@piagent/platform 1.6.1 → 1.8.0

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 (654) hide show
  1. package/CHANGELOG.md +71 -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/model-options.md +20 -15
  175. package/docs/operator-manual-vietnamese.md +12 -12
  176. package/docs/piagent-codex-benchmark-recovery-plan.md +1817 -0
  177. package/docs/piagent-core-evaluation-specs-2026-09-06.json +706 -0
  178. package/docs/piagent-core-improvement-plan.md +528 -0
  179. package/docs/piagent-core-offline-implementation-2026-09-06.md +219 -0
  180. package/docs/piagent-core-plan-checkpoint-2026-09-06.json +69 -0
  181. package/docs/piagent-core-reasoning-research-2026-09-06.md +225 -0
  182. package/docs/piagent-core-research-sources-2026-09-06.json +416 -0
  183. package/docs/piagent-open-session-wait.md +19 -0
  184. package/docs/piagent-runtime-benchmark-execution-plan.md +685 -0
  185. package/docs/publishing-for-teams.md +3 -0
  186. package/docs/quality-benchmark.md +128 -8
  187. package/docs/quickstart-vietnamese.md +2 -2
  188. package/docs/release-install-policy.md +38 -11
  189. package/docs/research/harness-next-evidence-review.md +113 -0
  190. package/docs/team-onboarding.md +7 -7
  191. package/docs/third-party-neutrality.md +7 -0
  192. package/docs/usage-observability.md +1 -1
  193. package/docs-site/architecture.html +1 -1
  194. package/docs-site/benchmark.html +1 -1
  195. package/docs-site/commands.html +1 -1
  196. package/docs-site/content/en/runtime.html +1 -1
  197. package/docs-site/content/en/whats-new.html +23 -0
  198. package/docs-site/content/runtime.html +2 -2
  199. package/docs-site/content/whats-new.html +23 -0
  200. package/docs-site/context-engine.html +1 -1
  201. package/docs-site/en/architecture.html +1 -1
  202. package/docs-site/en/benchmark.html +1 -1
  203. package/docs-site/en/commands.html +1 -1
  204. package/docs-site/en/context-engine.html +1 -1
  205. package/docs-site/en/index.html +1 -1
  206. package/docs-site/en/mcp-approval.html +1 -1
  207. package/docs-site/en/mcp-auth.html +1 -1
  208. package/docs-site/en/mcp-servers.html +1 -1
  209. package/docs-site/en/mcp.html +1 -1
  210. package/docs-site/en/permissions.html +1 -1
  211. package/docs-site/en/profiles.html +1 -1
  212. package/docs-site/en/quickstart.html +1 -1
  213. package/docs-site/en/runtime.html +2 -2
  214. package/docs-site/en/scope.html +1 -1
  215. package/docs-site/en/security.html +1 -1
  216. package/docs-site/en/team.html +1 -1
  217. package/docs-site/en/whats-new.html +24 -1
  218. package/docs-site/en/workflows.html +1 -1
  219. package/docs-site/index.html +1 -1
  220. package/docs-site/mcp-approval.html +1 -1
  221. package/docs-site/mcp-auth.html +1 -1
  222. package/docs-site/mcp-servers.html +1 -1
  223. package/docs-site/mcp.html +1 -1
  224. package/docs-site/permissions.html +1 -1
  225. package/docs-site/profiles.html +1 -1
  226. package/docs-site/quickstart.html +1 -1
  227. package/docs-site/runtime.html +3 -3
  228. package/docs-site/scope.html +1 -1
  229. package/docs-site/security.html +1 -1
  230. package/docs-site/team.html +1 -1
  231. package/docs-site/whats-new.html +24 -1
  232. package/docs-site/workflows.html +1 -1
  233. package/evals/architecture-conformance-v1/lane.json +18 -0
  234. package/evals/architecture-conformance-v1/runner.mjs +77 -0
  235. package/evals/fixtures/approved-host-contracts.invalid.json +43 -0
  236. package/evals/fixtures/approved-host-contracts.valid.json +43 -0
  237. package/evals/fixtures/benchmark-attempt-outcome.invalid.json +42 -0
  238. package/evals/fixtures/benchmark-attempt-outcome.valid.json +42 -0
  239. package/evals/fixtures/benchmark-independent-verification-plan.invalid.json +7 -0
  240. package/evals/fixtures/benchmark-independent-verification-plan.valid.json +21 -0
  241. package/evals/fixtures/composite-criterion.invalid.json +10 -0
  242. package/evals/fixtures/composite-criterion.valid.json +33 -0
  243. package/evals/fixtures/contract-family-library.invalid.json +42 -0
  244. package/evals/fixtures/contract-family-library.valid.json +41 -0
  245. package/evals/fixtures/contract-selection-recipe.invalid.json +29 -0
  246. package/evals/fixtures/contract-selection-recipe.valid.json +29 -0
  247. package/evals/fixtures/host-contract-plan.invalid.json +42 -0
  248. package/evals/fixtures/host-contract-plan.valid.json +41 -0
  249. package/evals/fixtures/scoped-project-verification-receipt-v2.invalid.json +42 -0
  250. package/evals/fixtures/scoped-project-verification-receipt-v2.valid.json +42 -0
  251. package/evals/fixtures/scoped-project-verification-receipt-v3.invalid.json +57 -0
  252. package/evals/fixtures/scoped-project-verification-receipt-v3.valid.json +57 -0
  253. package/evals/fixtures/scoped-verification-receipt-v1.invalid.json +40 -0
  254. package/evals/fixtures/scoped-verification-receipt-v1.valid.json +40 -0
  255. package/evals/fs5-artifact-archive/README.md +32 -0
  256. package/evals/fs5-artifact-archive/manifest.json +148 -0
  257. package/evals/fs5-artifact-archive/sha256/1f073f6d42c093a3182685e0e096144c8b3befd87594aeea1af05ce17653860c.txt +380 -0
  258. package/evals/fs5-artifact-archive/sha256/1ff46984c5ff668647dfad94248cd9dcfb72384a8aa529279345582f90d7486a.txt +331 -0
  259. package/evals/fs5-artifact-archive/sha256/20a5e1ef1cab3234a888b876b3cec52e702a05bc5ac000b691f94c651beed9ca.txt +800 -0
  260. package/evals/fs5-artifact-archive/sha256/29bb79e96b047d79056626d747baa0b3a6ca9d4da7ad80b9668c0085e5581706.txt +116 -0
  261. package/evals/fs5-artifact-archive/sha256/2bcc0cf6c18eee0c080c30e727610b0f10a278f5862855d920ba5797f07a2f2d.txt +252 -0
  262. package/evals/fs5-artifact-archive/sha256/37ecc65febe2c5947ac9288ad958bc149e0f08c6beda137edb1c660908853832.txt +324 -0
  263. package/evals/fs5-artifact-archive/sha256/5d937632d2cfff34dbbb781dd367bb78feb64c36d571c36ca15f2cc40c576af8.txt +142 -0
  264. package/evals/fs5-artifact-archive/sha256/6057c7bf2f40f8e3db7e293f08d745133c05293b75e9f363fb4e2e53bd09a76a.txt +244 -0
  265. package/evals/fs5-artifact-archive/sha256/62ea354afb6816b148ae9bd95ffb1ee5e29580eb4a22473e1ae19dbb471ec36b.txt +930 -0
  266. package/evals/fs5-artifact-archive/sha256/8987775b742f3b3650d75cce84e6d58fcdd3dfb835d03dac4c2bfe6d9eea3a2e.txt +903 -0
  267. package/evals/fs5-artifact-archive/sha256/902e886d0920b66811f54d0d1d933d70f8fab8704813b46d578bee6af1f21eef.txt +942 -0
  268. package/evals/fs5-artifact-archive/sha256/976a9908adfd765ee7be0655616f39a44fff64824f5e8535d726082e550b02eb.txt +910 -0
  269. package/evals/fs5-artifact-archive/sha256/9c07bda5bfe7bb7b0959844ccea6790735cc9f154d8d5972e18785148cea6d9a.txt +101 -0
  270. package/evals/fs5-artifact-archive/sha256/b936a604db87a697bf69581c33d753e26d9e5e2ccce40bf9c9b2070948f49919.txt +931 -0
  271. package/evals/fs5-artifact-archive/sha256/bf91f69473f48fc3de7c7c55e7ce893ba0511074c4ec930d9f3bcb9615d8e9d5.txt +950 -0
  272. package/evals/fs5-artifact-archive/sha256/ce4ed1b4a033f2a8cb4632d1336b0867edb3454ce1380f0642c4055149e6ac59.txt +784 -0
  273. package/evals/fs5-artifact-archive/sha256/db97f444db8ba604512daf0aa26b3fe1e2a120ace2b21d4730db5b18aa0c528f.txt +950 -0
  274. package/evals/fs5-causal-arm.v1.json +9 -5
  275. package/evals/fs5-pilot-protocol.v1.json +1 -1
  276. package/evals/fs5-pilot-protocol.v4.json +4 -4
  277. package/evals/fs5-pilot-protocol.v5.json +6 -5
  278. package/evals/harness-next/development-corpus.mjs +106 -0
  279. package/evals/harness-next/resource-budget-probe.mjs +38 -0
  280. package/evals/harness-next/run-calibration.mjs +83 -0
  281. package/evals/harness-next/run-resource-probe.mjs +55 -0
  282. package/evals/harness-next/selected-development-corpus.mjs +27 -0
  283. package/evals/harness-next.v1.json +51 -0
  284. package/evals/long-horizon-v1/README.md +19 -0
  285. package/evals/long-horizon-v1/runner.mjs +116 -48
  286. package/evals/long-horizon-v1/worker.mjs +77 -30
  287. package/evals/private-holdout-v1/CURRENT_BOUNDARY.md +53 -0
  288. package/evals/private-holdout-v1/README.md +5 -2
  289. package/evals/private-holdout-v1/public-exposure.v2.json +273 -0
  290. package/package.json +9 -4
  291. package/packages/piagent-core/README.md +2 -2
  292. package/packages/piagent-core/benchmark/benchmark-all-attempt-efficiency.js +55 -0
  293. package/packages/piagent-core/benchmark/benchmark-artifact-binding.js +49 -0
  294. package/packages/piagent-core/benchmark/benchmark-bootstrap.js +265 -14
  295. package/packages/piagent-core/benchmark/benchmark-budget-governor.js +271 -0
  296. package/packages/piagent-core/benchmark/benchmark-campaign.js +592 -0
  297. package/packages/piagent-core/benchmark/benchmark-candidate.js +8 -3
  298. package/packages/piagent-core/benchmark/benchmark-checkpoint.js +67 -0
  299. package/packages/piagent-core/benchmark/benchmark-claim-restrictions.js +269 -2
  300. package/packages/piagent-core/benchmark/benchmark-cli.js +120 -3
  301. package/packages/piagent-core/benchmark/benchmark-codex-events.js +235 -0
  302. package/packages/piagent-core/benchmark/benchmark-codex-outcome.js +20 -0
  303. package/packages/piagent-core/benchmark/benchmark-codex-relative-efficiency.js +2 -1
  304. package/packages/piagent-core/benchmark/benchmark-codex-rollout.js +521 -0
  305. package/packages/piagent-core/benchmark/benchmark-codex.js +104 -8
  306. package/packages/piagent-core/benchmark/benchmark-comparison.js +284 -1
  307. package/packages/piagent-core/benchmark/benchmark-core.js +107 -107
  308. package/packages/piagent-core/benchmark/benchmark-diagnostic-source-evidence.js +40 -0
  309. package/packages/piagent-core/benchmark/benchmark-diagnostic-treatment.js +104 -0
  310. package/packages/piagent-core/benchmark/benchmark-evaluator-v3.js +226 -0
  311. package/packages/piagent-core/benchmark/benchmark-execution-guard.js +23 -1
  312. package/packages/piagent-core/benchmark/benchmark-forensics.js +63 -50
  313. package/packages/piagent-core/benchmark/benchmark-goal-assessment.js +193 -0
  314. package/packages/piagent-core/benchmark/benchmark-goal-report.js +87 -0
  315. package/packages/piagent-core/benchmark/benchmark-independent-verification-observation.js +64 -0
  316. package/packages/piagent-core/benchmark/benchmark-matrix-summary.js +141 -0
  317. package/packages/piagent-core/benchmark/benchmark-matrix.js +1 -0
  318. package/packages/piagent-core/benchmark/benchmark-normalized-cost.js +62 -0
  319. package/packages/piagent-core/benchmark/benchmark-pi-home.js +1 -0
  320. package/packages/piagent-core/benchmark/benchmark-preflight.js +69 -8
  321. package/packages/piagent-core/benchmark/benchmark-process.js +23 -9
  322. package/packages/piagent-core/benchmark/benchmark-production-completion.js +84 -0
  323. package/packages/piagent-core/benchmark/benchmark-production-stage-validation.js +209 -0
  324. package/packages/piagent-core/benchmark/benchmark-provider-free-evidence.js +162 -45
  325. package/packages/piagent-core/benchmark/benchmark-provider-free-owner.js +24 -0
  326. package/packages/piagent-core/benchmark/benchmark-provider-wire.js +378 -0
  327. package/packages/piagent-core/benchmark/benchmark-record-validation.d.ts +2 -0
  328. package/packages/piagent-core/benchmark/benchmark-record-validation.js +269 -3
  329. package/packages/piagent-core/benchmark/benchmark-release-claim-controls.js +132 -0
  330. package/packages/piagent-core/benchmark/benchmark-report.js +79 -4
  331. package/packages/piagent-core/benchmark/benchmark-resume-recovery.js +280 -10
  332. package/packages/piagent-core/benchmark/benchmark-runner-policy.js +88 -1
  333. package/packages/piagent-core/benchmark/benchmark-runtime-identity.js +50 -1
  334. package/packages/piagent-core/benchmark/benchmark-runtime.js +102 -10
  335. package/packages/piagent-core/benchmark/benchmark-safety-evidence.js +74 -0
  336. package/packages/piagent-core/benchmark/benchmark-service-tier.js +189 -0
  337. package/packages/piagent-core/benchmark/benchmark-stage-diagnostic.js +144 -128
  338. package/packages/piagent-core/benchmark/benchmark-suite-identity.js +264 -1
  339. package/packages/piagent-core/benchmark/benchmark-suite-runtime.js +4 -0
  340. package/packages/piagent-core/benchmark/benchmark-suite.js +211 -4
  341. package/packages/piagent-core/benchmark/benchmark-summary-support.js +30 -2
  342. package/packages/piagent-core/benchmark/benchmark-timing-contract.js +189 -0
  343. package/packages/piagent-core/benchmark/benchmark-timing-diagnostics.js +110 -139
  344. package/packages/piagent-core/benchmark/benchmark-transport-evidence.js +203 -0
  345. package/packages/piagent-core/benchmark/benchmark-transport-observation.js +147 -0
  346. package/packages/piagent-core/benchmark/benchmark-tree-identity.js +2 -1
  347. package/packages/piagent-core/benchmark/benchmark-usage.js +395 -33
  348. package/packages/piagent-core/benchmark/benchmark-workflow.js +136 -38
  349. package/packages/piagent-core/benchmark/fs5-pilot-protocol.js +21 -2
  350. package/packages/piagent-core/capabilities/runtime-integrity.js +3 -0
  351. package/packages/piagent-core/extensions/acceptance-api-baseline.js +24 -0
  352. package/packages/piagent-core/extensions/acceptance-assessment.js +153 -0
  353. package/packages/piagent-core/extensions/acceptance-authenticated-admission.d.ts +41 -0
  354. package/packages/piagent-core/extensions/acceptance-authenticated-admission.js +120 -0
  355. package/packages/piagent-core/extensions/acceptance-behavior-proof.js +264 -70
  356. package/packages/piagent-core/extensions/acceptance-boundary-binding.js +31 -0
  357. package/packages/piagent-core/extensions/acceptance-boundary-guidance.js +214 -0
  358. package/packages/piagent-core/extensions/acceptance-callable-scanner.js +244 -0
  359. package/packages/piagent-core/extensions/acceptance-clause-group.js +45 -0
  360. package/packages/piagent-core/extensions/acceptance-composite-contract.js +334 -0
  361. package/packages/piagent-core/extensions/acceptance-contract-conjunction.js +78 -0
  362. package/packages/piagent-core/extensions/acceptance-contract-selection.js +161 -0
  363. package/packages/piagent-core/extensions/acceptance-contract-semantics.js +211 -211
  364. package/packages/piagent-core/extensions/acceptance-declared-api.js +126 -0
  365. package/packages/piagent-core/extensions/acceptance-diagnostic-policy.ts +14 -0
  366. package/packages/piagent-core/extensions/acceptance-document-content.js +73 -0
  367. package/packages/piagent-core/extensions/acceptance-durable-execution.d.ts +59 -0
  368. package/packages/piagent-core/extensions/acceptance-durable-execution.js +139 -0
  369. package/packages/piagent-core/extensions/acceptance-error-classes.js +1 -1
  370. package/packages/piagent-core/extensions/acceptance-evidence-store.js +223 -0
  371. package/packages/piagent-core/extensions/acceptance-executable-evidence.js +490 -88
  372. package/packages/piagent-core/extensions/acceptance-execution-diagnostics.js +34 -0
  373. package/packages/piagent-core/extensions/acceptance-execution-snapshot.js +174 -0
  374. package/packages/piagent-core/extensions/acceptance-executor/.dockerignore +21 -0
  375. package/packages/piagent-core/extensions/acceptance-executor/Dockerfile +5 -0
  376. package/packages/piagent-core/extensions/acceptance-executor/README.md +103 -0
  377. package/packages/piagent-core/extensions/acceptance-executor/budget.mjs +54 -0
  378. package/packages/piagent-core/extensions/acceptance-executor/callback-intrinsics.mjs +126 -0
  379. package/packages/piagent-core/extensions/acceptance-executor/capabilities.mjs +127 -0
  380. package/packages/piagent-core/extensions/acceptance-executor/guest.mjs +285 -0
  381. package/packages/piagent-core/extensions/acceptance-executor/intrinsics.mjs +133 -0
  382. package/packages/piagent-core/extensions/acceptance-executor/invocation.mjs +39 -0
  383. package/packages/piagent-core/extensions/acceptance-executor/jobs.mjs +29 -0
  384. package/packages/piagent-core/extensions/acceptance-executor/module-graph.mjs +70 -0
  385. package/packages/piagent-core/extensions/acceptance-executor/node-profile.mjs +398 -0
  386. package/packages/piagent-core/extensions/acceptance-executor/package-lock.json +40 -0
  387. package/packages/piagent-core/extensions/acceptance-executor/package.json +10 -0
  388. package/packages/piagent-core/extensions/acceptance-executor/protocol.mjs +248 -0
  389. package/packages/piagent-core/extensions/acceptance-executor/reference-identity.mjs +61 -0
  390. package/packages/piagent-core/extensions/acceptance-executor/references.mjs +14 -0
  391. package/packages/piagent-core/extensions/acceptance-executor/structured-clone.mjs +53 -0
  392. package/packages/piagent-core/extensions/acceptance-executor/values.mjs +97 -0
  393. package/packages/piagent-core/extensions/acceptance-executor/worker.mjs +75 -0
  394. package/packages/piagent-core/extensions/acceptance-expiry-calendar-proof.js +162 -0
  395. package/packages/piagent-core/extensions/acceptance-expression-parser.js +155 -0
  396. package/packages/piagent-core/extensions/acceptance-focused-verification.js +40 -0
  397. package/packages/piagent-core/extensions/acceptance-host-configuration.js +337 -0
  398. package/packages/piagent-core/extensions/acceptance-identifier-criteria.js +25 -0
  399. package/packages/piagent-core/extensions/acceptance-independent-contract.js +209 -0
  400. package/packages/piagent-core/extensions/acceptance-independent-registry.js +106 -0
  401. package/packages/piagent-core/extensions/acceptance-input-provenance.js +40 -0
  402. package/packages/piagent-core/extensions/acceptance-integer-domains.js +322 -0
  403. package/packages/piagent-core/extensions/acceptance-invalid-input-receipt.js +20 -0
  404. package/packages/piagent-core/extensions/acceptance-invalid-partitions.js +26 -0
  405. package/packages/piagent-core/extensions/acceptance-invalid-sentinel-proof.js +477 -0
  406. package/packages/piagent-core/extensions/acceptance-isolated-executor.js +254 -0
  407. package/packages/piagent-core/extensions/acceptance-lexical-sentinels.js +130 -0
  408. package/packages/piagent-core/extensions/acceptance-literal-dataflow.js +735 -0
  409. package/packages/piagent-core/extensions/acceptance-local-binding.js +106 -0
  410. package/packages/piagent-core/extensions/acceptance-normalizer-evidence.js +719 -0
  411. package/packages/piagent-core/extensions/acceptance-obligations.js +75 -0
  412. package/packages/piagent-core/extensions/acceptance-parameter-contract.js +756 -0
  413. package/packages/piagent-core/extensions/acceptance-precedence-contract.js +572 -0
  414. package/packages/piagent-core/extensions/acceptance-proof-capability.js +26 -0
  415. package/packages/piagent-core/extensions/acceptance-receipt.js +139 -152
  416. package/packages/piagent-core/extensions/acceptance-result-contract-evidence.js +159 -0
  417. package/packages/piagent-core/extensions/acceptance-source-origin.js +214 -0
  418. package/packages/piagent-core/extensions/acceptance-state-machine-evidence.js +64 -8
  419. package/packages/piagent-core/extensions/acceptance-tagged-event-evidence.js +685 -0
  420. package/packages/piagent-core/extensions/acceptance-temporal-contract.js +756 -0
  421. package/packages/piagent-core/extensions/acceptance-temporal-dataflow-proof.js +618 -0
  422. package/packages/piagent-core/extensions/acceptance-test-binding-evidence.js +20 -2
  423. package/packages/piagent-core/extensions/acceptance-test-bindings.js +86 -0
  424. package/packages/piagent-core/extensions/acceptance-tuple-shapes.js +12 -0
  425. package/packages/piagent-core/extensions/context-engine.js +11 -6
  426. package/packages/piagent-core/extensions/criterion-context-pack.js +96 -4
  427. package/packages/piagent-core/extensions/failure-types.ts +2 -2
  428. package/packages/piagent-core/extensions/guard-types.ts +5 -1
  429. package/packages/piagent-core/extensions/piagent-guard.ts +197 -155
  430. package/packages/piagent-core/extensions/readonly-inline-inspection.ts +378 -0
  431. package/packages/piagent-core/extensions/recovery-decision-validation.ts +92 -0
  432. package/packages/piagent-core/extensions/shell-write-targets.js +15 -14
  433. package/packages/piagent-core/extensions/task-contract-view.js +4 -4
  434. package/packages/piagent-core/extensions/task-digest-migration.js +2 -2
  435. package/packages/piagent-core/extensions/task-journal.js +48 -3
  436. package/packages/piagent-core/extensions/task-runtime-audit.js +8 -2
  437. package/packages/piagent-core/extensions/task-state.js +74 -38
  438. package/packages/piagent-core/extensions/task-terminal-disposition.js +10 -0
  439. package/packages/piagent-core/extensions/trajectory-contract.ts +5 -0
  440. package/packages/piagent-core/extensions/verification-intelligence.js +60 -8
  441. package/packages/piagent-core/extensions/workflow-catalog.ts +98 -0
  442. package/packages/piagent-core/extensions/workspace-revision.js +48 -0
  443. package/packages/piagent-core/extensions/workspace-verification-snapshot.js +13 -0
  444. package/packages/piagent-core/package.json +3 -3
  445. package/packages/piagent-core/policies/base-policy.json +1 -0
  446. package/packages/piagent-core/runtime/context/adaptive-planner.ts +7 -3
  447. package/packages/piagent-core/runtime/hooks/agent-start-hook.ts +60 -59
  448. package/packages/piagent-core/runtime/hooks/agent-start-types.ts +48 -0
  449. package/packages/piagent-core/runtime/hooks/completion-hook.ts +76 -83
  450. package/packages/piagent-core/runtime/hooks/input-hook.ts +92 -7
  451. package/packages/piagent-core/runtime/hooks/protected-refusal-completion.ts +86 -0
  452. package/packages/piagent-core/runtime/hooks/session-hooks.ts +5 -2
  453. package/packages/piagent-core/runtime/hooks/session-start-hook.ts +50 -18
  454. package/packages/piagent-core/runtime/hooks/tool-result-hook.ts +53 -20
  455. package/packages/piagent-core/runtime/hooks/tool-result-metadata.ts +9 -0
  456. package/packages/piagent-core/runtime/hooks/tool-result-navigation.ts +36 -0
  457. package/packages/piagent-core/runtime/hooks/tool-result-retrieval.ts +42 -0
  458. package/packages/piagent-core/runtime/input/chat-images.ts +1 -1
  459. package/packages/piagent-core/runtime/inspection/activity-event-adapter.ts +1 -1
  460. package/packages/piagent-core/runtime/inspection/same-process-bridge-proof.ts +1 -1
  461. package/packages/piagent-core/runtime/inspection/task-handoff-history.ts +2 -2
  462. package/packages/piagent-core/runtime/inspection/tool-failure-classification.ts +49 -8
  463. package/packages/piagent-core/runtime/inspection/webui-snapshot.ts +61 -15
  464. package/packages/piagent-core/runtime/inspection/workspace-file-reader.ts +2 -132
  465. package/packages/piagent-core/runtime/model/model-route-policy.ts +33 -13
  466. package/packages/piagent-core/runtime/model/model-route-types.ts +10 -6
  467. package/packages/piagent-core/runtime/model/provider-wire-fingerprint.ts +20 -0
  468. package/packages/piagent-core/runtime/model/service-tier-runtime.ts +242 -0
  469. package/packages/piagent-core/runtime/policy/source-worktree-transaction.ts +75 -14
  470. package/packages/piagent-core/runtime/product/activity-inspector.ts +6 -3
  471. package/packages/piagent-core/runtime/product/edit-verifier-timing.ts +26 -0
  472. package/packages/piagent-core/runtime/product/efficiency-metrics.ts +12 -7
  473. package/packages/piagent-core/runtime/product/operator-projections.ts +8 -3
  474. package/packages/piagent-core/runtime/product/task-usage-hooks.ts +78 -0
  475. package/packages/piagent-core/runtime/product/task-usage.ts +104 -0
  476. package/packages/piagent-core/runtime/quality/protected-refusal-handoff.ts +35 -0
  477. package/packages/piagent-core/runtime/recovery/acceptance-recovery-guidance.ts +21 -0
  478. package/packages/piagent-core/runtime/recovery/continuation-budget.ts +11 -2
  479. package/packages/piagent-core/runtime/recovery/diagnostic-delivery.ts +61 -0
  480. package/packages/piagent-core/runtime/recovery/handoff-projection.ts +192 -57
  481. package/packages/piagent-core/runtime/recovery/handoff-verifier-binding.ts +24 -0
  482. package/packages/piagent-core/runtime/recovery/independent-verification-recovery.ts +79 -0
  483. package/packages/piagent-core/runtime/recovery/recovery-policy.ts +27 -39
  484. package/packages/piagent-core/runtime/recovery/resume-state.ts +8 -5
  485. package/packages/piagent-core/runtime/registration/activity-inspector-command.ts +2 -1
  486. package/packages/piagent-core/runtime/registration/context-commands.ts +12 -8
  487. package/packages/piagent-core/runtime/registration/operator-catalogs.ts +4 -9
  488. package/packages/piagent-core/runtime/registration/policy-tools.ts +5 -3
  489. package/packages/piagent-core/runtime/registration/session-commands.ts +9 -7
  490. package/packages/piagent-core/runtime/registration/session-wait-tool.ts +54 -0
  491. package/packages/piagent-core/runtime/registration/task-completion-tools.ts +27 -11
  492. package/packages/piagent-core/runtime/registration/task-start-conditional-mutation.ts +45 -0
  493. package/packages/piagent-core/runtime/registration/task-start-guidance.ts +5 -2
  494. package/packages/piagent-core/runtime/registration/task-start-tool.ts +23 -23
  495. package/packages/piagent-core/runtime/registration/workflow-commands.ts +64 -93
  496. package/packages/piagent-core/runtime/session/completion-signals.ts +23 -0
  497. package/packages/piagent-core/runtime/session/message-signals.ts +6 -22
  498. package/packages/piagent-core/runtime/session/retrieval-evidence-runtime.ts +188 -0
  499. package/packages/piagent-core/runtime/session/runtime-state.ts +29 -35
  500. package/packages/piagent-core/runtime/session/system-prompt.ts +3 -3
  501. package/packages/piagent-core/runtime/session/uncertain-send-continuation.ts +231 -0
  502. package/packages/piagent-core/runtime/session/usage.ts +4 -3
  503. package/packages/piagent-core/runtime/tools/apply-patch-tool.ts +6 -6
  504. package/packages/piagent-core/runtime/tools/phase-tool-runtime.ts +1 -1
  505. package/packages/piagent-core/runtime/tools/phase-tools.ts +3 -2
  506. package/packages/piagent-core/runtime/tools/session-wait.ts +42 -0
  507. package/packages/piagent-core/runtime/tools/tool-groups.ts +10 -3
  508. package/packages/piagent-core/runtime/trajectory/trajectory-runtime.ts +192 -5
  509. package/packages/piagent-core/runtime/trajectory/trajectory-types.ts +3 -2
  510. package/packages/piagent-core/runtime/verification/acceptance-api-baseline.js +488 -0
  511. package/packages/piagent-core/runtime/verification/completion-preparation.ts +22 -0
  512. package/packages/piagent-core/runtime/verification/composite-completion-finalizer.ts +105 -0
  513. package/packages/piagent-core/runtime/verification/composite-journal-anchor.ts +181 -0
  514. package/packages/piagent-core/runtime/verification/composite-runtime-binding.ts +76 -0
  515. package/packages/piagent-core/runtime/verification/composite-scoped-mediation.ts +318 -0
  516. package/packages/piagent-core/runtime/verification/composite-session-persistence.ts +175 -0
  517. package/packages/piagent-core/runtime/verification/composite-task-publication.ts +273 -0
  518. package/packages/piagent-core/runtime/verification/exact-verifier-reuse.ts +11 -11
  519. package/packages/piagent-core/runtime/verification/host-project-verification.ts +71 -0
  520. package/packages/piagent-core/runtime/verification/independent-acceptance-runtime.ts +500 -0
  521. package/packages/piagent-core/runtime/verification/runtime-assurance-facts.ts +500 -0
  522. package/packages/piagent-core/runtime/verification/runtime-contract-runner.ts +107 -0
  523. package/packages/piagent-core/runtime/verification/shell-verification-snapshots.ts +56 -0
  524. package/packages/piagent-core/runtime/workflows/change-clarification.ts +61 -0
  525. package/packages/piagent-core/runtime/workflows/input-routing.ts +21 -14
  526. package/packages/piagent-core/runtime/workflows/runtime-intake-compaction.ts +44 -7
  527. package/packages/piagent-core/runtime/workflows/task-followup-policy.ts +163 -0
  528. package/packages/piagent-core/runtime/workflows/task-intake.ts +221 -242
  529. package/packages/piagent-core/runtime/workflows/task-repair-intent.ts +67 -0
  530. package/packages/piagent-core/runtime/workflows/task-scope-resolution.ts +83 -0
  531. package/packages/piagent-core/runtime/workflows/webui-runtime-command.ts +5 -2
  532. package/packages/piagent-core/runtime/workflows/webui-workflow.ts +39 -32
  533. package/packages/piagent-core/runtime/workflows/workflow-follow-up.ts +33 -0
  534. package/packages/piagent-core/security/workspace-file-reader.ts +132 -0
  535. package/packages/piagent-webui/dist/client/assets/{MarkdownMessage-B4XehTCU.js → MarkdownMessage-DMU7Xe_O.js} +1 -1
  536. package/packages/piagent-webui/dist/client/assets/index-Cui9AAV_.js +176 -0
  537. package/packages/piagent-webui/dist/client/assets/{index-BV-_03yl.css → index-DVahUVWy.css} +1 -1
  538. package/packages/piagent-webui/dist/client/index.html +2 -2
  539. package/packages/piagent-webui/extension/dispatch-user-entry.ts +13 -0
  540. package/packages/piagent-webui/extension/piagent-webui.ts +10 -2
  541. package/packages/piagent-webui/extension/review-controller.ts +29 -6
  542. package/packages/piagent-webui/extension/same-session-bridge.ts +6 -7
  543. package/packages/piagent-webui/extension/session-stream-adapter.ts +60 -11
  544. package/packages/piagent-webui/extension/source-mutation-controller.ts +29 -3
  545. package/packages/piagent-webui/extension/source-open-controller.ts +29 -7
  546. package/packages/piagent-webui/extension/source-revert-controller.ts +26 -3
  547. package/packages/piagent-webui/gateway/gateway-events.ts +22 -5
  548. package/packages/piagent-webui/gateway/gateway-service.ts +7 -2
  549. package/packages/piagent-webui/gateway/gateway-session-stream.ts +141 -19
  550. package/packages/piagent-webui/gateway/rpc-ui-context.ts +34 -9
  551. package/packages/piagent-webui/gateway/runtime-command-controller.ts +46 -6
  552. package/packages/piagent-webui/gateway/runtime-session-controls.ts +17 -2
  553. package/packages/piagent-webui/gateway/session-catalog.ts +9 -3
  554. package/packages/piagent-webui/gateway/session-command-controller.ts +29 -9
  555. package/packages/piagent-webui/gateway/session-effective-options.ts +2 -9
  556. package/packages/piagent-webui/gateway/session-inspection-registry.ts +14 -6
  557. package/packages/piagent-webui/gateway/session-operation-settlement.ts +28 -0
  558. package/packages/piagent-webui/gateway/session-prompt-dispatch.ts +37 -0
  559. package/packages/piagent-webui/gateway/session-runtime-close.ts +36 -0
  560. package/packages/piagent-webui/gateway/session-runtime-factory.ts +183 -6
  561. package/packages/piagent-webui/gateway/session-runtime-ownership.ts +48 -0
  562. package/packages/piagent-webui/gateway/session-runtime-state.ts +24 -0
  563. package/packages/piagent-webui/gateway/session-runtime-supervisor.ts +114 -114
  564. package/packages/piagent-webui/server/core-inspection-provider.ts +73 -14
  565. package/packages/piagent-webui/server/loopback-server.ts +9 -7
  566. package/packages/piagent-webui/server/terminal-delivery-receipt.ts +144 -0
  567. package/packages/piagent-webui/server/transcript-projection.ts +65 -12
  568. package/packages/piagent-webui/shared/message-correlation.ts +25 -0
  569. package/schemas/approved-host-contracts.schema.json +973 -0
  570. package/schemas/benchmark-attempt-outcome.schema.json +303 -0
  571. package/schemas/benchmark-independent-verification-plan.schema.json +34 -0
  572. package/schemas/composite-criterion.schema.json +788 -0
  573. package/schemas/contract-family-library.schema.json +28 -0
  574. package/schemas/contract-selection-recipe.schema.json +31 -0
  575. package/schemas/failure-evidence.schema.json +2 -2
  576. package/schemas/helper-request.schema.json +16 -4
  577. package/schemas/host-contract-plan.schema.json +25 -0
  578. package/schemas/piagent-webui/gateway-protocol-v1.schema.json +16 -2
  579. package/schemas/piagent-webui/runtime-command-v1.schema.json +26 -3
  580. package/schemas/piagent-webui/runtime-event-v2.schema.json +32 -1
  581. package/schemas/piagent-webui/session-live-state-v1.schema.json +2 -0
  582. package/schemas/piagent-webui/snapshot-v1.schema.json +3 -0
  583. package/schemas/piagent-webui/source-change-v1.schema.json +12 -0
  584. package/schemas/piagent-webui/transcript-v1.schema.json +2 -1
  585. package/schemas/public-contracts/codex-mcp-metadata-v1.schema.json +206 -0
  586. package/schemas/public-contracts/codex-mcp-metadata-v2.schema.json +228 -0
  587. package/schemas/public-contracts/codex-mcp-metadata-v3.schema.json +233 -0
  588. package/schemas/public-contracts/revocation-v1.schema.json +87 -0
  589. package/schemas/public-contracts/subscription-contract-v1.schema.json +76 -0
  590. package/schemas/scoped-project-verification-receipt-v2.schema.json +83 -0
  591. package/schemas/scoped-project-verification-receipt-v3.schema.json +295 -0
  592. package/schemas/scoped-verification-receipt-v1.schema.json +76 -0
  593. package/schemas/task-contract.schema.json +30 -0
  594. package/scripts/approve-independent-verification.mjs +57 -0
  595. package/scripts/benchmark-arm-observer.mjs +320 -0
  596. package/scripts/benchmark-budget-ipc.mjs +34 -0
  597. package/scripts/benchmark-budget-launcher.mjs +91 -0
  598. package/scripts/benchmark-budget-processes.mjs +133 -0
  599. package/scripts/benchmark-budget-publication.mjs +228 -0
  600. package/scripts/benchmark-budget-runtime.mjs +136 -0
  601. package/scripts/benchmark-candidate-ablation.mjs +265 -0
  602. package/scripts/benchmark-codex-journey.mjs +406 -0
  603. package/scripts/benchmark-independent-verification.mjs +628 -0
  604. package/scripts/benchmark-journey-outcome.mjs +92 -0
  605. package/scripts/benchmark-production-v3-human-review.mjs +192 -0
  606. package/scripts/benchmark-public-input-catalog.mjs +158 -0
  607. package/scripts/benchmark-registered-plan-author.mjs +355 -0
  608. package/scripts/benchmark-registered-plan-recipe-support.mjs +85 -0
  609. package/scripts/benchmark-registered-plan-recipes-backend.mjs +165 -0
  610. package/scripts/benchmark-registered-plan-recipes-data.mjs +164 -0
  611. package/scripts/benchmark-registered-plan-recipes-platform.mjs +165 -0
  612. package/scripts/benchmark-runner-completion.mjs +133 -0
  613. package/scripts/benchmark-runner-configuration.mjs +159 -0
  614. package/scripts/benchmark-runner-core.mjs +217 -388
  615. package/scripts/benchmark-runner-finalization.mjs +140 -27
  616. package/scripts/benchmark-runner-provider-boundary.mjs +44 -0
  617. package/scripts/benchmark-runner-schedule.mjs +256 -0
  618. package/scripts/benchmark-runner-support.mjs +136 -4
  619. package/scripts/benchmark-runner.mjs +95 -18
  620. package/scripts/benchmark-scoped-broker-custody.mjs +483 -0
  621. package/scripts/benchmark-scoped-docs-verifier.mjs +203 -0
  622. package/scripts/benchmark-scoped-frozen-qualification.mjs +676 -0
  623. package/scripts/benchmark-scoped-pi-router.mjs +196 -0
  624. package/scripts/benchmark-scoped-project-verification-receipt.mjs +49 -0
  625. package/scripts/benchmark-scoped-project-verifier.mjs +461 -0
  626. package/scripts/benchmark-scoped-session-custody.mjs +599 -0
  627. package/scripts/benchmark-scoped-session-factory.mjs +342 -0
  628. package/scripts/benchmark-scoped-tool-broker.mjs +523 -0
  629. package/scripts/benchmark-scoped-verification-supervisor.mjs +773 -0
  630. package/scripts/benchmark-session-evaluator.mjs +169 -0
  631. package/scripts/benchmark-session-provider-boundary.mjs +62 -0
  632. package/scripts/benchmark-session-usage.mjs +28 -0
  633. package/scripts/benchmark-session.mjs +343 -50
  634. package/scripts/benchmark-webui-journey.mjs +678 -0
  635. package/scripts/check-public-wording.mjs +46 -0
  636. package/scripts/configure-model-scope.sh +6 -2
  637. package/scripts/fs4-readiness-evaluation.mjs +7 -1
  638. package/scripts/install-global.sh +2 -2
  639. package/scripts/model-route-evaluation-core.ts +1 -0
  640. package/scripts/phase-tools-evaluation.mjs +1 -0
  641. package/scripts/pi-auto.sh +7 -1
  642. package/scripts/pi-model-route.mjs +1 -1
  643. package/scripts/pi-usage-history.mjs +201 -11
  644. package/scripts/piagent-cli.mjs +8 -0
  645. package/scripts/private-holdout-readiness.mjs +7 -1
  646. package/scripts/public-evaluation-exposure.mjs +140 -0
  647. package/scripts/select-independent-verification.mjs +62 -0
  648. package/scripts/setup.sh +2 -2
  649. package/scripts/typescript-loader.mjs +135 -1
  650. package/scripts/update-global.mjs +6 -7
  651. package/scripts/verify-local.sh +37 -56
  652. package/scripts/verify-release-identity.mjs +5 -3
  653. package/templates/global/settings.json +5 -1
  654. package/packages/piagent-webui/dist/client/assets/index-CgOEyRPG.js +0 -176
@@ -5,8 +5,12 @@ import {
5
5
  buildWebUiRuntimeCommand,
6
6
  isWebUiRuntimeActionId
7
7
  } from "../../piagent-core/runtime/workflows/webui-runtime-command.ts";
8
+ import {
9
+ PIAGENT_SERVICE_TIER_RECEIPT_ENTRY_TYPE,
10
+ parseServiceTierReceipt
11
+ } from "../../piagent-core/runtime/model/service-tier-runtime.ts";
8
12
  import type { Catalog } from "../contracts/generated/session-catalog-v1.ts";
9
- import type { Action, Command, Receipt } from "../contracts/generated/runtime-command-v1.ts";
13
+ import type { Action, Command, Output, Receipt } from "../contracts/generated/runtime-command-v1.ts";
10
14
  import { GatewayEventStore } from "./gateway-events.ts";
11
15
  import { SessionRuntimeSupervisor } from "./session-runtime-supervisor.ts";
12
16
 
@@ -15,6 +19,9 @@ const REVISION = /^[A-Za-z0-9][A-Za-z0-9._:~-]{0,159}$/;
15
19
  const MAX_REPLAYS = 200;
16
20
 
17
21
  type Replay = { digest: string; receipt: Receipt };
22
+ type FastControlAssessment = Pick<Receipt, "state" | "resultCode" | "reasonCode">;
23
+
24
+ const FAST_ACTIONS = new Set<Action>(["runtime.fast-status", "runtime.fast-on", "runtime.fast-off"]);
18
25
 
19
26
  function record(value: unknown): value is Record<string, unknown> {
20
27
  return Boolean(value) && typeof value === "object" && !Array.isArray(value);
@@ -41,6 +48,33 @@ function commandDigest(command: Command): string {
41
48
  argument: command.argument, confirmed: command.confirmed })).digest("hex");
42
49
  }
43
50
 
51
+ function normalizedOutputs(value: unknown): Receipt["outputs"] {
52
+ if (!Array.isArray(value)) return [];
53
+ return value.slice(0, 8).map((item) => {
54
+ if (!record(item)) return item;
55
+ const { details: _details, ...base } = item;
56
+ if (item.customType !== PIAGENT_SERVICE_TIER_RECEIPT_ENTRY_TYPE) return base;
57
+ const details = parseServiceTierReceipt(item.details);
58
+ return details ? { ...base, details } : base;
59
+ }) as Receipt["outputs"];
60
+ }
61
+
62
+ function assessFastControl(action: Action, outputs: readonly Output[]): FastControlAssessment | null {
63
+ if (!FAST_ACTIONS.has(action)) return null;
64
+ const candidates = outputs.filter((output) => output.customType === PIAGENT_SERVICE_TIER_RECEIPT_ENTRY_TYPE);
65
+ if (candidates.length === 0) return { state: "uncertain", resultCode: "effect-unknown", reasonCode: "fast-service-tier-receipt-missing" };
66
+ if (candidates.length !== 1) return { state: "uncertain", resultCode: "effect-unknown", reasonCode: "fast-service-tier-receipt-ambiguous" };
67
+ const details = parseServiceTierReceipt(candidates[0]?.details);
68
+ if (!details) return { state: "uncertain", resultCode: "effect-unknown", reasonCode: "fast-service-tier-receipt-invalid" };
69
+ if (action === "runtime.fast-on" && !details.enabled || action === "runtime.fast-off" && details.enabled) {
70
+ return { state: "uncertain", resultCode: "effect-unknown", reasonCode: "fast-service-tier-receipt-state-mismatch" };
71
+ }
72
+ if (details.reasonCode === "provider-not-supported") {
73
+ return { state: "rejected", resultCode: "unavailable", reasonCode: "provider-not-supported" };
74
+ }
75
+ return { state: "settled", resultCode: "completed", reasonCode: null };
76
+ }
77
+
44
78
  export class RuntimeCommandController {
45
79
  readonly #catalog: () => Promise<Catalog>;
46
80
  readonly #runtimes: SessionRuntimeSupervisor;
@@ -89,11 +123,17 @@ export class RuntimeCommandController {
89
123
  const result = await this.#runtimes.runRuntimeCommand(command.sessionRef, built.command);
90
124
  const after = await this.#readyCatalog();
91
125
  const afterRow = after.sessions.find((session) => session.sessionRef === command.sessionRef);
92
- const zeroTurnViolation = built.spec.effect === "read-only" && result.modelCallObserved;
93
- const receipt = this.#receipt(command, zeroTurnViolation ? "uncertain" : "settled",
94
- zeroTurnViolation ? "effect-unknown" : "completed",
95
- zeroTurnViolation ? "read-only-command-started-model-call" : null,
96
- afterRow?.sessionRevision ?? null, result.modelCallObserved, result.outputs as Receipt["outputs"]);
126
+ const outputs = normalizedOutputs(result.outputs);
127
+ const fastControl = assessFastControl(command.action, outputs);
128
+ const zeroTurnExpected = built.spec.effect === "read-only" || built.spec.effect === "session-setting";
129
+ const zeroTurnViolation = zeroTurnExpected && result.modelCallObserved;
130
+ const zeroTurnReason = built.spec.effect === "session-setting"
131
+ ? "session-setting-command-started-model-call"
132
+ : "read-only-command-started-model-call";
133
+ const receipt = this.#receipt(command, zeroTurnViolation ? "uncertain" : fastControl?.state ?? "settled",
134
+ zeroTurnViolation ? "effect-unknown" : fastControl?.resultCode ?? "completed",
135
+ zeroTurnViolation ? zeroTurnReason : fastControl?.reasonCode ?? null,
136
+ afterRow?.sessionRevision ?? null, result.modelCallObserved, outputs);
97
137
  this.#events.publish("session.changed", { catalogRevision: after.catalogRevision, session: afterRow ?? row });
98
138
  return this.#settle(command, digest, receipt);
99
139
  } catch (error) {
@@ -1,4 +1,17 @@
1
1
  import { redactSensitiveText } from "../../piagent-core/security/sensitive-data.js";
2
+ import {
3
+ PIAGENT_SERVICE_TIER_RECEIPT_ENTRY_TYPE,
4
+ parseServiceTierReceipt,
5
+ type ServiceTierReceipt
6
+ } from "../../piagent-core/runtime/model/service-tier-runtime.ts";
7
+
8
+ type RuntimeCommandOutput = {
9
+ customType: string;
10
+ content: string;
11
+ truncated: boolean;
12
+ redacted: boolean;
13
+ details?: ServiceTierReceipt;
14
+ };
2
15
 
3
16
  export async function executePermissionCommand(session: any,
4
17
  permissionMode: "read-only" | "workspace-write" | "trusted-full-access"): Promise<void> {
@@ -12,7 +25,7 @@ export async function executePermissionCommand(session: any,
12
25
  }
13
26
 
14
27
  export async function executeRuntimeCommand(session: any, command: string): Promise<{
15
- outputs: Array<{ customType: string; content: string; truncated: boolean; redacted: boolean }>;
28
+ outputs: RuntimeCommandOutput[];
16
29
  modelCallObserved: boolean;
17
30
  }> {
18
31
  const before = Array.isArray(session.messages) ? session.messages.length : 0;
@@ -22,8 +35,10 @@ export async function executeRuntimeCommand(session: any, command: string): Prom
22
35
  const source = typeof message.content === "string" ? message.content : JSON.stringify(message.content ?? "");
23
36
  const bounded = source.slice(0, 12_000), redacted = redactSensitiveText(bounded);
24
37
  const rawType = String(message.customType ?? "runtime-output");
38
+ const details = rawType === PIAGENT_SERVICE_TIER_RECEIPT_ENTRY_TYPE
39
+ ? parseServiceTierReceipt(message.details) : null;
25
40
  return { customType: /^[A-Za-z0-9._-]{1,120}$/.test(rawType) ? rawType : "runtime-output", content: redacted.text,
26
- truncated: source.length > bounded.length, redacted: redacted.redacted };
41
+ truncated: source.length > bounded.length, redacted: redacted.redacted, ...(details ? { details } : {}) };
27
42
  });
28
43
  return { outputs, modelCallObserved: added.some((message: any) => message?.role === "assistant") };
29
44
  }
@@ -2,6 +2,7 @@ import { createHmac } from "node:crypto";
2
2
  import path from "node:path";
3
3
 
4
4
  import { redactSensitiveText } from "../../piagent-core/security/sensitive-data.js";
5
+ import { workflowCommandPattern } from "../../piagent-core/runtime/workflows/webui-workflow.ts";
5
6
  import type { Catalog, SessionRow } from "../contracts/generated/session-catalog-v1.ts";
6
7
  import type { MetadataSnapshot, SessionMetadata } from "./session-metadata-store.ts";
7
8
  import { projectRefForCwd, sessionRefForPath } from "../ownership/session-refs.ts";
@@ -54,7 +55,12 @@ function display(value: unknown, maximum: number, fallback: string): string {
54
55
  return (clean || fallback).slice(0, maximum);
55
56
  }
56
57
 
57
- const INTERNAL_FRESH_TRANSITION = /(?:^|\b)(?:\/fresh\s+(?:task|scout|be-to-fe)|(?:task|scout|be-to-fe):)\b[\s\S]{0,180}\bRead task intake from \.pi\/task-inbox\//i;
58
+ const WORKFLOW_PATTERN = workflowCommandPattern({ aliases: false });
59
+ const INTERNAL_FRESH_TRANSITION = new RegExp(
60
+ `(?:^|\\b)(?:\\/fresh\\s+(?:${WORKFLOW_PATTERN})|(?:${WORKFLOW_PATTERN}):)\\b[\\s\\S]{0,180}\\bRead task intake from \\.pi\\/task-inbox\\/`, "i"
61
+ );
62
+ const LEADING_WORKFLOW = new RegExp(`^\\/(?:fresh\\s+)?(?:${WORKFLOW_PATTERN})\\s+`, "i");
63
+ const LEADING_WORKFLOW_LABEL = new RegExp(`^(?:${WORKFLOW_PATTERN}):\\s*`, "i");
58
64
 
59
65
  function atWordBoundary(value: string, maximum: number): string {
60
66
  if (value.length <= maximum) return value;
@@ -78,8 +84,8 @@ export function projectedSessionTitle(info: Pick<PiSessionInfo, "name" | "firstM
78
84
  const internal = INTERNAL_FRESH_TRANSITION.test(source) || INTERNAL_FRESH_TRANSITION.test(first)
79
85
  || /\.pi\/task-inbox\//i.test(source);
80
86
  const cleaned = display(source, 500, "")
81
- .replace(/^\/(?:fresh\s+)?(?:task|scout|be-to-fe)\s+/i, "")
82
- .replace(/^(?:task|scout|be-to-fe):\s*/i, "")
87
+ .replace(LEADING_WORKFLOW, "")
88
+ .replace(LEADING_WORKFLOW_LABEL, "")
83
89
  .replace(/^[-#*>\s]+/, "")
84
90
  // Keep underscores: the shared redactor deliberately emits
85
91
  // [REDACTED_SECRET], and title cleanup must not mutate that proof marker.
@@ -6,6 +6,7 @@ import { buildWebUiWorkflowCommand, isWebUiWorkflowId, type WebUiWorkflowId } fr
6
6
  import { GatewayEventStore } from "./gateway-events.ts";
7
7
  import { SessionCommandStore, type SessionAction, type SessionCommandIdentity } from "./session-command-store.ts";
8
8
  import { SessionRuntimeSupervisor } from "./session-runtime-supervisor.ts";
9
+ import type { SessionSendResult } from "./session-prompt-dispatch.ts";
9
10
  import { SessionMetadataStore } from "./session-metadata-store.ts";
10
11
 
11
12
  const REF = /^[A-Za-z0-9][A-Za-z0-9._~-]{0,159}$/;
@@ -125,9 +126,9 @@ export class SessionCommandController {
125
126
  }
126
127
 
127
128
  async #execute(command: SessionCommand): Promise<Receipt> {
128
- const current = await this.#readyCatalog(), row = command.sessionRef ? current.sessions.find((item) => item.sessionRef === command.sessionRef) : null;
129
129
  const replay = this.#store.lookup(command);
130
130
  if (replay.state === "settled") return { ...replay.receipt!, deduplicated: true };
131
+ const current = await this.#readyCatalog(), row = command.sessionRef ? current.sessions.find((item) => item.sessionRef === command.sessionRef) : null;
131
132
  if (replay.state === "conflict") return this.#rejected(command, current, row, "invalid-command", "idempotency-payload-mismatch");
132
133
  if (replay.state === "unavailable") return this.#rejected(command, current, row, "unavailable", "session-command-journal-unavailable");
133
134
  if (replay.state === "pending") {
@@ -148,7 +149,9 @@ export class SessionCommandController {
148
149
  catch { return this.#rejected(command, current, row, "unavailable", "session-command-admission-failed"); }
149
150
  let targetSessionRef = command.sessionRef;
150
151
  let attachmentReservation: { commit(): void; release(): void } | null = null;
151
- let sendAttempted = false;
152
+ let sendAttempted = false, dispatchCommitted = false;
153
+ let deferredDispatch: Pick<SessionSendResult, "dispatch" | "cancel"> | null = null;
154
+ let observedCatalog = current, observedRow = row;
152
155
  try {
153
156
  let resultCode: "created" | "acquired" | "released" | "started" | "queued" | "steered" | "aborted" | "model-changed" | "thinking-changed" | "permission-changed"
154
157
  | "renamed" | "pinned" | "unpinned" | "archived" | "unarchived" | "forked" | "no-change",
@@ -172,12 +175,15 @@ export class SessionCommandController {
172
175
  }
173
176
  const created = await this.#readyCatalog(), createdRow = created.sessions.find((item) => item.sessionRef === targetSessionRef);
174
177
  if (!createdRow) throw new Error("session-catalog-refresh-failed");
178
+ observedCatalog = created; observedRow = createdRow;
175
179
  if (command.payload.deferInitialMessage === true) resultCode = "created";
176
180
  else {
177
181
  const message = buildWebUiWorkflowCommand((command.payload.workflow as WebUiWorkflowId | undefined) ?? null,
178
182
  String(command.payload.message));
179
183
  const sent = await this.#runtimes.send(targetSessionRef, { delivery: "new-operation", message,
180
- expectedOperationRef: null }, createdRow.sessionRevision);
184
+ messageRequestId: String(command.payload.messageRequestId), expectedOperationRef: null },
185
+ createdRow.sessionRevision, { deferDispatch: true });
186
+ deferredDispatch = sent;
181
187
  resultCode = sent.resultCode === "started" ? "started" : sent.resultCode; operationRef = sent.operationRef;
182
188
  }
183
189
  } else if (command.action === "session.acquire") { await this.#runtimes.acquire(command.sessionRef!); resultCode = "acquired"; }
@@ -186,6 +192,7 @@ export class SessionCommandController {
186
192
  await this.#runtimes.acquire(command.sessionRef!);
187
193
  const acquired = await this.#readyCatalog(), acquiredRow = acquired.sessions.find((item) => item.sessionRef === command.sessionRef);
188
194
  if (!acquiredRow) throw new Error("session-catalog-refresh-failed");
195
+ observedCatalog = acquired; observedRow = acquiredRow;
189
196
  const payload = command.payload as { delivery: "new-operation" | "follow-up" | "steer"; message: string;
190
197
  expectedOperationRef: string | null; messageRequestId: string; attachmentRefs: string[]; workflow?: WebUiWorkflowId };
191
198
  // Staged refs are claimed once, here, before the runtime is asked to
@@ -198,9 +205,10 @@ export class SessionCommandController {
198
205
  if (!this.#prepareAttachments && payload.attachmentRefs.length > 0) throw new Error("session-attachment-unavailable");
199
206
  sendAttempted = true;
200
207
  const message = buildWebUiWorkflowCommand(payload.workflow ?? null, prepared.text);
201
- const sent = await this.#runtimes.send(command.sessionRef!, { ...payload, message, images: prepared.images },
202
- acquiredRow.sessionRevision);
203
- prepared.commit();
208
+ const sent = await this.#runtimes.send(command.sessionRef!, { ...payload, message,
209
+ messageRequestId: payload.messageRequestId, images: prepared.images },
210
+ acquiredRow.sessionRevision, { deferDispatch: true });
211
+ deferredDispatch = sent;
204
212
  resultCode = sent.resultCode; operationRef = sent.operationRef;
205
213
  } else if (command.action === "session.abort") {
206
214
  operationRef = String(command.payload.operationRef);
@@ -239,15 +247,27 @@ export class SessionCommandController {
239
247
  }
240
248
  const after = await this.#readyCatalog(), afterRow = after.sessions.find((item) => item.sessionRef === targetSessionRef);
241
249
  if (!afterRow) throw new Error("session-catalog-refresh-failed");
250
+ observedCatalog = after; observedRow = afterRow;
242
251
  const receipt = this.#settled(command, after, afterRow, resultCode, operationRef);
243
252
  try { this.#store.settle(command, receipt, this.#now()); }
244
- catch { return this.#uncertain(command, after, afterRow, "session-command-settlement-not-durable"); }
253
+ catch {
254
+ deferredDispatch?.cancel?.(); try { attachmentReservation?.release(); } catch { /* staged refs remain unavailable to this command */ }
255
+ return this.#uncertain(command, after, afterRow, "session-command-settlement-not-durable");
256
+ }
257
+ if (attachmentReservation) {
258
+ try { attachmentReservation.commit(); } catch { /* prepared bytes remain dispatchable; the one-shot registry fails closed */ }
259
+ attachmentReservation = null;
260
+ }
245
261
  this.#events.publish("session.changed", { catalogRevision: after.catalogRevision, session: afterRow });
262
+ dispatchCommitted = true; deferredDispatch?.dispatch?.();
246
263
  return receipt;
247
264
  } catch (cause) {
248
- const after = await this.#readyCatalog(), afterRow = after.sessions.find((item) => item.sessionRef === targetSessionRef) ?? row;
265
+ if (!dispatchCommitted) deferredDispatch?.cancel?.();
266
+ let after = observedCatalog;
267
+ try { after = await this.#readyCatalog(); } catch { /* the last proven catalog remains authoritative */ }
268
+ const afterRow = after.sessions.find((item) => item.sessionRef === targetSessionRef) ?? observedRow;
249
269
  const message = cause instanceof Error ? cause.message : "session-command-failed";
250
- const noDispatchEffect = command.action === "session.send" && (!sendAttempted
270
+ const noDispatchEffect = command.action === "session.send" && (!sendAttempted || Boolean(deferredDispatch && !dispatchCommitted)
251
271
  || /^(session-runtime-unavailable|session-operation-conflict|session-owner-conflict|session-runtime-busy)$/.test(message));
252
272
  if (attachmentReservation) {
253
273
  try { if (noDispatchEffect) attachmentReservation.release(); else attachmentReservation.commit(); }
@@ -25,15 +25,8 @@ export function effectiveThinkingLevel(session: any): string | null {
25
25
  return THINKING_LEVELS.has(level) ? level : null;
26
26
  }
27
27
 
28
- export async function configureSessionOptions(session: any, modelRef: string | null,
29
- thinkingLevel: string): Promise<EffectiveSessionOptions> {
30
- if (modelRef) {
31
- const models = session.modelRuntime?.getAvailableSnapshot?.() ?? [];
32
- const model = models.find((value: any) => webUiModelRef(String(value.provider ?? ""),
33
- String(value.id ?? value.modelId ?? "")) === modelRef);
34
- try { if (model) await session.setModel(model); } catch { /* Canonical read-back below decides the effect. */ }
35
- }
36
- try { await session.setThinkingLevel(thinkingLevel); } catch { /* Canonical read-back below decides the effect. */ }
28
+ export function inspectEffectiveSessionOptions(session: any, modelRef: string | null,
29
+ thinkingLevel: string): EffectiveSessionOptions {
37
30
  const effectiveModel = effectiveModelRef(session), effectiveThinking = effectiveThinkingLevel(session);
38
31
  if (!effectiveModel) return { state: "unknown", modelRef: null,
39
32
  thinkingLevel: effectiveThinking, reasonCode: "session-model-effect-unknown" };
@@ -108,11 +108,13 @@ export class SessionInspectionRegistry {
108
108
  readonly #models?: ModelRuntime;
109
109
  readonly #projects?: ProjectSource;
110
110
  readonly #mcpAuth?: McpAuthBroker;
111
+ readonly #operationLiveness: ((sessionRef: string) => "idle" | "running" | "unknown") | null;
111
112
  readonly #providers = new Map<string, { modifiedAt: string; provider: CoreInspectionProvider }>();
112
113
 
113
114
  constructor(options: { gatewayInstanceRef: string; host: PiHost; key: Buffer; packageRoot: string; agentDir?: string; models?: ModelRuntime; projects?: ProjectSource;
114
115
  mcpAuth?: McpAuthBroker; listSessions?: () => Promise<PiSessionInfo[]>;
115
- openLiveSession?: (sessionRef: string) => ReturnType<PiHost["SessionManager"]["open"]> | null }) {
116
+ openLiveSession?: (sessionRef: string) => ReturnType<PiHost["SessionManager"]["open"]> | null;
117
+ operationLiveness?: (sessionRef: string) => "idle" | "running" | "unknown" }) {
116
118
  this.#gatewayInstanceRef = options.gatewayInstanceRef;
117
119
  this.#host = options.host;
118
120
  this.#listSessions = options.listSessions ?? (() => options.host.SessionManager.listAll());
@@ -123,6 +125,7 @@ export class SessionInspectionRegistry {
123
125
  this.#models = options.models;
124
126
  this.#projects = options.projects;
125
127
  this.#mcpAuth = options.mcpAuth;
128
+ this.#operationLiveness = options.operationLiveness ?? null;
126
129
  }
127
130
 
128
131
  #connectionRef(server: { scope: string; origin: string; name: string }): string {
@@ -214,10 +217,9 @@ export class SessionInspectionRegistry {
214
217
  return [{ modelRef: webUiModelRef(provider, modelId), provider, modelId, displayName: safeName(value.name ?? modelId),
215
218
  reasoning: value.reasoning, imageInput: inputs ? inputs.includes("image") : null, thinkingLevels }];
216
219
  });
217
- // Keep the WebUI new-session baseline deterministic and visible. If the
218
- // preferred model is not authenticated, the host's own default remains the
219
- // safe fallback instead of silently substituting another catalog entry.
220
- const defaultModel = models.find((value) => value.provider === "openai-codex" && value.modelId === "gpt-5.6-sol");
220
+ // Only offer a default present in the authenticated host catalog. Retain
221
+ // the previous Sol baseline for accounts that have not received GPT-6 yet.
222
+ const defaultModel = ["gpt-6-sol", "gpt-5.6-sol"].map((id) => models.find((value) => value.provider === "openai-codex" && value.modelId === id)).find(Boolean);
221
223
  return { schemaVersion: 1, version: "piagent-session-creation-options-v1", generatedAt: new Date().toISOString(),
222
224
  projects: [...projects.values()].slice(0, 200), models,
223
225
  defaultModelRef: defaultModel?.modelRef ?? null, defaultThinkingLevel: defaultModel ? "high" : null,
@@ -255,7 +257,12 @@ export class SessionInspectionRegistry {
255
257
  eventStore,
256
258
  task: () => safeRead(() => activeSessionTask(info.cwd, info.id), undefined),
257
259
  activityEvents: () => safeRead(() => readContextTelemetry(info.cwd, { limit: 5_000 }), []),
258
- sessionEntries: () => entries,
260
+ // A provider can outlive the turn that created it (notably while a new
261
+ // session has not reached the host index yet). Read the branch at request
262
+ // time so transcript/activity projection cannot freeze at the pre-turn
263
+ // snapshot while the live manager continues appending entries.
264
+ sessionEntries: () => safeRead(() => manager.getBranch(), []),
265
+ sessionFile: () => info.path,
259
266
  protectedPaths: () => runtimeProtectedPaths(this.#packageRoot, info.cwd),
260
267
  // The document workspace lists the project plus whatever the operator
261
268
  // granted in the profile. Without this it silently shows the project only,
@@ -269,6 +276,7 @@ export class SessionInspectionRegistry {
269
276
  contextUsage: () => contextUsage,
270
277
  model: () => model,
271
278
  thinkingLevel: () => context.thinkingLevel,
279
+ operationLiveness: () => this.#operationLiveness?.(sessionRef) ?? "unknown",
272
280
  approvalProjection: () => piApprovalBroker.projection(info.cwd, info.id),
273
281
  approvalDetail: (approvalRef) => piApprovalBroker.detail(info.cwd, info.id, approvalRef)
274
282
  });
@@ -0,0 +1,28 @@
1
+ import { settleIndependentWebUiOperation } from "../../piagent-core/extensions/acceptance-independent-registry.js";
2
+ import type { activeSessionTask } from "../../piagent-core/extensions/task-state.js";
3
+ import type { GatewaySessionStream } from "./gateway-session-stream.ts";
4
+ import type { ScopedBrokerRouter } from "./session-runtime-factory.ts";
5
+
6
+ type SettlementInput = { sessionId: string; operationRef: string; messageRequestId: string | null;
7
+ manager: any; stream: Pick<GatewaySessionStream, "markBlocked" | "markError"> };
8
+
9
+ export function createSessionOperationSettlement(options: {
10
+ scopedBrokerRouter?: ScopedBrokerRouter;
11
+ compositeSettlementEvidence?(): unknown | Promise<unknown>;
12
+ }) {
13
+ const router = options.scopedBrokerRouter;
14
+ const evidence = options.compositeSettlementEvidence ?? (router ? () => router.settlementEvidence()
15
+ : () => { throw new Error("composite-settlement-evidence-unavailable"); });
16
+ return async (cwd: string, task: ReturnType<typeof activeSessionTask>, input: SettlementInput) => {
17
+ const { sessionId, operationRef, messageRequestId, manager, stream } = input;
18
+ const settlement = await settleIndependentWebUiOperation(cwd, task,
19
+ { operationRef, messageRequestId, manager, evidence });
20
+ if (settlement.status === "blocked") stream.markBlocked("composite-settlement-blocked");
21
+ if (settlement.status === "not-applicable" && task?.trace?.outcome === "completed"
22
+ && router?.discardUnusedSettlement && messageRequestId) {
23
+ try { router.discardUnusedSettlement({ sessionId, operationRef, messageRequestId }); }
24
+ catch { stream.markError("scoped-operation-retirement-failed"); }
25
+ }
26
+ return settlement;
27
+ };
28
+ }
@@ -0,0 +1,37 @@
1
+ import { GatewaySessionStream } from "./gateway-session-stream.ts";
2
+ import { waitForOperationStart } from "./session-operation-start.ts";
3
+
4
+ export type LaunchedSessionPrompt = {
5
+ started: Promise<void>;
6
+ completion: Promise<void>;
7
+ };
8
+ export type SessionSendResult = { resultCode: "started" | "queued" | "steered"; operationRef: string;
9
+ dispatch?: () => void; cancel?: () => boolean };
10
+
11
+ /**
12
+ * Launches host prompt work on the next event-loop turn. Gateway command
13
+ * admission can therefore persist and return its exact operation receipt
14
+ * before expensive synchronous Pi hooks begin, while direct runtime callers
15
+ * can still await the observed agent start through `started`.
16
+ */
17
+ export function launchSessionPrompt(options: {
18
+ stream: GatewaySessionStream;
19
+ invoke(): Promise<void> | void;
20
+ finish(): Promise<void>;
21
+ canInvoke(): boolean;
22
+ }): LaunchedSessionPrompt {
23
+ const prompt = new Promise<void>((resolve, reject) => {
24
+ setImmediate(() => {
25
+ if (!options.canInvoke()) { reject(new Error("session-operation-cancelled-before-dispatch")); return; }
26
+ try { Promise.resolve(options.invoke()).then(resolve, reject); }
27
+ catch (error) { reject(error); }
28
+ });
29
+ });
30
+ const observed = waitForOperationStart(options.stream, prompt);
31
+ const lifecycle = observed.then(async (mode) => {
32
+ if (mode === "deferred") await options.stream.settled();
33
+ else await Promise.race([options.stream.settled(), prompt]);
34
+ });
35
+ const completion = lifecycle.then(options.finish, () => { options.stream.markError(); return options.finish(); });
36
+ return { started: observed.then(() => undefined), completion };
37
+ }
@@ -0,0 +1,36 @@
1
+ import { boundedResult } from "./session-operation-watchdog.ts";
2
+
3
+ type ClosingRuntime = {
4
+ operationRef: string | null;
5
+ settling: boolean;
6
+ completion: Promise<void> | null;
7
+ stream: { markError(reasonCode: string): void } | null;
8
+ };
9
+
10
+ /** Drain ordinary settlement, then revoke authority for work that cannot finish. */
11
+ export async function closeSessionRuntimes<T extends ClosingRuntime>(options: {
12
+ opening: ReadonlyMap<string, Promise<unknown>>;
13
+ active: ReadonlyMap<string, T>;
14
+ terminationTimeoutMs: number;
15
+ abort(sessionRef: string, operationRef: string): Promise<void>;
16
+ release(sessionRef: string): Promise<unknown>;
17
+ quarantine(sessionRef: string, operationRef: string, active: T, reasonCode: string): Promise<void>;
18
+ }): Promise<void> {
19
+ await Promise.allSettled([...options.opening.values()]);
20
+ await Promise.allSettled([...options.active.entries()].filter(([, active]) => active.operationRef)
21
+ .map(([sessionRef, active]) => options.abort(sessionRef, active.operationRef!)));
22
+ await Promise.allSettled([...options.active.entries()].map(async ([sessionRef, active]) => {
23
+ // Normal settlement owns its completion promise and cannot be aborted.
24
+ if (active.settling && active.completion) {
25
+ await boundedResult(active.completion, options.terminationTimeoutMs);
26
+ }
27
+ if (options.active.get(sessionRef) !== active) return;
28
+ if (active.operationRef && active.stream) {
29
+ const reasonCode = "session-runtime-shutdown-incomplete";
30
+ active.stream.markError(reasonCode);
31
+ await options.quarantine(sessionRef, active.operationRef, active, reasonCode);
32
+ return;
33
+ }
34
+ await options.release(sessionRef);
35
+ }));
36
+ }