@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
@@ -35,11 +35,19 @@ function timestamp(value: unknown): value is string {
35
35
  if (typeof value !== "string" || !/^\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}\.\d{3}Z$/.test(value)) return false;
36
36
  const parsed = Date.parse(value); return Number.isFinite(parsed) && new Date(parsed).toISOString() === value;
37
37
  }
38
+ function nullableRevision(value: unknown): value is string | null { return value === null || typeof value === "string" && REVISION.test(value); }
38
39
  function wireIdentity(value: BridgeIdentity): BridgeIdentity { return { projectRef: value.projectRef, runtimeInstanceId: value.runtimeInstanceId,
39
40
  sessionRef: value.sessionRef, taskId: value.taskId, taskRunId: value.taskRunId, agentOperationId: null, toolCallId: null }; }
40
41
  function wireRevisions(value: Record<string, any>): BridgeRevisions { return { runtimeRevision: value.runtimeRevision, taskRevision: value.taskRevision,
41
42
  controlRevision: value.controlRevision, workspaceRevision: value.workspaceRevision, indexRevision: value.indexRevision,
42
43
  approvalRevision: value.approvalRevision, sessionOptionRevision: value.sessionOptionRevision, queueRevision: value.queueRevision }; }
44
+ function replayIdentityDisposition(record: Pick<SourceMutationEvidenceRecord, "projectRef" | "runtimeInstanceId" | "sessionRef" | "taskId" | "taskRunId">,
45
+ identity: BridgeIdentity): "task-identity-mismatch" | "runtime-or-session-replaced" | null {
46
+ if (record.taskId !== identity.taskId || record.taskRunId !== identity.taskRunId) return "task-identity-mismatch";
47
+ if (record.projectRef !== identity.projectRef || record.runtimeInstanceId !== identity.runtimeInstanceId || record.sessionRef !== identity.sessionRef)
48
+ return "runtime-or-session-replaced";
49
+ return null;
50
+ }
43
51
 
44
52
  function receiptFromRecord(record: SourceMutationEvidenceRecord, deduplicated: boolean): SourceMutationReceipt {
45
53
  if (record.action === "source.revert") throw new Error("source-mutation-evidence-action-mismatch");
@@ -109,8 +117,9 @@ export class SourceMutationController {
109
117
  if (prior.length) {
110
118
  if (prior.some((record) => record.commandId !== command.commandId || record.idempotencyKeyDigest !== keyDigest || record.actionDigest !== command.actionDigest
111
119
  || record.action !== command.action)) return this.#reject(command, "idempotency-payload-mismatch", "idempotency-payload-mismatch", false);
112
- if (prior.some((record) => record.runtimeInstanceId !== identity.runtimeInstanceId || record.sessionRef !== identity.sessionRef))
113
- return this.#reject(command, "identity-mismatch", "runtime-or-session-replaced", false);
120
+ const dispositions = prior.map((record) => replayIdentityDisposition(record, identity)).filter(Boolean);
121
+ if (dispositions.length) return this.#reject(command, "identity-mismatch",
122
+ dispositions.includes("task-identity-mismatch") ? "task-identity-mismatch" : "runtime-or-session-replaced", false);
114
123
  return receiptFromRecord(prior.at(-1)!, true);
115
124
  }
116
125
  const now = this.#now().getTime();
@@ -122,6 +131,20 @@ export class SourceMutationController {
122
131
  let resolved: { projection: SourceMutationProjection; authority: SourceMutationAuthority | null };
123
132
  try { resolved = await this.#resolve(command.action, command.payload.fileRef); }
124
133
  catch { return this.#reject(command, "capability-unavailable", "mutation-preview-unavailable", true); }
134
+ const rebound = this.#bridge.snapshot();
135
+ if (rebound.state !== "ready" || !rebound.identity || !rebound.revisions)
136
+ return this.#reject(command, "resync-required", "mutation-binding-unavailable-after-preview", true);
137
+ const reboundIdentity = wireIdentity(rebound.identity), reboundRevisions = wireRevisions(rebound.revisions);
138
+ if (!same(reboundIdentity, identity)) {
139
+ const taskChanged = reboundIdentity.taskId !== identity.taskId || reboundIdentity.taskRunId !== identity.taskRunId;
140
+ return this.#reject(command, "identity-mismatch", taskChanged ? "task-identity-mismatch" : "runtime-or-session-replaced", false);
141
+ }
142
+ if (rebound.taskState === "terminal") return this.#reject(command, "terminal-task", "task-terminal-after-preview", false);
143
+ if (rebound.liveness !== "idle") return this.#reject(command, "capability-unavailable", "agent-not-idle-after-preview", true);
144
+ if (command.expectedRevisions.runtimeRevision !== reboundRevisions.runtimeRevision
145
+ || command.expectedRevisions.taskRevision !== reboundRevisions.taskRevision
146
+ || command.expectedRevisions.controlRevision !== reboundRevisions.controlRevision)
147
+ return this.#reject(command, "stale-revision", "stale-task-control-revision-after-preview", true);
125
148
  if (!resolved.authority || resolved.projection.state !== "ready" || !resolved.projection.target)
126
149
  return this.#reject(command, "capability-unavailable", resolved.projection.reasonCode ?? "mutation-preview-unavailable", true);
127
150
  const target = resolved.projection.target;
@@ -185,8 +208,11 @@ export class SourceMutationController {
185
208
  || command.identity.agentOperationId !== null || command.identity.toolCallId !== null) return "invalid-command-identity";
186
209
  const preimageKeys = [...REVISION_KEYS, "workspacePreimage", "indexPreimage", "patchPreimage"];
187
210
  if (!exactKeys(command.expectedRevisions, preimageKeys) || !REVISION.test(command.expectedRevisions.runtimeRevision)
188
- || !REVISION.test(command.expectedRevisions.taskRevision ?? "") || !REVISION.test(command.expectedRevisions.workspaceRevision ?? "")
211
+ || !REVISION.test(command.expectedRevisions.taskRevision ?? "") || !REVISION.test(command.expectedRevisions.controlRevision ?? "")
212
+ || !REVISION.test(command.expectedRevisions.workspaceRevision ?? "")
189
213
  || !REVISION.test(command.expectedRevisions.indexRevision ?? "") || !WORKSPACE.test(command.expectedRevisions.workspacePreimage)
214
+ || ![command.expectedRevisions.approvalRevision, command.expectedRevisions.sessionOptionRevision,
215
+ command.expectedRevisions.queueRevision].every(nullableRevision)
190
216
  || !DIGEST.test(command.expectedRevisions.indexPreimage) || !DIGEST.test(command.expectedRevisions.patchPreimage)) return "invalid-mutation-authority";
191
217
  if (!exactKeys(command.payload, ["fileRef", "hunkRefs", "contentDigest"]) || !REF.test(command.payload.fileRef)
192
218
  || !Array.isArray(command.payload.hunkRefs) || command.payload.hunkRefs.length > 128
@@ -33,6 +33,13 @@ function revisions(value: Record<string, any>, record?: SourceHandoffRecord): Br
33
33
  taskRevision: record?.taskRevision ?? value.taskRevision, controlRevision: value.controlRevision,
34
34
  workspaceRevision: record?.workspaceRevision ?? value.workspaceRevision, indexRevision: value.indexRevision,
35
35
  approvalRevision: value.approvalRevision, sessionOptionRevision: value.sessionOptionRevision, queueRevision: value.queueRevision }; }
36
+ function replayIdentityDisposition(record: Pick<SourceHandoffRecord, "projectRef" | "runtimeInstanceId" | "sessionRef" | "taskId" | "taskRunId">,
37
+ current: BridgeIdentity): "task-identity-mismatch" | "runtime-or-session-replaced" | null {
38
+ if (record.taskId !== current.taskId || record.taskRunId !== current.taskRunId) return "task-identity-mismatch";
39
+ if (record.projectRef !== current.projectRef || record.runtimeInstanceId !== current.runtimeInstanceId || record.sessionRef !== current.sessionRef)
40
+ return "runtime-or-session-replaced";
41
+ return null;
42
+ }
36
43
 
37
44
  export class SourceOpenController {
38
45
  readonly #bridge: SameSessionPiBridge; readonly #root: string; readonly #resolve: (fileRef: string) => Promise<SourceOpenAuthority | null>;
@@ -55,8 +62,9 @@ export class SourceOpenController {
55
62
  if (prior.length) {
56
63
  if (prior.some((record) => record.commandId !== command.commandId || record.idempotencyKeyDigest !== keyDigest || record.actionDigest !== command.actionDigest))
57
64
  return this.#reject(command, "idempotency-payload-mismatch", "idempotency-payload-mismatch", false);
58
- if (prior.some((record) => record.runtimeInstanceId !== currentIdentity.runtimeInstanceId || record.sessionRef !== currentIdentity.sessionRef))
59
- return this.#reject(command, "identity-mismatch", "runtime-or-session-replaced", false);
65
+ const dispositions = prior.map((record) => replayIdentityDisposition(record, currentIdentity)).filter(Boolean);
66
+ if (dispositions.length) return this.#reject(command, "identity-mismatch",
67
+ dispositions.includes("task-identity-mismatch") ? "task-identity-mismatch" : "runtime-or-session-replaced", false);
60
68
  const last = prior.at(-1)!; if (last.phase !== "requested") return this.#receipt(last, currentRevisions, true);
61
69
  const uncertain = this.#append(command, currentIdentity, keyDigest, last, "uncertain", "effect-unknown", "source-open-outcome-unknown");
62
70
  return this.#receipt(uncertain, currentRevisions, true);
@@ -64,9 +72,22 @@ export class SourceOpenController {
64
72
  const now = this.#now().getTime(); if (Date.parse(command.requestedAt) > now + 30_000) return this.#reject(command, "invalid-command", "requested-at-in-future", false);
65
73
  if (now > Date.parse(command.expiresAt)) return this.#reject(command, "expired", "command-expired", false);
66
74
  if (!same(command.identity, currentIdentity)) return this.#reject(command, "identity-mismatch", "identity-mismatch", false);
67
- if (command.expectedRevisions.runtimeRevision !== currentRevisions.runtimeRevision || command.expectedRevisions.taskRevision !== currentRevisions.taskRevision)
68
- return this.#reject(command, "stale-revision", "stale-source-open-revision", true);
75
+ if (command.expectedRevisions.runtimeRevision !== currentRevisions.runtimeRevision || command.expectedRevisions.taskRevision !== currentRevisions.taskRevision
76
+ || command.expectedRevisions.controlRevision !== currentRevisions.controlRevision)
77
+ return this.#reject(command, "stale-revision", "stale-task-control-revision", true);
69
78
  let authority: SourceOpenAuthority | null; try { authority = await this.#resolve(command.payload.fileRef); } catch { authority = null; }
79
+ const rebound = this.#bridge.snapshot();
80
+ if (rebound.state !== "ready" || !rebound.identity || !rebound.revisions)
81
+ return this.#reject(command, "resync-required", "source-open-binding-unavailable-after-target-resolution", true);
82
+ const reboundIdentity = identity(rebound.identity), reboundRevisions = revisions(rebound.revisions);
83
+ if (!same(reboundIdentity, currentIdentity)) {
84
+ const taskChanged = reboundIdentity.taskId !== currentIdentity.taskId || reboundIdentity.taskRunId !== currentIdentity.taskRunId;
85
+ return this.#reject(command, "identity-mismatch", taskChanged ? "task-identity-mismatch" : "runtime-or-session-replaced", false);
86
+ }
87
+ if (command.expectedRevisions.runtimeRevision !== reboundRevisions.runtimeRevision
88
+ || command.expectedRevisions.taskRevision !== reboundRevisions.taskRevision
89
+ || command.expectedRevisions.controlRevision !== reboundRevisions.controlRevision)
90
+ return this.#reject(command, "stale-revision", "stale-task-control-revision-after-target-resolution", true);
70
91
  if (!authority) return this.#reject(command, "capability-unavailable", "source-open-target-unavailable", true);
71
92
  if (authority.target.taskRevision !== command.expectedRevisions.taskRevision || authority.target.workspaceRevision !== command.expectedRevisions.workspaceRevision)
72
93
  return this.#reject(command, "stale-revision", "source-open-target-stale", true);
@@ -110,9 +131,10 @@ export class SourceOpenController {
110
131
  || command.identity.agentOperationId !== null || command.identity.toolCallId !== null) return "invalid-command-identity";
111
132
  if (!exact(command.expectedRevisions, [...REVISION_KEYS, "workspacePreimage", "indexPreimage", "patchPreimage"])
112
133
  || !REVISION.test(command.expectedRevisions.runtimeRevision) || !REVISION.test(command.expectedRevisions.taskRevision ?? "")
113
- || !REVISION.test(command.expectedRevisions.workspaceRevision ?? "") || command.expectedRevisions.workspacePreimage !== null
114
- || ![command.expectedRevisions.controlRevision, command.expectedRevisions.indexRevision, command.expectedRevisions.approvalRevision,
115
- command.expectedRevisions.sessionOptionRevision, command.expectedRevisions.queueRevision].every(nullableRevision)
134
+ || !REVISION.test(command.expectedRevisions.controlRevision ?? "") || !REVISION.test(command.expectedRevisions.workspaceRevision ?? "")
135
+ || command.expectedRevisions.workspacePreimage !== null
136
+ || ![command.expectedRevisions.indexRevision, command.expectedRevisions.approvalRevision, command.expectedRevisions.sessionOptionRevision,
137
+ command.expectedRevisions.queueRevision].every(nullableRevision)
116
138
  || command.expectedRevisions.indexPreimage !== null || command.expectedRevisions.patchPreimage !== null) return "invalid-source-open-authority";
117
139
  if (!exact(command.payload, ["fileRef", "line", "column"]) || !REF.test(command.payload.fileRef)
118
140
  || command.payload.line !== null && (!Number.isInteger(command.payload.line) || command.payload.line < 1 || command.payload.line > 100_000_000)
@@ -38,6 +38,13 @@ function wireIdentity(value: BridgeIdentity): BridgeIdentity { return { projectR
38
38
  function wireRevisions(value: Record<string, any>): BridgeRevisions { return { runtimeRevision: value.runtimeRevision, taskRevision: value.taskRevision,
39
39
  controlRevision: value.controlRevision, workspaceRevision: value.workspaceRevision, indexRevision: value.indexRevision,
40
40
  approvalRevision: value.approvalRevision, sessionOptionRevision: value.sessionOptionRevision, queueRevision: value.queueRevision }; }
41
+ function replayIdentityDisposition(record: Pick<SourceMutationEvidenceRecord, "projectRef" | "runtimeInstanceId" | "sessionRef" | "taskId" | "taskRunId">,
42
+ identity: BridgeIdentity): "task-identity-mismatch" | "runtime-or-session-replaced" | null {
43
+ if (record.taskId !== identity.taskId || record.taskRunId !== identity.taskRunId) return "task-identity-mismatch";
44
+ if (record.projectRef !== identity.projectRef || record.runtimeInstanceId !== identity.runtimeInstanceId || record.sessionRef !== identity.sessionRef)
45
+ return "runtime-or-session-replaced";
46
+ return null;
47
+ }
41
48
  function receipt(record: SourceMutationEvidenceRecord, deduplicated: boolean): SourceRevertReceipt {
42
49
  const identity = { projectRef: record.projectRef, runtimeInstanceId: record.runtimeInstanceId, sessionRef: record.sessionRef,
43
50
  taskId: record.taskId, taskRunId: record.taskRunId, agentOperationId: null, toolCallId: null } as BridgeIdentity;
@@ -90,8 +97,9 @@ export class SourceRevertController {
90
97
  if (prior.length) {
91
98
  if (prior.some((record) => record.commandId !== command.commandId || record.idempotencyKeyDigest !== keyDigest || record.actionDigest !== command.actionDigest
92
99
  || record.action !== command.action)) return this.#reject(command, "idempotency-payload-mismatch", "idempotency-payload-mismatch", false);
93
- if (prior.some((record) => record.runtimeInstanceId !== identity.runtimeInstanceId || record.sessionRef !== identity.sessionRef))
94
- return this.#reject(command, "identity-mismatch", "runtime-or-session-replaced", false);
100
+ const dispositions = prior.map((record) => replayIdentityDisposition(record, identity)).filter(Boolean);
101
+ if (dispositions.length) return this.#reject(command, "identity-mismatch",
102
+ dispositions.includes("task-identity-mismatch") ? "task-identity-mismatch" : "runtime-or-session-replaced", false);
95
103
  return receipt(prior.at(-1)!, true);
96
104
  }
97
105
  const now = this.#now().getTime();
@@ -103,10 +111,25 @@ export class SourceRevertController {
103
111
  let resolved: { projection: SourceRevertProjection; authority: SourceRevertAuthority | null };
104
112
  try { resolved = await this.#resolve(command.payload.fileRef, command.payload.hunkRefs); }
105
113
  catch { return this.#reject(command, "capability-unavailable", "revert-preview-unavailable", true); }
114
+ const rebound = this.#bridge.snapshot();
115
+ if (rebound.state !== "ready" || !rebound.identity || !rebound.revisions)
116
+ return this.#reject(command, "resync-required", "revert-binding-unavailable-after-preview", true);
117
+ const reboundIdentity = wireIdentity(rebound.identity), reboundRevisions = wireRevisions(rebound.revisions);
118
+ if (!same(reboundIdentity, identity)) {
119
+ const taskChanged = reboundIdentity.taskId !== identity.taskId || reboundIdentity.taskRunId !== identity.taskRunId;
120
+ return this.#reject(command, "identity-mismatch", taskChanged ? "task-identity-mismatch" : "runtime-or-session-replaced", false);
121
+ }
122
+ if (rebound.taskState === "terminal") return this.#reject(command, "terminal-task", "task-terminal-after-preview", false);
123
+ if (rebound.liveness !== "idle") return this.#reject(command, "capability-unavailable", "agent-not-idle-after-preview", true);
124
+ if (command.expectedRevisions.runtimeRevision !== reboundRevisions.runtimeRevision
125
+ || command.expectedRevisions.taskRevision !== reboundRevisions.taskRevision
126
+ || command.expectedRevisions.controlRevision !== reboundRevisions.controlRevision)
127
+ return this.#reject(command, "stale-revision", "stale-task-control-revision-after-preview", true);
106
128
  if (!resolved.authority || resolved.projection.state !== "ready" || !resolved.projection.target)
107
129
  return this.#reject(command, "capability-unavailable", resolved.projection.reasonCode ?? "revert-preview-unavailable", true);
108
130
  const target = resolved.projection.target;
109
- if (!this.#matches(command, target) || now > Date.parse(target.expiresAt)) return this.#reject(command, "stale-revision", "revert-preview-stale", true);
131
+ if (!this.#matches(command, target) || this.#now().getTime() > Date.parse(target.expiresAt))
132
+ return this.#reject(command, "stale-revision", "revert-preview-stale", true);
110
133
  const before = { ...bridgeRevisions, workspaceRevision: target.workspaceRevision, indexRevision: target.indexRevision };
111
134
  let requested: SourceMutationEvidenceRecord;
112
135
  try { requested = appendSourceMutationEvidence({ projectRoot: this.#projectRoot, projectRef: identity.projectRef, runtimeInstanceId: identity.runtimeInstanceId,
@@ -14,10 +14,14 @@ export type GatewayProtocolEvent = {
14
14
  payload: Record<string, unknown>;
15
15
  };
16
16
 
17
+ export type GatewayTaskStatus = "pending" | "completed" | "refused" | "failed" | "unknown";
18
+
17
19
  export type GatewayTerminalSettlement = {
18
20
  sessionRef: string;
19
21
  operationRef: string;
22
+ messageRequestId?: string;
20
23
  settlement: "blocked" | "aborted" | "error" | "unknown";
24
+ taskStatus: GatewayTaskStatus;
21
25
  reasonCode: string;
22
26
  settledAt: string;
23
27
  sequence: number;
@@ -30,7 +34,12 @@ const SETTLEMENT_RETENTION_COUNT = 100;
30
34
  const SETTLEMENT_RETENTION_MS = 60 * 60_000;
31
35
  const MAX_SETTLEMENT_DECISION_COUNT = 10_000;
32
36
  type GatewaySettlementDecision = { settledAt: string; settlement: "completed" | GatewayTerminalSettlement["settlement"];
33
- reasonCode: string | null };
37
+ taskStatus: GatewayTaskStatus; reasonCode: string | null };
38
+
39
+ function settlementTaskStatus(value: unknown): GatewayTaskStatus {
40
+ return ["pending", "completed", "refused", "failed", "unknown"].includes(String(value))
41
+ ? value as GatewayTaskStatus : "unknown";
42
+ }
34
43
 
35
44
  function settlementReason(settlement: GatewayTerminalSettlement["settlement"], value: unknown): string {
36
45
  if (typeof value === "string" && REASON_CODE_PATTERN.test(value) && !redactSensitiveText(value).redacted) return value;
@@ -84,7 +93,8 @@ export class GatewayEventStore {
84
93
  if (indexed) return indexed;
85
94
  const retained = this.#settlements.find((item) => item.sessionRef === sessionRef && item.operationRef === operationRef);
86
95
  if (!retained) return undefined;
87
- const restored = { settledAt: retained.settledAt, settlement: retained.settlement, reasonCode: retained.reasonCode };
96
+ const restored = { settledAt: retained.settledAt, settlement: retained.settlement,
97
+ taskStatus: retained.taskStatus, reasonCode: retained.reasonCode };
88
98
  this.#settlementDecisions.set(key, restored);
89
99
  return restored;
90
100
  }
@@ -100,11 +110,16 @@ export class GatewayEventStore {
100
110
  if (this.#settlementDecision(sessionRef, operationRef)) return;
101
111
  this.#settlementDecisions.set(key, { settledAt: event.generatedAt,
102
112
  settlement: settlement as GatewaySettlementDecision["settlement"],
113
+ taskStatus: settlementTaskStatus(event.payload.taskStatus),
103
114
  reasonCode: settlement === "completed" ? null : String(event.payload.reasonCode) });
104
- const retained = { sessionRef, operationRef, settledAt: event.generatedAt, sequence: event.sequence };
115
+ const messageRequestId = typeof event.payload.messageRequestId === "string" && OPAQUE_REF_PATTERN.test(event.payload.messageRequestId)
116
+ ? event.payload.messageRequestId : undefined;
117
+ const retained = { sessionRef, operationRef, ...(messageRequestId ? { messageRequestId } : {}),
118
+ settledAt: event.generatedAt, sequence: event.sequence };
105
119
  if (settlement !== "completed") {
106
120
  const outcome = settlement as GatewayTerminalSettlement["settlement"];
107
- this.#settlements.push({ ...retained, settlement: outcome, reasonCode: settlementReason(outcome, event.payload.reasonCode) });
121
+ this.#settlements.push({ ...retained, settlement: outcome, taskStatus: settlementTaskStatus(event.payload.taskStatus),
122
+ reasonCode: settlementReason(outcome, event.payload.reasonCode) });
108
123
  }
109
124
  this.#pruneSettlements(now);
110
125
  }
@@ -112,6 +127,7 @@ export class GatewayEventStore {
112
127
  publish(kind: GatewayEventKind, payload: Record<string, unknown>, now = new Date()): GatewayProtocolEvent {
113
128
  if (!Number.isFinite(now.getTime())) throw new Error("gateway-event-time-invalid");
114
129
  if (kind === "operation.settled") {
130
+ payload = { ...payload, taskStatus: settlementTaskStatus(payload.taskStatus) };
115
131
  if (payload.settlement === "completed") payload = { ...payload, reasonCode: null };
116
132
  else {
117
133
  const settlement = ["blocked", "aborted", "error", "unknown"].includes(String(payload.settlement))
@@ -123,7 +139,8 @@ export class GatewayEventStore {
123
139
  if (typeof sessionRef === "string" && OPAQUE_REF_PATTERN.test(sessionRef)
124
140
  && typeof operationRef === "string" && OPAQUE_REF_PATTERN.test(operationRef)) {
125
141
  const canonical = this.#settlementDecision(sessionRef, operationRef);
126
- if (canonical) payload = { ...payload, settlement: canonical.settlement, reasonCode: canonical.reasonCode };
142
+ if (canonical) payload = { ...payload, settlement: canonical.settlement,
143
+ taskStatus: canonical.taskStatus, reasonCode: canonical.reasonCode };
127
144
  }
128
145
  }
129
146
  this.#sequence += 1;
@@ -32,6 +32,7 @@ import { ProjectRegistry } from "./project-registry.ts";
32
32
  import { pickNativeProjectFolders } from "./native-project-picker.ts";
33
33
  import { ProviderAuthBroker } from "./provider-auth-broker.ts";
34
34
  import { McpAuthBroker } from "./mcp-auth-broker.ts";
35
+ import type { ScopedBrokerRouter } from "./session-runtime-factory.ts";
35
36
 
36
37
  function unavailable(reasonCode: string) {
37
38
  return { status: "unavailable" as const, version: null, reasonCode };
@@ -80,6 +81,8 @@ export async function startPiagentGateway(options: {
80
81
  packageRoot: string;
81
82
  expectedPiVersion: string;
82
83
  agentDir?: string;
84
+ staticRoot?: string;
85
+ scopedBrokerRouter?: ScopedBrokerRouter;
83
86
  }): Promise<{ descriptor: GatewayDescriptor; wait(): Promise<void>; close(): Promise<void> }> {
84
87
  const state = gatewayProfileState(options.agentDir);
85
88
  process.env.PI_CODING_AGENT_DIR = state.agentDir;
@@ -87,7 +90,7 @@ export async function startPiagentGateway(options: {
87
90
  const metadata = new SessionMetadataStore(state.root, key);
88
91
  const projects = new ProjectRegistry(state.root, key);
89
92
  const gatewayInstanceRef = `gateway_${process.pid}_${randomBytes(24).toString("base64url")}`;
90
- const staticRoot = path.join(options.packageRoot, "packages", "piagent-webui", "dist", "client");
93
+ const staticRoot = options.staticRoot ?? path.join(options.packageRoot, "packages", "piagent-webui", "dist", "client");
91
94
  if (!fs.existsSync(path.join(staticRoot, "index.html"))) throw new Error("gateway-webui-build-missing");
92
95
 
93
96
  let descriptor: GatewayDescriptor | null = null;
@@ -145,6 +148,7 @@ export async function startPiagentGateway(options: {
145
148
  runtimes = new SessionRuntimeSupervisor({
146
149
  gatewayInstanceRef, key, leases, listSessions: () => host.SessionManager.listAll(),
147
150
  host, agentDir: state.agentDir, packageRoot: options.packageRoot, modelRuntime: inspectionModels, events,
151
+ scopedBrokerRouter: options.scopedBrokerRouter,
148
152
  resolveProject: (projectRef) => projects.resolve(projectRef)
149
153
  });
150
154
  const readCatalog = () => buildSessionCatalog({
@@ -182,7 +186,8 @@ export async function startPiagentGateway(options: {
182
186
  events, command: commands });
183
187
  const inspections = new SessionInspectionRegistry({ gatewayInstanceRef, host, key, packageRoot: options.packageRoot, agentDir: state.agentDir,
184
188
  models: inspectionModels, projects, mcpAuth, listSessions: () => runtimes!.listSessions(),
185
- openLiveSession: (sessionRef) => runtimes!.liveSessionManager(sessionRef) });
189
+ openLiveSession: (sessionRef) => runtimes!.liveSessionManager(sessionRef),
190
+ operationLiveness: (sessionRef) => runtimes!.currentOperation(sessionRef) ? "running" : "idle" });
186
191
  // Staged bytes live beside the inspection projection they were checked
187
192
  // against, so both read the same session through the same registry.
188
193
  attachments = new SessionAttachmentRegistry({
@@ -2,7 +2,8 @@ import { randomBytes } from "node:crypto";
2
2
 
3
3
  import { redactSensitiveText } from "../../piagent-core/security/sensitive-data.js";
4
4
  import { hasVisibleText } from "../shared/text-visibility.ts";
5
- import { GatewayEventStore } from "./gateway-events.ts";
5
+ import type { TerminalDeliveryReceipt } from "../server/terminal-delivery-receipt.ts";
6
+ import { GatewayEventStore, type GatewayTaskStatus } from "./gateway-events.ts";
6
7
  import { SessionOperationLifecycle, type SessionOperationObservation,
7
8
  type SessionOperationPhase, type SessionOperationRetryPolicyOptions } from "./session-operation-lifecycle.ts";
8
9
 
@@ -17,6 +18,14 @@ const COMPLETION_GATE = /^\[Piagent completion gate: (CONTINUING|NOT APPROVED)\]
17
18
  export type GatewayOperationSettlement = "completed" | "blocked" | "aborted" | "error" | "unknown";
18
19
  type Settlement = { outcome: GatewayOperationSettlement; reasonCode: string | null };
19
20
 
21
+ function projectedTaskStatus(taskOutcome: string | null, override: GatewayTaskStatus | null): GatewayTaskStatus {
22
+ if (override) return override;
23
+ if (taskOutcome === "pending") return "pending";
24
+ if (taskOutcome === "completed") return "completed";
25
+ if (["blocked", "partial", "failed"].includes(String(taskOutcome))) return "failed";
26
+ return "unknown";
27
+ }
28
+
20
29
  function clean(value: unknown): string {
21
30
  return String(value ?? "").replace(/\u001b(?:\[[0-?]*[ -/]*[@-~]|[@-_])/g, "")
22
31
  .replace(/[\u0000-\u0008\u000b\u000c\u000e-\u001f\u007f]/g, " ");
@@ -26,6 +35,66 @@ function safeTool(value: unknown): string {
26
35
  if (safe.redacted) return "redacted-tool";
27
36
  return safe.text.replace(/[^A-Za-z0-9._:-]/g, "-").replace(/^[^A-Za-z0-9]+/, "").slice(0, 160) || "tool";
28
37
  }
38
+
39
+ const DIRECT_FILE_KEYS = ["file", "filePath", "file_path", "filename", "targetFile", "target_file"] as const;
40
+ const PATH_KEYS = ["path", "targetPath", "target_path"] as const;
41
+ const FILE_MUTATION_OR_READ_TOOL = /(?:^|[._:-])(read|write|edit|patch|apply)(?:[._:-]|$)/i;
42
+ const FILE_SEARCH_TOOL = /(?:^|[._:-])(grep|search|find|glob|list|ls)(?:[._:-]|$)/i;
43
+ const FILE_LIKE_BASENAME = /(?:^\.[^./\\]+$)|(?:\.[A-Za-z0-9][A-Za-z0-9+_-]{0,15}$)|^(?:Dockerfile|Containerfile|Makefile|README|LICENSE|CHANGELOG)$/i;
44
+ const PATCH_FILE_HEADER = /^\*\*\* (?:Update|Add|Delete) File:\s*(.+)$/m;
45
+
46
+ function record(value: unknown): Record<string, unknown> | null {
47
+ return value !== null && typeof value === "object" && !Array.isArray(value) ? value as Record<string, unknown> : null;
48
+ }
49
+
50
+ function firstString(records: readonly Record<string, unknown>[], keys: readonly string[]): string | null {
51
+ for (const source of records) for (const key of keys) {
52
+ const value = source[key];
53
+ if (typeof value === "string" && value.trim()) return value;
54
+ if (Array.isArray(value)) {
55
+ const first = value.find((item) => typeof item === "string" && item.trim());
56
+ if (typeof first === "string") return first;
57
+ }
58
+ }
59
+ return null;
60
+ }
61
+
62
+ function safeBasename(value: unknown): string | null {
63
+ if (typeof value !== "string") return null;
64
+ const source = clean(value).trim();
65
+ if (!source || source.includes("\n") || source.includes("\r") || /(?:^|\s)(?:https?|data):\/\//i.test(source)) return null;
66
+ const parts = source.replace(/\\/g, "/").replace(/\/+$/, "").split("/");
67
+ const basename = (parts.at(-1) ?? "").trim().replace(/^["']|["']$/g, "");
68
+ if (!basename || basename === "." || basename === "..") return null;
69
+ const projected = redactSensitiveText(basename);
70
+ if (projected.redacted || projected.text !== basename || /[\u0000-\u001f\u007f/\\]/.test(basename)) return null;
71
+ return basename.slice(0, 160);
72
+ }
73
+
74
+ /**
75
+ * Returns only a redaction-checked basename suitable for the compact chat
76
+ * progress row. Raw arguments and full paths remain confined to the runtime;
77
+ * the Activity inspector keeps its independent, richer projection.
78
+ */
79
+ export function safeToolFileLabel(toolName: unknown, args: unknown): string | null {
80
+ const tool = safeTool(toolName), direct = record(args); if (!direct) return null;
81
+ const nested = record(direct.args), input = record(direct.input);
82
+ const records = [direct, ...(nested ? [nested] : []), ...(input ? [input] : [])];
83
+ const explicitFile = firstString(records, DIRECT_FILE_KEYS);
84
+ if (explicitFile) return safeBasename(explicitFile);
85
+ const path = firstString(records, PATH_KEYS);
86
+ if (path) {
87
+ const basename = safeBasename(path);
88
+ if (basename && (FILE_MUTATION_OR_READ_TOOL.test(tool)
89
+ || FILE_SEARCH_TOOL.test(tool) && FILE_LIKE_BASENAME.test(basename))) return basename;
90
+ }
91
+ if (FILE_MUTATION_OR_READ_TOOL.test(tool)) {
92
+ const patch = firstString(records, ["patch"]);
93
+ const headerPath = typeof patch === "string" ? PATCH_FILE_HEADER.exec(patch)?.[1] : null;
94
+ if (headerPath) return safeBasename(headerPath);
95
+ }
96
+ return null;
97
+ }
29
98
  function ref(prefix: string): string { return `${prefix}_${randomBytes(18).toString("base64url")}`; }
30
99
  function messageText(message: any): string {
31
100
  if (typeof message?.content === "string") return message.content;
@@ -69,6 +138,7 @@ export function runtimeRestartReasonCode(event: unknown): "runtime-restart-requi
69
138
  export class GatewaySessionStream {
70
139
  readonly sessionRef: string;
71
140
  readonly operationRef: string;
141
+ readonly messageRequestId: string | null;
72
142
  readonly #events: GatewayEventStore;
73
143
  #messageRef: string | null = null;
74
144
  #messageSequence = 0;
@@ -83,7 +153,12 @@ export class GatewaySessionStream {
83
153
  #lifecycleTerminationReasonCode: string | null = null;
84
154
  #settleListeners: Array<() => void> = [];
85
155
  readonly #toolRefs = new Map<string, string>();
156
+ readonly #toolLabels = new Map<string, string>();
157
+ readonly #toolFileLabels = new Map<string, string | null>();
86
158
  readonly #lifecycle: SessionOperationLifecycle;
159
+ readonly #commitTerminalDeliveryReceipt: ((message: unknown) => TerminalDeliveryReceipt | undefined) | undefined;
160
+ #terminalDeliveryMessage: unknown;
161
+ #receiptOnly = true;
87
162
 
88
163
  get runtimeRestartRequired(): boolean { return this.#runtimeRestartRequired; }
89
164
  get lifecycleTerminationReasonCode(): string | null { return this.#lifecycleTerminationReasonCode; }
@@ -91,6 +166,9 @@ export class GatewaySessionStream {
91
166
  get retryAbortRequired(): boolean { return this.#lifecycle.retryAbortRequired; }
92
167
 
93
168
  markAborted(reasonCode = "operation-aborted"): void { this.#forcedSettlement = { outcome: "aborted", reasonCode }; }
169
+ markBlocked(reasonCode = "operation-blocked"): void {
170
+ if (!this.#forcedSettlement) this.#forcedSettlement = { outcome: "blocked", reasonCode };
171
+ }
94
172
  markError(reasonCode = "operation-failed"): void {
95
173
  if (!this.#forcedSettlement) this.#forcedSettlement = { outcome: "error", reasonCode };
96
174
  }
@@ -102,10 +180,17 @@ export class GatewaySessionStream {
102
180
  if (!wasSettled) for (const resolve of this.#settleListeners.splice(0)) resolve();
103
181
  }
104
182
 
105
- constructor(options: { sessionRef: string; operationRef: string; events: GatewayEventStore;
106
- retryPolicy?: SessionOperationRetryPolicyOptions }) {
107
- this.sessionRef = options.sessionRef; this.operationRef = options.operationRef; this.#events = options.events;
183
+ constructor(options: { sessionRef: string; operationRef: string; messageRequestId?: string | null; events: GatewayEventStore;
184
+ retryPolicy?: SessionOperationRetryPolicyOptions;
185
+ commitTerminalDeliveryReceipt?: (message: unknown) => TerminalDeliveryReceipt | undefined }) {
186
+ this.sessionRef = options.sessionRef; this.operationRef = options.operationRef;
187
+ this.messageRequestId = options.messageRequestId ?? null; this.#events = options.events;
108
188
  this.#lifecycle = new SessionOperationLifecycle({ operationRef: options.operationRef, retryPolicy: options.retryPolicy });
189
+ this.#commitTerminalDeliveryReceipt = options.commitTerminalDeliveryReceipt;
190
+ }
191
+
192
+ #correlation(): { messageRequestId?: string } {
193
+ return this.messageRequestId ? { messageRequestId: this.messageRequestId } : {};
109
194
  }
110
195
 
111
196
  started(): Promise<void> {
@@ -122,6 +207,9 @@ export class GatewaySessionStream {
122
207
  const wasStarted = this.#lifecycle.started, wasSettled = this.#lifecycle.hostSettled;
123
208
  const observation = this.#lifecycle.observe(event);
124
209
  if (!observation.accepted) return observation;
210
+ if (event?.type === "agent_start" || event?.message?.role === "assistant" || event?.assistantMessageEvent
211
+ || String(event?.type ?? "").startsWith("tool_execution_")) this.#receiptOnly = false;
212
+ if (event?.type === "message_end" && event.message?.role === "custom") this.#terminalDeliveryMessage = event.message;
125
213
  if (!wasStarted && this.#lifecycle.started) for (const resolve of this.#startListeners.splice(0)) resolve();
126
214
  if (!wasSettled && this.#lifecycle.hostSettled) for (const resolve of this.#settleListeners.splice(0)) resolve();
127
215
  if (observation.retry === "abort") this.markError(observation.reasonCode ?? "automatic-retry-blocked");
@@ -129,6 +217,7 @@ export class GatewaySessionStream {
129
217
  return observation;
130
218
  }
131
219
  if (event?.type === "agent_settled") {
220
+ this.#settleActiveTools("operation-settled-before-tool-result");
132
221
  return observation;
133
222
  }
134
223
  if (event?.type === "message_start" && event.message?.role === "assistant") {
@@ -146,22 +235,35 @@ export class GatewaySessionStream {
146
235
  this.#flush(true); this.#lastMessageRef = this.#messageRef; this.#settlement = assistantSettlement(event.message); return observation;
147
236
  }
148
237
  if (event?.type === "tool_execution_start" && typeof event.toolCallId === "string") {
149
- const toolCallRef = ref("tool"); this.#toolRefs.set(event.toolCallId, toolCallRef);
150
- this.#events.publish("tool.started", { sessionRef: this.sessionRef, operationRef: this.operationRef, toolCallRef,
151
- toolLabel: safeTool(event.toolName), isError: null, reasonCode: null });
238
+ const toolCallRef = ref("tool"), toolLabel = safeTool(event.toolName), fileLabel = safeToolFileLabel(event.toolName, event.args);
239
+ this.#toolRefs.set(event.toolCallId, toolCallRef); this.#toolLabels.set(event.toolCallId, toolLabel);
240
+ this.#toolFileLabels.set(event.toolCallId, fileLabel);
241
+ this.#events.publish("tool.started", { sessionRef: this.sessionRef, operationRef: this.operationRef, ...this.#correlation(), toolCallRef,
242
+ toolLabel, fileLabel, isError: null, reasonCode: null });
152
243
  return observation;
153
244
  }
154
245
  if (event?.type === "tool_execution_end" && typeof event.toolCallId === "string") {
155
246
  const toolCallRef = this.#toolRefs.get(event.toolCallId); if (!toolCallRef) return observation;
156
247
  const reasonCode = runtimeRestartReasonCode(event);
157
248
  if (reasonCode) this.#runtimeRestartRequired = true;
158
- this.#events.publish("tool.completed", { sessionRef: this.sessionRef, operationRef: this.operationRef, toolCallRef,
159
- toolLabel: safeTool(event.toolName), isError: event.isError === true, reasonCode });
160
- this.#toolRefs.delete(event.toolCallId);
249
+ this.#events.publish("tool.completed", { sessionRef: this.sessionRef, operationRef: this.operationRef, ...this.#correlation(), toolCallRef,
250
+ toolLabel: safeTool(event.toolName), fileLabel: this.#toolFileLabels.get(event.toolCallId) ?? null,
251
+ isError: event.isError === true, reasonCode });
252
+ this.#toolRefs.delete(event.toolCallId); this.#toolLabels.delete(event.toolCallId); this.#toolFileLabels.delete(event.toolCallId);
161
253
  }
254
+ if (event?.type === "turn_end") this.#settleActiveTools("turn-ended-before-tool-result");
162
255
  return observation;
163
256
  }
164
257
 
258
+ #settleActiveTools(reasonCode: string): void {
259
+ for (const [toolCallId, toolCallRef] of this.#toolRefs) {
260
+ this.#events.publish("tool.completed", { sessionRef: this.sessionRef, operationRef: this.operationRef, ...this.#correlation(), toolCallRef,
261
+ toolLabel: this.#toolLabels.get(toolCallId) ?? "tool", fileLabel: this.#toolFileLabels.get(toolCallId) ?? null,
262
+ isError: true, reasonCode });
263
+ }
264
+ this.#toolRefs.clear(); this.#toolLabels.clear(); this.#toolFileLabels.clear();
265
+ }
266
+
165
267
  #flush(final: boolean): void {
166
268
  if (!this.#messageRef || !this.#buffer || this.#eventCount >= MAX_EVENTS) {
167
269
  if (this.#eventCount >= MAX_EVENTS) this.#truncated = true;
@@ -186,37 +288,57 @@ export class GatewaySessionStream {
186
288
  #emit(value: string): void {
187
289
  for (let offset = 0; offset < value.length; offset += MAX_DELTA) {
188
290
  if (this.#eventCount >= MAX_EVENTS) { this.#truncated = true; break; }
189
- this.#events.publish("message.delta", { sessionRef: this.sessionRef, operationRef: this.operationRef,
291
+ this.#events.publish("message.delta", { sessionRef: this.sessionRef, operationRef: this.operationRef, ...this.#correlation(),
190
292
  messageRef: this.#messageRef!, messageSequence: this.#messageSequence++, delta: value.slice(offset, offset + MAX_DELTA) });
191
293
  this.#eventCount += 1;
192
294
  }
193
295
  }
194
296
 
195
- complete(sessionRevision: string | null, taskOutcome: string | null = null): void {
297
+ complete(sessionRevision: string | null, taskOutcome: string | null = null,
298
+ taskStatusOverride: GatewayTaskStatus | null = null): void {
196
299
  if (!this.#lifecycle.markTerminal()) return;
300
+ const receipt = this.#receiptOnly && !this.#forcedSettlement && !this.#runtimeRestartRequired
301
+ && sessionRevision !== null && taskOutcome !== null && taskOutcome !== "pending"
302
+ ? this.#commitTerminalDeliveryReceipt?.(this.#terminalDeliveryMessage) : undefined;
303
+ if (receipt) {
304
+ // The canonical transcript renders a custom receipt. Never stream it as
305
+ // assistant text or open a model turn just to supply a final message.
306
+ this.#lastMessageRef = ref("receipt");
307
+ const diagnostic = receipt.details.diagnosticDelivery as { mode?: string; qualityClaim?: string } | undefined;
308
+ const diagnosticDelivered = taskOutcome === "completed" && receipt.details.completionApproved === false
309
+ && diagnostic?.mode === "diagnostic" && diagnostic.qualityClaim === "withheld";
310
+ const completedTask = receipt.details.outcome === "completed" && (receipt.details.completionApproved || diagnosticDelivered)
311
+ && receipt.details.gateDecision === "pass";
312
+ // Delivery of an already refused task is a completed operation, never
313
+ // approval to perform the refused action. Both the committed receipt
314
+ // and the supervisor's current native task projection must agree.
315
+ const refusedTask = taskOutcome === "blocked" && taskStatusOverride === "refused"
316
+ && receipt.details.outcome === "blocked" && receipt.details.completionApproved === false
317
+ && receipt.details.gateDecision === "fail";
318
+ this.#settlement = completedTask || refusedTask ? { outcome: "completed", reasonCode: null }
319
+ : { outcome: "blocked", reasonCode: "terminal-receipt-not-approved" };
320
+ }
197
321
  this.#flush(true);
322
+ this.#settleActiveTools("operation-settled-before-tool-result");
198
323
  let settlement = this.#forcedSettlement ?? (this.#runtimeRestartRequired
199
324
  ? { outcome: "unknown" as const, reasonCode: "runtime-restart-required" }
200
325
  : this.#settlement);
201
326
  const messageRef = this.#lastMessageRef ?? this.#messageRef;
327
+ const taskStatus = projectedTaskStatus(taskOutcome, taskStatusOverride);
202
328
  if (settlement.outcome === "completed" && !messageRef) {
203
329
  settlement = { outcome: "unknown", reasonCode: "assistant-message-unavailable" };
204
- } else if (settlement.outcome === "completed" && taskOutcome === "pending") {
205
- // Provider completion is not governed-task completion. The durable task
206
- // contract is authoritative, so withhold both success settlement and the
207
- // durable assistant-success event until that contract becomes terminal.
208
- settlement = { outcome: "unknown", reasonCode: "task-completion-pending" };
209
330
  } else if (settlement.outcome === "completed" && sessionRevision === null) {
210
331
  // Streaming text is not durable success until the canonical session
211
332
  // projection confirms it. Still terminate the UI, but do not show draft.
212
333
  settlement = { outcome: "unknown", reasonCode: "session-projection-unavailable" };
213
334
  }
214
335
  if (settlement.outcome === "completed" && messageRef && sessionRevision) {
215
- this.#events.publish("message.completed", { sessionRef: this.sessionRef, operationRef: this.operationRef,
336
+ this.#events.publish("message.completed", { sessionRef: this.sessionRef, operationRef: this.operationRef, ...this.#correlation(),
216
337
  messageRef, sessionRevision, truncated: this.#truncated });
217
338
  }
218
- this.#events.publish("operation.settled", { sessionRef: this.sessionRef, operationRef: this.operationRef,
339
+ this.#events.publish("operation.settled", { sessionRef: this.sessionRef, operationRef: this.operationRef, ...this.#correlation(),
219
340
  messageRef, sessionRevision, settlement: settlement.outcome,
341
+ taskStatus,
220
342
  reasonCode: settlement.outcome === "completed" ? null : settlement.reasonCode ?? "operation-settlement-unknown" });
221
343
  }
222
344
  }
@@ -16,13 +16,38 @@ export function rpcUiContext(): object {
16
16
  getThinkingBorderColor: () => plain,
17
17
  getBashModeBorderColor: () => plain
18
18
  });
19
- return new Proxy({}, {
20
- get(_target, property) {
21
- if (property === GATEWAY_RUNTIME_UI_MARKER) return true;
22
- if (property === "theme") return theme;
23
- if (property === "confirm") return () => new Promise<boolean>(() => undefined);
24
- if (property === "select" || property === "input") return async () => undefined;
25
- return () => undefined;
26
- }
27
- });
19
+ const noop = () => undefined;
20
+ // Pi copies this context when wrapping prompts. Keep methods and the ownership
21
+ // marker enumerable so notification delivery and gateway detection survive.
22
+ return {
23
+ [GATEWAY_RUNTIME_UI_MARKER]: true,
24
+ theme,
25
+ confirm: () => new Promise<boolean>(() => undefined),
26
+ select: async () => undefined,
27
+ input: async () => undefined,
28
+ editor: async () => undefined,
29
+ custom: async () => undefined,
30
+ notify: noop,
31
+ onTerminalInput: () => noop,
32
+ setStatus: noop,
33
+ setWorkingMessage: noop,
34
+ setWorkingVisible: noop,
35
+ setWorkingIndicator: noop,
36
+ setHiddenThinkingLabel: noop,
37
+ setWidget: noop,
38
+ setFooter: noop,
39
+ setHeader: noop,
40
+ setTitle: noop,
41
+ pasteToEditor: noop,
42
+ setEditorText: noop,
43
+ getEditorText: () => "",
44
+ addAutocompleteProvider: noop,
45
+ setEditorComponent: noop,
46
+ getEditorComponent: noop,
47
+ getAllThemes: () => [],
48
+ getTheme: noop,
49
+ setTheme: () => ({ success: false, error: "UI not available" }),
50
+ getToolsExpanded: () => false,
51
+ setToolsExpanded: noop
52
+ };
28
53
  }