMarkdown(raw) {
715
+ return String(raw).replaceAll("|", "\\|");
716
+ }
717
+
718
+ function csvEscape(raw) {
719
+ const s = String(raw ?? "");
720
+ if (!/[",\n]/.test(s)) return s;
721
+ return `"${s.replaceAll("\"", "\"\"")}"`;
722
+ }
723
+
724
+ function renderCsv(report) {
725
+ const columns = [
726
+ "session_id", "name", "cwd", "kind", "first_timestamp", "last_timestamp",
727
+ "provider", "model", "thinking_level", "user_messages", "assistant_messages",
728
+ "tool_calls", "tool_results", "input_tokens", "output_tokens", "cache_read_tokens",
729
+ "cache_write_tokens", "reasoning_tokens", "total_tokens", "cost_usd", "file"
730
+ ];
731
+ const rows = [columns.join(",")];
732
+ for (const s of report.sessions) {
733
+ rows.push([
734
+ s.id, s.name, s.cwd, s.isSubagent ? "subagent" : "main", iso(s.firstTimestamp), iso(s.lastTimestamp),
735
+ s.provider, s.modelId, s.thinkingLevel, s.messages.user, s.messages.assistant,
736
+ s.messages.toolCalls, s.messages.toolResults, s.tokens.input, s.tokens.output, s.tokens.cacheRead,
737
+ s.tokens.cacheWrite, s.tokens.reasoning, s.tokens.total, s.tokens.cost.toFixed(6), s.file
738
+ ].map(csvEscape).join(","));
739
+ }
740
+ return `${rows.join("\n")}\n`;
741
+ }
742
+
743
+ function toJson(report) {
744
+ return JSON.stringify(report, (_key, value) => value instanceof Date ? value.toISOString() : value, 2) + "\n";
745
+ }
746
+
747
+ function render(report, options) {
748
+ switch (options.format) {
749
+ case "json":
750
+ return toJson(report);
751
+ case "csv":
752
+ return renderCsv(report);
753
+ case "markdown":
754
+ return renderMarkdown(report, options.limit);
755
+ default:
756
+ return renderTable(report, options.limit);
757
+ }
758
+ }
759
+
760
+ export {
761
+ buildReport,
762
+ loadReport,
763
+ parseArgs,
764
+ render,
765
+ summarizeSession,
766
+ walkJsonl
767
+ };
768
+
769
+ async function main() {
770
+ const options = parseArgs(process.argv.slice(2));
771
+ if (options.help) {
772
+ process.stdout.write(USAGE);
773
+ return;
774
+ }
775
+ const report = loadReport(options);
776
+ process.stdout.write(render(report, options));
777
+ }
778
+
779
+ if (import.meta.url === pathToFileURL(process.argv[1] || "").href) {
780
+ main().catch((error) => {
781
+ console.error(`FAIL: ${error?.message ?? error}`);
782
+ process.exit(1);
783
+ });
784
+ }