@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
@@ -0,0 +1,242 @@
1
+ import type { ExtensionAPI, ExtensionContext } from "@earendil-works/pi-coding-agent";
2
+
3
+ export const PIAGENT_FAST_MODE_STATE_ENTRY_TYPE = "piagent-fast-mode-state";
4
+ export const PIAGENT_SERVICE_TIER_RECEIPT_ENTRY_TYPE = "piagent-service-tier-receipt";
5
+ export const PIAGENT_FAST_SERVICE_TIER = "fast" as const;
6
+ export const PIAGENT_FAST_PROVIDER_SERVICE_TIER = "priority" as const;
7
+
8
+ type FastModeSource = "default" | "environment" | "session-command" | "session-state";
9
+
10
+ export type ServiceTierReceipt = {
11
+ schemaVersion: 1;
12
+ enabled: boolean;
13
+ mode: "standard" | "fast";
14
+ source: FastModeSource;
15
+ requestedServiceTier: "fast" | null;
16
+ observedRequestServiceTier: "priority" | null;
17
+ providerResponseServiceTier: null;
18
+ providerResponseEvidence: "unavailable-host-api";
19
+ provider: string | null;
20
+ modelId: string | null;
21
+ applied: boolean;
22
+ reasonCode: "fast-mode-disabled" | "provider-not-supported" | "fast-request-pending" | "fast-request-observed";
23
+ };
24
+
25
+ type SessionState = {
26
+ enabled: boolean;
27
+ source: FastModeSource;
28
+ observedRequestServiceTier: "priority" | null;
29
+ provider: string | null;
30
+ modelId: string | null;
31
+ observationPersisted: boolean;
32
+ };
33
+
34
+ type StateEntry = { type?: unknown; customType?: unknown; data?: unknown };
35
+
36
+ const SERVICE_TIER_RECEIPT_KEYS = [
37
+ "schemaVersion", "enabled", "mode", "source", "requestedServiceTier", "observedRequestServiceTier",
38
+ "providerResponseServiceTier", "providerResponseEvidence", "provider", "modelId", "applied", "reasonCode"
39
+ ] as const;
40
+ const FAST_MODE_SOURCES = new Set<FastModeSource>(["default", "environment", "session-command", "session-state"]);
41
+
42
+ function record(value: unknown): Record<string, unknown> | null {
43
+ return value !== null && typeof value === "object" && !Array.isArray(value) ? value as Record<string, unknown> : null;
44
+ }
45
+
46
+ function sessionId(ctx: ExtensionContext): string {
47
+ return ctx.sessionManager.getSessionId();
48
+ }
49
+
50
+ function environmentOverride(value: unknown): boolean | null {
51
+ if (typeof value !== "string" || !value.trim()) return null;
52
+ const normalized = value.trim().toLowerCase();
53
+ if (["1", "true", "yes", "on", "fast"].includes(normalized)) return true;
54
+ if (["0", "false", "no", "off", "standard", "default"].includes(normalized)) return false;
55
+ return null;
56
+ }
57
+
58
+ function restoredState(ctx: ExtensionContext): boolean | null {
59
+ let entries: StateEntry[] = [];
60
+ try { entries = ctx.sessionManager.getBranch() as StateEntry[]; } catch { return null; }
61
+ for (let index = entries.length - 1; index >= 0; index -= 1) {
62
+ const entry = entries[index], data = record(entry.data);
63
+ if (entry.type !== "custom" || entry.customType !== PIAGENT_FAST_MODE_STATE_ENTRY_TYPE || data?.schemaVersion !== 1) continue;
64
+ if (typeof data.enabled === "boolean") return data.enabled;
65
+ }
66
+ return null;
67
+ }
68
+
69
+ function isOpenAiCodex(provider: unknown): boolean {
70
+ return String(provider ?? "") === "openai-codex";
71
+ }
72
+
73
+ function boundedIdentity(value: unknown): value is string | null {
74
+ return value === null || typeof value === "string" && value.length <= 512 && !value.includes("\0");
75
+ }
76
+
77
+ export function parseServiceTierReceipt(value: unknown): ServiceTierReceipt | null {
78
+ const candidate = record(value);
79
+ if (!candidate || Object.keys(candidate).length !== SERVICE_TIER_RECEIPT_KEYS.length
80
+ || !SERVICE_TIER_RECEIPT_KEYS.every((key) => Object.hasOwn(candidate, key))) return null;
81
+ if (candidate.schemaVersion !== 1 || typeof candidate.enabled !== "boolean"
82
+ || !["standard", "fast"].includes(String(candidate.mode))
83
+ || !FAST_MODE_SOURCES.has(candidate.source as FastModeSource)
84
+ || ![null, PIAGENT_FAST_SERVICE_TIER].includes(candidate.requestedServiceTier as null | "fast")
85
+ || ![null, PIAGENT_FAST_PROVIDER_SERVICE_TIER].includes(candidate.observedRequestServiceTier as null | "priority")
86
+ || candidate.providerResponseServiceTier !== null || candidate.providerResponseEvidence !== "unavailable-host-api"
87
+ || !boundedIdentity(candidate.provider) || !boundedIdentity(candidate.modelId)
88
+ || typeof candidate.applied !== "boolean"
89
+ || !["fast-mode-disabled", "provider-not-supported", "fast-request-pending", "fast-request-observed"].includes(String(candidate.reasonCode))) return null;
90
+
91
+ const receipt = candidate as ServiceTierReceipt;
92
+ const expectedMode = receipt.enabled ? "fast" : "standard";
93
+ const expectedRequested = receipt.enabled ? PIAGENT_FAST_SERVICE_TIER : null;
94
+ const supported = isOpenAiCodex(receipt.provider);
95
+ const expectedReason = !receipt.enabled ? "fast-mode-disabled" : !supported ? "provider-not-supported"
96
+ : receipt.observedRequestServiceTier === PIAGENT_FAST_PROVIDER_SERVICE_TIER ? "fast-request-observed" : "fast-request-pending";
97
+ const expectedApplied = receipt.enabled && supported
98
+ && receipt.observedRequestServiceTier === PIAGENT_FAST_PROVIDER_SERVICE_TIER;
99
+ if (receipt.mode !== expectedMode || receipt.requestedServiceTier !== expectedRequested
100
+ || receipt.reasonCode !== expectedReason || receipt.applied !== expectedApplied
101
+ || (!receipt.enabled || !supported) && receipt.observedRequestServiceTier !== null) return null;
102
+ return { ...receipt };
103
+ }
104
+
105
+ export class ServiceTierRuntime {
106
+ readonly #environmentValue: string | undefined;
107
+ readonly #states = new Map<string, SessionState>();
108
+
109
+ constructor(options: { environmentValue?: string } = {}) {
110
+ this.#environmentValue = options.environmentValue;
111
+ }
112
+
113
+ restore(ctx: ExtensionContext): ServiceTierReceipt {
114
+ const environment = environmentOverride(this.#environmentValue);
115
+ const restored = environment === null ? restoredState(ctx) : null;
116
+ const enabled = environment ?? restored ?? false;
117
+ const source: FastModeSource = environment !== null ? "environment" : restored !== null ? "session-state" : "default";
118
+ this.#states.set(sessionId(ctx), {
119
+ enabled, source, observedRequestServiceTier: null, provider: null, modelId: null, observationPersisted: false
120
+ });
121
+ return this.receipt(ctx);
122
+ }
123
+
124
+ forget(ctx: ExtensionContext): void {
125
+ this.#states.delete(sessionId(ctx));
126
+ }
127
+
128
+ set(pi: ExtensionAPI, ctx: ExtensionContext, enabled: boolean): ServiceTierReceipt {
129
+ const state: SessionState = {
130
+ enabled, source: "session-command", observedRequestServiceTier: null, provider: null, modelId: null, observationPersisted: false
131
+ };
132
+ this.#states.set(sessionId(ctx), state);
133
+ pi.appendEntry(PIAGENT_FAST_MODE_STATE_ENTRY_TYPE, {
134
+ schemaVersion: 1,
135
+ enabled,
136
+ requestedServiceTier: enabled ? PIAGENT_FAST_SERVICE_TIER : null,
137
+ source: state.source
138
+ });
139
+ return this.receipt(ctx);
140
+ }
141
+
142
+ receipt(ctx: ExtensionContext): ServiceTierReceipt {
143
+ const state = this.#state(ctx);
144
+ const provider = typeof ctx.model?.provider === "string" ? ctx.model.provider : null;
145
+ const modelId = typeof ctx.model?.id === "string" ? ctx.model.id : null;
146
+ const supported = isOpenAiCodex(provider);
147
+ const observationMatchesCurrentModel = state.observedRequestServiceTier === PIAGENT_FAST_PROVIDER_SERVICE_TIER
148
+ && state.provider === provider && state.modelId === modelId;
149
+ const observed = observationMatchesCurrentModel ? PIAGENT_FAST_PROVIDER_SERVICE_TIER : null;
150
+ return {
151
+ schemaVersion: 1,
152
+ enabled: state.enabled,
153
+ mode: state.enabled ? "fast" : "standard",
154
+ source: state.source,
155
+ requestedServiceTier: state.enabled ? PIAGENT_FAST_SERVICE_TIER : null,
156
+ observedRequestServiceTier: observed,
157
+ providerResponseServiceTier: null,
158
+ providerResponseEvidence: "unavailable-host-api",
159
+ provider,
160
+ modelId,
161
+ applied: state.enabled && supported && observed === PIAGENT_FAST_PROVIDER_SERVICE_TIER,
162
+ reasonCode: !state.enabled ? "fast-mode-disabled" : !supported ? "provider-not-supported"
163
+ : observed === PIAGENT_FAST_PROVIDER_SERVICE_TIER ? "fast-request-observed" : "fast-request-pending"
164
+ };
165
+ }
166
+
167
+ applyProviderRequest(pi: ExtensionAPI, ctx: ExtensionContext, payloadValue: unknown): {
168
+ payload: unknown;
169
+ changed: boolean;
170
+ receipt: ServiceTierReceipt;
171
+ observationChanged: boolean;
172
+ } {
173
+ const state = this.#state(ctx), payload = record(payloadValue);
174
+ if (!state.enabled || !isOpenAiCodex(ctx.model?.provider) || !payload) {
175
+ return { payload: payloadValue, changed: false, receipt: this.receipt(ctx), observationChanged: false };
176
+ }
177
+ const next = payload.service_tier === PIAGENT_FAST_PROVIDER_SERVICE_TIER
178
+ ? payloadValue
179
+ : { ...payload, service_tier: PIAGENT_FAST_PROVIDER_SERVICE_TIER };
180
+ const observationChanged = state.observedRequestServiceTier !== PIAGENT_FAST_PROVIDER_SERVICE_TIER
181
+ || state.provider !== String(ctx.model?.provider ?? "") || state.modelId !== String(ctx.model?.id ?? "");
182
+ state.observedRequestServiceTier = PIAGENT_FAST_PROVIDER_SERVICE_TIER;
183
+ state.provider = String(ctx.model?.provider ?? "") || null;
184
+ state.modelId = String(ctx.model?.id ?? "") || null;
185
+ const receipt = this.receipt(ctx);
186
+ if (observationChanged && !state.observationPersisted) {
187
+ pi.appendEntry(PIAGENT_SERVICE_TIER_RECEIPT_ENTRY_TYPE, receipt);
188
+ state.observationPersisted = true;
189
+ }
190
+ return { payload: next, changed: next !== payloadValue, receipt, observationChanged };
191
+ }
192
+
193
+ #state(ctx: ExtensionContext): SessionState {
194
+ const id = sessionId(ctx), found = this.#states.get(id);
195
+ if (found) return found;
196
+ this.restore(ctx);
197
+ return this.#states.get(id)!;
198
+ }
199
+ }
200
+
201
+ function receiptText(receipt: ServiceTierReceipt): string {
202
+ return [
203
+ `fastMode: ${receipt.mode}`,
204
+ `provider/model: ${receipt.provider && receipt.modelId ? `${receipt.provider}/${receipt.modelId}` : "unknown"}`,
205
+ `requestedServiceTier: ${receipt.requestedServiceTier ?? "no Piagent override (host/provider configuration applies)"}`,
206
+ `observedRequestServiceTier: ${receipt.observedRequestServiceTier ?? "not-observed"}`,
207
+ `providerMapping: Fast mode uses OpenAI service_tier=priority`,
208
+ `providerResponseServiceTier: unavailable (Pi host API does not expose response.service_tier)`,
209
+ `state: ${receipt.reasonCode}`,
210
+ "modelThinkingParity: unchanged",
211
+ "note: Fast service can consume provider quota or cost differently; Fast mode never changes model or thinking"
212
+ ].join("\n");
213
+ }
214
+
215
+ export function registerFastModeCommand(pi: ExtensionAPI, runtime: ServiceTierRuntime): void {
216
+ pi.registerCommand("fast", {
217
+ description: "Show or switch OpenAI Codex Fast service tier without changing model/thinking",
218
+ getArgumentCompletions: (prefix: string) => ["status", "on", "off"]
219
+ .filter((value) => value.startsWith(String(prefix ?? "").trim().toLowerCase()))
220
+ .map((value) => ({ value, label: value })),
221
+ handler: async (args, ctx) => {
222
+ const action = String(args ?? "").trim().toLowerCase() || "status";
223
+ if (!["status", "on", "off"].includes(action)) {
224
+ ctx.ui.notify("Usage: /fast status|on|off", "warning");
225
+ return;
226
+ }
227
+ let receipt = action === "status" ? runtime.receipt(ctx) : runtime.set(pi, ctx, action === "on");
228
+ if (action === "on" && receipt.reasonCode === "provider-not-supported") {
229
+ ctx.ui.notify("Fast mode is enabled for the session, but the active provider is not OpenAI Codex.", "warning");
230
+ } else {
231
+ ctx.ui.notify(`Piagent Fast override: ${receipt.enabled ? "on" : "off"}`, receipt.enabled ? "info" : "warning");
232
+ }
233
+ receipt = runtime.receipt(ctx);
234
+ pi.sendMessage({
235
+ customType: "piagent-service-tier-receipt",
236
+ content: receiptText(receipt),
237
+ display: true,
238
+ details: receipt
239
+ }, { triggerTurn: false });
240
+ }
241
+ });
242
+ }
@@ -6,25 +6,26 @@ import path from "node:path";
6
6
 
7
7
  import { collectExactGitPatchAuthority } from "../inspection/diff-projection.ts";
8
8
  import { collectGitStatusForPaths, runReadOnlyGit } from "../inspection/git-status-adapter.ts";
9
- import { localFilterDisableArgs } from "../inspection/git-filter-safety.ts";
10
9
  import { collectIndexPreimage, collectSelectedWorkspacePreimage } from "../inspection/source-mutation-projection.ts";
10
+ import { readWorkspaceFile } from "../inspection/workspace-file-reader.ts";
11
11
  import type { SourceRevertAuthority } from "../inspection/source-revert-projection.ts";
12
12
  import type { SourceIndexTransactionResult } from "./source-index-transaction.ts";
13
13
 
14
14
  function result(value: Omit<SourceIndexTransactionResult, "executor" | "directExecution">): SourceIndexTransactionResult {
15
15
  return { ...value, executor: "pi-guard", directExecution: false };
16
16
  }
17
- function environment(): NodeJS.ProcessEnv {
17
+ function environment(extra: NodeJS.ProcessEnv = {}): NodeJS.ProcessEnv {
18
18
  return { PATH: process.env.PATH, LC_ALL: "C", LANG: "C", GIT_CONFIG_NOSYSTEM: "1", GIT_CONFIG_GLOBAL: os.devNull,
19
19
  GIT_OPTIONAL_LOCKS: "0", GIT_PAGER: "cat", PAGER: "cat", GIT_TERMINAL_PROMPT: "0", GIT_ASKPASS: "",
20
- ...(process.platform === "win32" ? { SystemRoot: process.env.SystemRoot, ComSpec: process.env.ComSpec } : {}) };
20
+ ...(process.platform === "win32" ? { SystemRoot: process.env.SystemRoot, ComSpec: process.env.ComSpec } : {}), ...extra };
21
21
  }
22
- async function runGit(repoRoot: string, args: string[], input: Buffer, timeoutMs: number): Promise<void> {
23
- const filterArgs = await localFilterDisableArgs(repoRoot, timeoutMs);
22
+ async function runGitInTemporaryTree(treeRoot: string, args: string[], input: Buffer, timeoutMs: number): Promise<void> {
24
23
  const argv = ["--no-pager", "--literal-pathspecs", "-c", "color.ui=false", "-c", "core.fsmonitor=false", "-c", `core.hooksPath=${os.devNull}`,
25
- "-c", "diff.external=", "-c", "submodule.recurse=false", ...filterArgs, "-C", repoRoot, ...args];
24
+ "-c", "diff.external=", "-c", "submodule.recurse=false", "-C", treeRoot, ...args];
26
25
  await new Promise<void>((resolve, reject) => {
27
- const child = spawn("git", argv, { cwd: repoRoot, env: environment(), shell: false, stdio: ["pipe", "ignore", "pipe"], windowsHide: true });
26
+ const child = spawn("git", argv, { cwd: treeRoot,
27
+ env: environment({ GIT_CEILING_DIRECTORIES: path.dirname(treeRoot), GIT_ATTR_NOSYSTEM: "1" }),
28
+ shell: false, stdio: ["pipe", "ignore", "pipe"], windowsHide: true });
28
29
  let stderr = 0, settled = false; child.stderr.on("data", (chunk: Buffer) => { stderr += chunk.length; if (stderr > 64 * 1024) child.kill("SIGKILL"); });
29
30
  child.stdin.on("error", () => undefined); child.stdin.end(input);
30
31
  const timer = setTimeout(() => child.kill("SIGKILL"), timeoutMs); timer.unref?.();
@@ -52,12 +53,59 @@ async function restoreWholeFile(authority: SourceRevertAuthority, recheck: () =>
52
53
  finally { fs.rmSync(temporary, { force: true }); }
53
54
  const actual = fs.readFileSync(target); if (!actual.equals(bytes)) throw new Error("revert-postcondition-failed");
54
55
  }
55
- async function restoreSelectedHunk(authority: SourceRevertAuthority, hunkRef: string, timeoutMs: number): Promise<void> {
56
+ type PreparedSelectedHunk = { stagingPath: string; target: string; initialMode: number };
57
+ function inside(root: string, candidate: string): boolean {
58
+ const relative = path.relative(root, candidate);
59
+ return relative === "" || relative !== ".." && !relative.startsWith(`..${path.sep}`) && !path.isAbsolute(relative);
60
+ }
61
+ function cleanupScratchRoot(scratchRoot: string): void {
62
+ try { fs.rmSync(scratchRoot, { recursive: true, force: true }); } catch {}
63
+ }
64
+ function cleanupStagingFile(stagingPath: string | null): void {
65
+ if (stagingPath) try { fs.rmSync(stagingPath, { force: true }); } catch {}
66
+ }
67
+ async function prepareSelectedHunk(authority: SourceRevertAuthority, hunkRef: string, timeoutMs: number): Promise<PreparedSelectedHunk | null> {
56
68
  const block = authority.patchAuthority.hunks.find((hunk) => hunk.hunkRef === hunkRef);
57
69
  if (!block) throw new Error("revert-hunk-stale");
58
70
  const patch = Buffer.concat([authority.patchAuthority.header, block.bytes]);
59
71
  if (patch.length > 2 * 1024 * 1024) throw new Error("revert-patch-oversized");
60
- await runGit(authority.repoRoot, ["apply", "--reverse", "--recount", "--whitespace=nowarn", "-"], patch, timeoutMs);
72
+ const repoRoot = fs.realpathSync.native(authority.repoRoot), target = path.resolve(repoRoot, authority.repoPath);
73
+ const parent = fs.realpathSync.native(path.dirname(target));
74
+ if (parent !== path.dirname(target) || !inside(repoRoot, target)) throw new Error("revert-target-unsafe");
75
+ const current = fs.lstatSync(target);
76
+ if (!current.isFile() || current.isSymbolicLink() || fs.realpathSync.native(target) !== target) throw new Error("revert-target-unsafe");
77
+ const scratchRoot = fs.realpathSync.native(fs.mkdtempSync(path.join(os.tmpdir(), `piagent-revert-tree.${process.pid}.`)));
78
+ fs.chmodSync(scratchRoot, 0o700);
79
+ const scratchTarget = path.resolve(scratchRoot, ...authority.repoPath.split("/"));
80
+ let stagingPath: string | null = null;
81
+ try {
82
+ if (!inside(scratchRoot, scratchTarget)) throw new Error("revert-target-unsafe");
83
+ fs.mkdirSync(path.dirname(scratchTarget), { recursive: true, mode: 0o700 });
84
+ const descriptor = fs.openSync(scratchTarget,
85
+ fs.constants.O_WRONLY | fs.constants.O_CREAT | fs.constants.O_EXCL | (fs.constants.O_NOFOLLOW ?? 0), 0o600);
86
+ try {
87
+ fs.writeFileSync(descriptor, readWorkspaceFile(repoRoot, authority.repoPath, 16 * 1024 * 1024));
88
+ fs.fchmodSync(descriptor, current.mode & 0o777); fs.fsyncSync(descriptor);
89
+ } finally { fs.closeSync(descriptor); }
90
+ if (collectSelectedWorkspacePreimage(scratchRoot, [authority.repoPath]) !== authority.target.workspacePreimage) return null;
91
+ await runGitInTemporaryTree(scratchRoot, ["apply", "--reverse", "--recount", "--whitespace=nowarn", "-"], patch, timeoutMs);
92
+ const result = fs.lstatSync(scratchTarget);
93
+ if (!result.isFile() || result.isSymbolicLink() || fs.realpathSync.native(scratchTarget) !== scratchTarget)
94
+ throw new Error("revert-target-unsafe");
95
+ const candidate = readWorkspaceFile(scratchRoot, authority.repoPath, 16 * 1024 * 1024);
96
+ stagingPath = path.join(parent, `.${path.basename(target)}.piagent-revert.${process.pid}.${randomBytes(8).toString("hex")}`);
97
+ const stagingDescriptor = fs.openSync(stagingPath,
98
+ fs.constants.O_WRONLY | fs.constants.O_CREAT | fs.constants.O_EXCL | (fs.constants.O_NOFOLLOW ?? 0), 0o600);
99
+ try {
100
+ fs.writeFileSync(stagingDescriptor, candidate);
101
+ fs.fchmodSync(stagingDescriptor, result.mode & 0o777);
102
+ fs.fsyncSync(stagingDescriptor);
103
+ } finally { fs.closeSync(stagingDescriptor); }
104
+ return { stagingPath, target, initialMode: current.mode & 0o777 };
105
+ } catch (error) {
106
+ cleanupStagingFile(stagingPath);
107
+ throw error;
108
+ } finally { cleanupScratchRoot(scratchRoot); }
61
109
  }
62
110
 
63
111
  export async function executeGuardedSourceWorktreeRevert(options: { authority: SourceRevertAuthority; expectedIndexPreimage: string;
@@ -74,13 +122,26 @@ export async function executeGuardedSourceWorktreeRevert(options: { authority: S
74
122
  if (!options.recheck() || lockedIndex !== options.expectedIndexPreimage || lockedWorkspace !== options.expectedWorkspacePreimage)
75
123
  return result({ state: "rejected", reasonCode: "mutation-preimage-stale", ...before,
76
124
  afterIndexPreimage: lockedIndex, afterWorkspacePreimage: lockedWorkspace });
77
- const finalRecheck = async () => options.recheck()
78
- && await collectIndexPreimage(options.authority.repoRoot) === options.expectedIndexPreimage
79
- && collectSelectedWorkspacePreimage(options.authority.repoRoot, [options.authority.repoPath]) === options.expectedWorkspacePreimage;
125
+ const finalRecheck = async () => {
126
+ if (!options.recheck()) return false;
127
+ const index = await collectIndexPreimage(options.authority.repoRoot);
128
+ const workspace = collectSelectedWorkspacePreimage(options.authority.repoRoot, [options.authority.repoPath]);
129
+ return index === options.expectedIndexPreimage && workspace === options.expectedWorkspacePreimage && options.recheck();
130
+ };
80
131
  if (options.authority.target.hunkRefs.length) {
81
- if (!await finalRecheck()) return result({ state: "rejected", reasonCode: "mutation-preimage-stale", ...before,
132
+ const prepared = await prepareSelectedHunk(options.authority, options.authority.target.hunkRefs[0],
133
+ Math.max(100, Math.min(60_000, options.timeoutMs ?? 5_000)));
134
+ if (!prepared) return result({ state: "rejected", reasonCode: "mutation-preimage-stale", ...before,
82
135
  afterIndexPreimage: null, afterWorkspacePreimage: null });
83
- await restoreSelectedHunk(options.authority, options.authority.target.hunkRefs[0], Math.max(100, Math.min(60_000, options.timeoutMs ?? 5_000)));
136
+ try {
137
+ if (!await finalRecheck()) return result({ state: "rejected", reasonCode: "mutation-preimage-stale", ...before,
138
+ afterIndexPreimage: null, afterWorkspacePreimage: null });
139
+ const current = fs.lstatSync(prepared.target);
140
+ if (!current.isFile() || current.isSymbolicLink() || (current.mode & 0o777) !== prepared.initialMode)
141
+ return result({ state: "rejected", reasonCode: "mutation-preimage-stale", ...before,
142
+ afterIndexPreimage: null, afterWorkspacePreimage: null });
143
+ fs.renameSync(prepared.stagingPath, prepared.target);
144
+ } finally { cleanupStagingFile(prepared.stagingPath); }
84
145
  } else await restoreWholeFile(options.authority, finalRecheck);
85
146
  committed = true;
86
147
  const afterIndexPreimage = await collectIndexPreimage(options.authority.repoRoot);
@@ -57,7 +57,7 @@ function commandProjection(events: ActivityInspectorEvent[]) {
57
57
  const decision = decisions.get(call.toolCallId), result = results.get(call.toolCallId);
58
58
  const blocked = decision?.decision === "blocked";
59
59
  const recovered = Boolean(call.toolCallId && recoveries.has(call.toolCallId));
60
- const failed = !blocked && !recovered && !handledToolFailure(result?.reasonCode) && Boolean(result && failedToolResult(result));
60
+ const failed = !blocked && !recovered && !handledToolFailure(result?.reasonCode, result?.toolName ?? call.toolName) && Boolean(result && failedToolResult(result));
61
61
  return { toolCallId: call.toolCallId ?? "unknown", command: call.command ?? "[redacted or unavailable]",
62
62
  status: blocked ? "blocked" : !result ? "running" : failed ? "failed" : "passed",
63
63
  exitCode: result?.exitCode ?? null, exitCodeExact: result?.exitCodeExact === true,
@@ -114,9 +114,12 @@ export async function buildActivityInspector(input: {
114
114
  eventCursor?: string;
115
115
  resyncRequired?: boolean;
116
116
  eventReplay?: { eventRetentionCount: number; eventRetentionSeconds: number };
117
+ operationLiveness?: "idle" | "running" | "unknown";
117
118
  }) {
118
119
  const projection = await buildWebUiInspectionProjection(input);
119
120
  const events = projection.scopedEvents, snapshot = projection.snapshot;
121
+ const current = snapshot.activity.running.length > 0 ? input.current ?? []
122
+ : (input.current ?? []).filter((activity) => (activity.status ?? "running") !== "running");
120
123
  const calls = events.filter((event) => event.event === "tool_call"), results = events.filter((event) => event.event === "tool_result");
121
124
  const byName = Object.fromEntries([...new Set(calls.map((event) => event.toolName ?? "unknown"))].sort().map((name) => [name, calls.filter((event) => (event.toolName ?? "unknown") === name).length]));
122
125
  const commands = commandProjection(events), files = sourceFiles(projection, input.task), safety = safetyProjection(events);
@@ -124,9 +127,9 @@ export async function buildActivityInspector(input: {
124
127
  schemaVersion: 2, version: ACTIVITY_INSPECTOR_VERSION, snapshot,
125
128
  state: { taskId: snapshot.identity.taskId, taskRunId: snapshot.identity.taskRunId,
126
129
  phase: snapshot.session.operation.hostPhase.state === "known" && snapshot.session.operation.hostPhase.value !== "idle" ? snapshot.session.operation.hostPhase.value : null,
127
- outcome: snapshot.task?.outcome ?? "idle", running: snapshot.activity.running.length, current: input.current ?? [] },
130
+ outcome: snapshot.task?.outcome ?? "idle", running: snapshot.activity.running.length, current },
128
131
  criteria: snapshot.task?.criteria ?? [], sourceChanges: snapshot.sourceChanges, files,
129
- tools: { calls: calls.length, results: results.length, failed: results.filter((event) => failedToolResult(event) && !handledToolFailure(event.reasonCode)).length,
132
+ tools: { calls: calls.length, results: results.length, failed: results.filter((event) => failedToolResult(event) && !handledToolFailure(event.reasonCode, event.toolName)).length,
130
133
  blocked: events.filter((event) => event.event === "tool_decision" && event.decision === "blocked").length, byName,
131
134
  perToolTokens: null, perToolTokensReason: "Pi reports model usage by response/turn; built-in tool calls do not carry attributable model-token totals." },
132
135
  commands, safety,
@@ -0,0 +1,26 @@
1
+ import type { TaskContract } from "../../extensions/guard-types.ts";
2
+ import { latestObservedVerificationEvidence, verificationEvidenceProvesStableTree } from "../../extensions/verification-intelligence.js";
3
+
4
+ export type CompletionSource = { approved: boolean; treeDigest: string | null; revisionDigest: string | null };
5
+
6
+ // A report consumes an existing gate result. It cannot approve completion.
7
+ // This metric means first observed edit reaching the accepted final source,
8
+ // not proof that an individual hunk alone solved the task.
9
+ export function firstCorrectEditTiming(task: TaskContract, records: Record<string, any>[],
10
+ completeTelemetry: boolean, source?: CompletionSource) {
11
+ const unavailable = { timeToFirstCorrectEditMs: null, timeToFirstCorrectEditReason: "edit-to-current-verifier-and-completion-evidence-unavailable" };
12
+ if (!completeTelemetry || !source?.approved || !source.treeDigest || !source.revisionDigest
13
+ || task.trace.outcome !== "completed" || !task.verifyCommands.length) return unavailable;
14
+ const latest = latestObservedVerificationEvidence(task.verifyEvidence);
15
+ const verifiers = task.verifyCommands.map((command) => latest.get(command.trim()));
16
+ if (!verifiers.every((entry) => verificationEvidenceProvesStableTree(entry, source.treeDigest, source.revisionDigest))) return unavailable;
17
+ const startedAt = Date.parse(task.createdAt);
18
+ const earliestVerifier = Math.min(...verifiers.map((entry: any) => Date.parse(entry.observedAt)));
19
+ const candidates = records.filter((event) => event.event === "tool_result" && event.taskUsageVersion === 1
20
+ && event.taskId === task.taskId && event.taskRunId === task.taskRunId && event.sessionId === task.sessionId
21
+ && event.mutationObserved === true && event.isError === false && typeof event.toolCallId === "string"
22
+ && event.postWorkingTreeDigest === source.treeDigest && event.postWorkspaceRevisionDigest === source.revisionDigest)
23
+ .map((event) => Date.parse(event.recordedAt)).filter((time) => Number.isFinite(time) && time >= startedAt && time < earliestVerifier);
24
+ return candidates.length ? { timeToFirstCorrectEditMs: Math.min(...candidates) - startedAt,
25
+ timeToFirstCorrectEditReason: "observed-edit-to-accepted-source; same-tree-and-revision-verifiers" } : unavailable;
26
+ }
@@ -5,6 +5,8 @@ import type { TaskContract } from "../../extensions/guard-types.ts";
5
5
  import { readSolverShadowEvents } from "../solver/solver-shadow.ts";
6
6
  import { readTrajectoryStore } from "../trajectory/trajectory-store.ts";
7
7
  import type { TrajectoryPhase } from "../trajectory/trajectory-types.ts";
8
+ import { readTaskUsage } from "./task-usage.ts";
9
+ import type { CompletionSource } from "./edit-verifier-timing.ts";
8
10
 
9
11
  export const TASK_EFFICIENCY_METRICS_VERSION = "task-efficiency-v1" as const;
10
12
 
@@ -44,12 +46,14 @@ function phaseDurations(task: TaskContract, trajectory: ReturnType<typeof readTr
44
46
  export function buildTaskEfficiencyMetrics(
45
47
  cwd: string,
46
48
  task: TaskContract,
47
- input: { activeToolGroups?: string[]; exactUsage?: ExactUsage } = {}
49
+ input: { activeToolGroups?: string[]; exactUsage?: ExactUsage; completionSource?: CompletionSource } = {}
48
50
  ) {
49
51
  const sessionHash = hash(task.sessionId);
50
52
  const solverStore = readSolverShadowEvents(cwd);
51
- const solverEvent = [...solverStore.records].reverse().find((event) => event.sessionHash === sessionHash);
52
53
  const trajectory = readTrajectoryStore(cwd, task.taskRunId);
54
+ const recommendation = trajectory.status === "ok" ? trajectory.state?.recommendationRef : null;
55
+ const solverEvent = recommendation ? [...solverStore.records].reverse().find((event) => event.sessionHash === sessionHash
56
+ && event.featureHash === recommendation.featureHash && hash(JSON.stringify(event.decision)) === recommendation.decisionDigest) : undefined;
53
57
  const decision = solverEvent?.decision;
54
58
  const helperUsage = task.acceptanceReceipt?.helperUsage;
55
59
  const provenance = task.acceptanceReceipt?.provenance;
@@ -58,9 +62,10 @@ export function buildTaskEfficiencyMetrics(
58
62
  const relevantFileMs = relevantEvent && contextEvidence.length > 0 ? duration(task.createdAt, relevantEvent.observedAt) : null;
59
63
  const recommendedTools = decision?.toolGroups ?? [];
60
64
  const activeTools = [...new Set(input.activeToolGroups ?? [])].sort();
65
+ const usage = readTaskUsage(cwd, task, input.completionSource);
61
66
  const exactUsage = input.exactUsage && (input.exactUsage.tokens !== null || input.exactUsage.cost !== null)
62
67
  ? input.exactUsage
63
- : { tokens: null, cost: null, currency: null, source: "unavailable" as const };
68
+ : { tokens: usage.tokens, cost: null, currency: null, source: usage.tokens === null ? "unavailable" as const : "pi-runtime" as const, scope: usage.scope };
64
69
 
65
70
  return {
66
71
  schemaVersion: 1,
@@ -95,12 +100,11 @@ export function buildTaskEfficiencyMetrics(
95
100
  recommendedGroups: recommendedTools,
96
101
  activeGroups: activeTools,
97
102
  activeRecommendedGroups: activeTools.filter((group) => recommendedTools.includes(group)),
98
- actualInvocationCounts: null
103
+ actualInvocationCounts: usage.actualInvocationCounts
99
104
  },
100
105
  timing: {
101
106
  taskDurationMs: duration(task.createdAt, task.updatedAt),
102
- timeToFirstCorrectEditMs: null,
103
- timeToFirstCorrectEditReason: "no persisted edit-to-verifier causal marker"
107
+ ...usage.editTiming
104
108
  },
105
109
  verification: {
106
110
  attempts: task.verifyEvidence.length,
@@ -130,6 +134,7 @@ export function buildTaskEfficiencyMetrics(
130
134
  acceptance: task.acceptanceReceipt?.criteria.map((item) => ({ id: item.id, status: item.status })) ?? [],
131
135
  recoveryDisposition: provenance?.finalRecoveryDisposition ?? "unknown"
132
136
  },
133
- exactUsage
137
+ exactUsage,
138
+ usageAccounting: usage
134
139
  };
135
140
  }
@@ -1,7 +1,8 @@
1
+ import { currentWorkspaceRevisionDigest } from "../../extensions/workspace-revision.js";
1
2
  import type { TaskContract } from "../../extensions/guard-types.ts";
2
3
  import { replayTaskCheckpoints } from "../../extensions/task-journal.js";
3
4
  import { taskContractValidationErrors, taskDigestMigrationArchiveStatus } from "../../extensions/task-state.js";
4
- import { latestObservedVerificationEvidence, verificationEvidenceProvesStableTree } from "../../extensions/verification-intelligence.js";
5
+ import { latestObservedVerificationEvidence, verificationEvidenceProvesCurrentWorkspace } from "../../extensions/verification-intelligence.js";
5
6
  import {
6
7
  isCurrentWorkingTreeDigest,
7
8
  WORKING_TREE_DIGEST_ALGORITHM,
@@ -79,7 +80,9 @@ export function buildLiveTaskStatus(cwd: string, task: TaskContract | undefined,
79
80
  const currentTreeDigest = isCurrentWorkingTreeDigest(resume.currentTreeDigest) ? resume.currentTreeDigest : null;
80
81
  const migrationRefresh = task.workingTreeDigestMigration?.status === "verification-refresh-required";
81
82
  const latestVerifiers = latestObservedVerificationEvidence(task.verifyEvidence);
82
- return { schemaVersion: 1, version: LIVE_TASK_STATUS_VERSION, state: task.trace.outcome === "pending" ? resume.enforcementSafe ? "active" : "corrupt" : "terminal", task: { taskId: task.taskId, taskRunId: task.taskRunId, sessionHash: resume.sessionId === task.sessionId ? "matched" : "conflict", attempt: task.attempt, maxAttempts: task.maxAttempts, phase: input.trajectory?.phase ?? resume.phase, outcome: task.trace.outcome, pendingVerifiers: task.verifyCommands.filter((command) => !currentTreeDigest || !verificationEvidenceProvesStableTree(latestVerifiers.get(command.trim()), currentTreeDigest)), pendingReview: task.workPlan.some((step) => step.id === "review" && step.status !== "done") }, tree: { ...digestMigration(task), currentDigest: currentTreeDigest }, runtime: input.runtime ?? null, modelRouting: input.modelRoute ?? null, activeToolGroups: input.activeToolGroups ?? [], recovery: { classification: evidence.failureClassification?.category ?? null, action: evidence.recovery?.action ?? null }, helpers: input.helpers ?? task.acceptanceReceipt?.helperUsage?.helpers ?? [], checkpoint: latest ? { id: latest.checkpointId, phase: latest.phase, status: latest.status, sequence: latest.sequence } : null, resume: { decision: resume.decision, safe: resume.enforcementSafe, staleVerifier: resume.staleVerifierEvidence, authorityPolicy: resume.authorityPolicy }, blockers: resume.warnings, nextSafeAction: task.trace.outcome !== "pending" ? "none-terminal" : migrationRefresh ? "rerun-exact-verifier" : resume.enforcementSafe ? resume.staleVerifierEvidence ? "rerun-exact-verifier" : "continue-current-phase" : "inspect-handoff-and-recover", receipt: task.trace.outcome === "pending" ? null : buildCompletionReceiptView(task, { cwd, gate: input.completionGate, modelRoute: input.modelRoute }), efficiency: buildTaskEfficiencyMetrics(cwd, task, { activeToolGroups: input.activeToolGroups }) };
83
+ const receipt = task.trace.outcome === "pending" ? null : buildCompletionReceiptView(task, { cwd, gate: input.completionGate, modelRoute: input.modelRoute });
84
+ const completionSource = { approved: receipt?.completionApproved === true && resume.enforcementSafe, treeDigest: currentTreeDigest, revisionDigest: receipt?.completionApproved ? currentWorkspaceRevisionDigest(cwd) : null };
85
+ return { schemaVersion: 1, version: LIVE_TASK_STATUS_VERSION, state: task.trace.outcome === "pending" ? resume.enforcementSafe ? "active" : "corrupt" : "terminal", task: { taskId: task.taskId, taskRunId: task.taskRunId, sessionHash: resume.sessionId === task.sessionId ? "matched" : "conflict", attempt: task.attempt, maxAttempts: task.maxAttempts, phase: input.trajectory?.phase ?? resume.phase, outcome: task.trace.outcome, pendingVerifiers: task.verifyCommands.filter((command) => !currentTreeDigest || !verificationEvidenceProvesCurrentWorkspace(latestVerifiers.get(command.trim()), currentTreeDigest, cwd)), pendingReview: task.workPlan.some((step) => step.id === "review" && step.status !== "done") }, tree: { ...digestMigration(task), currentDigest: currentTreeDigest }, runtime: input.runtime ?? null, modelRouting: input.modelRoute ?? null, activeToolGroups: input.activeToolGroups ?? [], recovery: { classification: evidence.failureClassification?.category ?? null, action: evidence.recovery?.action ?? null }, helpers: input.helpers ?? task.acceptanceReceipt?.helperUsage?.helpers ?? [], checkpoint: latest ? { id: latest.checkpointId, phase: latest.phase, status: latest.status, sequence: latest.sequence } : null, resume: { decision: resume.decision, safe: resume.enforcementSafe, staleVerifier: resume.staleVerifierEvidence, authorityPolicy: resume.authorityPolicy }, blockers: resume.warnings, nextSafeAction: task.trace.outcome !== "pending" ? "none-terminal" : migrationRefresh ? "rerun-exact-verifier" : resume.decision === "blocked" ? "inspect-handoff-and-recover" : resume.enforcementSafe ? resume.staleVerifierEvidence ? "rerun-exact-verifier" : "continue-current-phase" : "inspect-handoff-and-recover", receipt, efficiency: buildTaskEfficiencyMetrics(cwd, task, { activeToolGroups: input.activeToolGroups, completionSource }) };
83
86
  }
84
87
 
85
88
  export function formatLiveTaskStatus(view: ReturnType<typeof buildLiveTaskStatus>): string { if (!view.task) return `task: none\nmodelRoute: ${view.modelRouting ? `${view.modelRouting.disposition}/${view.modelRouting.capabilityBand}` : "none"}\nnext: start-or-resume-task`; const task = view.task; return [`task: ${task.taskId} (${task.taskRunId}); attempt=${task.attempt}/${task.maxAttempts}`, `state: ${view.state}; phase=${task.phase ?? "unknown"}; outcome=${task.outcome}`, `modelRoute: ${view.modelRouting ? `${view.modelRouting.disposition}; band=${view.modelRouting.capabilityBand}; target=${view.modelRouting.modelId ?? "none"}:${view.modelRouting.effort ?? "none"}; enforced=${view.modelRouting.enforced}` : "none"}`, `pending: verify=${task.pendingVerifiers.join(",") || "none"}; review=${task.pendingReview}`, `recovery: ${view.recovery.classification ?? "none"}/${view.recovery.action ?? "none"}`, `helpers: ${view.helpers.length}; checkpoint=${view.checkpoint?.id ?? "none"}`, `resume: ${view.resume.decision}; safe=${view.resume.safe}; staleVerifier=${view.resume.staleVerifier}; authority=${view.resume.authorityPolicy.disposition}/${view.resume.authorityPolicy.reason}`, `blockers: ${view.blockers.join("; ") || "none"}`, `next: ${view.nextSafeAction}`].join("\n"); }
@@ -92,7 +95,9 @@ export function buildCompletionReceiptView(task: TaskContract, input: { cwd?: st
92
95
  const recovery = task.acceptanceReceipt?.provenance;
93
96
  const helpers = task.acceptanceReceipt?.helperUsage;
94
97
  const latestVerifiers = latestObservedVerificationEvidence(task.verifyEvidence);
95
- const acceptanceSatisfied = task.acceptanceReceipt?.criteria.every((item) => item.status === "satisfied") === true;
98
+ const acceptanceRequired = task.changeMode === "source-change" && task.changedFiles.length > 0;
99
+ const acceptanceSatisfied = !acceptanceRequired
100
+ || task.acceptanceReceipt?.criteria.every((item) => item.status === "satisfied") === true;
96
101
  const treeEvidenceCurrent = taskTreeEvidenceIsCurrent(input.cwd, task, baselineDigest, currentDigest);
97
102
  const gateTreeCurrent = isCurrentWorkingTreeDigest(gate?.currentWorkingTreeDigest)
98
103
  && gate?.currentWorkingTreeDigest === currentDigest;
@@ -0,0 +1,78 @@
1
+ import crypto from "node:crypto";
2
+ import type { ExtensionAPI, ExtensionContext } from "@earendil-works/pi-coding-agent";
3
+ import type { TaskContract } from "../../extensions/guard-types.ts";
4
+
5
+ type Request = { requestId: string; taskId?: string; taskRunId?: string; sessionId: string; providerRequests: number };
6
+
7
+ // Metadata only, written to the existing telemetry stream. SDK turns are not
8
+ // provider billing receipts; missing terminals and internal retries stay unknown.
9
+ export function registerTaskUsageHooks(pi: ExtensionAPI, input: {
10
+ activeTask: (ctx: ExtensionContext) => TaskContract | undefined;
11
+ telemetry: (ctx: ExtensionContext, event: Record<string, unknown>) => void;
12
+ }): void {
13
+ const requests = new Map<string, Request>();
14
+ const messages = new WeakMap<object, Request>();
15
+ const calls = new Map<string, { request: Request; invoked: boolean }>();
16
+ const key = (ctx: ExtensionContext) => `${ctx.cwd}\u0000${ctx.sessionManager.getSessionId()}`;
17
+ const emit = (ctx: ExtensionContext, request: Request, event: string, data = {}) => input.telemetry(ctx, {
18
+ ...request, event, taskUsageVersion: 1, ...data
19
+ });
20
+ pi.on("turn_start", (_event, ctx) => {
21
+ const task = input.activeTask(ctx);
22
+ const request: Request = { requestId: crypto.randomUUID(), taskId: task?.taskId,
23
+ taskRunId: task?.taskRunId, sessionId: ctx.sessionManager.getSessionId(), providerRequests: 0 };
24
+ requests.set(key(ctx), request);
25
+ emit(ctx, request, "task_usage_request");
26
+ });
27
+ pi.on("before_provider_request", (_event, ctx) => {
28
+ const request = requests.get(key(ctx));
29
+ if (request) {
30
+ request.providerRequests += 1;
31
+ emit(ctx, request, "task_usage_dispatch");
32
+ }
33
+ });
34
+ const observe = (message: any, ctx: ExtensionContext, allowBind: boolean) => {
35
+ if (message?.role !== "assistant") return;
36
+ const request = messages.get(message) ?? (allowBind ? requests.get(key(ctx)) : undefined);
37
+ if (!request) return;
38
+ messages.set(message, request);
39
+ // Only numeric usage metadata crosses this boundary. No content, arguments,
40
+ // headers, opaque reasoning, or error text is retained.
41
+ const usage = message.usage;
42
+ const numeric = (value: unknown) => typeof value === "number" && Number.isFinite(value) && value >= 0 ? value : null;
43
+ emit(ctx, request, "task_usage_terminal", {
44
+ usageSemantics: "request-cumulative", usageRevision: "terminal",
45
+ stopReason: ["stop", "length", "toolUse", "error", "aborted"].includes(message.stopReason) ? message.stopReason : "unknown",
46
+ usage: Object.fromEntries(["input", "output", "cacheRead", "cacheWrite", "totalTokens"].map((field) => [field, numeric(usage?.[field])])),
47
+ cost: numeric(usage?.cost?.total), costBasis: "sdk-rate-estimate", currency: "USD"
48
+ });
49
+ };
50
+ pi.on("message_end", (event, ctx) => observe(event.message, ctx, true));
51
+ // The pinned SDK emits the same final message at both boundaries. An unknown
52
+ // replay object must not inherit the identity of a newer task at turn_end.
53
+ pi.on("turn_end", (event, ctx) => observe(event.message, ctx, false));
54
+ const callKey = (ctx: ExtensionContext, id: string) => `${key(ctx)}\u0000${id}`;
55
+ pi.on("tool_execution_start", (event, ctx) => {
56
+ const request = requests.get(key(ctx));
57
+ if (!request) return;
58
+ calls.set(callKey(ctx, event.toolCallId), { request, invoked: false });
59
+ emit(ctx, request, "task_tool_attempted", { toolCallId: event.toolCallId, toolName: event.toolName });
60
+ });
61
+ // SDK start events also cover blocked/invalid/truncated calls. Only the
62
+ // afterToolCall-backed tool_result hook proves execute was actually entered.
63
+ pi.on("tool_result", (event, ctx) => {
64
+ const call = calls.get(callKey(ctx, event.toolCallId));
65
+ if (!call) return;
66
+ call.invoked = true;
67
+ emit(ctx, call.request, "task_tool_invoked", { toolCallId: event.toolCallId, toolName: event.toolName });
68
+ });
69
+ pi.on("tool_execution_end", (event, ctx) => {
70
+ const id = callKey(ctx, event.toolCallId), call = calls.get(id);
71
+ if (call && !call.invoked) emit(ctx, call.request, "task_tool_not_executed", { toolCallId: event.toolCallId, toolName: event.toolName });
72
+ calls.delete(id);
73
+ });
74
+ pi.on("session_shutdown", (_event, ctx) => {
75
+ requests.delete(key(ctx));
76
+ for (const id of calls.keys()) if (id.startsWith(`${key(ctx)}\u0000`)) calls.delete(id);
77
+ });
78
+ }