@piagent/platform 1.6.0 → 1.7.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (736) hide show
  1. package/CHANGELOG.md +104 -0
  2. package/README.md +24 -15
  3. package/README.vi.md +36 -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 +27 -3
  10. package/benchmarks/production-v1/suite.json +5 -1
  11. package/benchmarks/production-v2/grade.mjs +892 -0
  12. package/benchmarks/production-v2/project/README.md +7 -0
  13. package/benchmarks/production-v2/project/config/service.json +4 -0
  14. package/benchmarks/production-v2/project/data/audit-ledger.jsonl +1 -0
  15. package/benchmarks/production-v2/project/docs/ops.md +3 -0
  16. package/benchmarks/production-v2/project/logs/incident.log +2 -0
  17. package/benchmarks/production-v2/project/package.json +11 -0
  18. package/benchmarks/production-v2/project/scripts/check.mjs +21 -0
  19. package/benchmarks/production-v2/project/src/backend/auth.js +3 -0
  20. package/benchmarks/production-v2/project/src/backend/billing-window.js +4 -0
  21. package/benchmarks/production-v2/project/src/backend/cache.js +15 -0
  22. package/benchmarks/production-v2/project/src/backend/invoice.js +4 -0
  23. package/benchmarks/production-v2/project/src/backend/revocation-cache.js +3 -0
  24. package/benchmarks/production-v2/project/src/data/csv.js +3 -0
  25. package/benchmarks/production-v2/project/src/data/dedup.js +5 -0
  26. package/benchmarks/production-v2/project/src/data/migration.js +9 -0
  27. package/benchmarks/production-v2/project/src/data/ndjson-stream.js +5 -0
  28. package/benchmarks/production-v2/project/src/data/versioned-replay.js +8 -0
  29. package/benchmarks/production-v2/project/src/frontend/chat-events.js +7 -0
  30. package/benchmarks/production-v2/project/src/frontend/pagination.js +8 -0
  31. package/benchmarks/production-v2/project/src/frontend/request-lifecycle.js +15 -0
  32. package/benchmarks/production-v2/project/src/frontend/search-state.js +8 -0
  33. package/benchmarks/production-v2/project/src/frontend/unicode-search.js +7 -0
  34. package/benchmarks/production-v2/project/src/fullstack/contract-sync.js +10 -0
  35. package/benchmarks/production-v2/project/src/platform/args.js +10 -0
  36. package/benchmarks/production-v2/project/src/platform/config.js +7 -0
  37. package/benchmarks/production-v2/project/src/platform/workflow-session.js +9 -0
  38. package/benchmarks/production-v2/project/src/platform/workspace.js +13 -0
  39. package/benchmarks/production-v2/project/src/reliability/checkpoint.js +5 -0
  40. package/benchmarks/production-v2/project/src/reliability/expiry.js +5 -0
  41. package/benchmarks/production-v2/project/src/reliability/retry.js +15 -0
  42. package/benchmarks/production-v2/project/test/smoke.test.js +21 -0
  43. package/benchmarks/production-v2/project/vendor/README.md +1 -0
  44. package/benchmarks/production-v2/prompts/abort-reconnect-supersession.md +11 -0
  45. package/benchmarks/production-v2/prompts/backend-frontend-contract-sync.md +12 -0
  46. package/benchmarks/production-v2/prompts/billing-cutoff-clock-skew.md +11 -0
  47. package/benchmarks/production-v2/prompts/bounded-retry.md +7 -0
  48. package/benchmarks/production-v2/prompts/chunked-record-boundary.md +9 -0
  49. package/benchmarks/production-v2/prompts/cli-double-dash.md +7 -0
  50. package/benchmarks/production-v2/prompts/config-precedence.md +6 -0
  51. package/benchmarks/production-v2/prompts/destructive-history-refusal.md +3 -0
  52. package/benchmarks/production-v2/prompts/expiry-boundary.md +7 -0
  53. package/benchmarks/production-v2/prompts/idempotent-replay-conflict.md +11 -0
  54. package/benchmarks/production-v2/prompts/incident-diagnosis.md +4 -0
  55. package/benchmarks/production-v2/prompts/invoice-rounding.md +9 -0
  56. package/benchmarks/production-v2/prompts/journeys/config-precedence-scout.md +3 -0
  57. package/benchmarks/production-v2/prompts/journeys/expiry-boundary-scout.md +3 -0
  58. package/benchmarks/production-v2/prompts/journeys/quoted-csv-scout.md +3 -0
  59. package/benchmarks/production-v2/prompts/journeys/recover-after-uncertain-send.md +4 -0
  60. package/benchmarks/production-v2/prompts/journeys/repository-prompt-injection-scout.md +4 -0
  61. package/benchmarks/production-v2/prompts/journeys/resumable-checkpoint-partial-failure-scout.md +4 -0
  62. package/benchmarks/production-v2/prompts/journeys/stable-dedup-scout.md +3 -0
  63. package/benchmarks/production-v2/prompts/journeys/stale-search-response-scout.md +3 -0
  64. package/benchmarks/production-v2/prompts/journeys/tenant-cache-isolation-scout.md +4 -0
  65. package/benchmarks/production-v2/prompts/journeys/verify-current-contract.md +4 -0
  66. package/benchmarks/production-v2/prompts/journeys/workflow-switch-same-session-scout.md +4 -0
  67. package/benchmarks/production-v2/prompts/pagination-boundary.md +7 -0
  68. package/benchmarks/production-v2/prompts/protected-env-refusal.md +2 -0
  69. package/benchmarks/production-v2/prompts/quoted-csv.md +6 -0
  70. package/benchmarks/production-v2/prompts/reconnect-chat-event-order.md +19 -0
  71. package/benchmarks/production-v2/prompts/repository-prompt-injection.md +5 -0
  72. package/benchmarks/production-v2/prompts/resumable-checkpoint-partial-failure.md +11 -0
  73. package/benchmarks/production-v2/prompts/revoked-session-cache.md +10 -0
  74. package/benchmarks/production-v2/prompts/schema-migration.md +6 -0
  75. package/benchmarks/production-v2/prompts/stable-dedup.md +6 -0
  76. package/benchmarks/production-v2/prompts/stale-search-response.md +6 -0
  77. package/benchmarks/production-v2/prompts/tenant-cache-isolation.md +6 -0
  78. package/benchmarks/production-v2/prompts/tenant-role-authorization.md +6 -0
  79. package/benchmarks/production-v2/prompts/unicode-search.md +6 -0
  80. package/benchmarks/production-v2/prompts/workflow-switch-same-session.md +30 -0
  81. package/benchmarks/production-v2/prompts/workspace-order.md +7 -0
  82. package/benchmarks/production-v2/spend-control.v1.json +115 -0
  83. package/benchmarks/production-v2/suite.json +1009 -0
  84. package/benchmarks/production-v2/variant.mjs +169 -0
  85. package/benchmarks/production-v3/CONTRACT-REVISION-20260922.md +31 -0
  86. package/benchmarks/production-v3/calibration/corpus.v1.json +63 -0
  87. package/benchmarks/production-v3/calibration/human-review-rubric.v1.json +36 -0
  88. package/benchmarks/production-v3/grade.mjs +1017 -0
  89. package/benchmarks/production-v3/project/README.md +7 -0
  90. package/benchmarks/production-v3/project/config/service.json +4 -0
  91. package/benchmarks/production-v3/project/data/audit-ledger.jsonl +1 -0
  92. package/benchmarks/production-v3/project/docs/ops.md +3 -0
  93. package/benchmarks/production-v3/project/logs/incident.log +2 -0
  94. package/benchmarks/production-v3/project/package.json +11 -0
  95. package/benchmarks/production-v3/project/scripts/check.mjs +21 -0
  96. package/benchmarks/production-v3/project/src/backend/auth.js +3 -0
  97. package/benchmarks/production-v3/project/src/backend/billing-window.js +4 -0
  98. package/benchmarks/production-v3/project/src/backend/cache.js +15 -0
  99. package/benchmarks/production-v3/project/src/backend/invoice.js +4 -0
  100. package/benchmarks/production-v3/project/src/backend/revocation-cache.js +12 -0
  101. package/benchmarks/production-v3/project/src/data/csv.js +3 -0
  102. package/benchmarks/production-v3/project/src/data/dedup.js +5 -0
  103. package/benchmarks/production-v3/project/src/data/migration.js +9 -0
  104. package/benchmarks/production-v3/project/src/data/ndjson-stream.js +5 -0
  105. package/benchmarks/production-v3/project/src/data/versioned-replay.js +8 -0
  106. package/benchmarks/production-v3/project/src/frontend/chat-events.js +7 -0
  107. package/benchmarks/production-v3/project/src/frontend/pagination.js +8 -0
  108. package/benchmarks/production-v3/project/src/frontend/request-lifecycle.js +23 -0
  109. package/benchmarks/production-v3/project/src/frontend/search-state.js +8 -0
  110. package/benchmarks/production-v3/project/src/frontend/unicode-search.js +7 -0
  111. package/benchmarks/production-v3/project/src/fullstack/contract-sync.js +10 -0
  112. package/benchmarks/production-v3/project/src/platform/args.js +10 -0
  113. package/benchmarks/production-v3/project/src/platform/config.js +7 -0
  114. package/benchmarks/production-v3/project/src/platform/workflow-session.js +9 -0
  115. package/benchmarks/production-v3/project/src/platform/workspace.js +13 -0
  116. package/benchmarks/production-v3/project/src/reliability/checkpoint.js +5 -0
  117. package/benchmarks/production-v3/project/src/reliability/expiry.js +14 -0
  118. package/benchmarks/production-v3/project/src/reliability/retry.js +15 -0
  119. package/benchmarks/production-v3/project/test/smoke.test.js +21 -0
  120. package/benchmarks/production-v3/project/vendor/README.md +1 -0
  121. package/benchmarks/production-v3/prompts/abort-reconnect-supersession.md +11 -0
  122. package/benchmarks/production-v3/prompts/backend-frontend-contract-sync.md +20 -0
  123. package/benchmarks/production-v3/prompts/billing-cutoff-clock-skew.md +11 -0
  124. package/benchmarks/production-v3/prompts/bounded-retry.md +7 -0
  125. package/benchmarks/production-v3/prompts/chunked-record-boundary.md +11 -0
  126. package/benchmarks/production-v3/prompts/cli-double-dash.md +7 -0
  127. package/benchmarks/production-v3/prompts/config-precedence.md +6 -0
  128. package/benchmarks/production-v3/prompts/destructive-history-refusal.md +9 -0
  129. package/benchmarks/production-v3/prompts/expiry-boundary.md +14 -0
  130. package/benchmarks/production-v3/prompts/idempotent-replay-conflict.md +13 -0
  131. package/benchmarks/production-v3/prompts/incident-diagnosis.md +4 -0
  132. package/benchmarks/production-v3/prompts/invoice-rounding.md +9 -0
  133. package/benchmarks/production-v3/prompts/journeys/config-precedence-scout.md +3 -0
  134. package/benchmarks/production-v3/prompts/journeys/expiry-boundary-scout.md +3 -0
  135. package/benchmarks/production-v3/prompts/journeys/quoted-csv-scout.md +3 -0
  136. package/benchmarks/production-v3/prompts/journeys/recover-after-uncertain-send.md +4 -0
  137. package/benchmarks/production-v3/prompts/journeys/repository-prompt-injection-scout.md +4 -0
  138. package/benchmarks/production-v3/prompts/journeys/resumable-checkpoint-partial-failure-scout.md +4 -0
  139. package/benchmarks/production-v3/prompts/journeys/stable-dedup-scout.md +3 -0
  140. package/benchmarks/production-v3/prompts/journeys/stale-search-response-scout.md +3 -0
  141. package/benchmarks/production-v3/prompts/journeys/tenant-cache-isolation-scout.md +4 -0
  142. package/benchmarks/production-v3/prompts/journeys/verify-current-contract.md +4 -0
  143. package/benchmarks/production-v3/prompts/journeys/workflow-switch-same-session-scout.md +4 -0
  144. package/benchmarks/production-v3/prompts/pagination-boundary.md +9 -0
  145. package/benchmarks/production-v3/prompts/protected-env-refusal.md +8 -0
  146. package/benchmarks/production-v3/prompts/quoted-csv.md +6 -0
  147. package/benchmarks/production-v3/prompts/reconnect-chat-event-order.md +20 -0
  148. package/benchmarks/production-v3/prompts/repository-prompt-injection.md +5 -0
  149. package/benchmarks/production-v3/prompts/resumable-checkpoint-partial-failure.md +15 -0
  150. package/benchmarks/production-v3/prompts/revoked-session-cache.md +10 -0
  151. package/benchmarks/production-v3/prompts/schema-migration.md +6 -0
  152. package/benchmarks/production-v3/prompts/stable-dedup.md +6 -0
  153. package/benchmarks/production-v3/prompts/stale-search-response.md +6 -0
  154. package/benchmarks/production-v3/prompts/tenant-cache-isolation.md +6 -0
  155. package/benchmarks/production-v3/prompts/tenant-role-authorization.md +6 -0
  156. package/benchmarks/production-v3/prompts/unicode-search.md +6 -0
  157. package/benchmarks/production-v3/prompts/workflow-switch-same-session.md +30 -0
  158. package/benchmarks/production-v3/prompts/workspace-order.md +7 -0
  159. package/benchmarks/production-v3/public-checks/abort-reconnect-supersession.js +34 -0
  160. package/benchmarks/production-v3/public-checks/expiry-boundary.js +29 -0
  161. package/benchmarks/production-v3/public-checks/revoked-session-cache.js +20 -0
  162. package/benchmarks/production-v3/spend-control.v1.json +121 -0
  163. package/benchmarks/production-v3/suite.json +1010 -0
  164. package/benchmarks/production-v3/variant.mjs +178 -0
  165. package/catalog/capabilities.json +16 -16
  166. package/docs/authority-manifest.md +1 -1
  167. package/docs/auto-delegation-policy.md +38 -73
  168. package/docs/benchmark-codex-mcp-metadata-contract-v1.md +159 -0
  169. package/docs/benchmark-codex-mcp-metadata-contract-v2.md +182 -0
  170. package/docs/benchmark-codex-mcp-metadata-contract-v3.md +194 -0
  171. package/docs/benchmark-public-contract-addendum-v2.md +85 -0
  172. package/docs/benchmark-public-contract-addendum-v3.md +116 -0
  173. package/docs/command-reference-vietnamese.md +30 -35
  174. package/docs/context-engine.md +5 -3
  175. package/docs/context-window-policy.md +37 -1
  176. package/docs/distribution-standard.md +9 -9
  177. package/docs/en/architecture.md +26 -13
  178. package/docs/languages.json +4 -4
  179. package/docs/operator-manual-vietnamese.md +26 -46
  180. package/docs/piagent-codex-benchmark-recovery-plan.md +1817 -0
  181. package/docs/piagent-core-evaluation-specs-2026-09-06.json +706 -0
  182. package/docs/piagent-core-improvement-plan.md +528 -0
  183. package/docs/piagent-core-offline-implementation-2026-09-06.md +219 -0
  184. package/docs/piagent-core-plan-checkpoint-2026-09-06.json +69 -0
  185. package/docs/piagent-core-reasoning-research-2026-09-06.md +225 -0
  186. package/docs/piagent-core-research-sources-2026-09-06.json +416 -0
  187. package/docs/piagent-open-session-wait.md +19 -0
  188. package/docs/piagent-runtime-benchmark-execution-plan.md +685 -0
  189. package/docs/publishing-for-teams.md +7 -0
  190. package/docs/quality-benchmark.md +299 -23
  191. package/docs/quickstart-vietnamese.md +2 -2
  192. package/docs/release-install-policy.md +32 -9
  193. package/docs/research/harness-next-evidence-review.md +113 -0
  194. package/docs/runtime-harness-standard.md +3 -3
  195. package/docs/security-threat-model.md +1 -1
  196. package/docs/subagent-orchestration-capabilities.md +19 -51
  197. package/docs/subagents-and-multiagent.md +21 -47
  198. package/docs/task-implementation-contract.md +3 -3
  199. package/docs/task-lifecycle-vietnamese.md +1 -1
  200. package/docs/team-onboarding.md +7 -7
  201. package/docs/third-party-neutrality.md +7 -0
  202. package/docs/usage-observability.md +1 -1
  203. package/docs/vi/architecture.md +26 -13
  204. package/docs-site/architecture.html +1 -1
  205. package/docs-site/benchmark.html +89 -95
  206. package/docs-site/commands.html +1 -1
  207. package/docs-site/content/benchmark.html +88 -94
  208. package/docs-site/content/en/benchmark.html +134 -90
  209. package/docs-site/content/en/index.html +19 -19
  210. package/docs-site/content/en/runtime.html +5 -5
  211. package/docs-site/content/en/whats-new.html +23 -1
  212. package/docs-site/content/en/workflows.html +2 -2
  213. package/docs-site/content/index.html +19 -19
  214. package/docs-site/content/runtime.html +6 -6
  215. package/docs-site/content/whats-new.html +23 -1
  216. package/docs-site/content/workflows.html +2 -2
  217. package/docs-site/context-engine.html +1 -1
  218. package/docs-site/en/architecture.html +1 -1
  219. package/docs-site/en/benchmark.html +135 -91
  220. package/docs-site/en/commands.html +1 -1
  221. package/docs-site/en/context-engine.html +1 -1
  222. package/docs-site/en/index.html +20 -20
  223. package/docs-site/en/mcp-approval.html +1 -1
  224. package/docs-site/en/mcp-auth.html +1 -1
  225. package/docs-site/en/mcp-servers.html +1 -1
  226. package/docs-site/en/mcp.html +1 -1
  227. package/docs-site/en/permissions.html +1 -1
  228. package/docs-site/en/profiles.html +1 -1
  229. package/docs-site/en/quickstart.html +1 -1
  230. package/docs-site/en/runtime.html +6 -6
  231. package/docs-site/en/scope.html +1 -1
  232. package/docs-site/en/security.html +1 -1
  233. package/docs-site/en/team.html +1 -1
  234. package/docs-site/en/whats-new.html +24 -2
  235. package/docs-site/en/workflows.html +3 -3
  236. package/docs-site/index.html +20 -20
  237. package/docs-site/mcp-approval.html +1 -1
  238. package/docs-site/mcp-auth.html +1 -1
  239. package/docs-site/mcp-servers.html +1 -1
  240. package/docs-site/mcp.html +1 -1
  241. package/docs-site/permissions.html +1 -1
  242. package/docs-site/profiles.html +1 -1
  243. package/docs-site/quickstart.html +1 -1
  244. package/docs-site/runtime.html +7 -7
  245. package/docs-site/scope.html +1 -1
  246. package/docs-site/security.html +1 -1
  247. package/docs-site/team.html +1 -1
  248. package/docs-site/whats-new.html +24 -2
  249. package/docs-site/workflows.html +3 -3
  250. package/evals/architecture-conformance-v1/lane.json +18 -0
  251. package/evals/architecture-conformance-v1/runner.mjs +77 -0
  252. package/evals/fixtures/approved-host-contracts.invalid.json +43 -0
  253. package/evals/fixtures/approved-host-contracts.valid.json +43 -0
  254. package/evals/fixtures/benchmark-assurance-evidence.invalid.json +1 -1
  255. package/evals/fixtures/benchmark-assurance-evidence.valid.json +1 -1
  256. package/evals/fixtures/benchmark-attempt-outcome.invalid.json +42 -0
  257. package/evals/fixtures/benchmark-attempt-outcome.valid.json +42 -0
  258. package/evals/fixtures/benchmark-independent-verification-plan.invalid.json +7 -0
  259. package/evals/fixtures/benchmark-independent-verification-plan.valid.json +21 -0
  260. package/evals/fixtures/codex-relative-efficiency-v1.invalid.json +4 -0
  261. package/evals/fixtures/codex-relative-efficiency-v1.valid.json +94 -0
  262. package/evals/fixtures/composite-criterion.invalid.json +10 -0
  263. package/evals/fixtures/composite-criterion.valid.json +33 -0
  264. package/evals/fixtures/contract-family-library.invalid.json +42 -0
  265. package/evals/fixtures/contract-family-library.valid.json +41 -0
  266. package/evals/fixtures/contract-selection-recipe.invalid.json +29 -0
  267. package/evals/fixtures/contract-selection-recipe.valid.json +29 -0
  268. package/evals/fixtures/helper-request.invalid.json +1 -1
  269. package/evals/fixtures/helper-request.valid.json +1 -1
  270. package/evals/fixtures/host-contract-plan.invalid.json +42 -0
  271. package/evals/fixtures/host-contract-plan.valid.json +41 -0
  272. package/evals/fixtures/piagent-webui/subagent-tree-v1.valid.json +7 -0
  273. package/evals/fixtures/scoped-project-verification-receipt-v2.invalid.json +42 -0
  274. package/evals/fixtures/scoped-project-verification-receipt-v2.valid.json +42 -0
  275. package/evals/fixtures/scoped-project-verification-receipt-v3.invalid.json +57 -0
  276. package/evals/fixtures/scoped-project-verification-receipt-v3.valid.json +57 -0
  277. package/evals/fixtures/scoped-verification-receipt-v1.invalid.json +40 -0
  278. package/evals/fixtures/scoped-verification-receipt-v1.valid.json +40 -0
  279. package/evals/fixtures/task-authority-snapshot.valid.json +4 -4
  280. package/evals/fs4-readiness-matrix.v1.json +2 -2
  281. package/evals/fs5-artifact-archive/README.md +32 -0
  282. package/evals/fs5-artifact-archive/manifest.json +148 -0
  283. package/evals/fs5-artifact-archive/sha256/1f073f6d42c093a3182685e0e096144c8b3befd87594aeea1af05ce17653860c.txt +380 -0
  284. package/evals/fs5-artifact-archive/sha256/1ff46984c5ff668647dfad94248cd9dcfb72384a8aa529279345582f90d7486a.txt +331 -0
  285. package/evals/fs5-artifact-archive/sha256/20a5e1ef1cab3234a888b876b3cec52e702a05bc5ac000b691f94c651beed9ca.txt +800 -0
  286. package/evals/fs5-artifact-archive/sha256/29bb79e96b047d79056626d747baa0b3a6ca9d4da7ad80b9668c0085e5581706.txt +116 -0
  287. package/evals/fs5-artifact-archive/sha256/2bcc0cf6c18eee0c080c30e727610b0f10a278f5862855d920ba5797f07a2f2d.txt +252 -0
  288. package/evals/fs5-artifact-archive/sha256/37ecc65febe2c5947ac9288ad958bc149e0f08c6beda137edb1c660908853832.txt +324 -0
  289. package/evals/fs5-artifact-archive/sha256/5d937632d2cfff34dbbb781dd367bb78feb64c36d571c36ca15f2cc40c576af8.txt +142 -0
  290. package/evals/fs5-artifact-archive/sha256/6057c7bf2f40f8e3db7e293f08d745133c05293b75e9f363fb4e2e53bd09a76a.txt +244 -0
  291. package/evals/fs5-artifact-archive/sha256/62ea354afb6816b148ae9bd95ffb1ee5e29580eb4a22473e1ae19dbb471ec36b.txt +930 -0
  292. package/evals/fs5-artifact-archive/sha256/8987775b742f3b3650d75cce84e6d58fcdd3dfb835d03dac4c2bfe6d9eea3a2e.txt +903 -0
  293. package/evals/fs5-artifact-archive/sha256/902e886d0920b66811f54d0d1d933d70f8fab8704813b46d578bee6af1f21eef.txt +942 -0
  294. package/evals/fs5-artifact-archive/sha256/976a9908adfd765ee7be0655616f39a44fff64824f5e8535d726082e550b02eb.txt +910 -0
  295. package/evals/fs5-artifact-archive/sha256/9c07bda5bfe7bb7b0959844ccea6790735cc9f154d8d5972e18785148cea6d9a.txt +101 -0
  296. package/evals/fs5-artifact-archive/sha256/b936a604db87a697bf69581c33d753e26d9e5e2ccce40bf9c9b2070948f49919.txt +931 -0
  297. package/evals/fs5-artifact-archive/sha256/bf91f69473f48fc3de7c7c55e7ce893ba0511074c4ec930d9f3bcb9615d8e9d5.txt +950 -0
  298. package/evals/fs5-artifact-archive/sha256/ce4ed1b4a033f2a8cb4632d1336b0867edb3454ce1380f0642c4055149e6ac59.txt +784 -0
  299. package/evals/fs5-artifact-archive/sha256/db97f444db8ba604512daf0aa26b3fe1e2a120ace2b21d4730db5b18aa0c528f.txt +950 -0
  300. package/evals/fs5-causal-arm.v1.json +10 -6
  301. package/evals/fs5-pilot-protocol.v1.json +3 -3
  302. package/evals/fs5-pilot-protocol.v2.json +8 -8
  303. package/evals/fs5-pilot-protocol.v3.json +8 -8
  304. package/evals/fs5-pilot-protocol.v4.json +7 -7
  305. package/evals/fs5-pilot-protocol.v5.json +10 -8
  306. package/evals/harness-next/development-corpus.mjs +106 -0
  307. package/evals/harness-next/resource-budget-probe.mjs +38 -0
  308. package/evals/harness-next/run-calibration.mjs +83 -0
  309. package/evals/harness-next/run-resource-probe.mjs +55 -0
  310. package/evals/harness-next/selected-development-corpus.mjs +27 -0
  311. package/evals/harness-next.v1.json +51 -0
  312. package/evals/long-horizon-v1/README.md +19 -0
  313. package/evals/long-horizon-v1/runner.mjs +116 -48
  314. package/evals/long-horizon-v1/worker.mjs +77 -30
  315. package/evals/private-holdout-v1/CURRENT_BOUNDARY.md +53 -0
  316. package/evals/private-holdout-v1/README.md +5 -2
  317. package/evals/private-holdout-v1/public-exposure.v1.json +1 -1
  318. package/evals/private-holdout-v1/public-exposure.v2.json +273 -0
  319. package/evals/runtime-conformance-v1/README.md +43 -0
  320. package/evals/runtime-conformance-v1/evaluate-paid-canary.mjs +198 -0
  321. package/evals/runtime-conformance-v1/lane.json +57 -0
  322. package/evals/runtime-conformance-v1/paid-canary.json +31 -0
  323. package/evals/runtime-conformance-v1/runner.mjs +456 -0
  324. package/package.json +9 -4
  325. package/packages/piagent-core/README.md +15 -6
  326. package/packages/piagent-core/benchmark/benchmark-all-attempt-efficiency.js +55 -0
  327. package/packages/piagent-core/benchmark/benchmark-artifact-binding.js +49 -0
  328. package/packages/piagent-core/benchmark/benchmark-bootstrap.js +265 -14
  329. package/packages/piagent-core/benchmark/benchmark-budget-governor.js +271 -0
  330. package/packages/piagent-core/benchmark/benchmark-campaign.js +592 -0
  331. package/packages/piagent-core/benchmark/benchmark-candidate.js +8 -3
  332. package/packages/piagent-core/benchmark/benchmark-checkpoint.js +67 -0
  333. package/packages/piagent-core/benchmark/benchmark-claim-restrictions.js +269 -2
  334. package/packages/piagent-core/benchmark/benchmark-cli.js +120 -3
  335. package/packages/piagent-core/benchmark/benchmark-codex-events.js +235 -0
  336. package/packages/piagent-core/benchmark/benchmark-codex-outcome.js +20 -0
  337. package/packages/piagent-core/benchmark/benchmark-codex-relative-efficiency.js +699 -0
  338. package/packages/piagent-core/benchmark/benchmark-codex-rollout.js +521 -0
  339. package/packages/piagent-core/benchmark/benchmark-codex.js +104 -8
  340. package/packages/piagent-core/benchmark/benchmark-comparison.js +284 -1
  341. package/packages/piagent-core/benchmark/benchmark-core.js +111 -110
  342. package/packages/piagent-core/benchmark/benchmark-diagnostic-source-evidence.js +40 -0
  343. package/packages/piagent-core/benchmark/benchmark-diagnostic-treatment.js +104 -0
  344. package/packages/piagent-core/benchmark/benchmark-evaluator-v3.js +226 -0
  345. package/packages/piagent-core/benchmark/benchmark-execution-guard.js +23 -1
  346. package/packages/piagent-core/benchmark/benchmark-forensics.js +67 -52
  347. package/packages/piagent-core/benchmark/benchmark-goal-assessment.js +193 -0
  348. package/packages/piagent-core/benchmark/benchmark-goal-report.js +87 -0
  349. package/packages/piagent-core/benchmark/benchmark-independent-verification-observation.js +64 -0
  350. package/packages/piagent-core/benchmark/benchmark-matrix-summary.js +141 -0
  351. package/packages/piagent-core/benchmark/benchmark-matrix.js +1 -0
  352. package/packages/piagent-core/benchmark/benchmark-normalized-cost.js +62 -0
  353. package/packages/piagent-core/benchmark/benchmark-pi-home.js +1 -0
  354. package/packages/piagent-core/benchmark/benchmark-preflight.js +70 -7
  355. package/packages/piagent-core/benchmark/benchmark-process.js +23 -9
  356. package/packages/piagent-core/benchmark/benchmark-production-completion.js +84 -0
  357. package/packages/piagent-core/benchmark/benchmark-production-stage-validation.js +209 -0
  358. package/packages/piagent-core/benchmark/benchmark-provider-free-evidence.js +462 -0
  359. package/packages/piagent-core/benchmark/benchmark-provider-free-owner.js +24 -0
  360. package/packages/piagent-core/benchmark/benchmark-provider-wire.js +378 -0
  361. package/packages/piagent-core/benchmark/benchmark-record-validation.d.ts +2 -0
  362. package/packages/piagent-core/benchmark/benchmark-record-validation.js +343 -8
  363. package/packages/piagent-core/benchmark/benchmark-release-claim-controls.js +132 -0
  364. package/packages/piagent-core/benchmark/benchmark-report.js +94 -5
  365. package/packages/piagent-core/benchmark/benchmark-resume-recovery.js +280 -10
  366. package/packages/piagent-core/benchmark/benchmark-runner-policy.js +88 -1
  367. package/packages/piagent-core/benchmark/benchmark-runtime-causal.js +244 -0
  368. package/packages/piagent-core/benchmark/benchmark-runtime-identity.js +50 -1
  369. package/packages/piagent-core/benchmark/benchmark-runtime.js +102 -10
  370. package/packages/piagent-core/benchmark/benchmark-safety-evidence.js +74 -0
  371. package/packages/piagent-core/benchmark/benchmark-service-tier.js +189 -0
  372. package/packages/piagent-core/benchmark/benchmark-stage-diagnostic.js +246 -33
  373. package/packages/piagent-core/benchmark/benchmark-suite-identity.js +264 -1
  374. package/packages/piagent-core/benchmark/benchmark-suite-runtime.js +4 -0
  375. package/packages/piagent-core/benchmark/benchmark-suite.js +232 -4
  376. package/packages/piagent-core/benchmark/benchmark-summary-support.js +33 -1
  377. package/packages/piagent-core/benchmark/benchmark-timing-contract.js +189 -0
  378. package/packages/piagent-core/benchmark/benchmark-timing-diagnostics.js +110 -139
  379. package/packages/piagent-core/benchmark/benchmark-transport-evidence.js +203 -0
  380. package/packages/piagent-core/benchmark/benchmark-transport-observation.js +147 -0
  381. package/packages/piagent-core/benchmark/benchmark-tree-identity.js +2 -1
  382. package/packages/piagent-core/benchmark/benchmark-usage.js +410 -37
  383. package/packages/piagent-core/benchmark/benchmark-workflow.js +136 -38
  384. package/packages/piagent-core/benchmark/fs5-pilot-protocol.js +21 -2
  385. package/packages/piagent-core/capabilities/runtime-integrity.js +6 -0
  386. package/packages/piagent-core/extensions/acceptance-api-baseline.js +24 -0
  387. package/packages/piagent-core/extensions/acceptance-assessment.js +153 -0
  388. package/packages/piagent-core/extensions/acceptance-authenticated-admission.d.ts +41 -0
  389. package/packages/piagent-core/extensions/acceptance-authenticated-admission.js +120 -0
  390. package/packages/piagent-core/extensions/acceptance-behavior-proof.js +1000 -0
  391. package/packages/piagent-core/extensions/acceptance-boundary-binding.js +31 -0
  392. package/packages/piagent-core/extensions/acceptance-boundary-guidance.js +214 -0
  393. package/packages/piagent-core/extensions/acceptance-callable-scanner.js +244 -0
  394. package/packages/piagent-core/extensions/acceptance-clause-group.js +45 -0
  395. package/packages/piagent-core/extensions/acceptance-composite-contract.js +334 -0
  396. package/packages/piagent-core/extensions/acceptance-contract-conjunction.js +78 -0
  397. package/packages/piagent-core/extensions/acceptance-contract-selection.js +161 -0
  398. package/packages/piagent-core/extensions/acceptance-contract-semantics.js +211 -211
  399. package/packages/piagent-core/extensions/acceptance-declared-api.js +126 -0
  400. package/packages/piagent-core/extensions/acceptance-diagnostic-policy.ts +14 -0
  401. package/packages/piagent-core/extensions/acceptance-document-content.js +73 -0
  402. package/packages/piagent-core/extensions/acceptance-durable-execution.d.ts +59 -0
  403. package/packages/piagent-core/extensions/acceptance-durable-execution.js +139 -0
  404. package/packages/piagent-core/extensions/acceptance-error-classes.js +1 -1
  405. package/packages/piagent-core/extensions/acceptance-evidence-store.js +223 -0
  406. package/packages/piagent-core/extensions/acceptance-executable-evidence.js +490 -88
  407. package/packages/piagent-core/extensions/acceptance-execution-diagnostics.js +34 -0
  408. package/packages/piagent-core/extensions/acceptance-execution-snapshot.js +174 -0
  409. package/packages/piagent-core/extensions/acceptance-executor/.dockerignore +21 -0
  410. package/packages/piagent-core/extensions/acceptance-executor/Dockerfile +5 -0
  411. package/packages/piagent-core/extensions/acceptance-executor/README.md +103 -0
  412. package/packages/piagent-core/extensions/acceptance-executor/budget.mjs +54 -0
  413. package/packages/piagent-core/extensions/acceptance-executor/callback-intrinsics.mjs +126 -0
  414. package/packages/piagent-core/extensions/acceptance-executor/capabilities.mjs +127 -0
  415. package/packages/piagent-core/extensions/acceptance-executor/guest.mjs +285 -0
  416. package/packages/piagent-core/extensions/acceptance-executor/intrinsics.mjs +133 -0
  417. package/packages/piagent-core/extensions/acceptance-executor/invocation.mjs +39 -0
  418. package/packages/piagent-core/extensions/acceptance-executor/jobs.mjs +29 -0
  419. package/packages/piagent-core/extensions/acceptance-executor/module-graph.mjs +70 -0
  420. package/packages/piagent-core/extensions/acceptance-executor/node-profile.mjs +398 -0
  421. package/packages/piagent-core/extensions/acceptance-executor/package-lock.json +40 -0
  422. package/packages/piagent-core/extensions/acceptance-executor/package.json +10 -0
  423. package/packages/piagent-core/extensions/acceptance-executor/protocol.mjs +248 -0
  424. package/packages/piagent-core/extensions/acceptance-executor/reference-identity.mjs +61 -0
  425. package/packages/piagent-core/extensions/acceptance-executor/references.mjs +14 -0
  426. package/packages/piagent-core/extensions/acceptance-executor/structured-clone.mjs +53 -0
  427. package/packages/piagent-core/extensions/acceptance-executor/values.mjs +97 -0
  428. package/packages/piagent-core/extensions/acceptance-executor/worker.mjs +75 -0
  429. package/packages/piagent-core/extensions/acceptance-expiry-calendar-proof.js +162 -0
  430. package/packages/piagent-core/extensions/acceptance-expression-parser.js +155 -0
  431. package/packages/piagent-core/extensions/acceptance-focused-verification.js +40 -0
  432. package/packages/piagent-core/extensions/acceptance-host-configuration.js +337 -0
  433. package/packages/piagent-core/extensions/acceptance-identifier-criteria.js +25 -0
  434. package/packages/piagent-core/extensions/acceptance-independent-contract.js +209 -0
  435. package/packages/piagent-core/extensions/acceptance-independent-registry.js +106 -0
  436. package/packages/piagent-core/extensions/acceptance-input-provenance.js +40 -0
  437. package/packages/piagent-core/extensions/acceptance-integer-domains.js +322 -0
  438. package/packages/piagent-core/extensions/acceptance-invalid-input-receipt.js +20 -0
  439. package/packages/piagent-core/extensions/acceptance-invalid-partitions.js +26 -0
  440. package/packages/piagent-core/extensions/acceptance-invalid-sentinel-proof.js +477 -0
  441. package/packages/piagent-core/extensions/acceptance-isolated-executor.js +254 -0
  442. package/packages/piagent-core/extensions/acceptance-lexical-sentinels.js +130 -0
  443. package/packages/piagent-core/extensions/acceptance-literal-dataflow.js +735 -0
  444. package/packages/piagent-core/extensions/acceptance-local-binding.js +106 -0
  445. package/packages/piagent-core/extensions/acceptance-normalizer-evidence.js +719 -0
  446. package/packages/piagent-core/extensions/acceptance-obligations.js +75 -0
  447. package/packages/piagent-core/extensions/acceptance-parameter-contract.js +756 -0
  448. package/packages/piagent-core/extensions/acceptance-precedence-contract.js +572 -0
  449. package/packages/piagent-core/extensions/acceptance-proof-capability.js +26 -0
  450. package/packages/piagent-core/extensions/acceptance-receipt.js +208 -180
  451. package/packages/piagent-core/extensions/acceptance-result-contract-evidence.js +159 -0
  452. package/packages/piagent-core/extensions/acceptance-source-origin.js +214 -0
  453. package/packages/piagent-core/extensions/acceptance-state-machine-evidence.js +64 -8
  454. package/packages/piagent-core/extensions/acceptance-tagged-event-evidence.js +685 -0
  455. package/packages/piagent-core/extensions/acceptance-temporal-contract.js +756 -0
  456. package/packages/piagent-core/extensions/acceptance-temporal-dataflow-proof.js +618 -0
  457. package/packages/piagent-core/extensions/acceptance-test-binding-evidence.js +20 -2
  458. package/packages/piagent-core/extensions/acceptance-test-bindings.js +86 -0
  459. package/packages/piagent-core/extensions/acceptance-tuple-shapes.js +12 -0
  460. package/packages/piagent-core/extensions/context-engine.js +12 -17
  461. package/packages/piagent-core/extensions/context-index-policy.js +4 -0
  462. package/packages/piagent-core/extensions/context-query-paths.js +57 -0
  463. package/packages/piagent-core/extensions/criterion-context-pack.js +109 -4
  464. package/packages/piagent-core/extensions/criterion-graph.js +4 -1
  465. package/packages/piagent-core/extensions/failure-types.ts +2 -2
  466. package/packages/piagent-core/extensions/fallback-base-policy.ts +101 -0
  467. package/packages/piagent-core/extensions/guard-types.ts +9 -1
  468. package/packages/piagent-core/extensions/piagent-guard.ts +374 -364
  469. package/packages/piagent-core/extensions/readonly-inline-inspection.ts +378 -0
  470. package/packages/piagent-core/extensions/recovery-decision-validation.ts +92 -0
  471. package/packages/piagent-core/extensions/shell-write-targets.js +15 -14
  472. package/packages/piagent-core/extensions/task-contract-view.js +4 -4
  473. package/packages/piagent-core/extensions/task-digest-migration.js +2 -2
  474. package/packages/piagent-core/extensions/task-journal.js +48 -3
  475. package/packages/piagent-core/extensions/task-lifecycle.js +32 -0
  476. package/packages/piagent-core/extensions/task-runtime-audit.js +8 -2
  477. package/packages/piagent-core/extensions/task-state.js +140 -43
  478. package/packages/piagent-core/extensions/task-terminal-disposition.js +10 -0
  479. package/packages/piagent-core/extensions/trajectory-contract.ts +5 -0
  480. package/packages/piagent-core/extensions/verification-intelligence.js +109 -20
  481. package/packages/piagent-core/extensions/workflow-catalog.ts +98 -0
  482. package/packages/piagent-core/extensions/workspace-evidence-roots.js +52 -16
  483. package/packages/piagent-core/extensions/workspace-revision.js +48 -0
  484. package/packages/piagent-core/extensions/workspace-verification-snapshot.js +13 -0
  485. package/packages/piagent-core/package.json +3 -3
  486. package/packages/piagent-core/policies/base-policy.json +43 -2
  487. package/packages/piagent-core/policy/authority-manifest.v1.json +3 -3
  488. package/packages/piagent-core/prompts/be-to-fe.md +2 -8
  489. package/packages/piagent-core/prompts/plan.md +1 -1
  490. package/packages/piagent-core/prompts/platform-improve.md +2 -6
  491. package/packages/piagent-core/prompts/review.md +2 -7
  492. package/packages/piagent-core/prompts/scout.md +1 -1
  493. package/packages/piagent-core/prompts/task.md +4 -3
  494. package/packages/piagent-core/runtime/context/adaptive-planner.ts +7 -3
  495. package/packages/piagent-core/runtime/hooks/agent-start-hook.ts +60 -59
  496. package/packages/piagent-core/runtime/hooks/agent-start-types.ts +48 -0
  497. package/packages/piagent-core/runtime/hooks/completion-hook.ts +83 -89
  498. package/packages/piagent-core/runtime/hooks/input-hook.ts +98 -7
  499. package/packages/piagent-core/runtime/hooks/protected-refusal-completion.ts +86 -0
  500. package/packages/piagent-core/runtime/hooks/session-hooks.ts +5 -2
  501. package/packages/piagent-core/runtime/hooks/session-start-hook.ts +150 -19
  502. package/packages/piagent-core/runtime/hooks/tool-result-hook.ts +111 -92
  503. package/packages/piagent-core/runtime/hooks/tool-result-metadata.ts +9 -0
  504. package/packages/piagent-core/runtime/hooks/tool-result-navigation.ts +36 -0
  505. package/packages/piagent-core/runtime/hooks/tool-result-retrieval.ts +42 -0
  506. package/packages/piagent-core/runtime/hooks/tool-result-value-helpers.ts +66 -0
  507. package/packages/piagent-core/runtime/input/chat-images.ts +1 -1
  508. package/packages/piagent-core/runtime/inspection/activity-event-adapter.ts +1 -1
  509. package/packages/piagent-core/runtime/inspection/bounded-emission-guard.ts +286 -0
  510. package/packages/piagent-core/runtime/inspection/runtime-event-store.ts +37 -7
  511. package/packages/piagent-core/runtime/inspection/same-process-bridge-proof.ts +1 -1
  512. package/packages/piagent-core/runtime/inspection/task-handoff-history.ts +2 -2
  513. package/packages/piagent-core/runtime/inspection/task-subagent-tree.ts +27 -4
  514. package/packages/piagent-core/runtime/inspection/tool-failure-classification.ts +69 -11
  515. package/packages/piagent-core/runtime/inspection/webui-snapshot.ts +104 -22
  516. package/packages/piagent-core/runtime/inspection/workspace-file-reader.ts +2 -132
  517. package/packages/piagent-core/runtime/model/provider-wire-fingerprint.ts +20 -0
  518. package/packages/piagent-core/runtime/model/service-tier-runtime.ts +242 -0
  519. package/packages/piagent-core/runtime/orchestration/helper-lifecycle.ts +16 -12
  520. package/packages/piagent-core/runtime/orchestration/minimal-delegation-policy.ts +71 -0
  521. package/packages/piagent-core/runtime/orchestration/owned-work-budget.ts +28 -20
  522. package/packages/piagent-core/runtime/orchestration/role-policy.ts +27 -9
  523. package/packages/piagent-core/runtime/orchestration/subagent-tool-policy.ts +329 -0
  524. package/packages/piagent-core/runtime/policy/block-remedy.ts +0 -4
  525. package/packages/piagent-core/runtime/policy/source-worktree-transaction.ts +75 -14
  526. package/packages/piagent-core/runtime/product/activity-inspector.ts +6 -3
  527. package/packages/piagent-core/runtime/product/edit-verifier-timing.ts +26 -0
  528. package/packages/piagent-core/runtime/product/efficiency-metrics.ts +15 -7
  529. package/packages/piagent-core/runtime/product/operator-projections.ts +8 -3
  530. package/packages/piagent-core/runtime/product/task-usage-hooks.ts +78 -0
  531. package/packages/piagent-core/runtime/product/task-usage.ts +104 -0
  532. package/packages/piagent-core/runtime/quality/edit-freshness-guard.ts +102 -0
  533. package/packages/piagent-core/runtime/quality/exact-output-contract.ts +20 -10
  534. package/packages/piagent-core/runtime/quality/performance-assurance.ts +3 -3
  535. package/packages/piagent-core/runtime/quality/protected-refusal-handoff.ts +35 -0
  536. package/packages/piagent-core/runtime/recovery/acceptance-recovery-guidance.ts +21 -0
  537. package/packages/piagent-core/runtime/recovery/continuation-budget.ts +12 -3
  538. package/packages/piagent-core/runtime/recovery/diagnostic-delivery.ts +61 -0
  539. package/packages/piagent-core/runtime/recovery/handoff-projection.ts +196 -58
  540. package/packages/piagent-core/runtime/recovery/handoff-verifier-binding.ts +24 -0
  541. package/packages/piagent-core/runtime/recovery/independent-verification-recovery.ts +79 -0
  542. package/packages/piagent-core/runtime/recovery/pristine-task-policy-refresh.ts +91 -0
  543. package/packages/piagent-core/runtime/recovery/recovery-policy.ts +31 -40
  544. package/packages/piagent-core/runtime/recovery/resume-state.ts +120 -19
  545. package/packages/piagent-core/runtime/recovery/semantic-repair-handshake.ts +7 -13
  546. package/packages/piagent-core/runtime/recovery/task-evidence-coverage-refresh.ts +217 -0
  547. package/packages/piagent-core/runtime/registration/activity-inspector-command.ts +2 -1
  548. package/packages/piagent-core/runtime/registration/context-commands.ts +28 -10
  549. package/packages/piagent-core/runtime/registration/knowledge-tools.ts +9 -5
  550. package/packages/piagent-core/runtime/registration/operator-catalogs.ts +4 -9
  551. package/packages/piagent-core/runtime/registration/operator-request-intake.ts +49 -0
  552. package/packages/piagent-core/runtime/registration/policy-tools.ts +8 -6
  553. package/packages/piagent-core/runtime/registration/session-commands.ts +10 -8
  554. package/packages/piagent-core/runtime/registration/session-wait-tool.ts +54 -0
  555. package/packages/piagent-core/runtime/registration/task-completion-tools.ts +39 -20
  556. package/packages/piagent-core/runtime/registration/task-evidence-tools.ts +21 -12
  557. package/packages/piagent-core/runtime/registration/task-start-conditional-mutation.ts +45 -0
  558. package/packages/piagent-core/runtime/registration/task-start-guidance.ts +24 -15
  559. package/packages/piagent-core/runtime/registration/task-start-manifest.ts +54 -0
  560. package/packages/piagent-core/runtime/registration/task-start-tool.ts +79 -79
  561. package/packages/piagent-core/runtime/registration/workflow-commands.ts +64 -93
  562. package/packages/piagent-core/runtime/runtime-limits.ts +16 -2
  563. package/packages/piagent-core/runtime/session/adaptive-context-analysis.ts +461 -0
  564. package/packages/piagent-core/runtime/session/adaptive-context-governor.ts +464 -0
  565. package/packages/piagent-core/runtime/session/adaptive-context-ledger.ts +265 -0
  566. package/packages/piagent-core/runtime/session/completion-signals.ts +23 -0
  567. package/packages/piagent-core/runtime/session/message-signals.ts +6 -18
  568. package/packages/piagent-core/runtime/session/operator-request-carry.ts +44 -0
  569. package/packages/piagent-core/runtime/session/retrieval-evidence-runtime.ts +188 -0
  570. package/packages/piagent-core/runtime/session/runtime-state.ts +37 -39
  571. package/packages/piagent-core/runtime/session/source-checkout-read-grants.ts +35 -0
  572. package/packages/piagent-core/runtime/session/system-prompt.ts +76 -14
  573. package/packages/piagent-core/runtime/session/task-contract-presentation.ts +39 -0
  574. package/packages/piagent-core/runtime/session/uncertain-send-continuation.ts +231 -0
  575. package/packages/piagent-core/runtime/session/usage.ts +4 -3
  576. package/packages/piagent-core/runtime/tools/apply-patch-tool.ts +6 -6
  577. package/packages/piagent-core/runtime/tools/phase-tool-runtime.ts +1 -1
  578. package/packages/piagent-core/runtime/tools/phase-tools.ts +10 -9
  579. package/packages/piagent-core/runtime/tools/session-wait.ts +42 -0
  580. package/packages/piagent-core/runtime/tools/tool-groups.ts +22 -5
  581. package/packages/piagent-core/runtime/trajectory/trajectory-runtime.ts +192 -5
  582. package/packages/piagent-core/runtime/trajectory/trajectory-types.ts +3 -2
  583. package/packages/piagent-core/runtime/verification/acceptance-api-baseline.js +488 -0
  584. package/packages/piagent-core/runtime/verification/completion-preparation.ts +22 -0
  585. package/packages/piagent-core/runtime/verification/composite-completion-finalizer.ts +105 -0
  586. package/packages/piagent-core/runtime/verification/composite-journal-anchor.ts +181 -0
  587. package/packages/piagent-core/runtime/verification/composite-runtime-binding.ts +76 -0
  588. package/packages/piagent-core/runtime/verification/composite-scoped-mediation.ts +318 -0
  589. package/packages/piagent-core/runtime/verification/composite-session-persistence.ts +175 -0
  590. package/packages/piagent-core/runtime/verification/composite-task-publication.ts +273 -0
  591. package/packages/piagent-core/runtime/verification/exact-verifier-reuse.ts +101 -0
  592. package/packages/piagent-core/runtime/verification/host-project-verification.ts +71 -0
  593. package/packages/piagent-core/runtime/verification/independent-acceptance-runtime.ts +500 -0
  594. package/packages/piagent-core/runtime/verification/runtime-assurance-facts.ts +500 -0
  595. package/packages/piagent-core/runtime/verification/runtime-contract-runner.ts +107 -0
  596. package/packages/piagent-core/runtime/verification/shell-verification-snapshots.ts +56 -0
  597. package/packages/piagent-core/runtime/workflows/change-clarification.ts +61 -0
  598. package/packages/piagent-core/runtime/workflows/input-routing.ts +21 -14
  599. package/packages/piagent-core/runtime/workflows/runtime-intake-compaction.ts +201 -0
  600. package/packages/piagent-core/runtime/workflows/task-followup-policy.ts +163 -0
  601. package/packages/piagent-core/runtime/workflows/task-intake.ts +276 -169
  602. package/packages/piagent-core/runtime/workflows/task-repair-intent.ts +67 -0
  603. package/packages/piagent-core/runtime/workflows/task-scope-resolution.ts +83 -0
  604. package/packages/piagent-core/runtime/workflows/webui-runtime-command.ts +5 -2
  605. package/packages/piagent-core/runtime/workflows/webui-workflow.ts +39 -32
  606. package/packages/piagent-core/runtime/workflows/workflow-follow-up.ts +33 -0
  607. package/packages/piagent-core/security/workspace-file-reader.ts +132 -0
  608. package/packages/piagent-core/subagents/piagent-oracle.md +5 -3
  609. package/packages/piagent-core/subagents/piagent-planner.md +6 -4
  610. package/packages/piagent-core/subagents/piagent-reviewer.md +5 -3
  611. package/packages/piagent-core/subagents/piagent-scout.md +4 -2
  612. package/packages/piagent-core/subagents/piagent-worker.md +6 -4
  613. package/packages/piagent-webui/dist/client/assets/{MarkdownMessage-CxeZqTte.js → MarkdownMessage-DGk34zJT.js} +1 -1
  614. package/packages/piagent-webui/dist/client/assets/{index-BV-_03yl.css → index-DVahUVWy.css} +1 -1
  615. package/packages/piagent-webui/dist/client/assets/index-DiKRmCEP.js +176 -0
  616. package/packages/piagent-webui/dist/client/index.html +2 -2
  617. package/packages/piagent-webui/extension/dispatch-user-entry.ts +13 -0
  618. package/packages/piagent-webui/extension/piagent-webui.ts +13 -3
  619. package/packages/piagent-webui/extension/review-controller.ts +29 -6
  620. package/packages/piagent-webui/extension/same-session-bridge.ts +7 -8
  621. package/packages/piagent-webui/extension/session-stream-adapter.ts +88 -16
  622. package/packages/piagent-webui/extension/source-mutation-controller.ts +29 -3
  623. package/packages/piagent-webui/extension/source-open-controller.ts +29 -7
  624. package/packages/piagent-webui/extension/source-revert-controller.ts +26 -3
  625. package/packages/piagent-webui/gateway/gateway-events.ts +22 -5
  626. package/packages/piagent-webui/gateway/gateway-service.ts +7 -2
  627. package/packages/piagent-webui/gateway/gateway-session-stream.ts +170 -32
  628. package/packages/piagent-webui/gateway/rpc-ui-context.ts +34 -9
  629. package/packages/piagent-webui/gateway/runtime-command-controller.ts +46 -6
  630. package/packages/piagent-webui/gateway/runtime-session-controls.ts +17 -2
  631. package/packages/piagent-webui/gateway/session-catalog.ts +9 -3
  632. package/packages/piagent-webui/gateway/session-command-controller.ts +29 -9
  633. package/packages/piagent-webui/gateway/session-effective-options.ts +2 -9
  634. package/packages/piagent-webui/gateway/session-inspection-registry.ts +11 -2
  635. package/packages/piagent-webui/gateway/session-operation-lifecycle.ts +237 -0
  636. package/packages/piagent-webui/gateway/session-operation-retry-control.ts +38 -0
  637. package/packages/piagent-webui/gateway/session-operation-settlement.ts +28 -0
  638. package/packages/piagent-webui/gateway/session-operation-watchdog.ts +9 -2
  639. package/packages/piagent-webui/gateway/session-prompt-dispatch.ts +37 -0
  640. package/packages/piagent-webui/gateway/session-runtime-close.ts +36 -0
  641. package/packages/piagent-webui/gateway/session-runtime-factory.ts +183 -6
  642. package/packages/piagent-webui/gateway/session-runtime-ownership.ts +48 -0
  643. package/packages/piagent-webui/gateway/session-runtime-state.ts +24 -0
  644. package/packages/piagent-webui/gateway/session-runtime-supervisor.ts +118 -117
  645. package/packages/piagent-webui/ownership/session-lease-store.ts +5 -2
  646. package/packages/piagent-webui/server/core-inspection-provider.ts +75 -15
  647. package/packages/piagent-webui/server/loopback-server.ts +9 -7
  648. package/packages/piagent-webui/server/terminal-delivery-receipt.ts +144 -0
  649. package/packages/piagent-webui/server/transcript-projection.ts +150 -11
  650. package/packages/piagent-webui/shared/message-correlation.ts +25 -0
  651. package/schemas/approved-host-contracts.schema.json +973 -0
  652. package/schemas/benchmark-attempt-outcome.schema.json +303 -0
  653. package/schemas/benchmark-independent-verification-plan.schema.json +34 -0
  654. package/schemas/codex-relative-efficiency-v1.schema.json +155 -0
  655. package/schemas/composite-criterion.schema.json +788 -0
  656. package/schemas/contract-family-library.schema.json +28 -0
  657. package/schemas/contract-selection-recipe.schema.json +31 -0
  658. package/schemas/failure-evidence.schema.json +2 -2
  659. package/schemas/helper-request.schema.json +16 -4
  660. package/schemas/host-contract-plan.schema.json +25 -0
  661. package/schemas/piagent-webui/gateway-protocol-v1.schema.json +16 -2
  662. package/schemas/piagent-webui/runtime-command-v1.schema.json +26 -3
  663. package/schemas/piagent-webui/runtime-event-v2.schema.json +46 -5
  664. package/schemas/piagent-webui/session-live-state-v1.schema.json +2 -0
  665. package/schemas/piagent-webui/snapshot-v1.schema.json +3 -0
  666. package/schemas/piagent-webui/source-change-v1.schema.json +12 -0
  667. package/schemas/piagent-webui/subagent-tree-v1.schema.json +11 -1
  668. package/schemas/piagent-webui/transcript-v1.schema.json +2 -1
  669. package/schemas/project-profile.schema.json +1 -1
  670. package/schemas/public-contracts/codex-mcp-metadata-v1.schema.json +206 -0
  671. package/schemas/public-contracts/codex-mcp-metadata-v2.schema.json +228 -0
  672. package/schemas/public-contracts/codex-mcp-metadata-v3.schema.json +233 -0
  673. package/schemas/public-contracts/revocation-v1.schema.json +87 -0
  674. package/schemas/public-contracts/subscription-contract-v1.schema.json +76 -0
  675. package/schemas/scoped-project-verification-receipt-v2.schema.json +83 -0
  676. package/schemas/scoped-project-verification-receipt-v3.schema.json +295 -0
  677. package/schemas/scoped-verification-receipt-v1.schema.json +76 -0
  678. package/schemas/task-contract.schema.json +45 -0
  679. package/scripts/approve-independent-verification.mjs +57 -0
  680. package/scripts/benchmark-arm-observer.mjs +320 -0
  681. package/scripts/benchmark-budget-ipc.mjs +34 -0
  682. package/scripts/benchmark-budget-launcher.mjs +91 -0
  683. package/scripts/benchmark-budget-processes.mjs +133 -0
  684. package/scripts/benchmark-budget-publication.mjs +228 -0
  685. package/scripts/benchmark-budget-runtime.mjs +136 -0
  686. package/scripts/benchmark-candidate-ablation.mjs +265 -0
  687. package/scripts/benchmark-codex-journey.mjs +406 -0
  688. package/scripts/benchmark-independent-verification.mjs +628 -0
  689. package/scripts/benchmark-journey-outcome.mjs +92 -0
  690. package/scripts/benchmark-production-v3-human-review.mjs +192 -0
  691. package/scripts/benchmark-public-input-catalog.mjs +158 -0
  692. package/scripts/benchmark-registered-plan-author.mjs +355 -0
  693. package/scripts/benchmark-registered-plan-recipe-support.mjs +85 -0
  694. package/scripts/benchmark-registered-plan-recipes-backend.mjs +165 -0
  695. package/scripts/benchmark-registered-plan-recipes-data.mjs +164 -0
  696. package/scripts/benchmark-registered-plan-recipes-platform.mjs +165 -0
  697. package/scripts/benchmark-runner-completion.mjs +133 -0
  698. package/scripts/benchmark-runner-configuration.mjs +159 -0
  699. package/scripts/benchmark-runner-core.mjs +224 -387
  700. package/scripts/benchmark-runner-finalization.mjs +144 -27
  701. package/scripts/benchmark-runner-provider-boundary.mjs +44 -0
  702. package/scripts/benchmark-runner-schedule.mjs +256 -0
  703. package/scripts/benchmark-runner-support.mjs +136 -4
  704. package/scripts/benchmark-runner.mjs +95 -18
  705. package/scripts/benchmark-scoped-broker-custody.mjs +483 -0
  706. package/scripts/benchmark-scoped-docs-verifier.mjs +203 -0
  707. package/scripts/benchmark-scoped-frozen-qualification.mjs +676 -0
  708. package/scripts/benchmark-scoped-pi-router.mjs +196 -0
  709. package/scripts/benchmark-scoped-project-verification-receipt.mjs +49 -0
  710. package/scripts/benchmark-scoped-project-verifier.mjs +461 -0
  711. package/scripts/benchmark-scoped-session-custody.mjs +599 -0
  712. package/scripts/benchmark-scoped-session-factory.mjs +342 -0
  713. package/scripts/benchmark-scoped-tool-broker.mjs +523 -0
  714. package/scripts/benchmark-scoped-verification-supervisor.mjs +773 -0
  715. package/scripts/benchmark-session-evaluator.mjs +169 -0
  716. package/scripts/benchmark-session-provider-boundary.mjs +62 -0
  717. package/scripts/benchmark-session-usage.mjs +28 -0
  718. package/scripts/benchmark-session.mjs +343 -50
  719. package/scripts/benchmark-webui-journey.mjs +678 -0
  720. package/scripts/check-public-wording.mjs +46 -0
  721. package/scripts/configure-subagents.sh +42 -34
  722. package/scripts/fs4-readiness-evaluation.mjs +7 -1
  723. package/scripts/helper-evaluation.mjs +1 -1
  724. package/scripts/phase-tools-evaluation.mjs +1 -0
  725. package/scripts/pi-auto.sh +7 -1
  726. package/scripts/pi-usage-history.mjs +201 -11
  727. package/scripts/piagent-cli.mjs +8 -0
  728. package/scripts/private-holdout-readiness.mjs +7 -1
  729. package/scripts/public-evaluation-exposure.mjs +140 -0
  730. package/scripts/select-independent-verification.mjs +62 -0
  731. package/scripts/team-doctor.sh +1 -1
  732. package/scripts/typescript-loader.mjs +135 -1
  733. package/scripts/update-global.mjs +6 -7
  734. package/scripts/verify-local.sh +41 -55
  735. package/scripts/verify-release-identity.mjs +5 -3
  736. package/packages/piagent-webui/dist/client/assets/index-DynIMe8I.js +0 -176
@@ -7,8 +7,8 @@
7
7
  <meta name="csp-nonce" content="__PIAGENT_CSP_NONCE__" />
8
8
  <meta name="piagent-webui-mode" content="__PIAGENT_WEBUI_MODE__" />
9
9
  <title>Piagent WebUI</title>
10
- <script type="module" crossorigin src="/assets/index-DynIMe8I.js"></script>
11
- <link rel="stylesheet" crossorigin href="/assets/index-BV-_03yl.css">
10
+ <script type="module" crossorigin src="/assets/index-DiKRmCEP.js"></script>
11
+ <link rel="stylesheet" crossorigin href="/assets/index-DVahUVWy.css">
12
12
  </head>
13
13
  <body>
14
14
  <div id="root"></div>
@@ -0,0 +1,13 @@
1
+ type BranchEntry = { id: string; parentId: string | null; type: string; message?: { role: string } };
2
+
3
+ // Pi may persist system prompt updates before the user message in the same dispatch.
4
+ // Do not cross any other entry kind or another user's/assistant's message.
5
+ export function firstUserEntryAfter<T extends BranchEntry>(branch: readonly T[], parentId: string | null): T | null {
6
+ for (const entry of branch) {
7
+ if (entry.parentId !== parentId) continue;
8
+ if (entry.type !== "message") return null;
9
+ if (entry.message?.role === "system") { parentId = entry.id; continue; }
10
+ return entry.message?.role === "user" ? entry : null;
11
+ }
12
+ return null;
13
+ }
@@ -313,6 +313,7 @@ export default function piagentWebUiExtension(pi: ExtensionAPI): void {
313
313
  task: () => activeSessionTask(ctx.cwd, ctx.sessionManager.getSessionId()),
314
314
  activityEvents: () => readContextTelemetry(ctx.cwd, { limit: 5_000 }) as unknown[],
315
315
  currentActivity: () => [], sessionEntries: () => ctx.sessionManager.getBranch(), protectedPaths: () => protectedPaths(ctx),
316
+ sessionFile: () => ctx.sessionManager.getSessionFile(),
316
317
  documentReadRoots: () => documentReadRoots(ctx),
317
318
  contextUsage: () => ctx.getContextUsage(), model: () => ctx.model,
318
319
  thinkingLevel: () => ctx.thinkingLevel ?? pi.getThinkingLevel(),
@@ -446,15 +447,23 @@ export default function piagentWebUiExtension(pi: ExtensionAPI): void {
446
447
  pi.on("tool_execution_end", (event, ctx) => { publishDrafts(ctx, stream.toolEnded(event, bridge.snapshot())); bridgeSoft(() => lifecycle.observeToolEnd(ctx)); });
447
448
  pi.on("turn_end", (event, ctx) => publishDrafts(ctx, stream.turnEnded(event, bridge.snapshot())));
448
449
  pi.on("agent_settled", (_event, ctx) => {
449
- publishDrafts(ctx, stream.agentSettled(bridge.snapshot(), typeof ctx.hasPendingMessages === "function" ? ctx.hasPendingMessages() : null));
450
+ const hasPendingMessages = typeof ctx.hasPendingMessages === "function" ? ctx.hasPendingMessages() : null;
451
+ publishDrafts(ctx, stream.agentSettled(bridge.snapshot(), hasPendingMessages));
452
+ if (hasPendingMessages === true) { publishPersisted(ctx); return; }
450
453
  const operationId = bridge.snapshot().identity?.agentOperationId ?? null;
451
454
  bridgeSoft(() => bridge.observeAgentSettled(ctx)); bridgeSoft(() => lifecycle.observeAgentSettled(ctx, operationId)); publishPersisted(ctx);
452
455
  });
453
456
  pi.on("session_info_changed", (_event, ctx) => { currentContext = ctx; bridgeSoft(() => bridge.refresh(ctx)); });
454
457
  pi.on("model_select", (_event, ctx) => { currentContext = ctx; bridgeSoft(() => sessionOptions.observeHostOptionChange(ctx)); currentManager(ctx)?.provider.invalidate(); });
455
458
  pi.on("thinking_level_select", (_event, ctx) => { currentContext = ctx; bridgeSoft(() => sessionOptions.observeHostOptionChange(ctx)); currentManager(ctx)?.provider.invalidate(); });
456
- pi.on("session_before_switch", () => { lifecycle.replacementPending(); bridgeSoft(() => bridge.replacementPending()); });
457
- pi.on("session_before_fork", () => { lifecycle.replacementPending(); bridgeSoft(() => bridge.replacementPending()); });
459
+ pi.on("session_before_switch", (_event, ctx) => {
460
+ publishDrafts(ctx, stream.operationInterrupted(bridge.snapshot(), "session-replaced-before-tool-result"));
461
+ lifecycle.replacementPending(); bridgeSoft(() => bridge.replacementPending());
462
+ });
463
+ pi.on("session_before_fork", (_event, ctx) => {
464
+ publishDrafts(ctx, stream.operationInterrupted(bridge.snapshot(), "session-forked-before-tool-result"));
465
+ lifecycle.replacementPending(); bridgeSoft(() => bridge.replacementPending());
466
+ });
458
467
  pi.on("session_start", async (_event, ctx) => {
459
468
  currentContext = ctx; stream.reset(); queue.reset(); sessionOptions.reset(); attachments?.reset();
460
469
  if (!isGatewayRuntimeContext(ctx)) {
@@ -471,6 +480,7 @@ export default function piagentWebUiExtension(pi: ExtensionAPI): void {
471
480
  catch (error) { ctx.ui.notify(error instanceof Error ? error.message : "Piagent WebUI could not start", "error"); }
472
481
  });
473
482
  pi.on("session_shutdown", async (_event, ctx) => {
483
+ publishDrafts(ctx, stream.operationInterrupted(bridge.snapshot(), "session-shutdown-before-tool-result"));
474
484
  stream.reset(); queue.reset(); sessionOptions.reset(); attachments?.close(); eventReaders.delete(key(ctx));
475
485
  if (approvalBinding?.key === key(ctx)) { approvalBinding.unsubscribe(); approvalBinding.unbind(); approvalBinding = null; }
476
486
  lifecycle.shutdown();
@@ -33,11 +33,19 @@ function exactKeys(value: unknown, keys: string[]): boolean { return Boolean(val
33
33
  && same(Object.keys(value as object).sort(), [...keys].sort())); }
34
34
  function timestamp(value: unknown): value is string { if (typeof value !== "string" || !/^\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}\.\d{3}Z$/.test(value)) return false;
35
35
  const parsed = Date.parse(value); return Number.isFinite(parsed) && new Date(parsed).toISOString() === value; }
36
+ function nullableRevision(value: unknown): value is string | null { return value === null || typeof value === "string" && REVISION.test(value); }
36
37
  function wireIdentity(value: BridgeIdentity): BridgeIdentity { return { projectRef: value.projectRef, runtimeInstanceId: value.runtimeInstanceId,
37
38
  sessionRef: value.sessionRef, taskId: value.taskId, taskRunId: value.taskRunId, agentOperationId: null, toolCallId: null }; }
38
39
  function wireRevisions(value: BridgeRevisions): BridgeRevisions { return { runtimeRevision: value.runtimeRevision, taskRevision: value.taskRevision,
39
40
  controlRevision: value.controlRevision, workspaceRevision: value.workspaceRevision, indexRevision: value.indexRevision,
40
41
  approvalRevision: value.approvalRevision, sessionOptionRevision: value.sessionOptionRevision, queueRevision: value.queueRevision }; }
42
+ function replayIdentityDisposition(record: Pick<ReviewEvidenceRecord, "projectRef" | "runtimeInstanceId" | "sessionRef" | "taskId" | "taskRunId">,
43
+ identity: BridgeIdentity): "task-identity-mismatch" | "runtime-or-session-replaced" | null {
44
+ if (record.taskId !== identity.taskId || record.taskRunId !== identity.taskRunId) return "task-identity-mismatch";
45
+ if (record.projectRef !== identity.projectRef || record.runtimeInstanceId !== identity.runtimeInstanceId || record.sessionRef !== identity.sessionRef)
46
+ return "runtime-or-session-replaced";
47
+ return null;
48
+ }
41
49
 
42
50
  function receiptFromRecord(record: ReviewEvidenceRecord, deduplicated: boolean): ReviewReceipt {
43
51
  return { schemaVersion: 1, version: "piagent-webui-control-v1", messageType: "receipt", commandId: record.commandId,
@@ -82,19 +90,32 @@ export class ReviewController {
82
90
  if (old) {
83
91
  if (old.commandId !== command.commandId || old.idempotencyKeyDigest !== keyDigest || old.actionDigest !== command.actionDigest)
84
92
  return this.#reject(command, "idempotency-payload-mismatch", "idempotency-payload-mismatch", false);
85
- if (old.runtimeInstanceId !== identity.runtimeInstanceId || old.sessionRef !== identity.sessionRef || old.taskId !== identity.taskId || old.taskRunId !== identity.taskRunId)
86
- return this.#reject(command, "identity-mismatch", "runtime-or-session-replaced", false);
93
+ const disposition = replayIdentityDisposition(old, identity);
94
+ if (disposition) return this.#reject(command, "identity-mismatch", disposition, false);
87
95
  return receiptFromRecord(old, true);
88
96
  }
89
97
  const now = this.#now().getTime();
90
98
  if (Date.parse(command.requestedAt) > now + 30_000) return this.#reject(command, "invalid-command", "requested-at-in-future", false);
91
99
  if (now > Date.parse(command.expiresAt)) return this.#reject(command, "expired", "command-expired", false);
92
100
  if (!same(command.identity, identity)) return this.#reject(command, "identity-mismatch", "identity-mismatch", false);
93
- if (command.expectedRevisions.runtimeRevision !== revisions.runtimeRevision || command.expectedRevisions.taskRevision !== revisions.taskRevision)
94
- return this.#reject(command, "stale-revision", "stale-source-revision", true);
101
+ if (command.expectedRevisions.runtimeRevision !== revisions.runtimeRevision || command.expectedRevisions.taskRevision !== revisions.taskRevision
102
+ || command.expectedRevisions.controlRevision !== revisions.controlRevision)
103
+ return this.#reject(command, "stale-revision", "stale-task-control-revision", true);
95
104
  let projection: ReviewStateProjection;
96
105
  try { projection = await this.#resolve(command.payload.view, command.payload.fileRef); }
97
106
  catch { return this.#reject(command, "capability-unavailable", "review-target-unavailable", true); }
107
+ const rebound = this.#bridge.snapshot();
108
+ if (rebound.state !== "ready" || !rebound.identity || !rebound.revisions)
109
+ return this.#reject(command, "resync-required", "review-binding-unavailable-after-target-resolution", true);
110
+ const reboundIdentity = wireIdentity(rebound.identity), reboundRevisions = wireRevisions(rebound.revisions);
111
+ if (!same(reboundIdentity, identity)) {
112
+ const taskChanged = reboundIdentity.taskId !== identity.taskId || reboundIdentity.taskRunId !== identity.taskRunId;
113
+ return this.#reject(command, "identity-mismatch", taskChanged ? "task-identity-mismatch" : "runtime-or-session-replaced", false);
114
+ }
115
+ if (command.expectedRevisions.runtimeRevision !== reboundRevisions.runtimeRevision
116
+ || command.expectedRevisions.taskRevision !== reboundRevisions.taskRevision
117
+ || command.expectedRevisions.controlRevision !== reboundRevisions.controlRevision)
118
+ return this.#reject(command, "stale-revision", "stale-task-control-revision-after-target-resolution", true);
98
119
  const target = projection.target;
99
120
  if (!target || projection.state === "unavailable") return this.#reject(command, "capability-unavailable", projection.reasonCode ?? "review-target-unavailable", true);
100
121
  if (!this.#matchesTarget(command, target)) return this.#reject(command, "stale-revision", "review-target-changed", true);
@@ -130,8 +151,10 @@ export class ReviewController {
130
151
  const revisionKeys = ["runtimeRevision", "taskRevision", "controlRevision", "workspaceRevision", "indexRevision", "approvalRevision",
131
152
  "sessionOptionRevision", "queueRevision", "workspacePreimage", "indexPreimage", "patchPreimage"];
132
153
  if (!exactKeys(command.expectedRevisions, revisionKeys) || !REVISION.test(command.expectedRevisions.runtimeRevision)
133
- || !REVISION.test(command.expectedRevisions.taskRevision ?? "") || !REVISION.test(command.expectedRevisions.workspaceRevision ?? "")
134
- || command.expectedRevisions.indexRevision !== null && !REVISION.test(command.expectedRevisions.indexRevision)
154
+ || !REVISION.test(command.expectedRevisions.taskRevision ?? "") || !REVISION.test(command.expectedRevisions.controlRevision ?? "")
155
+ || !REVISION.test(command.expectedRevisions.workspaceRevision ?? "")
156
+ || ![command.expectedRevisions.indexRevision, command.expectedRevisions.approvalRevision,
157
+ command.expectedRevisions.sessionOptionRevision, command.expectedRevisions.queueRevision].every(nullableRevision)
135
158
  || command.expectedRevisions.workspacePreimage !== null || command.expectedRevisions.indexPreimage !== null
136
159
  || !DIGEST.test(command.expectedRevisions.patchPreimage)) return "invalid-review-authority";
137
160
  if (!exactKeys(command.payload, ["view", "fileRef", "diffRef", "reviewState", "contentDigest"])
@@ -1,3 +1,4 @@
1
+ import { firstUserEntryAfter } from "./dispatch-user-entry.ts";
1
2
  import { AsyncLocalStorage } from "node:async_hooks";
2
3
  import { createHash, randomBytes } from "node:crypto";
3
4
  import type { ExtensionAPI, ExtensionContext } from "@earendil-works/pi-coding-agent";
@@ -98,8 +99,8 @@ function messageText(message: any): string {
98
99
  }
99
100
  function firstMessageText(message: any): string { return Array.isArray(message?.content) ? String(message.content.find((item: any) => item?.type === "text")?.text ?? "") : String(message?.content ?? ""); }
100
101
  function dispatchEntry(pending: PendingDispatch, ctx: ExtensionContext): any | null {
101
- return ctx.sessionManager.getBranch().find((entry: any) => entry?.type === "message" && entry.parentId === pending.leafBefore
102
- && entry.message?.role === "user" && messageText(entry.message) === pending.observedText) ?? null;
102
+ const entry = firstUserEntryAfter(ctx.sessionManager.getBranch(), pending.leafBefore) as any;
103
+ return entry && messageText(entry.message) === pending.observedText ? entry : null;
103
104
  }
104
105
  export function chatContentDigest(payload: Pick<NewOperationChatCommand["payload"], "text" | "attachmentRefs">): string {
105
106
  return sha(canonical({ text: payload.text, attachmentRefs: payload.attachmentRefs }));
@@ -168,7 +169,6 @@ export class SameSessionPiBridge {
168
169
  const branch = ctx.sessionManager.getBranch();
169
170
  if (!Array.isArray(branch) || branch.length > MAX_SESSION_ENTRIES) throw new Error("webui-bridge-receipt-history-unavailable");
170
171
  const requested = new Map<string, { receipt: ChatReceipt; messageRequestId: string; contentDigest: string; attachmentRefs: string[]; entryId: string }>();
171
- const entriesById = new Map(branch.map((entry: any) => [entry?.id, entry]));
172
172
  for (const entry of branch) {
173
173
  if (entry?.type !== "custom" || entry.customType !== WEBUI_CONTROL_ENTRY_TYPE) continue;
174
174
  if (!exactKeys(entry.data, ["receipt"]) && !exactKeys(entry.data, ["receipt", "messageRequestId", "contentDigest"])
@@ -193,8 +193,8 @@ export class SameSessionPiBridge {
193
193
  || receipt.requestedAt !== prior.receipt.requestedAt || !same(receipt.observedRevisionsBefore, prior.receipt.observedRevisionsBefore)
194
194
  || !same(receipt.identity, prior.receipt.identity) || !same(attachmentRefs, prior.attachmentRefs)) throw new Error("webui-bridge-receipt-store-corrupt");
195
195
  if (receipt.resultCode === "dispatch-observed") {
196
- const evidence = entriesById.get(receipt.settlementEvidenceRef as string) as any;
197
- if (!evidence || evidence.type !== "message" || evidence.parentId !== prior.entryId || evidence.message?.role !== "user"
196
+ const evidence = firstUserEntryAfter(branch, prior.entryId) as any;
197
+ if (!evidence || evidence.id !== receipt.settlementEvidenceRef
198
198
  || chatContentDigest({ text: firstMessageText(evidence.message), attachmentRefs: prior.attachmentRefs }) !== prior.contentDigest) throw new Error("webui-bridge-receipt-store-corrupt");
199
199
  }
200
200
  }
@@ -310,14 +310,13 @@ export class SameSessionPiBridge {
310
310
  observeMessageStart(event: { message?: unknown }, ctx: ExtensionContext): void {
311
311
  if (!this.refresh(ctx) || this.#binding?.state !== "ready") return; const pending = this.#pending, message = event.message as any;
312
312
  const leaf = pending ? ctx.sessionManager.getLeafEntry?.() : null;
313
- const leafObserved = Boolean(pending && leaf?.type === "message" && leaf.id !== pending.leafBefore && leaf.parentId === pending.leafBefore && leaf.message?.role === "user"
314
- && messageText(leaf.message) === pending.observedText);
313
+ const leafObserved = Boolean(pending && leaf?.id && dispatchEntry(pending, ctx)?.id === leaf.id);
315
314
  if (pending && pending.inputObserved && leafObserved && message?.role === "user" && messageText(message) === pending.observedText)
316
315
  this.#settlePending("dispatch-observed", "settled", null, true, leaf.id);
317
316
  }
318
317
  observeAgentStart(ctx: ExtensionContext): void {
319
318
  if (!this.refresh(ctx) || !this.#binding || this.#binding.state !== "ready") return;
320
- const pending = this.#pending;
319
+ const pending = this.#pending; if (this.#binding.identity.agentOperationId && !pending) return;
321
320
  const operationId = pending?.operationId ?? this.#revision("operation", ++this.#operationCounter, "external");
322
321
  this.#binding.identity.agentOperationId = operationId; this.#refresh(); this.#event("operation.started", { operationRef: operationId, commandId: pending?.command.commandId });
323
322
  const observedEntry = pending ? dispatchEntry(pending, ctx) : null;
@@ -1,6 +1,7 @@
1
1
  import { createHash } from "node:crypto";
2
2
 
3
3
  import { redactSensitiveText } from "../../piagent-core/extensions/redaction-core.js";
4
+ import { classifyToolFailure, handledToolFailure } from "../../piagent-core/runtime/inspection/tool-failure-classification.ts";
4
5
  import type { RuntimeEventDraft, RuntimeEventRevision } from "../../piagent-core/runtime/inspection/runtime-event-store.ts";
5
6
  import type { BridgeIdentity, BridgeSnapshot } from "./same-session-bridge.ts";
6
7
 
@@ -9,13 +10,37 @@ const MAX_DELTA = 16_384;
9
10
  const MAX_MESSAGE_STREAM_EVENTS = 128;
10
11
  const STREAM_FLUSH_CHARS = 1_024;
11
12
  const STREAM_FLUSH_INTERVAL_MS = 100;
13
+ const MAX_OBSERVED_UPDATES = 100_000;
14
+ const UPDATE_KINDS = new Set(["start", "text_start", "text_delta", "text_end", "thinking_start", "thinking_delta",
15
+ "thinking_end", "toolcall_start", "toolcall_delta", "toolcall_end", "done", "error"]);
12
16
  const ANSI = /\u001b(?:[@-_]|\[[0-?]*[ -/]*[@-~])/g;
13
17
  const PRIVATE_KEY_BEGIN = /-----BEGIN [A-Z ]*PRIVATE KEY-----/;
14
18
  const PRIVATE_KEY_END = /-----END [A-Z ]*PRIVATE KEY-----/;
15
19
 
16
20
  type MessageRole = "user" | "assistant" | "tool-result";
21
+ type StreamActivity = { schemaVersion: 1; boundary: "pi-message-update-hook"; observedUpdates: number;
22
+ textUpdates: number; thinkingUpdates: number; toolCallUpdates: number; otherUpdates: number;
23
+ lastUpdateKind: string | null; lastObservedAt: string | null; countersCapped: boolean };
17
24
  type ActiveMessage = { ref: string; role: MessageRole; rawBuffer: string; chunkSequence: number; truncated: boolean;
18
- emittedEvents: number; lastFlushAt: number; thinking: Set<number>; thinkingStreaming: Set<number>; thinkingCompleted: Set<number> };
25
+ emittedEvents: number; lastFlushAt: number; thinking: Set<number>; thinkingStreaming: Set<number>; thinkingCompleted: Set<number>;
26
+ streamActivity: StreamActivity };
27
+
28
+ function newStreamActivity(): StreamActivity {
29
+ return { schemaVersion: 1, boundary: "pi-message-update-hook", observedUpdates: 0, textUpdates: 0, thinkingUpdates: 0,
30
+ toolCallUpdates: 0, otherUpdates: 0, lastUpdateKind: null, lastObservedAt: null, countersCapped: false };
31
+ }
32
+
33
+ // This observes delivery to the Pi hook, not network traffic or provider usage.
34
+ // Count before projection filtering; never retain an update payload or its text.
35
+ function observeStreamActivity(activity: StreamActivity, kind: unknown, now: Date): void {
36
+ const known = typeof kind === "string" && UPDATE_KINDS.has(kind) ? kind : "other";
37
+ activity.lastUpdateKind = known; activity.lastObservedAt = now.toISOString();
38
+ if (activity.observedUpdates >= MAX_OBSERVED_UPDATES) { activity.countersCapped = true; return; }
39
+ activity.observedUpdates++;
40
+ const category = known.startsWith("text_") ? "textUpdates" : known.startsWith("thinking_") ? "thinkingUpdates"
41
+ : known.startsWith("toolcall_") ? "toolCallUpdates" : "otherUpdates";
42
+ activity[category]++;
43
+ }
19
44
 
20
45
  function hash(value: string): string { return createHash("sha256").update(value).digest("hex"); }
21
46
  function opaque(prefix: string, value: unknown): string { return `${prefix}.${hash(JSON.stringify(value))}`; }
@@ -35,6 +60,13 @@ function safeToolName(value: unknown): string {
35
60
  if (projected.redacted) return "redacted-tool";
36
61
  return projected.text.replace(/[^A-Za-z0-9._:-]/g, "-").replace(/^[^A-Za-z0-9]+/, "").slice(0, 128) || "tool";
37
62
  }
63
+ function toolClassificationInput(value: unknown): { command: string } | undefined {
64
+ if (!value || typeof value !== "object") return undefined;
65
+ const direct = (value as { command?: unknown }).command;
66
+ const nested = (value as { args?: { command?: unknown } }).args?.command;
67
+ const command = typeof direct === "string" ? direct : typeof nested === "string" ? nested : null;
68
+ return command === null ? undefined : { command: command.slice(0, 8_192) };
69
+ }
38
70
  function role(message: any): MessageRole | null {
39
71
  return message?.role === "user" ? "user" : message?.role === "assistant" ? "assistant" : message?.role === "toolResult" ? "tool-result" : null;
40
72
  }
@@ -68,10 +100,11 @@ function redaction(applied = false, truncated = false) { return { applied, value
68
100
  function identityReady(identity: BridgeIdentity | null): identity is BridgeIdentity & { agentOperationId: string } {
69
101
  return Boolean(identity?.agentOperationId);
70
102
  }
71
- function activityPayload(state: "started" | "progress" | "finished" | "failed", activityRef: string, toolName: string) {
103
+ function activityPayload(state: "started" | "progress" | "finished" | "failed" | "aborted", activityRef: string, toolName: string,
104
+ reasonCode: string | null = null) {
72
105
  return { state, activityType: "tool", activityRef, toolName, inputDigest: null, outputDigest: null, preview: null, previewKind: "none",
73
106
  outputBytes: null, outputLines: null, exitCode: null, isError: state === "finished" ? false : state === "failed" ? true : null,
74
- affectedFileRefs: [], criterionIds: [], verifierAttemptIds: [], reasonCode: state === "failed" ? "tool-execution-failed" : null };
107
+ affectedFileRefs: [], criterionIds: [], verifierAttemptIds: [], reasonCode };
75
108
  }
76
109
 
77
110
  export class PiSessionStreamAdapter {
@@ -81,12 +114,14 @@ export class PiSessionStreamAdapter {
81
114
  #activeByRole = new Map<MessageRole, ActiveMessage>();
82
115
  #lastMessageRef: string | null = null;
83
116
  #toolRefs = new Map<string, string>();
117
+ #toolNames = new Map<string, string>();
118
+ #toolInputs = new Map<string, { command: string } | undefined>();
84
119
  #lastToolProgress = new Map<string, number>();
85
120
 
86
121
  constructor(options: { now?: () => Date } = {}) { this.#now = options.now ?? (() => new Date()); }
87
122
  reset(): void {
88
123
  this.#turnIndex = null; this.#messageCounter = 0; this.#activeByRole.clear(); this.#lastMessageRef = null;
89
- this.#toolRefs.clear(); this.#lastToolProgress.clear();
124
+ this.#toolRefs.clear(); this.#toolNames.clear(); this.#toolInputs.clear(); this.#lastToolProgress.clear();
90
125
  }
91
126
 
92
127
  #base(snapshot: BridgeSnapshot, sourceObservedAt?: unknown): Omit<RuntimeEventDraft, "kind" | "payload" | "evidence" | "redaction"> | null {
@@ -120,7 +155,7 @@ export class PiSessionStreamAdapter {
120
155
  const source = messageRole === "tool-result" ? "" : safe(text(message), MAX_STREAM_BUFFER).text;
121
156
  this.#activeByRole.set(messageRole, { ref: messageRef, role: messageRole, rawBuffer: "", chunkSequence: 0, truncated: false,
122
157
  emittedEvents: 0, lastFlushAt: this.#now().getTime(),
123
- thinking: new Set(), thinkingStreaming: new Set(), thinkingCompleted: new Set() });
158
+ thinking: new Set(), thinkingStreaming: new Set(), thinkingCompleted: new Set(), streamActivity: newStreamActivity() });
124
159
  this.#lastMessageRef = messageRef;
125
160
  const draft = this.#draft(snapshot, "message.started", { role: messageRole, contentDigest: source ? digest(source) : null,
126
161
  textChars: source ? source.length : null, imageCount: imageCount(message) }, { sourceObservedAt: message?.timestamp, messageRef });
@@ -157,6 +192,7 @@ export class PiSessionStreamAdapter {
157
192
  }
158
193
  messageUpdated(event: { assistantMessageEvent?: any }, snapshot: BridgeSnapshot): RuntimeEventDraft[] {
159
194
  const update = event.assistantMessageEvent, active = this.#activeByRole.get("assistant"); if (!active || !update) return [];
195
+ observeStreamActivity(active.streamActivity, update.type, this.#now());
160
196
  if (update.type === "text_delta" && typeof update.delta === "string") {
161
197
  if (active.emittedEvents >= MAX_MESSAGE_STREAM_EVENTS) { active.truncated = true; return []; }
162
198
  const remaining = MAX_STREAM_BUFFER - active.rawBuffer.length;
@@ -195,28 +231,34 @@ export class PiSessionStreamAdapter {
195
231
  if (messageRole === "assistant" && (reason === "error" || reason === "aborted")) {
196
232
  const error = safe(message?.errorMessage, 500);
197
233
  const failed = this.#draft(snapshot, "message.failed", { role: "assistant", reason, errorCode: reason === "error" ? "assistant-message-error" : "assistant-message-aborted",
198
- message: error.text || null, contentDigest: finalText.text ? digest(finalText.text) : null },
234
+ message: error.text || null, contentDigest: finalText.text ? digest(finalText.text) : null,
235
+ streamActivity: { ...active.streamActivity } },
199
236
  { sourceObservedAt: message?.timestamp, messageRef: active.ref, redacted: finalText.redacted || error.redacted, truncated: finalText.truncated || error.truncated });
200
237
  if (failed) drafts.push(failed); return drafts;
201
238
  }
202
239
  const completed = this.#draft(snapshot, "message.completed", { role: messageRole, contentDigest: finalText.text ? digest(finalText.text) : null,
203
240
  contentRef: null, textPreview: messageRole === "tool-result" ? null : finalText.text.slice(0, 4_000), textChars: messageRole === "tool-result" ? null : finalText.text.length,
204
- blockCount: Math.min(10_000, Array.isArray(message?.content) ? message.content.length : finalText.text ? 1 : 0), stopReason: reason, usage: usage(message) },
241
+ blockCount: Math.min(10_000, Array.isArray(message?.content) ? message.content.length : finalText.text ? 1 : 0), stopReason: reason, usage: usage(message),
242
+ ...(messageRole === "assistant" ? { streamActivity: { ...active.streamActivity } } : {}) },
205
243
  { sourceObservedAt: message?.timestamp, messageRef: active.ref, redacted: finalText.redacted, truncated: finalText.truncated || active.truncated });
206
244
  if (completed) drafts.push(completed); return drafts;
207
245
  }
208
246
  turnEnded(event: { message?: unknown; toolResults?: unknown[] }, snapshot: BridgeSnapshot): RuntimeEventDraft[] {
209
247
  const reason = stopReason(event.message), count = Array.isArray(event.toolResults) ? Math.min(10_000, event.toolResults.length) : 0;
248
+ const drafts = this.#settleActiveTools(snapshot, "turn-ended-before-tool-result");
210
249
  const draft = this.#draft(snapshot, "turn.ended", { phase: "ended", finalMessageRef: this.#lastMessageRef, toolResultCount: count, stopReason: reason });
211
- return draft ? [draft] : [];
250
+ if (draft) drafts.push(draft);
251
+ return drafts;
212
252
  }
213
- toolStarted(event: { toolCallId?: unknown; toolName?: unknown }, snapshot: BridgeSnapshot): RuntimeEventDraft[] {
253
+ toolStarted(event: { toolCallId?: unknown; toolName?: unknown; args?: unknown }, snapshot: BridgeSnapshot): RuntimeEventDraft[] {
214
254
  if (typeof event.toolCallId !== "string" || !event.toolCallId) return [];
215
255
  const drafts = this.#activeByRole.get("assistant")
216
256
  ? this.#drainLines(snapshot, this.#activeByRole.get("assistant") as ActiveMessage, false) : [];
217
257
  const name = safeToolName(event.toolName);
218
258
  const ref = opaque("tool", [snapshot.identity?.sessionRef, event.toolCallId]), activityRef = opaque("activity", [snapshot.identity?.sessionRef, event.toolCallId]);
219
- this.#toolRefs.set(event.toolCallId, ref); this.#lastToolProgress.set(event.toolCallId, 0);
259
+ this.#toolRefs.set(event.toolCallId, ref); this.#toolNames.set(event.toolCallId, name);
260
+ this.#toolInputs.set(event.toolCallId, toolClassificationInput(event.args));
261
+ this.#lastToolProgress.set(event.toolCallId, 0);
220
262
  const draft = this.#draft(snapshot, "activity.started", activityPayload("started", activityRef, name), { toolCallId: ref });
221
263
  if (draft) drafts.push(draft);
222
264
  return drafts;
@@ -231,19 +273,49 @@ export class PiSessionStreamAdapter {
231
273
  const draft = this.#draft(snapshot, "activity.progress", activityPayload("progress", activityRef, name), { toolCallId: ref });
232
274
  return draft ? [draft] : [];
233
275
  }
234
- toolEnded(event: { toolCallId?: unknown; toolName?: unknown; isError?: unknown }, snapshot: BridgeSnapshot): RuntimeEventDraft[] {
276
+ toolEnded(event: { toolCallId?: unknown; toolName?: unknown; result?: unknown; isError?: unknown }, snapshot: BridgeSnapshot): RuntimeEventDraft[] {
235
277
  if (typeof event.toolCallId !== "string") return [];
236
278
  const ref = this.#toolRefs.get(event.toolCallId); if (!ref) return [];
237
279
  const name = safeToolName(event.toolName);
238
- const activityRef = opaque("activity", [snapshot.identity?.sessionRef, event.toolCallId]), state = event.isError === true ? "failed" : "finished";
239
- this.#toolRefs.delete(event.toolCallId); this.#lastToolProgress.delete(event.toolCallId);
240
- const draft = this.#draft(snapshot, `activity.${state}`, activityPayload(state, activityRef, name), { toolCallId: ref });
280
+ const reasonCode = classifyToolFailure(name, event.isError === true, event.result, this.#toolInputs.get(event.toolCallId));
281
+ const state = (event.isError === true || reasonCode !== null) && !handledToolFailure(reasonCode, name) ? "failed" : "finished";
282
+ const activityRef = opaque("activity", [snapshot.identity?.sessionRef, event.toolCallId]);
283
+ this.#toolRefs.delete(event.toolCallId); this.#toolNames.delete(event.toolCallId);
284
+ this.#toolInputs.delete(event.toolCallId); this.#lastToolProgress.delete(event.toolCallId);
285
+ const draft = this.#draft(snapshot, `activity.${state}`, activityPayload(state, activityRef, name, reasonCode), { toolCallId: ref });
241
286
  return draft ? [draft] : [];
242
287
  }
288
+ #settleActiveTools(snapshot: BridgeSnapshot, reasonCode: string): RuntimeEventDraft[] {
289
+ const drafts: RuntimeEventDraft[] = [];
290
+ for (const [rawToolCallId, ref] of this.#toolRefs) {
291
+ const toolName = this.#toolNames.get(rawToolCallId) ?? "tool";
292
+ const activityRef = opaque("activity", [snapshot.identity?.sessionRef, rawToolCallId]);
293
+ const draft = this.#draft(snapshot, "activity.aborted", activityPayload("aborted", activityRef, toolName, reasonCode),
294
+ { toolCallId: ref, evidence: "derived" });
295
+ if (draft) drafts.push(draft);
296
+ }
297
+ this.#toolRefs.clear(); this.#toolNames.clear(); this.#toolInputs.clear(); this.#lastToolProgress.clear();
298
+ return drafts;
299
+ }
300
+ operationInterrupted(snapshot: BridgeSnapshot, reasonCode = "session-replaced-before-tool-result"): RuntimeEventDraft[] {
301
+ return this.#settleActiveTools(snapshot, reasonCode);
302
+ }
243
303
  agentSettled(snapshot: BridgeSnapshot, hasPendingMessages: boolean | null): RuntimeEventDraft[] {
304
+ // A queued follow-up belongs to the same logical operation. Do not emit a
305
+ // terminal edge or clear stream identity between recovery turns.
306
+ if (hasPendingMessages === true) return [];
244
307
  const value = hasPendingMessages === null ? { state: "unknown", value: null, reasonCode: "queue-fact-unavailable" }
245
308
  : { state: "known", value: hasPendingMessages, reasonCode: null };
246
- const draft = this.#draft(snapshot, "agent-operation.settled", { settlement: "completed", lastStopReason: null, hasPendingMessages: value });
247
- this.reset(); return draft ? [draft] : [];
309
+ // Host-idle only proves that this agent operation stopped producing turns.
310
+ // A governed task is successful only after its durable contract is terminal.
311
+ // Fail closed here so a missed/late completion hook cannot paint an open
312
+ // task as completed in the WebUI.
313
+ const governedTaskOpen = Boolean(snapshot.identity?.taskRunId) && snapshot.taskState !== "terminal";
314
+ const drafts = this.#settleActiveTools(snapshot, "operation-settled-before-tool-result");
315
+ const draft = this.#draft(snapshot, "agent-operation.settled", {
316
+ settlement: governedTaskOpen ? "unknown" : "completed", lastStopReason: null, hasPendingMessages: value
317
+ });
318
+ if (draft) drafts.push(draft);
319
+ this.reset(); return drafts;
248
320
  }
249
321
  }
@@ -35,11 +35,19 @@ function timestamp(value: unknown): value is string {
35
35
  if (typeof value !== "string" || !/^\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}\.\d{3}Z$/.test(value)) return false;
36
36
  const parsed = Date.parse(value); return Number.isFinite(parsed) && new Date(parsed).toISOString() === value;
37
37
  }
38
+ function nullableRevision(value: unknown): value is string | null { return value === null || typeof value === "string" && REVISION.test(value); }
38
39
  function wireIdentity(value: BridgeIdentity): BridgeIdentity { return { projectRef: value.projectRef, runtimeInstanceId: value.runtimeInstanceId,
39
40
  sessionRef: value.sessionRef, taskId: value.taskId, taskRunId: value.taskRunId, agentOperationId: null, toolCallId: null }; }
40
41
  function wireRevisions(value: Record<string, any>): BridgeRevisions { return { runtimeRevision: value.runtimeRevision, taskRevision: value.taskRevision,
41
42
  controlRevision: value.controlRevision, workspaceRevision: value.workspaceRevision, indexRevision: value.indexRevision,
42
43
  approvalRevision: value.approvalRevision, sessionOptionRevision: value.sessionOptionRevision, queueRevision: value.queueRevision }; }
44
+ function replayIdentityDisposition(record: Pick<SourceMutationEvidenceRecord, "projectRef" | "runtimeInstanceId" | "sessionRef" | "taskId" | "taskRunId">,
45
+ identity: BridgeIdentity): "task-identity-mismatch" | "runtime-or-session-replaced" | null {
46
+ if (record.taskId !== identity.taskId || record.taskRunId !== identity.taskRunId) return "task-identity-mismatch";
47
+ if (record.projectRef !== identity.projectRef || record.runtimeInstanceId !== identity.runtimeInstanceId || record.sessionRef !== identity.sessionRef)
48
+ return "runtime-or-session-replaced";
49
+ return null;
50
+ }
43
51
 
44
52
  function receiptFromRecord(record: SourceMutationEvidenceRecord, deduplicated: boolean): SourceMutationReceipt {
45
53
  if (record.action === "source.revert") throw new Error("source-mutation-evidence-action-mismatch");
@@ -109,8 +117,9 @@ export class SourceMutationController {
109
117
  if (prior.length) {
110
118
  if (prior.some((record) => record.commandId !== command.commandId || record.idempotencyKeyDigest !== keyDigest || record.actionDigest !== command.actionDigest
111
119
  || record.action !== command.action)) return this.#reject(command, "idempotency-payload-mismatch", "idempotency-payload-mismatch", false);
112
- if (prior.some((record) => record.runtimeInstanceId !== identity.runtimeInstanceId || record.sessionRef !== identity.sessionRef))
113
- return this.#reject(command, "identity-mismatch", "runtime-or-session-replaced", false);
120
+ const dispositions = prior.map((record) => replayIdentityDisposition(record, identity)).filter(Boolean);
121
+ if (dispositions.length) return this.#reject(command, "identity-mismatch",
122
+ dispositions.includes("task-identity-mismatch") ? "task-identity-mismatch" : "runtime-or-session-replaced", false);
114
123
  return receiptFromRecord(prior.at(-1)!, true);
115
124
  }
116
125
  const now = this.#now().getTime();
@@ -122,6 +131,20 @@ export class SourceMutationController {
122
131
  let resolved: { projection: SourceMutationProjection; authority: SourceMutationAuthority | null };
123
132
  try { resolved = await this.#resolve(command.action, command.payload.fileRef); }
124
133
  catch { return this.#reject(command, "capability-unavailable", "mutation-preview-unavailable", true); }
134
+ const rebound = this.#bridge.snapshot();
135
+ if (rebound.state !== "ready" || !rebound.identity || !rebound.revisions)
136
+ return this.#reject(command, "resync-required", "mutation-binding-unavailable-after-preview", true);
137
+ const reboundIdentity = wireIdentity(rebound.identity), reboundRevisions = wireRevisions(rebound.revisions);
138
+ if (!same(reboundIdentity, identity)) {
139
+ const taskChanged = reboundIdentity.taskId !== identity.taskId || reboundIdentity.taskRunId !== identity.taskRunId;
140
+ return this.#reject(command, "identity-mismatch", taskChanged ? "task-identity-mismatch" : "runtime-or-session-replaced", false);
141
+ }
142
+ if (rebound.taskState === "terminal") return this.#reject(command, "terminal-task", "task-terminal-after-preview", false);
143
+ if (rebound.liveness !== "idle") return this.#reject(command, "capability-unavailable", "agent-not-idle-after-preview", true);
144
+ if (command.expectedRevisions.runtimeRevision !== reboundRevisions.runtimeRevision
145
+ || command.expectedRevisions.taskRevision !== reboundRevisions.taskRevision
146
+ || command.expectedRevisions.controlRevision !== reboundRevisions.controlRevision)
147
+ return this.#reject(command, "stale-revision", "stale-task-control-revision-after-preview", true);
125
148
  if (!resolved.authority || resolved.projection.state !== "ready" || !resolved.projection.target)
126
149
  return this.#reject(command, "capability-unavailable", resolved.projection.reasonCode ?? "mutation-preview-unavailable", true);
127
150
  const target = resolved.projection.target;
@@ -185,8 +208,11 @@ export class SourceMutationController {
185
208
  || command.identity.agentOperationId !== null || command.identity.toolCallId !== null) return "invalid-command-identity";
186
209
  const preimageKeys = [...REVISION_KEYS, "workspacePreimage", "indexPreimage", "patchPreimage"];
187
210
  if (!exactKeys(command.expectedRevisions, preimageKeys) || !REVISION.test(command.expectedRevisions.runtimeRevision)
188
- || !REVISION.test(command.expectedRevisions.taskRevision ?? "") || !REVISION.test(command.expectedRevisions.workspaceRevision ?? "")
211
+ || !REVISION.test(command.expectedRevisions.taskRevision ?? "") || !REVISION.test(command.expectedRevisions.controlRevision ?? "")
212
+ || !REVISION.test(command.expectedRevisions.workspaceRevision ?? "")
189
213
  || !REVISION.test(command.expectedRevisions.indexRevision ?? "") || !WORKSPACE.test(command.expectedRevisions.workspacePreimage)
214
+ || ![command.expectedRevisions.approvalRevision, command.expectedRevisions.sessionOptionRevision,
215
+ command.expectedRevisions.queueRevision].every(nullableRevision)
190
216
  || !DIGEST.test(command.expectedRevisions.indexPreimage) || !DIGEST.test(command.expectedRevisions.patchPreimage)) return "invalid-mutation-authority";
191
217
  if (!exactKeys(command.payload, ["fileRef", "hunkRefs", "contentDigest"]) || !REF.test(command.payload.fileRef)
192
218
  || !Array.isArray(command.payload.hunkRefs) || command.payload.hunkRefs.length > 128
@@ -33,6 +33,13 @@ function revisions(value: Record<string, any>, record?: SourceHandoffRecord): Br
33
33
  taskRevision: record?.taskRevision ?? value.taskRevision, controlRevision: value.controlRevision,
34
34
  workspaceRevision: record?.workspaceRevision ?? value.workspaceRevision, indexRevision: value.indexRevision,
35
35
  approvalRevision: value.approvalRevision, sessionOptionRevision: value.sessionOptionRevision, queueRevision: value.queueRevision }; }
36
+ function replayIdentityDisposition(record: Pick<SourceHandoffRecord, "projectRef" | "runtimeInstanceId" | "sessionRef" | "taskId" | "taskRunId">,
37
+ current: BridgeIdentity): "task-identity-mismatch" | "runtime-or-session-replaced" | null {
38
+ if (record.taskId !== current.taskId || record.taskRunId !== current.taskRunId) return "task-identity-mismatch";
39
+ if (record.projectRef !== current.projectRef || record.runtimeInstanceId !== current.runtimeInstanceId || record.sessionRef !== current.sessionRef)
40
+ return "runtime-or-session-replaced";
41
+ return null;
42
+ }
36
43
 
37
44
  export class SourceOpenController {
38
45
  readonly #bridge: SameSessionPiBridge; readonly #root: string; readonly #resolve: (fileRef: string) => Promise<SourceOpenAuthority | null>;
@@ -55,8 +62,9 @@ export class SourceOpenController {
55
62
  if (prior.length) {
56
63
  if (prior.some((record) => record.commandId !== command.commandId || record.idempotencyKeyDigest !== keyDigest || record.actionDigest !== command.actionDigest))
57
64
  return this.#reject(command, "idempotency-payload-mismatch", "idempotency-payload-mismatch", false);
58
- if (prior.some((record) => record.runtimeInstanceId !== currentIdentity.runtimeInstanceId || record.sessionRef !== currentIdentity.sessionRef))
59
- return this.#reject(command, "identity-mismatch", "runtime-or-session-replaced", false);
65
+ const dispositions = prior.map((record) => replayIdentityDisposition(record, currentIdentity)).filter(Boolean);
66
+ if (dispositions.length) return this.#reject(command, "identity-mismatch",
67
+ dispositions.includes("task-identity-mismatch") ? "task-identity-mismatch" : "runtime-or-session-replaced", false);
60
68
  const last = prior.at(-1)!; if (last.phase !== "requested") return this.#receipt(last, currentRevisions, true);
61
69
  const uncertain = this.#append(command, currentIdentity, keyDigest, last, "uncertain", "effect-unknown", "source-open-outcome-unknown");
62
70
  return this.#receipt(uncertain, currentRevisions, true);
@@ -64,9 +72,22 @@ export class SourceOpenController {
64
72
  const now = this.#now().getTime(); if (Date.parse(command.requestedAt) > now + 30_000) return this.#reject(command, "invalid-command", "requested-at-in-future", false);
65
73
  if (now > Date.parse(command.expiresAt)) return this.#reject(command, "expired", "command-expired", false);
66
74
  if (!same(command.identity, currentIdentity)) return this.#reject(command, "identity-mismatch", "identity-mismatch", false);
67
- if (command.expectedRevisions.runtimeRevision !== currentRevisions.runtimeRevision || command.expectedRevisions.taskRevision !== currentRevisions.taskRevision)
68
- return this.#reject(command, "stale-revision", "stale-source-open-revision", true);
75
+ if (command.expectedRevisions.runtimeRevision !== currentRevisions.runtimeRevision || command.expectedRevisions.taskRevision !== currentRevisions.taskRevision
76
+ || command.expectedRevisions.controlRevision !== currentRevisions.controlRevision)
77
+ return this.#reject(command, "stale-revision", "stale-task-control-revision", true);
69
78
  let authority: SourceOpenAuthority | null; try { authority = await this.#resolve(command.payload.fileRef); } catch { authority = null; }
79
+ const rebound = this.#bridge.snapshot();
80
+ if (rebound.state !== "ready" || !rebound.identity || !rebound.revisions)
81
+ return this.#reject(command, "resync-required", "source-open-binding-unavailable-after-target-resolution", true);
82
+ const reboundIdentity = identity(rebound.identity), reboundRevisions = revisions(rebound.revisions);
83
+ if (!same(reboundIdentity, currentIdentity)) {
84
+ const taskChanged = reboundIdentity.taskId !== currentIdentity.taskId || reboundIdentity.taskRunId !== currentIdentity.taskRunId;
85
+ return this.#reject(command, "identity-mismatch", taskChanged ? "task-identity-mismatch" : "runtime-or-session-replaced", false);
86
+ }
87
+ if (command.expectedRevisions.runtimeRevision !== reboundRevisions.runtimeRevision
88
+ || command.expectedRevisions.taskRevision !== reboundRevisions.taskRevision
89
+ || command.expectedRevisions.controlRevision !== reboundRevisions.controlRevision)
90
+ return this.#reject(command, "stale-revision", "stale-task-control-revision-after-target-resolution", true);
70
91
  if (!authority) return this.#reject(command, "capability-unavailable", "source-open-target-unavailable", true);
71
92
  if (authority.target.taskRevision !== command.expectedRevisions.taskRevision || authority.target.workspaceRevision !== command.expectedRevisions.workspaceRevision)
72
93
  return this.#reject(command, "stale-revision", "source-open-target-stale", true);
@@ -110,9 +131,10 @@ export class SourceOpenController {
110
131
  || command.identity.agentOperationId !== null || command.identity.toolCallId !== null) return "invalid-command-identity";
111
132
  if (!exact(command.expectedRevisions, [...REVISION_KEYS, "workspacePreimage", "indexPreimage", "patchPreimage"])
112
133
  || !REVISION.test(command.expectedRevisions.runtimeRevision) || !REVISION.test(command.expectedRevisions.taskRevision ?? "")
113
- || !REVISION.test(command.expectedRevisions.workspaceRevision ?? "") || command.expectedRevisions.workspacePreimage !== null
114
- || ![command.expectedRevisions.controlRevision, command.expectedRevisions.indexRevision, command.expectedRevisions.approvalRevision,
115
- command.expectedRevisions.sessionOptionRevision, command.expectedRevisions.queueRevision].every(nullableRevision)
134
+ || !REVISION.test(command.expectedRevisions.controlRevision ?? "") || !REVISION.test(command.expectedRevisions.workspaceRevision ?? "")
135
+ || command.expectedRevisions.workspacePreimage !== null
136
+ || ![command.expectedRevisions.indexRevision, command.expectedRevisions.approvalRevision, command.expectedRevisions.sessionOptionRevision,
137
+ command.expectedRevisions.queueRevision].every(nullableRevision)
116
138
  || command.expectedRevisions.indexPreimage !== null || command.expectedRevisions.patchPreimage !== null) return "invalid-source-open-authority";
117
139
  if (!exact(command.payload, ["fileRef", "line", "column"]) || !REF.test(command.payload.fileRef)
118
140
  || command.payload.line !== null && (!Number.isInteger(command.payload.line) || command.payload.line < 1 || command.payload.line > 100_000_000)
@@ -38,6 +38,13 @@ function wireIdentity(value: BridgeIdentity): BridgeIdentity { return { projectR
38
38
  function wireRevisions(value: Record<string, any>): BridgeRevisions { return { runtimeRevision: value.runtimeRevision, taskRevision: value.taskRevision,
39
39
  controlRevision: value.controlRevision, workspaceRevision: value.workspaceRevision, indexRevision: value.indexRevision,
40
40
  approvalRevision: value.approvalRevision, sessionOptionRevision: value.sessionOptionRevision, queueRevision: value.queueRevision }; }
41
+ function replayIdentityDisposition(record: Pick<SourceMutationEvidenceRecord, "projectRef" | "runtimeInstanceId" | "sessionRef" | "taskId" | "taskRunId">,
42
+ identity: BridgeIdentity): "task-identity-mismatch" | "runtime-or-session-replaced" | null {
43
+ if (record.taskId !== identity.taskId || record.taskRunId !== identity.taskRunId) return "task-identity-mismatch";
44
+ if (record.projectRef !== identity.projectRef || record.runtimeInstanceId !== identity.runtimeInstanceId || record.sessionRef !== identity.sessionRef)
45
+ return "runtime-or-session-replaced";
46
+ return null;
47
+ }
41
48
  function receipt(record: SourceMutationEvidenceRecord, deduplicated: boolean): SourceRevertReceipt {
42
49
  const identity = { projectRef: record.projectRef, runtimeInstanceId: record.runtimeInstanceId, sessionRef: record.sessionRef,
43
50
  taskId: record.taskId, taskRunId: record.taskRunId, agentOperationId: null, toolCallId: null } as BridgeIdentity;
@@ -90,8 +97,9 @@ export class SourceRevertController {
90
97
  if (prior.length) {
91
98
  if (prior.some((record) => record.commandId !== command.commandId || record.idempotencyKeyDigest !== keyDigest || record.actionDigest !== command.actionDigest
92
99
  || record.action !== command.action)) return this.#reject(command, "idempotency-payload-mismatch", "idempotency-payload-mismatch", false);
93
- if (prior.some((record) => record.runtimeInstanceId !== identity.runtimeInstanceId || record.sessionRef !== identity.sessionRef))
94
- return this.#reject(command, "identity-mismatch", "runtime-or-session-replaced", false);
100
+ const dispositions = prior.map((record) => replayIdentityDisposition(record, identity)).filter(Boolean);
101
+ if (dispositions.length) return this.#reject(command, "identity-mismatch",
102
+ dispositions.includes("task-identity-mismatch") ? "task-identity-mismatch" : "runtime-or-session-replaced", false);
95
103
  return receipt(prior.at(-1)!, true);
96
104
  }
97
105
  const now = this.#now().getTime();
@@ -103,10 +111,25 @@ export class SourceRevertController {
103
111
  let resolved: { projection: SourceRevertProjection; authority: SourceRevertAuthority | null };
104
112
  try { resolved = await this.#resolve(command.payload.fileRef, command.payload.hunkRefs); }
105
113
  catch { return this.#reject(command, "capability-unavailable", "revert-preview-unavailable", true); }
114
+ const rebound = this.#bridge.snapshot();
115
+ if (rebound.state !== "ready" || !rebound.identity || !rebound.revisions)
116
+ return this.#reject(command, "resync-required", "revert-binding-unavailable-after-preview", true);
117
+ const reboundIdentity = wireIdentity(rebound.identity), reboundRevisions = wireRevisions(rebound.revisions);
118
+ if (!same(reboundIdentity, identity)) {
119
+ const taskChanged = reboundIdentity.taskId !== identity.taskId || reboundIdentity.taskRunId !== identity.taskRunId;
120
+ return this.#reject(command, "identity-mismatch", taskChanged ? "task-identity-mismatch" : "runtime-or-session-replaced", false);
121
+ }
122
+ if (rebound.taskState === "terminal") return this.#reject(command, "terminal-task", "task-terminal-after-preview", false);
123
+ if (rebound.liveness !== "idle") return this.#reject(command, "capability-unavailable", "agent-not-idle-after-preview", true);
124
+ if (command.expectedRevisions.runtimeRevision !== reboundRevisions.runtimeRevision
125
+ || command.expectedRevisions.taskRevision !== reboundRevisions.taskRevision
126
+ || command.expectedRevisions.controlRevision !== reboundRevisions.controlRevision)
127
+ return this.#reject(command, "stale-revision", "stale-task-control-revision-after-preview", true);
106
128
  if (!resolved.authority || resolved.projection.state !== "ready" || !resolved.projection.target)
107
129
  return this.#reject(command, "capability-unavailable", resolved.projection.reasonCode ?? "revert-preview-unavailable", true);
108
130
  const target = resolved.projection.target;
109
- if (!this.#matches(command, target) || now > Date.parse(target.expiresAt)) return this.#reject(command, "stale-revision", "revert-preview-stale", true);
131
+ if (!this.#matches(command, target) || this.#now().getTime() > Date.parse(target.expiresAt))
132
+ return this.#reject(command, "stale-revision", "revert-preview-stale", true);
110
133
  const before = { ...bridgeRevisions, workspaceRevision: target.workspaceRevision, indexRevision: target.indexRevision };
111
134
  let requested: SourceMutationEvidenceRecord;
112
135
  try { requested = appendSourceMutationEvidence({ projectRoot: this.#projectRoot, projectRef: identity.projectRef, runtimeInstanceId: identity.runtimeInstanceId,