@sema-agent/core 7.58.0 → 7.60.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (519) hide show
  1. package/CHANGELOG.md +1340 -0
  2. package/README.md +6 -4
  3. package/dist/agents/agent-definition.d.ts +7 -0
  4. package/dist/agents/agent-definition.js +12 -0
  5. package/dist/agents/agent-transcript-tool.d.ts +2 -1
  6. package/dist/agents/agent-transcript-tool.js +3 -1
  7. package/dist/agents/cc-peer-directory.js +8 -8
  8. package/dist/agents/child-model-seat.d.ts +4 -2
  9. package/dist/agents/child-model-seat.js +9 -4
  10. package/dist/agents/cumulative-stats.js +0 -1
  11. package/dist/agents/inherited-tool-pool.d.ts +26 -1
  12. package/dist/agents/inherited-tool-pool.js +22 -0
  13. package/dist/agents/peer-directory.d.ts +9 -0
  14. package/dist/agents/peer-directory.js +6 -2
  15. package/dist/agents/retain-ledger.d.ts +11 -1
  16. package/dist/agents/retain-ledger.js +13 -2
  17. package/dist/agents/roster-store.d.ts +1 -1
  18. package/dist/agents/roster-store.js +3 -1
  19. package/dist/agents/send-message-tool.d.ts +8 -19
  20. package/dist/agents/send-message-tool.js +32 -50
  21. package/dist/agents/subagent.d.ts +51 -17
  22. package/dist/agents/subagent.js +150 -112
  23. package/dist/agents/tool-filter.d.ts +1 -3
  24. package/dist/agents/tool-filter.js +4 -12
  25. package/dist/bench/metrics.d.ts +3 -5
  26. package/dist/bench/metrics.js +0 -1
  27. package/dist/brain/anthropic.js +161 -45
  28. package/dist/brain/degrading.js +3 -1
  29. package/dist/brain/errors.d.ts +9 -3
  30. package/dist/brain/errors.js +7 -1
  31. package/dist/brain/failover.js +2 -1
  32. package/dist/brain/open-responses.js +3 -3
  33. package/dist/brain/openai.js +103 -53
  34. package/dist/brain/provider-dialect.d.ts +11 -0
  35. package/dist/brain/provider-dialect.js +5 -2
  36. package/dist/brain/reasoning-request.js +5 -4
  37. package/dist/brain/reasoning.d.ts +19 -0
  38. package/dist/brain/reasoning.js +17 -2
  39. package/dist/brain/request-route.d.ts +4 -1
  40. package/dist/brain/request-route.js +4 -1
  41. package/dist/brain/route-adjudicator.d.ts +84 -18
  42. package/dist/brain/route-adjudicator.js +90 -22
  43. package/dist/brain/route-conformance.js +1 -1
  44. package/dist/brain/status-sink.d.ts +7 -1
  45. package/dist/brain/status-sink.js +6 -0
  46. package/dist/brain/stream-engine.js +84 -49
  47. package/dist/brain/stream-shared.d.ts +30 -0
  48. package/dist/brain/stream-shared.js +34 -1
  49. package/dist/brain/timeout.d.ts +65 -6
  50. package/dist/brain/timeout.js +111 -11
  51. package/dist/core/a2a.d.ts +15 -5
  52. package/dist/core/a2a.js +34 -19
  53. package/dist/core/account-row.d.ts +61 -0
  54. package/dist/core/account-row.js +51 -0
  55. package/dist/core/advisor-verdict.d.ts +2 -1
  56. package/dist/core/advisor-verdict.js +1 -1
  57. package/dist/core/agent-definition.d.ts +20 -0
  58. package/dist/core/allow-entry.d.ts +20 -6
  59. package/dist/core/allow-entry.js +17 -7
  60. package/dist/core/approval-object.d.ts +29 -15
  61. package/dist/core/approval-object.js +47 -19
  62. package/dist/core/artifact-host.d.ts +0 -2
  63. package/dist/core/artifact-host.js +0 -1
  64. package/dist/core/auto-compaction.js +1 -1
  65. package/dist/core/auto-mode-defaults.d.ts +8 -0
  66. package/dist/core/auto-mode-defaults.js +1 -0
  67. package/dist/core/auto-mode-rebuild.d.ts +2 -1
  68. package/dist/core/auto-mode.d.ts +39 -4
  69. package/dist/core/auto-mode.js +63 -12
  70. package/dist/core/background-agent-store.d.ts +26 -4
  71. package/dist/core/background-agent-store.js +3 -2
  72. package/dist/core/background-shell.d.ts +12 -1
  73. package/dist/core/background-shell.js +12 -0
  74. package/dist/core/checkpoint-requirements.d.ts +13 -1
  75. package/dist/core/checkpoint-requirements.js +3 -1
  76. package/dist/core/checkpoint-store.d.ts +66 -25
  77. package/dist/core/checkpoint-store.js +22 -9
  78. package/dist/core/child-clamp-seats.d.ts +29 -0
  79. package/dist/core/child-clamp-seats.js +56 -1
  80. package/dist/core/compliance.d.ts +3 -3
  81. package/dist/core/compliance.js +1 -1
  82. package/dist/core/credential-echo.d.ts +52 -0
  83. package/dist/core/credential-echo.js +65 -0
  84. package/dist/core/credential-vocabulary.d.ts +13 -3
  85. package/dist/core/credential-vocabulary.js +12 -4
  86. package/dist/core/cross-boundary-taint.d.ts +4 -5
  87. package/dist/core/effective-path-target.d.ts +8 -2
  88. package/dist/core/effective-path-target.js +8 -3
  89. package/dist/core/engine-notice.d.ts +156 -23
  90. package/dist/core/engine-notice.js +19 -5
  91. package/dist/core/frame-settlement.d.ts +5 -0
  92. package/dist/core/frame-settlement.js +4 -0
  93. package/dist/core/gate-fold.js +8 -6
  94. package/dist/core/gate-lanes.js +47 -12
  95. package/dist/core/gate-outcome.d.ts +32 -3
  96. package/dist/core/gate-outcome.js +19 -1
  97. package/dist/core/governance-codes.d.ts +17 -1
  98. package/dist/core/governance-codes.js +25 -0
  99. package/dist/core/hands-band-screen.js +0 -1
  100. package/dist/core/hands-band.d.ts +0 -3
  101. package/dist/core/hooks.d.ts +70 -59
  102. package/dist/core/hooks.js +125 -80
  103. package/dist/core/image-admission.d.ts +106 -52
  104. package/dist/core/image-admission.js +184 -56
  105. package/dist/core/image-downsample.d.ts +36 -69
  106. package/dist/core/image-downsample.js +111 -73
  107. package/dist/core/integer-option.d.ts +21 -0
  108. package/dist/core/integer-option.js +13 -0
  109. package/dist/core/locked-config.d.ts +59 -5
  110. package/dist/core/locked-config.js +82 -50
  111. package/dist/core/mailbox-store.d.ts +45 -1
  112. package/dist/core/mailbox-store.js +53 -9
  113. package/dist/core/mcp-call-meta.d.ts +2 -2
  114. package/dist/core/mcp-call-meta.js +9 -10
  115. package/dist/core/mcp-failure.d.ts +122 -20
  116. package/dist/core/mcp-failure.js +73 -5
  117. package/dist/core/mcp-host-seat.d.ts +92 -0
  118. package/dist/core/mcp-host-seat.js +131 -0
  119. package/dist/core/mcp-probe.d.ts +7 -2
  120. package/dist/core/mcp-probe.js +7 -2
  121. package/dist/core/mcp-server-link.d.ts +22 -1
  122. package/dist/core/mcp-server-link.js +25 -1
  123. package/dist/core/mcp-server-spec.d.ts +38 -9
  124. package/dist/core/mcp-server-spec.js +55 -21
  125. package/dist/core/mcp.d.ts +27 -12
  126. package/dist/core/mcp.js +105 -76
  127. package/dist/core/memory-admission.d.ts +23 -6
  128. package/dist/core/memory-admission.js +19 -9
  129. package/dist/core/memory-engine/consolidation-driver.d.ts +2 -4
  130. package/dist/core/memory-engine/consolidation-driver.js +3 -3
  131. package/dist/core/memory-engine/consolidation.d.ts +4 -0
  132. package/dist/core/memory-engine/consolidation.js +1 -0
  133. package/dist/core/memory-engine/content-origin.d.ts +13 -0
  134. package/dist/core/memory-engine/content-origin.js +21 -3
  135. package/dist/core/memory-engine/delegation-settlement.js +5 -12
  136. package/dist/core/memory-engine/distiller.js +2 -2
  137. package/dist/core/memory-engine/dual-root.d.ts +50 -1
  138. package/dist/core/memory-engine/dual-root.js +44 -1
  139. package/dist/core/memory-engine/engine.d.ts +67 -14
  140. package/dist/core/memory-engine/engine.js +384 -212
  141. package/dist/core/memory-engine/export-bundle.d.ts +23 -0
  142. package/dist/core/memory-engine/export-bundle.js +23 -4
  143. package/dist/core/memory-engine/file-backend.d.ts +128 -55
  144. package/dist/core/memory-engine/file-backend.js +290 -117
  145. package/dist/core/memory-engine/index.d.ts +3 -2
  146. package/dist/core/memory-engine/index.js +2 -1
  147. package/dist/core/memory-engine/layout.d.ts +59 -11
  148. package/dist/core/memory-engine/layout.js +190 -40
  149. package/dist/core/memory-engine/memory-backend-contract.d.ts +1 -1
  150. package/dist/core/memory-engine/memory-backend-contract.js +39 -2
  151. package/dist/core/memory-engine/scope-contract.d.ts +13 -3
  152. package/dist/core/memory-engine/scope-contract.js +2 -2
  153. package/dist/core/memory-engine/sync-client.d.ts +51 -9
  154. package/dist/core/memory-engine/sync-client.js +107 -13
  155. package/dist/core/memory-engine/sync-report.d.ts +103 -0
  156. package/dist/core/memory-engine/sync-report.js +119 -0
  157. package/dist/core/memory-engine/sync.d.ts +8 -1
  158. package/dist/core/memory-engine/types.d.ts +6 -0
  159. package/dist/core/mode-policy.d.ts +9 -1
  160. package/dist/core/mode-policy.js +11 -1
  161. package/dist/core/model-receipt.d.ts +39 -0
  162. package/dist/core/model-receipt.js +139 -0
  163. package/dist/core/mods/agent-stations.d.ts +65 -0
  164. package/dist/core/mods/agent-stations.js +105 -0
  165. package/dist/core/mods/classic-events.d.ts +5 -1
  166. package/dist/core/mods/kernel/dispatch.d.ts +5 -1
  167. package/dist/core/mods/kernel/dispatch.js +6 -5
  168. package/dist/core/mods/kernel/events.d.ts +11 -4
  169. package/dist/core/mods/kernel/events.js +11 -7
  170. package/dist/core/mods/kernel/receive.d.ts +5 -0
  171. package/dist/core/mods/kernel/receive.js +10 -0
  172. package/dist/core/mods/kernel/registration.d.ts +4 -0
  173. package/dist/core/mods/kernel/registration.js +3 -0
  174. package/dist/core/mods/lifecycle-stations.d.ts +41 -0
  175. package/dist/core/mods/lifecycle-stations.js +35 -0
  176. package/dist/core/mods/stations.d.ts +73 -60
  177. package/dist/core/mods/stations.js +64 -40
  178. package/dist/core/mods/tool-call.d.ts +10 -4
  179. package/dist/core/mods/tool-call.js +28 -12
  180. package/dist/core/mounted-mcp.d.ts +5 -0
  181. package/dist/core/mounted-mcp.js +2 -2
  182. package/dist/core/namespaced-rule-name.d.ts +7 -1
  183. package/dist/core/namespaced-rule-name.js +13 -2
  184. package/dist/core/oracle-isolation.js +1 -1
  185. package/dist/core/own-property-receipt.d.ts +3 -0
  186. package/dist/core/own-property-receipt.js +3 -0
  187. package/dist/core/parent-spec-seat.d.ts +8 -3
  188. package/dist/core/parent-spec-seat.js +2 -0
  189. package/dist/core/park-selfcheck.js +14 -8
  190. package/dist/core/peer-seat.d.ts +133 -16
  191. package/dist/core/peer-seat.js +319 -42
  192. package/dist/core/permission-rule-consent.d.ts +0 -21
  193. package/dist/core/permission-rule-consent.js +11 -6
  194. package/dist/core/permission-rule-model.d.ts +20 -2
  195. package/dist/core/permission-rule-model.js +8 -7
  196. package/dist/core/present-plan-tool.d.ts +11 -3
  197. package/dist/core/present-plan-tool.js +2 -2
  198. package/dist/core/protocol-http-auth.d.ts +14 -2
  199. package/dist/core/protocol-http-auth.js +16 -7
  200. package/dist/core/read-face-seats.d.ts +7 -5
  201. package/dist/core/read-face-seats.js +4 -1
  202. package/dist/core/redaction-marks.d.ts +4 -3
  203. package/dist/core/redaction-marks.js +2 -2
  204. package/dist/core/remote-env.d.ts +146 -98
  205. package/dist/core/remote-env.js +113 -14
  206. package/dist/core/resume-startup-cause.js +1 -0
  207. package/dist/core/retention.js +37 -12
  208. package/dist/core/retired-keys.d.ts +6 -1
  209. package/dist/core/retired-keys.js +4 -1
  210. package/dist/core/rewrite-rejudge.d.ts +13 -5
  211. package/dist/core/rewrite-rejudge.js +9 -5
  212. package/dist/core/roles.d.ts +69 -3
  213. package/dist/core/roles.js +167 -80
  214. package/dist/core/rule-owner.d.ts +3 -0
  215. package/dist/core/runner/announce-once-ledger.d.ts +5 -4
  216. package/dist/core/runner/announce-once-ledger.js +4 -2
  217. package/dist/core/runner/assemble-result.d.ts +0 -5
  218. package/dist/core/runner/assemble-result.js +0 -1
  219. package/dist/core/runner/call-hook-stations.d.ts +17 -10
  220. package/dist/core/runner/call-hook-stations.js +23 -12
  221. package/dist/core/runner/child-tree.d.ts +7 -8
  222. package/dist/core/runner/child-tree.js +6 -6
  223. package/dist/core/runner/classifier-ladder.d.ts +10 -2
  224. package/dist/core/runner/classifier-ladder.js +39 -8
  225. package/dist/core/runner/construction-doors.js +5 -0
  226. package/dist/core/runner/contracts.d.ts +82 -57
  227. package/dist/core/runner/credential-origin-seat.js +4 -3
  228. package/dist/core/runner/cross-boundary-lane.d.ts +3 -5
  229. package/dist/core/runner/cross-boundary-lane.js +11 -2
  230. package/dist/core/runner/cwd-alignment-probe.d.ts +20 -0
  231. package/dist/core/runner/cwd-alignment-probe.js +21 -0
  232. package/dist/core/runner/decide-continuation.d.ts +2 -2
  233. package/dist/core/runner/env-capability-census.d.ts +15 -20
  234. package/dist/core/runner/env-capability-census.js +25 -34
  235. package/dist/core/runner/file-history-request.d.ts +2 -2
  236. package/dist/core/runner/file-history-request.js +0 -8
  237. package/dist/core/runner/gate-exit.d.ts +87 -11
  238. package/dist/core/runner/gate-exit.js +30 -8
  239. package/dist/core/runner/git-status-frame.d.ts +5 -2
  240. package/dist/core/runner/git-status-frame.js +3 -4
  241. package/dist/core/runner/image-removed-seat.d.ts +12 -0
  242. package/dist/core/runner/image-removed-seat.js +19 -0
  243. package/dist/core/runner/image.d.ts +7 -2
  244. package/dist/core/runner/image.js +14 -8
  245. package/dist/core/runner/inline-image.d.ts +10 -8
  246. package/dist/core/runner/inline-image.js +11 -9
  247. package/dist/core/runner/memory-consolidation-driver.js +5 -5
  248. package/dist/core/runner/park-commit.d.ts +1 -1
  249. package/dist/core/runner/park-commit.js +3 -0
  250. package/dist/core/runner/permission-rule-lanes.js +1 -1
  251. package/dist/core/runner/prepare-ask-lane.d.ts +1 -1
  252. package/dist/core/runner/prepare-ask-lane.js +2 -1
  253. package/dist/core/runner/prepare-boundary-parks.d.ts +2 -2
  254. package/dist/core/runner/prepare-boundary-parks.js +5 -5
  255. package/dist/core/runner/prepare-caps-and-workflow.d.ts +7 -5
  256. package/dist/core/runner/prepare-caps-and-workflow.js +6 -3
  257. package/dist/core/runner/prepare-config-doors.d.ts +6 -1
  258. package/dist/core/runner/prepare-config-doors.js +16 -9
  259. package/dist/core/runner/prepare-context-lane.js +2 -0
  260. package/dist/core/runner/prepare-delegation-surface.d.ts +9 -3
  261. package/dist/core/runner/prepare-delegation-surface.js +5 -4
  262. package/dist/core/runner/prepare-file-history.d.ts +10 -12
  263. package/dist/core/runner/prepare-file-history.js +3 -12
  264. package/dist/core/runner/prepare-gate-stations.d.ts +4 -3
  265. package/dist/core/runner/prepare-gate-stations.js +4 -3
  266. package/dist/core/runner/prepare-hands-readface.d.ts +7 -1
  267. package/dist/core/runner/prepare-hands-readface.js +19 -6
  268. package/dist/core/runner/prepare-inherited-gate.d.ts +6 -7
  269. package/dist/core/runner/prepare-inherited-gate.js +22 -31
  270. package/dist/core/runner/prepare-memory-engine-session.js +1 -1
  271. package/dist/core/runner/prepare-memory.d.ts +2 -2
  272. package/dist/core/runner/prepare-memory.js +12 -19
  273. package/dist/core/runner/prepare-park-ask.d.ts +4 -6
  274. package/dist/core/runner/prepare-park-ask.js +5 -9
  275. package/dist/core/runner/prepare-policy-chain.d.ts +8 -7
  276. package/dist/core/runner/prepare-policy-chain.js +136 -121
  277. package/dist/core/runner/prepare-project-context.d.ts +2 -2
  278. package/dist/core/runner/prepare-project-context.js +10 -7
  279. package/dist/core/runner/prepare-prompt-inputs.d.ts +5 -2
  280. package/dist/core/runner/prepare-prompt-inputs.js +10 -20
  281. package/dist/core/runner/prepare-protocol-tools.d.ts +5 -3
  282. package/dist/core/runner/prepare-protocol-tools.js +17 -2
  283. package/dist/core/runner/prepare-run-refs.d.ts +2 -3
  284. package/dist/core/runner/prepare-run-refs.js +4 -1
  285. package/dist/core/runner/prepare-suspend-saga.d.ts +0 -2
  286. package/dist/core/runner/prepare-suspend-saga.js +5 -7
  287. package/dist/core/runner/prepare-task.js +44 -28
  288. package/dist/core/runner/prepare-turn-wiring.d.ts +2 -0
  289. package/dist/core/runner/prepare-turn-wiring.js +7 -2
  290. package/dist/core/runner/prepare-wiring-manifest.d.ts +5 -4
  291. package/dist/core/runner/prepare-wiring-manifest.js +34 -19
  292. package/dist/core/runner/prepare-workspace-restore.js +3 -3
  293. package/dist/core/runner/reply-pricing.d.ts +30 -0
  294. package/dist/core/runner/reply-pricing.js +19 -0
  295. package/dist/core/runner/resume-admission.js +4 -2
  296. package/dist/core/runner/resume-apply.js +6 -1
  297. package/dist/core/runner/resume-checkpoint-screen.js +4 -4
  298. package/dist/core/runner/resume-policy-outcome.js +5 -5
  299. package/dist/core/runner/retired-runner-deps.js +4 -1
  300. package/dist/core/runner/role-seat-defaulted.d.ts +44 -0
  301. package/dist/core/runner/role-seat-defaulted.js +40 -0
  302. package/dist/core/runner/run-attachment-seats.js +3 -3
  303. package/dist/core/runner/run-brain-sinks.js +3 -0
  304. package/dist/core/runner/run-git-lane.d.ts +3 -1
  305. package/dist/core/runner/run-git-lane.js +1 -1
  306. package/dist/core/runner/run-harness-handlers.d.ts +3 -1
  307. package/dist/core/runner/run-harness-handlers.js +3 -1
  308. package/dist/core/runner/run-identity-wiring.d.ts +2 -5
  309. package/dist/core/runner/run-identity-wiring.js +6 -5
  310. package/dist/core/runner/run-leg.d.ts +2 -0
  311. package/dist/core/runner/run-leg.js +63 -56
  312. package/dist/core/runner/run-notification-lane.d.ts +3 -1
  313. package/dist/core/runner/run-notification-lane.js +16 -10
  314. package/dist/core/runner/run-settle-and-teardown.js +5 -1
  315. package/dist/core/runner/run-stop-and-final-verify.d.ts +4 -0
  316. package/dist/core/runner/run-stop-and-final-verify.js +14 -6
  317. package/dist/core/runner/run-terminal-adoption.d.ts +2 -5
  318. package/dist/core/runner/run-terminal-adoption.js +9 -3
  319. package/dist/core/runner/run-turn-boundary.js +89 -77
  320. package/dist/core/runner/runtask.d.ts +8 -5
  321. package/dist/core/runner/runtask.js +21 -13
  322. package/dist/core/runner/side-call-account.d.ts +1 -1
  323. package/dist/core/runner/side-call-account.js +3 -18
  324. package/dist/core/runner/skills-seat.js +2 -2
  325. package/dist/core/runner/sub-agent-deny-lists.d.ts +24 -0
  326. package/dist/core/runner/sub-agent-deny-lists.js +20 -7
  327. package/dist/core/runner/synthetic-tools.js +2 -2
  328. package/dist/core/runner/tool-disclosure.js +0 -2
  329. package/dist/core/runner/tool-end-body.js +2 -2
  330. package/dist/core/runner/tool-face-overlay.d.ts +1 -1
  331. package/dist/core/runner/tool-face-overlay.js +26 -19
  332. package/dist/core/runner/tool-occurrence.d.ts +1 -1
  333. package/dist/core/runner/tool-occurrence.js +2 -2
  334. package/dist/core/runner/tool-result-media-seat.d.ts +9 -1
  335. package/dist/core/runner/tool-result-media-seat.js +17 -1
  336. package/dist/core/runner/turn-attachments.d.ts +29 -27
  337. package/dist/core/runner/turn-attachments.js +27 -42
  338. package/dist/core/runner-deps.d.ts +32 -19
  339. package/dist/core/secret-env.d.ts +35 -4
  340. package/dist/core/secret-env.js +65 -3
  341. package/dist/core/sensitive-path-policy.js +3 -3
  342. package/dist/core/session-policy-store.d.ts +12 -3
  343. package/dist/core/session-policy-store.js +29 -10
  344. package/dist/core/session-store.d.ts +5 -7
  345. package/dist/core/session-store.js +50 -11
  346. package/dist/core/side-query.d.ts +11 -1
  347. package/dist/core/side-query.js +9 -0
  348. package/dist/core/skill-grant-binding.d.ts +31 -2
  349. package/dist/core/skill-grant-binding.js +31 -1
  350. package/dist/core/skill-grant.d.ts +17 -5
  351. package/dist/core/skill-grant.js +24 -6
  352. package/dist/core/store-contracts/background-agent-store-contract.js +27 -0
  353. package/dist/core/store-contracts/checkpoint-store-contract.js +3 -2
  354. package/dist/core/store-contracts/mailbox-store-contract.js +31 -1
  355. package/dist/core/store-contracts/tool-result-store-contract.d.ts +1 -1
  356. package/dist/core/store-contracts/tool-result-store-contract.js +37 -0
  357. package/dist/core/stub-env.d.ts +5 -5
  358. package/dist/core/stub-env.js +11 -1
  359. package/dist/core/swappable-deps.d.ts +17 -43
  360. package/dist/core/swappable-deps.js +15 -5
  361. package/dist/core/synthetic-read-verdict.d.ts +1 -1
  362. package/dist/core/synthetic-read-verdict.js +2 -1
  363. package/dist/core/task-event.d.ts +50 -10
  364. package/dist/core/task-notification.d.ts +88 -5
  365. package/dist/core/task-notification.js +146 -17
  366. package/dist/core/task-registry-agent.d.ts +18 -12
  367. package/dist/core/task-registry-agent.js +19 -6
  368. package/dist/core/task-registry-monitor.d.ts +7 -0
  369. package/dist/core/task-registry-monitor.js +28 -1
  370. package/dist/core/task-registry-shared.d.ts +43 -3
  371. package/dist/core/task-registry-shared.js +21 -7
  372. package/dist/core/task-registry-workflow.js +1 -1
  373. package/dist/core/task-registry.d.ts +50 -35
  374. package/dist/core/task-registry.js +64 -23
  375. package/dist/core/task-result.d.ts +3 -12
  376. package/dist/core/task-spec.d.ts +15 -7
  377. package/dist/core/task-stream.d.ts +1 -1
  378. package/dist/core/terminal-cause.d.ts +10 -0
  379. package/dist/core/tool-catalog-entries.js +1 -1
  380. package/dist/core/tool-check-reading.d.ts +1 -0
  381. package/dist/core/tool-check-reading.js +3 -3
  382. package/dist/core/tool-face-reader.d.ts +7 -6
  383. package/dist/core/tool-face-reader.js +4 -3
  384. package/dist/core/tool-name-aliases.d.ts +19 -0
  385. package/dist/core/tool-name-aliases.js +54 -2
  386. package/dist/core/tool-policy.d.ts +21 -4
  387. package/dist/core/tool-policy.js +1 -1
  388. package/dist/core/tool-registry.d.ts +5 -5
  389. package/dist/core/tool-registry.js +2 -2
  390. package/dist/core/tool-result-store.d.ts +88 -5
  391. package/dist/core/tool-result-store.js +51 -12
  392. package/dist/core/tool-roster.d.ts +112 -5
  393. package/dist/core/tool-roster.js +42 -1
  394. package/dist/core/tool-spec.d.ts +23 -5
  395. package/dist/core/trace.d.ts +0 -1
  396. package/dist/core/types.d.ts +5 -5
  397. package/dist/core/types.js +2 -2
  398. package/dist/core/untrusted-text.js +1 -1
  399. package/dist/core/user-attachment.d.ts +10 -5
  400. package/dist/core/user-attachment.js +15 -7
  401. package/dist/core/wiring-manifest.d.ts +6 -3
  402. package/dist/engine/compaction/compaction.js +29 -3
  403. package/dist/engine/execution-env/node-execution-env.d.ts +11 -5
  404. package/dist/engine/execution-env/node-execution-env.js +21 -13
  405. package/dist/engine/harness/agent-harness.d.ts +26 -3
  406. package/dist/engine/harness/agent-harness.js +49 -16
  407. package/dist/engine/harness/messages.d.ts +3 -1
  408. package/dist/engine/harness/messages.js +3 -2
  409. package/dist/engine/harness/types.d.ts +119 -28
  410. package/dist/engine/llm/rejected-images.d.ts +46 -0
  411. package/dist/engine/llm/rejected-images.js +54 -0
  412. package/dist/engine/llm/tool-turns.d.ts +2 -1
  413. package/dist/engine/llm/types.d.ts +18 -7
  414. package/dist/engine/loop/agent-loop.d.ts +20 -12
  415. package/dist/engine/loop/agent-loop.js +205 -36
  416. package/dist/engine/loop/types.d.ts +12 -3
  417. package/dist/engine/lsp/node-lsp-manager.d.ts +3 -3
  418. package/dist/engine/lsp/node-lsp-manager.js +32 -13
  419. package/dist/engine/session/path-to-root.d.ts +38 -0
  420. package/dist/engine/session/path-to-root.js +39 -0
  421. package/dist/engine/session/storage-base.js +4 -21
  422. package/dist/index.d.ts +35 -21
  423. package/dist/index.js +26 -17
  424. package/dist/internal/harness-types.d.ts +1 -1
  425. package/dist/internal/harness.d.ts +1 -0
  426. package/dist/internal/llm.d.ts +1 -0
  427. package/dist/orchestration/run-workflow-tool.d.ts +9 -3
  428. package/dist/orchestration/run-workflow-tool.js +59 -29
  429. package/dist/orchestration/workflow-governance.d.ts +1 -1
  430. package/dist/orchestration/workflow-governance.js +2 -1
  431. package/dist/orchestration/workflow-observe.d.ts +17 -1
  432. package/dist/orchestration/workflow-observe.js +20 -0
  433. package/dist/orchestration/workflow-types.d.ts +32 -7
  434. package/dist/orchestration/workflow.d.ts +4 -0
  435. package/dist/orchestration/workflow.js +762 -491
  436. package/dist/prompts/default.d.ts +1 -10
  437. package/dist/prompts/default.js +0 -2
  438. package/dist/shell/counting-pass.d.ts +50 -0
  439. package/dist/shell/counting-pass.js +502 -0
  440. package/dist/shell/lex.d.ts +4 -2
  441. package/dist/shell/lex.js +93 -34
  442. package/dist/shell/shell-words.d.ts +1 -1
  443. package/dist/shell/shell-words.js +7 -6
  444. package/dist/stores/file/background-agent-store.d.ts +1 -1
  445. package/dist/stores/file/background-agent-store.js +3 -1
  446. package/dist/stores/file/checkpoint-store.d.ts +127 -2
  447. package/dist/stores/file/checkpoint-store.js +19 -14
  448. package/dist/stores/file/file-history-store.d.ts +1 -1
  449. package/dist/stores/file/file-history-store.js +3 -1
  450. package/dist/stores/file/index.d.ts +4 -3
  451. package/dist/stores/file/index.js +17 -8
  452. package/dist/stores/file/mailbox-store.d.ts +1 -1
  453. package/dist/stores/file/mailbox-store.js +9 -11
  454. package/dist/stores/file/memory-store.d.ts +1 -1
  455. package/dist/stores/file/memory-store.js +3 -1
  456. package/dist/stores/file/session-policy-store.d.ts +1 -1
  457. package/dist/stores/file/session-policy-store.js +3 -1
  458. package/dist/stores/file/session-store.d.ts +1 -1
  459. package/dist/stores/file/session-store.js +3 -1
  460. package/dist/stores/file/shared-ledger.d.ts +26 -0
  461. package/dist/stores/file/shared-ledger.js +38 -9
  462. package/dist/stores/file/tool-result-store.d.ts +20 -2
  463. package/dist/stores/file/tool-result-store.js +74 -6
  464. package/dist/stores/file/usage-window-store.d.ts +1 -1
  465. package/dist/stores/file/usage-window-store.js +3 -1
  466. package/dist/stores/file/workflow-journal-store.d.ts +1 -11
  467. package/dist/stores/file/workflow-journal-store.js +3 -2
  468. package/dist/tools/fs/containment-face.d.ts +5 -0
  469. package/dist/tools/fs/encoding.d.ts +13 -4
  470. package/dist/tools/fs/encoding.js +17 -2
  471. package/dist/tools/fs/fs-bash-background-text.d.ts +7 -0
  472. package/dist/tools/fs/fs-bash-background-text.js +1 -0
  473. package/dist/tools/fs/fs-bash.d.ts +14 -4
  474. package/dist/tools/fs/fs-bash.js +29 -35
  475. package/dist/tools/fs/fs-pdf.d.ts +2 -2
  476. package/dist/tools/fs/fs-pdf.js +26 -21
  477. package/dist/tools/fs/fs-read.d.ts +5 -4
  478. package/dist/tools/fs/fs-read.js +35 -18
  479. package/dist/tools/fs/fs-search-tools.js +4 -1
  480. package/dist/tools/fs/fs-shared.d.ts +11 -13
  481. package/dist/tools/fs/fs-shared.js +11 -7
  482. package/dist/tools/fs/fs-write.js +5 -5
  483. package/dist/tools/fs/index.d.ts +14 -8
  484. package/dist/tools/fs/index.js +13 -5
  485. package/dist/tools/fs/program-read-model.d.ts +13 -0
  486. package/dist/tools/fs/program-read-model.js +15 -0
  487. package/dist/tools/fs/program-read-model.rows.js +1 -1
  488. package/dist/tools/fs/read-deny.d.ts +13 -10
  489. package/dist/tools/fs/read-deny.js +25 -5
  490. package/dist/tools/fs/read-floor.d.ts +62 -0
  491. package/dist/tools/fs/read-floor.js +58 -0
  492. package/dist/tools/fs/removal-targets.d.ts +9 -4
  493. package/dist/tools/fs/removal-targets.js +17 -9
  494. package/dist/tools/fs/search.d.ts +3 -3
  495. package/dist/tools/fs/search.js +4 -4
  496. package/dist/tools/fs/shell-read-check.d.ts +1 -1
  497. package/dist/tools/fs/shell-read-check.js +2 -1
  498. package/dist/tools/fs/shell-read-judge.d.ts +16 -8
  499. package/dist/tools/fs/shell-read-judge.js +10 -6
  500. package/dist/tools/fs/shell-read-plan.d.ts +6 -1
  501. package/dist/tools/fs/shell-read-plan.js +22 -15
  502. package/dist/tools/fs/shell-read-start.js +1 -0
  503. package/dist/tools/fs/shell-reentry-text.d.ts +15 -1
  504. package/dist/tools/fs/shell-reentry-text.js +25 -3
  505. package/dist/tools/monitor.d.ts +9 -4
  506. package/dist/tools/monitor.js +17 -8
  507. package/dist/tools/scheduler-tools.js +42 -4
  508. package/dist/tools/web.d.ts +0 -10
  509. package/dist/tools/web.js +1 -4
  510. package/package.json +5 -2
  511. package/test/export-surface.snapshot.json +161 -21
  512. package/wire-changes/core-stores-sql@7.59.0.json +29 -0
  513. package/wire-changes/core-stores-sql@7.60.0.json +30 -0
  514. package/wire-changes/core@7.59.0.json +2459 -0
  515. package/wire-changes/core@7.60.0.json +3274 -0
  516. package/wire-changes/mods-kernel@7.59.0.json +29 -0
  517. package/wire-changes/mods-kernel@7.60.0.json +62 -0
  518. package/wire-changes/primitives@7.59.0.json +29 -0
  519. package/wire-changes/primitives@7.60.0.json +30 -0
package/CHANGELOG.md CHANGED
@@ -1,3 +1,1343 @@
1
+ ## 7.60.0 — 2026-10-28
2
+
3
+ ### One image admission, one processor, one table: every image a run hands the model is compressed toward 512,000 bytes, and the bytes that ship must state a size within 2000 px (#2987, C-R974 S1 — Q3, Q5's hard rows, Q6, addenda 1 / 2 / 3; behavior change; behavior narrowing open to veto; divergence candidates D-2987-a and D-2987-b, open to veto; hard-breaking: ③ the processor seats and ④ the constant exports, three sentences each below; closed sets: `MediaRefusalReason` +2 `dimensions_unreadable` / `dimensions_too_large`, `EngineNoticeCode` +1 `config.image_processor_unavailable` (operator, engine); exports: −4 root `MCP_IMAGE_MAX_BASE64`, `IMAGE_TARGET_RAW_SIZE`, `sharpImageResizer`, type `McpImageResizer`, +1 root `IMAGE_ADMISSION`) (@server @client-core @cli @test)
4
+
5
+ - **Before.** The limits lived in five places (the single-image base64 limit, the 3.75 MB / 2000-px frame, the 5 / 64 MiB read caps, the 20 MiB URL body, the request budget). Two processor seats — `RunnerDeps.mcpImageResizer` (MCP) and `RunnerDeps.hands.readImageDownsampler` (Read, PDF pages) — each with its own sharp detection. The processor stopped at the frame ladder, so a 1206 × 2622 phone screenshot of 2.9 MB reached the model as a 920 × 2000 PNG of about 2 MB. PDF pages called the processor directly behind their own 5 MiB gate; MCP resized only above 5 MiB of base64 and spilled to disk otherwise; a Bash data URI over 5 MiB of base64 fell to the text path; the tool-result seat, `TaskSpec.images` and attachments admitted with no processor; a URL image was not admitted at all. Bytes with no processor went out on their base64 length alone, whatever their pixel size.
6
+ - **Now.** One table, `IMAGE_ADMISSION` (root export, frozen): `hard` = `formats` (the four header rows), `readBytes` (`withoutProcessor` 5 MiB, `withProcessor` 64 MiB), `urlBytes` 20 MiB, `rejectEdge` 2000, `imageBase64Chars` 5 MiB; `soft` = `frameEdge` 2000, `rawBytes` 3,932,160, `targetBytes` 512,000. `admitImageBytes(bytes, mime, processor)` takes the processor as a required `ImageDownsampler | "none"`: the hard checks, then the processor — the frame ladder, then the byte-target step (Claude Code 2.1.309's form: a non-JPEG tries JPEG quality 90 first, then at most five binary probes over [1, 89] — [1, 90] for a JPEG — keep the best quality at or under the target, else the smallest bytes seen; a failure of the step keeps the ladder's bytes) — then the base64 limit. What ships is the processor's bytes when they are one of the four formats, their type read off them (Claude Code's form; a processor's declared type is not taken), else the bytes as they came; whatever produced them, the bytes that ship must state their pixel size in their own header within `rejectEdge`, else `dimensions_unreadable` / `dimensions_too_large` (one sentence, `unprocessedSizeClause`, for every carrier) — a processor that hands the bytes back unchanged does not carry a frame past the edge. A processor runs once over bytes: what it shipped is remembered by digest (per processor, its latest 256), and those bytes admitted again — by the tool-result seat after a carrier, through the tool door's rebuilt block — take every hard check and no second encode, so a best-effort image (no quality reaches the target) is not re-encoded at each admission. Every carrier admits with the run's processor: Read, and a notebook's image outputs in the reader the Read tool and an attachment share (a refused one is a marker naming the refusal); PDF pages (their own gate retired); a Bash data URI (any size up to the read cap is admitted and compressed); MCP (its own resize retired — only bytes still over a hard limit after processing are spilled, any other refusal stays an image block for the seat's placeholder); the tool-result seat (every result's images); `TaskSpec.images` inline and by URL (the URL arm now passes the admission); attachments. An image inside the frame and at or under the target passes byte-identical, so the seat's second admission of what a carrier sent changes nothing. The seat is `RunnerDeps.imageProcessor?: ImageDownsampler | false` — a processor; `false` = none; absent = detect sharp once per process (`resolveImageProcessor`, one cache); any other value is refused at `new Runner` (`config.invalid_argument`). With no processor, the run's first image over the soft target — wherever it entered — emits `config.image_processor_unavailable` once (`detail: { bytes, targetBytes, sessionId, runId }`).
7
+ - **Hard-breaking ③ — the processor seats (open to veto).** *What changed*: `RunnerDeps.mcpImageResizer` and `RunnerDeps.hands.readImageDownsampler` are retired (refused at `new Runner` with `config.deps_retired_key`, naming the replacement); `sharpImageResizer` and `McpImageResizer` are removed; `materializeMcpTools`' fourth parameter is the seat's value (`ImageDownsampler | false`). *Who is hurt*: a deployment or test fixture that injected either seat — @test's `readImageDownsampler: false` fixture (repro core548); @server where it wires either. *Compensation*: set `RunnerDeps.imageProcessor` once — `false` keeps the no-sharp fixture's shape, an `ImageDownsampler` replaces either seat, `createImageDownsampler` / `sharpImageDownsampler` are unchanged.
8
+ - **Hard-breaking ④ — the constant exports (open to veto).** *What changed*: the root exports `MCP_IMAGE_MAX_BASE64` and `IMAGE_TARGET_RAW_SIZE` are removed. *Who is hurt*: @server — the entry check `src/http/admission.ts:9 / :178` no longer compiles; @client-core — `scripts/gen-upstream-tables.mjs:586 / :1045` (the tenth batch) reads a missing name. *Compensation*: read `IMAGE_ADMISSION.hard.imageBase64Chars` (the same 5,242,880 characters) and `IMAGE_ADMISSION.soft.rawBytes`; the other rows are named above; one pick-up car each.
9
+ - **Behavior narrowing, open to veto.** *Who needs it*: a deployment with no image processor — an image over 2000 px went out as it was, and the provider resized it or refused the request; the admission now refuses it at the source, naming its size and the limit (Claude Code 2.1.309's fallback arm refuses the same). *Who is hurt*: a deployment without sharp, or with `imageProcessor: false`, that sends images over 2000 px — Read, an attachment, a task image, a PDF page rendered over 2000 px; an MCP image of that size is spilled to disk with its path instead of inlined. A JPEG whose header states no frame size is refused when it would go on unprocessed. *Compensation*: install the optional `sharp`, or set a processor.
10
+ - **Behavior change.** An image the processor brings toward the target arrives as a JPEG (a PNG screenshot's mime changes).
11
+ - **Divergence candidate D-2987-a, open to veto (C-R974 addendum 2).** Every JPEG encode of the processor — the ladder's rungs and the byte-target step, one place — merges an alpha channel onto white first. Claude Code 2.1.309's `N$o` encodes with no merge, and sharp then drops the alpha onto black: a transparent PNG whose marks are dark (a line drawing, a logo, a diagram) became an all-black JPEG; the old ladder did this only above 3.75 MB, the byte-target step would have done it above 512,000 B. A frame with no alpha is unchanged.
12
+ - **Divergence candidate D-2987-b, open to veto (C-R974 addendum 3).** The processor auto-orients before re-encoding (Claude Code 2.1.309 does not): a frame whose header asks to be shown turned (EXIF orientation 2–8) is turned upright once, at the top of the processor, before its size is read or anything re-encodes it, so the frame, the dimensions and the coordinate caption are the upright ones and the tag is gone. Without it a re-encode drops the tag and a phone photo reaches the model sideways — the old ladder re-encoded only above 3.75 MB, the byte-target step re-encodes every photo above 512,000 B. A frame with orientation 1 or none is unchanged.
13
+ - **Not in this change (with #2988).** The request-budget rows (`mediaByteCapBytes` → `mediaBudgetBase64Chars`), the retirement of `onMediaStripped` and the rename `context.image_removed` → `context.media_removed`.
14
+ - **Consumers.** @server: the entry check reads `IMAGE_ADMISSION.hard.imageBase64Chars`; a deployment that wires a processor sets `RunnerDeps.imageProcessor`; `config.image_processor_unavailable` joins the operator notices. @client-core: the generated table reads `IMAGE_ADMISSION.hard.imageBase64Chars`. @cli: no code reads either name; L-2122 waits on #2988. @test: fixtures move to `imageProcessor: false`; the two refusal words and the notice code are new.
15
+ - **Pins.** `test/backlog2987-image-admission-one-processor.test.ts` (a deterministic encoder under the production pipeline: the 2.9 MB screenshot leaves Read, an attachment, an inline and a URL task image, MCP, a Bash data URI, a rendered PDF page and the tool-result seat at or under 700,000 base64 characters; over the old 5 MiB pass-through it is compressed by MCP and Bash, not spilled or dropped; a second admission is byte-identical, and a best-effort image passes MCP or Read then the seat with no second encode while other bytes in the same block are encoded; a pass-through processor does not carry a 3000 × 10 frame past the edge; a processor's mislabeled or non-image answer is not shipped; every JPEG encode merges alpha onto white; a notebook's image output is framed by the reader, refused as a marker with no processor, and rides byte-identical inside the frame and the target; no processor and a 2622-px edge ⇒ `dimensions_too_large` by every carrier, a failing or throwing processor alike; no frame header ⇒ `dimensions_unreadable`; `false` ⇒ none, the bytes ride unchanged and one notice per run across two tool results and a task image, a processor ⇒ none; the retired seats refused by name, a bad seat value refused; type pins). `test/image-downsample-real-sharp.test.ts` (real sharp: the byte-target step on a noisy frame and a screenshot-shaped PNG; a fixed point; a transparent PNG's dark marks survive the JPEG encode; an EXIF-orientation-6 camera JPEG comes out upright with no tag, orientation 1 untouched; a best-effort black-and-white frame passes the seat unchanged). Re-pinned: `test/read-image-pipeline.test.ts`, `test/mcp-content-mapping.test.ts`, `test/backlog1719-media-admission-one-seat.test.ts`, `test/backlog1829-engine-below-tools.test.ts`, `test/backlog2874-image-reject.test.ts` and `test/backlog1795-time-cut-consumers.test.ts` (their JPEG fixtures state a frame size), the toolkit-option fixtures (`imageProcessor: false`), `test/public-knob-liveness.test.ts`, `test/export-surface.snapshot.json` (−4 +1), `test/skip-baseline.json` (the real-sharp names), `test/backlog1279-child-undrained-to-parent.test.ts` (109 codes).
16
+ - **A tool result with no media awaits nothing.** The tool-result seat reads the executed call's content once (`toolResultMediaBlocks`) and resolves the run's processor only when a block claims media (`admitToolResultMediaBlocks`); a call whose result carries none no longer waits on the absent seat's one-time detection of sharp (a dynamic import — real I/O on the first tool call of a process, which a fake clock cannot advance). Behavior unchanged. Pins: `test/walltime-telemetry.test.ts` (件2a, both cells) and `test/backlog258-cycle-seq-outline.test.ts` ((1b) F3), red on the stage merge and green again. (#3078)
17
+
18
+ ### A resumed approval runs exactly the bytes the person approved: the resumed leg's PreToolUse re-screen no longer hands back the bytes that run (#2975; approval integrity, aligned with Claude Code; behavior change open to veto; hard-breaking: none; closed sets: none; exports: none) (@server @test @cli)
19
+
20
+ - **Before.** On a durable resume the approved call's `classic.PreToolUse` re-screen supplied the bytes that ran. For an unedited approval the hook's rewrite ran in place of the bytes the person approved, so a rewriter that is not idempotent (one that appends a mark on every application) was applied twice: `tool_start.args` named one form and the tool ran another. A rewrite of the person's edit refused the call instead (since 7.56.0, #1022).
21
+ - **Now.** The re-screen has three outlets: a refusal (deny, crash, silence, an answer outside the contract, a person-only ask under a policy decide), an ask (the decide already answered it) and `additionalContext`. Its `updatedInput` is not adopted, for an edited approval and an unedited one alike. What runs is the decide's approved bytes — the edit, or the call as shown — and `tool_start.args` is those bytes, held as one engine-owned frozen value: a host that writes into the start frame's `args` while the re-screen waits no longer moves what runs. The tool itself gets a detached copy, as a live call does, so a preparation that edits its input in place still works. This is Claude Code 2.1.309's form for a permission answer that arrives after its turn: the call passes the PreToolUse hooks again, and the answer's own input is what runs. Nothing announces the dropped form.
22
+ - **Who needs it / who is affected / compensation (open to veto).** *Needs it:* approval integrity — the executed bytes are the approved bytes. *Affected:* a deployment whose PreToolUse handler enforces by rewriting instead of denying on the resumed leg. That covers a rewrite it would add only after the park (a handler installed or changed between park and resume) and a rewrite of a person's edit: neither reaches the call any more. The handler can still deny or ask there. A rewriter that is not idempotent is no longer applied twice. *Compensation:* the park leg's screening already folded its rewrite into the bytes the person approved, so an idempotent rewriter sees no change, and a resume-time guard keeps its deny. *Not taken:* refusing the call whenever the re-screen hands back other bytes. That would refuse every resume under a rewriter that is not idempotent, and a re-issue could never converge.
23
+ - **Consumers.** @server (mods hosts): a PreToolUse rewrite that must hold for a resumed approval belongs on the park leg, or becomes a deny. @test: K9b's construction now runs the call once on the approved bytes (sha256 of the executed args = of the row's = of `tool_start.args`). @cli: nothing new to render.
24
+ - **Pins.** `test/backlog2975-approved-bytes.test.ts` (a fresh Runner redeems a park: a non-idempotent and an idempotent rewriter both run the approved bytes once; a person's edit runs as written; deny / ask / context outlets beside a dropped rewrite; a write into `tool_start.args` during the re-screen does not land; an in-place preparation runs on a resumed call; controls with no handler and a watching one). Re-pinned to the new reading: `test/backlog1294-approved-call-hooks.test.ts` ③c ③d ⑨b ⑪, `test/backlog1022-rewrite-recheck.test.ts` R10 R12, `test/backlog1974b-keyless-and-unreadable-resume.test.ts` ⑥, `test/backlog1328-parked-skill-rescreen-swap.test.ts` ① ③ ④ ④b ⑥ ⑦ ⑧b ⑨ ⑨b (a re-screen's skill swap is dropped and the card's skill activates), `test/backlog1391-approval-object.test.ts` B6 SH-h K-sur K-sur-key (the approved destination / command / value runs).
25
+
26
+ ### A memory sync round's peer is the identity its transport carries, and a cursor taken against another peer is refused before it can delete local entries (#2968, P1 data loss; C-R969 addendum 1 Q3 A; hard-breaking: `MemorySyncTransport` is `{ peer, post }` and `SyncMemoryScopeOptions.peer` with its `"central"` default is gone, three sentences below, open to veto; closed sets: none — the refusal reuses `config.memory_sync_cursor_mismatch`; exports: none added or removed, `MemorySyncTransport` changes kind type → interface) (@server @client-core)
27
+
28
+ - **Before.** `syncMemoryScope` took its peer from `opts.peer`, else the cursor's own `peer`, else the constant `"central"`, and the transport was a bare `(path, body)` function with no identity. A deployment that never sets `peer` (the server's sync runner) bound every cursor to `"central"`, whichever center its transport reached. Re-point the transport (a new `MEMORY_SYNC_URL`, another engine declaration) and keep the layer's cursor: the judge compared the cursor's peer with a value read from that same cursor and passed; the new center, not holding the baselined entries, read each as deleted centrally and unchanged on the device and answered it in `serverDeletes`; the client CAS-checked it against the same baseline and deleted the entry locally — every entry left unchanged on the device since the last round was gone. Reproduced against an in-process twin of the server's round (the same pure reconcile it imports): request `{ peer: "central", baseRevs: { X }, entries: [], deletes: [] }` → `serverDeletes: [X]` → local delete, empty local plane; the same center as control deletes nothing.
29
+ - **Now.** `MemorySyncTransport` = `{ readonly peer: string; post(path, body): Promise<unknown> }`. `peer` is the identity of the peer `post` reaches (for a central sync API, its base URL in the deployment's normalized form) and the round's only source of its peer: the request's `peer`, the returned cursor's `peer`, and the value a persisted cursor is judged against. A cursor for another peer — or one with no peer binding — is refused with `config.memory_sync_cursor_mismatch` (the judge and code that already covered scope and store) before any read or request, and the sentence names two ways out: (1) re-bind — run the layer over a transport that reaches the peer the cursor was taken against, or, only when this transport reaches that same peer under a new identity, set the cursor's peer to that identity; (2) drop the cursor to accept re-convergence from an empty baseline — entries changed on both sides are reported as conflicts, and entries deleted on one side since the last round come back from the other side. The round never re-binds or drops a cursor itself. A transport without a non-empty string `peer` — the previous function shape included — is refused before any read or request (`memory sync: transport.peer must be a non-empty string …`). The identities are compared raw and quoted through the same-principal egress pass, in the refusal and in the protocol error a response cursor answering for another peer raises: a URL credential reads `[redacted-credentials]`.
30
+ - **Hard-breaking, open to veto (@server).** *What changed*: `MemorySyncTransport` is an object with `peer` and `post` (it was a function); `SyncMemoryScopeOptions.peer` and the `"central"` default are gone, so a round's peer can no longer come from anywhere but its transport. *Who is hurt*: @server — `createMemorySyncTransport` (`src/memory-sync-client.ts`) returns the function shape and must return `{ peer, post }`; the `peer` option of `createMemorySyncRunner` and `DEFAULT_MEMORY_SYNC_PEER` lose their reader; both wiring sites (`src/boot/stores.ts` and `src/run-local.ts`, which never set a peer) build the transport; every cursor file persisted under `sync-cursors/central/…` holds `peer: "central"`, which the new identity will not equal; the central side's request `peer` (`parseMemorySyncRequest`, at most 190 characters) changes from `"central"` to the identity, so the `sync_cursors` rows it writes get a new partition value. *Compensation*: the compile error names every call site; the identity is the deployment's to mint (normalize the base URL once, keep credentials out of it — it rides the wire, the cursor and refusal sentences); an existing `"central"` cursor must still be handed to the round — looking it up only under the new identity's path and finding nothing runs an empty-baseline round nobody chose, the reconvergence C-R967 refuses to run unasked — where it is refused with the two ways out, so the operator re-binds it after confirming it was taken against the current center, or deletes it; nothing re-binds it automatically (that would stamp a cursor of unknown origin as this center's and run the deletion path above). The runner's own downgrade of a peer-mismatched cursor to a first round (`loadCursor` → `cursorOf`) goes for the same reason. If the C-R967 operator step ships in the same window, it can be one step.
31
+ - **Consumers.** @server as above. @client-core: no `syncMemoryScope` / `MemorySyncTransport` reference in its source (origin/main `45c9279`); the refusal it may relay through the server keeps its code. @cli: one test string names `syncMemoryScope` in prose; no call.
32
+ - **Known limit (stated).** The identity is what the deployment says it is. One center reachable under two identities (an alias host, a port change) reads as two peers and is refused — the safe side, answered by way out (1). A different center behind the same identity (a central store wiped and rebuilt at the same address) reads as the same peer, and the deletion path above stays open for that form until a center can state its own incarnation. The engine takes `peer` at its word (read once per round): a transport whose `peer` names one center while its `post` delivers to another passes the judge, and that other center's `serverDeletes` land — the judge cannot see where `post` sends. The backstop is the server's local-side door, which compares the engine the session goes to with the peer its sync leg reaches (`config.memory_sync_target_mismatch`, design/2897 §2); the engine adds no second reading of the identity (KL-2968-a).
33
+ - **Pins.** `test/backlog2968-sync-peer-binding.test.ts` (a real File store, an in-process twin of the server's round per center): same peer runs and carries the identity on the wire and in the cursor; another peer is refused with zero requests and zero backend calls, the entry stays, both ways out in the sentence; way out (2) pushes and deletes nothing; way out (1) back to the cursor's peer runs, and the same center under a new identity runs once the cursor's peer is rewritten; an upgrade cursor holding `"central"` against an empty new center is refused with zero requests; a cursor with no peer binding is refused, never run as an empty baseline; a transport with no identity (function shape, missing, empty, non-string) is refused before any read; an identity carrying a URL credential is never quoted — in a cursor it is another peer, on the transport it is refused before any request (#3020), in a response cursor it is a protocol error; a response cursor answering for another peer stays a protocol error with no local write. `test/memory-sync-client.test.ts`, `test/backlog2836-sync-cursor-store-binding.test.ts`, `test/backlog2818-typed-scope-sync-store.test.ts`, `test/backlog2956-memory-sync-store-root-export.test.ts` and `test/design336-origin-carriage.test.ts` build `{ peer, post }`; the liveness row for the retired option is gone (`test/public-knob-liveness.test.ts`); `test/export-surface.snapshot.json` records the kind change.
34
+
35
+ ### A memory sync round is reported one way — `memorySyncReportOf` folds each layer's calls to one outcome, a refusal carries its axis and its two ways out from one table, and `memoryTravelGapOf` reads what did not travel from that report alone (#2969, design/2897 S1, C-R969 and its addenda 1 / 2; behavior change: the cursor and store refusal sentences are worded from one table — codes unchanged, the error gains `detail` and `outs`; hard-breaking: none; closed sets: the round report's own refusal codes, two — `config.memory_sync_target_mismatch` (new, worded by the engine) and `session_active` (the server's code and sentence); exports: +2 root values `memorySyncReportOf`, `memoryTravelGapOf`, +2 root types `MemorySyncLayerReport`, `MemorySyncRoundReport`) (@server @client-core @cli)
36
+
37
+ - **Before.** A host that ran a round read `syncMemoryScope`'s result by hand: `ok` (which is true on a clean round that `maxPushEntries` or the server's `pull.limit` cut — `ok` does not look at truncation), the two truncation flags, `conflicts`, and for a refusal the error's `code` and a sentence whose ways out lived only in its words (one function per axis, and a separate sentence for the store plane). Each consumer would have assembled its own reading of "did this layer arrive".
38
+ - **Now.** `memorySyncReportOf({ target, sourceRevision, scopes, layers | refused })` is the one reading. `layers` lists each `syncMemoryScope` call as it settled — `{ scope, result }` or `{ scope, thrown }`, a layer continued after a cut once per call, in order — and each layer folds to one outcome: `synced` (the last call `ok` and nothing withheld for a later call), `incomplete` (`why: "conflicts"` with the last call's conflicts, or `why: "truncated"` — a half cut under a limit, or an entry the server advanced in its cursor without delivering it, which the call keeps at its old baseline and discloses in `warnings`; `ok` looks at neither), `refused` (a refusal the engine minted: `code`, `detail`, `outs`, `message`), `failed` (anything else thrown — a transport or protocol error, a bad argument — as `error`). Every text in the report the engine did not mint — a conflict's `reason` from the server, an applied entry's `slug`, a warning, a thrown message — passes the same-principal egress pass; ids and revs are carried as they are; a name the report carries verbatim (`target`, `sourceRevision`, a scope key, the `reaches` of a target mismatch) that holds a credential is refused with a `TypeError` that does not quote it — the same test as the transport's `peer` (#3020). The last call decides the outcome; all four outcomes carry `applied` (`{ server, local }`, verbatim) — what every call of the layer returned as landed, in order — so a layer continued and then refused or failed still reports what its earlier calls landed (a refused call lands nothing; a call that throws returns no account of what it landed, #3047). A replay lands nothing and is still `synced`: whether a layer arrived is its outcome, never its counts. `refused` instead of `layers` is the round's own refusal, no layer run: `{ code: "config.memory_sync_target_mismatch", reaches }` (the machine's sync reaches another peer than the engine the session goes to — the engine words the sentence) or `{ code: "session_active", message }` (the host's code and sentence). `memoryTravelGapOf(report, readAtDestination?)` lists every scope of the read set that did not arrive whole, once, with why — `refused`, `failed`, `incomplete`, `not_replicated` (no layer run for it), `not_read_at_destination` (synced, and the destination's read set, when given, does not hold it); it reads the report alone, so a party that knows no scope key (the shell) can call it on the report it was handed, and the same report after a JSON round trip gives the same gap. The quarantine is a control-plane directory with no key: no sync round reads it and no read set holds it, so its bytes reach neither the wire nor the gap.
39
+ - **Refusal detail.** `config.memory_sync_cursor_mismatch` gains `detail: { axis: "scope" | "peer" | "store", found: "absent" | "other" }` (`absent` = the cursor carries no binding on that axis — for the store, a cursor written before cursors carried one; `other` = it carries another value) and `outs` (the two ways out); `config.memory_sync_store` gains `outs`, no `detail`. Ways out come from one table, one row per thing a round is refused on: the way out depends on the axis alone — absent and other are answered alike, which of the two it was rides `detail.found` — so the store axis's pair (rebind / clear, #2957) answers both the moved store and the legacy cursor, and the peer axis's pair (re-bind / drop, #2968) both a cursor for another peer and one with none. The sentence is `memory sync: <what was found> — <why>; two ways out: (1) <outs[0]>; (2) <outs[1]>`, built from the same row, so `outs` and the sentence cannot disagree. A value it quotes that holds a credential (a URL with userinfo, or anything the same-principal pass would redact) is quoted as `[redacted-credentials]` alone, never as a cleaned copy — a password with a quote in it is spelled out by no redaction pattern; the protocol error for a response cursor that answers for another peer quotes the same way. Only a refusal the engine minted reads as `refused`: an error that merely carries one of the codes (a transport's own) is `failed`.
40
+ - **Consumers.** @server (S-2540): the local-side door answers with `memorySyncReportOf`'s report — the read set = the session's latest settled leg's `effectiveMemoryScopes` (`done.result`; mounted → its scope keys, none / memoryless → `[]`), never the normalized `TaskSpec.memory` (which says what the session asked to read, not what it reads; C-R969 addendum 3), one entry per call per layer, the target mismatch and the in-flight fence as `refused`; the executor's own pre-judgment of a cursor without a store and its sentences can read `detail` / `outs` instead (S2). @client-core (CC-1052): map `why` from the report — `incomplete` → `memory_sync_conflict`, `failed` → `transport`, plus `refused` and the round's refusal; one message for `config.memory_sync_target_mismatch` in the catalog. @cli (L-2149 ①′): the fork step's memory sync reads the report's outcomes and `memoryTravelGapOf(report)` — "these layers are not readable in the cloud yet" — and treats `refused` / `failed` as failure (relay `outs`, no retry for `refused`).
41
+ - **Exports (S-2540 D-8, #3046, C-R991).** +1 root value `carriesCredential` — the one test the sync client and its report already use for a name a sync carries verbatim (a peer identity, a target, a scope key): a URL with any userinfo (a bare user name included), or anything the same-principal egress pass would redact. @server judges `target` / layer names / the centre's `peer` with it and never re-derives it (the consumer smoke list records it). The in-flight fence code is `session_active` — the host's own 409 code — not `conflict.session_active` (design/2897 r1 :52 named a code the host never had; renamed before this member ever shipped, so no reader changes).
42
+ - **Known limits (stated).** The report is as true as the calls the host hands it: a call left out is a layer reported `not_replicated`, never a guess. `localOnly` (the `userlocal:` layer, design/2897 S3) is not in this report yet; it lands with the key family (#2970).
43
+ - **Pins.** `test/backlog2969-sync-round-report.test.ts` (a real File store, real `syncMemoryScope`, transports that count their calls): a clean round cut by `maxPushEntries` is `incomplete/truncated`, the continued round `synced`, the two calls fold to one layer holding both rounds' `applied`; a replay is `synced` with nothing applied; server conflicts are `incomplete/conflicts`; a server that advances its cursor past an entry it did not deliver is `incomplete/truncated` and the scope stays in the gap; a server's conflict reason, an applied slug and a warning carrying a token reach the report redacted; a credential-bearing target, scope key or `reaches` (a quote inside the password included) is refused unquoted; a cursor with no store binding is refused with zero requests and zero store reads, `{ store, absent }`, its outs verbatim in the sentence and equal to the moved store's; a moved store is `{ store, other }`; all six axis × found pairs read their axis's row; the store plane has its own row; a transport error carrying a refusal code is `failed`; a credential in a cursor's peer is redacted in the sentence; a gate-failing file is quarantined and the round's request carries none of its bytes; every way a call settles has one outcome, a secret in a thrown message is erased; the gap over a refused round, a scope with no layer, each outcome and a destination read set, and after a JSON round trip; the round's refusals and bad inputs (`TypeError`), and a report the engine could not have minted (a layer with no scope, outside the read set, listed twice, an outcome it does not read) refused by the gap; the two root values are the leaf's. `test/backlog2957-sync-store-move-refused.test.ts` builds `{ peer, post }` and keeps its scope-axis cell (the peer axis is #2968's); `test/export-surface.snapshot.json` +4; `scripts/consumer-smoke-probe.mjs` decides the two values.
44
+
45
+ ### A memory sync transport's `peer` is refused when it is only whitespace or carries a credential — userinfo in a URL, or anything the same-principal egress pass would redact (#3020, with #2968; behavior narrowing, open to veto; hard-breaking: none beyond #2968's transport shape; closed sets: none — the transport refusal keeps its sentence family, no code; exports: none) (@server)
46
+
47
+ - **Before.** The transport's `peer` was checked for a non-empty string only. `" "` ran a round (one request, a cursor bound to `" "`); `https://ops:<password>@host` rode the request body and the persisted cursor verbatim, though the contract says the identity is never a credential.
48
+ - **Now.** Before any read or request, a `peer` that is only whitespace is refused with the empty peer's sentence (`memory sync: transport.peer must be a non-empty string …`), and one that carries a credential — a URL with userinfo (a bare user name included: the slot is the credential's) or anything the same-principal egress pass would redact (a token in a query, a vendor key) — is refused with `memory sync: transport.peer [redacted-credentials] carries a credential …` — the identity itself is never quoted. The identity is refused, never cleaned: a cleaned identity is another identity, and the cursor taken under it would not be this peer's. A plain base URL, one with a port or a path, and an opaque identity run as before.
49
+ - **Behavior narrowing, three sentences (open to veto).** *Who needs it*: anyone whose sync cursors and request bodies are stored or logged — the identity rides both verbatim. *Who is hurt*: a deployment that builds the transport's identity from a URL with credentials in it (@server `createMemorySyncTransport`, which #2968 already asks to mint the identity from the normalized base URL without credentials). *Compensation*: the refusal comes before any read or request, so nothing is sent or written; the fix is the identity's mint, once.
50
+ - **Pins.** `test/backlog2969-sync-round-report.test.ts` (#3020 / #2968 follow-up cells): `" "`, `"\t\n"` and `""` are refused alike with zero requests and zero store reads; a URL with a password, one with a bare user name, a non-http URL with userinfo, and a URL with a token in its query are refused, quoted redacted, zero requests; a password holding a quote is absent from the refusal on the transport, the cursor refusal and the response-cursor protocol error alike; a plain base URL, one with a port and a path, `"central"` and an opaque `engine:` identity run and bind the cursor to themselves. `test/backlog2968-sync-peer-binding.test.ts`: a credential-bearing identity on the transport is refused — also against a cursor holding the same string.
51
+
52
+ ### A `tool.call` extension that fails once the tool ran withholds the result, and a `classic.PostToolUse` raise that cannot hand the result to its handlers withholds it too — kernel contract 6, `tool.call`'s failure column is `refuse` (#2675; security narrowing, open to veto; hard-breaking: yes, open to veto; closed sets: none; exports: none) (@server @dgx @test)
53
+
54
+ - **Before.** `tool.call`'s failure column `refuse_until_run` refused a call while the body had not started and, once it had, let the body's real result through with the failure reported beside it: a checking extension that crashed, overran or broke the form after `next(e)` left the result it was there to check in the model's context. `classic.PostToolUse` swallowed a dispatch that could not be raised for a result and passed the original result on (the gap CC 2.1.299 closed: "a tool that already ran has its result withheld").
55
+ - **Now.** A `refuse` station refuses on a handler's failure whether or not its bottom has started, and the dispatch's `failed` outcome carries `bottomStarted: true` when it had (as `aborted` already did); `refuse_until_run` stays the column of a go-ahead bottom (`agent.spawn`, whose subagent starts only after the station answers). `tool.call` — column `refuse` — withholds the result of a body that ran: `isError`, "a tool.call extension … after this call to "<tool>" ran: a hook that checks this result could not hand back its answer, so the result is withheld — the call ran (fail-closed)", the gate's allow and no `notExecuted` (the call ran; it stays in the PostToolBatch; the failure is on `onError`). The call settles only once its body has (a handler that failed while its `next(e)` still runs waits for it), and a frame the result carried (a ReadInbox message) is withheld with it — never handed over again through another lane. A body that threw produced no result: its throw stands. The withholding sentence names the extension and what it did, never the bus's reason (a thrown message may quote the result) — the reason is on `onError`. `classic.PostToolUse` withholds a result the bus refused to raise or whose dispatch broke (content replaced, `details` emptied); a PostToolUse handler that fails is still skipped and the result kept.
56
+ - **Hard-breaking (open to veto).** *What changed:* `MOD_EVENT_TABLE["tool.call"].station.failure` is `refuse` (was `refuse_until_run`); `KERNEL_CONTRACT_VERSION` is 6 (one number with the 7.60.0 kernel additions); a result a late-failing `tool.call` extension let through is withheld. *Who needs it:* a deployment whose `tool.call` or PostToolUse extension checks results (a redactor, a data-loss filter) — its failure no longer hands the model what it was checking. *Who is hurt and how to migrate:* code that reads `tool.call`'s row as `"refuse_until_run"` reads `"refuse"`; a host pinning the contract number re-pins 6; an extension that answered `{ deny }` or `{ result }` after `next(e)` now sees the result withheld instead of standing — answer before `next`, or hand on what came up.
57
+ - **Consumers.** @server @dgx: the mods-kernel package face (the `tool.call` row, the contract number); no wire key changes. @test: a fixture reading a result through a late-failing `tool.call` handler re-pins to the withheld sentence.
58
+ - **Pins.** `test/backlog2675-withheld-after-run.test.ts` (real Runner: a throw, an overrun, a non-data answer and a late deny after the body ⇒ withheld, ran once, allow + no `notExecuted`, named on `onError`; before the body ⇒ refused; a body's throw stands; a PostToolUse raise that could not be given the result ⇒ withheld; a throwing PostToolUse handler ⇒ skipped; the call stays in the batch); re-pinned `test/mods-tool-call.test.ts` C2 / C3 / C7 / L5 and `test/mods-kernel-wave2.test.ts` (failure answers, contract 6).
59
+
60
+ ### Edit, Write and NotebookEdit refuse a file that is not valid UTF-8 instead of rewriting its every undecodable byte as U+FFFD (#2676; security narrowing — data loss, open to veto; hard-breaking: none; closed sets: none; exports: none)
61
+
62
+ - **Before.** Encoding detection was BOM-only and a body read as UTF-8 was never checked: a Windows-1252, Shift-JIS or GBK file decoded with each legacy byte as U+FFFD, Read showed it (when under its binary threshold), and an Edit — or a Write over it, or a NotebookEdit — wrote the whole file back as UTF-8 with every such byte replaced by EF BF BD (the loss CC 2.1.296 / 2.1.299 closed by refusing).
63
+ - **Now.** The decoder's one "cannot be written back byte for byte" fact (`malformed`, which already named a truncated UTF-16 body) has a second why, `not_utf8`, and every write lane refuses on it: "… is not valid UTF-8: it uses an older encoding such as Windows-1252, Shift-JIS or GBK, or it is binary. … nothing was written", with the ways out (a shell command that keeps the encoding, asking the user to convert, or delete-then-Write). Read is unchanged; a valid file that holds U+FFFD itself edits as before.
64
+ - **Consumers.** none to change; a host that relied on an Edit converting a legacy file to UTF-8 now gets the refusal.
65
+ - **Pins.** `test/backlog2676-not-utf8-write-lanes.test.ts` (Windows-1252 Edit / Write / NotebookEdit refused with the bytes untouched; Shift-JIS and a BOM'd file with a legacy byte; controls; the decoder's fact).
66
+
67
+ ### An `rm` that walks into a mounted drive is a protected removal — `dangerous_delete` for the folder that holds drives, a drive, or an entry at a drive's top level (#2674; security narrowing, open to veto; hard-breaking: none; closed sets: none — the kind is the card's sentence, not a wire word; exports: none)
68
+
69
+ - **Before.** The protected locations were the filesystem root, a home, a workspace directory or its ancestor, and a top-level directory: `rm -rf /Volumes/USB`, `rm -rf /mnt/c/Users`, `rm -rf /media/alice/USB` and `rm /Volumes/USB/*` were ordinary removals an allow rule cleared and bypassPermissions released (the gap CC 2.1.304 closed).
70
+ - **Now.** The one dangerous-removal predicate reads one more kind for an `rm` that walks what it removes — a recursive option in its row's reading of the argv, or an operand naming every entry of a directory: `/Volumes` and `/System/Volumes` down to `<drive>/<entry>`, `/mnt/<drive>` and `/mnt/<letter>/<entry>`, `/media` and `/run/media` down to `<user>/<drive>/<entry>` (folded case, POSIX paths). The card says "would remove a mounted drive, the folder that holds mounted drives, or a file or folder at the top level of a drive … if you meant something deeper inside, name it by its full path"; it is the `dangerous_delete` class every protected removal mints, which no mode, rule or classifier answers. A non-walking `rm` of one file, `rmdir`, and anything deeper read as before.
71
+ - **Consumers.** none to change (one more sentence on an existing class).
72
+ - **Pins.** `test/backlog2674-mounted-drive-removal.test.ts` (real Runner: nine walking removals × three modes × with and without `Bash(rm:*)`; five controls an allow clears; an unsettled program word read as `rm`; the predicate's roots, walk condition, folds and win-form paths; `/Volumes` itself stays top-level).
73
+
74
+ ### A background fork started without a description shows no part of a credential from its prompt (#2678; security narrowing, open to veto; hard-breaking: none; closed sets: none; exports: none)
75
+
76
+ - **Before.** The description shown for a background fork (`Agent` with `subagent_type: "fork"`) with an empty description was the prompt's first 180 characters, cut BEFORE any redaction: a cut that landed inside a key left a head shorter than any recognizer reads (`AKIA` + nine characters), and that head passed every later pass into the registry row, the launch receipt and the completion notification (the gap CC 2.1.299 closed for teammates).
77
+ - **Now.** The description — or, with none, the prompt — goes through the same-principal egress pass first and is cut after it, the law every bounded operator line already keeps. A clean prompt reads exactly as before.
78
+ - **Consumers.** none to change.
79
+ - **Pins.** `test/backlog2678-fork-description-redaction.test.ts` (empty and blank descriptions over a prompt whose head ends inside a key: no piece of the key in the receipt, the registry row or the notification; a clean prompt and a given description read as before).
80
+
81
+ ### The no-run MCP probe judges a declared endpoint by the deployment's host floor — `McpProbeOptions.hostFloor` (#2677; security narrowing, open to veto; hard-breaking: none; closed sets: none; exports: none — a member of an exported type) (@server)
82
+
83
+ - **Before.** `probeMcpServers` had no floor seat: it dialed every declared http endpoint with no floor, so a health check contacted a server a task under `RunnerDeps.hostFloor` refuses to contact (the gap CC 2.1.299 closed for `claude mcp list` / `get`).
84
+ - **Now.** `McpProbeOptions.hostFloor` takes the declaration `RunnerDeps.hostFloor` takes, compiled by the one compile (a declaration that does not read is refused `config.host_floor_invalid` before anything is dialed) and judged at the one dial: a refused endpoint is never contacted and its row reads `failed` / `invalid_config` with the sentence a task's warning carries. A delegated leg's servers already read the run's floor (the child's spec goes through the same preparation).
85
+ - **Consumers.** @server: pass the deployment's `hostFloor` to `probeMcpServers` wherever the health check runs beside a Runner that declares one — absent, the probe still dials with no floor.
86
+ - **Pins.** `test/backlog2677-probe-host-floor.test.ts` (deny and `only` never contact the endpoint; controls contact it; a floor that does not read is refused before any dial; the probe's row and the Runner's warning carry one sentence).
87
+
88
+ ### A key behind an escaped quote names its value: JSON written inside a shell string is redacted like the bare object (#2679; security narrowing, open to veto; hard-breaking: none; closed sets: none; exports: none)
89
+
90
+ - **Before.** The keyword row of the credential vocabulary took a bare `"` / `'` on either side of its separator, so `echo "{\"secret\":\"…\"}"` and `bash -c "echo {\"api_key\":\"…\"}"` left `redactSecrets` / `redactHostLeaks` — and every engine face that reads the same-principal pass — with the value in clear, while the bare `{"secret": "…"}` was redacted (the gap CC 2.1.296 closed for shared transcripts and debug logs).
91
+ - **Now.** The quotes around the separator are read the way the `Authorization` header row reads them — a quote behind up to 63 backslashes — so an escaped key, at any depth the escape reading follows, names its value as the bare key does. No new row and no second reading of an escape.
92
+ - **Consumers.** none to change; a host that pins the exact output of `redactSecrets` over such text now sees the value replaced by the marker.
93
+ - **Pins.** `test/backlog2679-escaped-key-quote.test.ts` (three shell-string forms × both tiers; two escape layers; controls; a key with no value then a keyed value; idempotence).
94
+
95
+ ### The peer seat is told what the engine keeps — `RunnerDeps.peerSeat` is one object `{ answer, hold, release }`, and the engine holds each credential reference from the moment it takes a declaration naming it until nothing it keeps can dial with it (#2864; hard-breaking, open to veto)
96
+
97
+ - **Before.** `PeerSeat` was a function `(question, signal) => Promise<PeerAnswer>`: the seat was asked per send and told nothing else. A deployment that registers a credential behind the opaque reference it hands the engine could not know when the engine had stopped needing it — a session-scoped background child, a retained child revived later in the process and a cascade's rungs all kept asking after the call that handed the declaration had returned, so the registration's lifetime could only be guessed (by session deletion, by run id, by process lifetime) and each guess was wrong for one of them.
98
+ - **Now.** The seat is `{ answer(question, signal), hold(ref), release(ref) }`, all three required. An engine object that keeps a declaration it can still dial from calls `hold(ref)` once for each distinct reference its declarations name and `release(ref)` once for each it held. Two keepers: a run leg — from the `runTaskStream` call (fresh, resumed, child, cascade rung and repair attempt alike) until its stream has settled, through the failure backstop included — and a retained child's ledger entry — from its registration until it leaves the ledger (eviction, abandonment, replacement; a row not taken releases at once). A taker holds within the call its giver is alive for, so a reference handed along is never unheld in between. The engine counts nothing (two keepers of one reference are two holds); a checkpoint and a background-agent row keep no declaration. A leg's `mcp` / `a2a` lists are received once, at the call, by the door's own receipt (each list read once, each entry a frozen copy — a getter on a declaration now runs at the call, once, not after the leg's session lock): the leg holds what that copy names and its door judges that same copy, so what a leg holds and what it dials are one read. A `hold` that throws or whose promise rejects is reported (`onError`, phase `config`, classification `peer-seat-hold`) in the engine's words and gets no `release`; a `hold` that returns a promise is held once it fulfils (a release due before then is sent then). Scope: this process; no registration across processes is assumed.
99
+ - **Hard-breaking (open to veto).** *What changed:* `PeerSeat` is an interface of three required members instead of a function type, everywhere it is taken (`RunnerDeps.peerSeat`, `materializeMcpTools`' twelfth parameter, `materializeA2aTools`' fifth, `probeMcpServers({ peerSeat })`); a function — the 7.59.0 form — is refused there (`config.invalid_argument`, naming the form); `RunnerSelfSeat` gains a required `holdDeclarations(declared)`. *Who is hurt:* a deployment that implements the seat (tsc red; at run time `new Runner` refuses), and a hand-built `RunnerSelfSeat` (tsc red). *What to do:* put the former function under `answer` and add `hold` / `release` (no-ops for a seat that keeps no per-reference state; a counting seat reads "more holds than releases" as "the engine still has a consumer"); a hand-built `RunnerSelfSeat` forwards `holdDeclarations` to the Runner it stands for — one cast past the compiler without it is refused by name (`config.invalid_argument`, naming `holdDeclarations`) where a retained child's row would take its hold, not left to a bare TypeError from the middle of the delegation.
100
+ - **Consumers.** @server: the desk's seat object gains its engine half (`hold` / `release` on each entry's engine count; a hold of a dropped entry should not revive it); the transitional reading by session deletion retires. The server hands the seat over at four places, each a bare function until now and refused (`config.invalid_argument`) once the object is required (read from the server's tree before its 7.60.0 pickup): `src/boot/runner-deps.ts:748` (`RunnerDeps.peerSeat`), `src/capabilities/mcp-probe.ts:503` (`probeMcpServers`), `src/http/routes/sessions.ts:350` (`materializeMcpTools`, twelfth parameter) and `:420` (`materializeA2aTools`, fifth). @client-core / @cli: no code (the seat is the server's). @test: the hold / release pairing is observable only through a seat; the four shapes are pinned below.
101
+ - **Exports.** `PeerSeat` kind type → interface; +2 root types `PeerDeclarations` / `PeerDeclarationHold`; `RunnerSelfSeat` +`holdDeclarations`.
102
+ - **Pins.** `test/backlog2864-peer-seat-hold.test.ts` (S: the form refused at the reader and the four entrances; K: once per distinct reference, data members only, a throwing or rejecting hold gets no release and is reported (through the seat `new Runner` keeps too), no engine-side count; L: the host lets go right after the call and the leg's sends are answered and nothing is asked after its release — and the counterexample, a host that lets go before the call: the leg's late hold does not revive the entry and the send is refused; a door-refused leg still pairs; a getter declaration received and held at the call, a named-session leg included, the getter run once; C: a cascade's rungs pair each; B: a session-scoped background child keeps the entry after the spawning leg lets go — and the counterexample, a parent that lets go before the spawn; R: a retained entry holds across its legs and releases on eviction / abandonment / replacement / refusal — and the counterexample, an entry registered after its giver let go; H: a hand-built seat without `holdDeclarations` is refused by name at both retention registrations, and one that answers is asked once per row).
103
+ - **Known limits.** A session-scoped background child spawned by a tool call that outlives its own aborted leg may take its hold after that leg released (the abort orphan window) — KL-2864-a; a leg whose prepare never settles never releases — KL-2864-b; a connection whose close outlived the bounded teardown may reconnect once after the release (the seat then refuses — nothing is sent) — KL-2864-c; an asynchronous `hold` whose promise never settles keeps its release waiting — KL-2864-d.
104
+
105
+ ### Four more extension events are raised: `agent.offer` and `agent.spawn` around a delegation, `turn.complete` and `session.end` as a model turn and a leg end (#2457, #2458; mods wave three S2 / S3a, design/1398 waves 2–3, C-R704 / C-R940; behavior change only where a handler hangs on one of them; hard-breaking: the kernel contract 5 → 6 — a mod package written against 5 is refused, three sentences below, open to veto; closed sets: four station rows; exports: none) (@server @cli @client-core @test)
106
+
107
+ - **Before.** The four names were in `MOD_EVENT_TABLE` with no station: a registration that selected only them was refused at admission, and a workflow script's agent had no extension seat of its own (#1500 named this gap).
108
+ - **Now — `agent.offer`.** One offered set, read by the agent listing a leg renders (its first request and each boundary that owes the frame) and by the type a delegation names at dispatch. A handler answers `next(e)` or `{ isOffered: false }`: the type is left off the listing, and a delegation naming it anyway is refused before anything is allocated (`Sub-agent not started: subagent_type "<type>" is not offered here …`). `e = { agent, description, provider }`, `agent` and `provider` pinned; a failing handler is skipped and the type stays offered (the reference: "A hook that fails passes it through"). A revival is no offer. Known limit: the delegation tool's own description (the cached tool block, which names the built-in `fork` lane) and its refusal of an unknown type (which lists every accepted type) are fixed before any extension runs and do not read the offered set.
109
+ - **Now — `agent.spawn`.** Raised on the delegating leg's bus once the delegation tool has judged the call and before anything is allocated for the child, and for each agent a workflow script spawns (on the host leg's bus, `workflow: true`). Every member is pinned — the spawn's identity, its parent's, the identity envelope and every argument the gate judged (`prompt`, `description`, `subagentType`, `model`, `background`, `cwd`, `name`, …) — so the only answer is a refusal: `{ deny: <reason> }` instead of `next` starts no child and the model reads which extension refused; `next(e)` lets it start and comes up as `{ model }` (the model the delegation named, `inherit` when the child's own prepare decides). Failure column `refuse_until_run`: a throw, a timeout, a changed member or an answer of its own refuses while the child has not been let start; once it has, the spawn stands and the failure is reported on `onError` (`phase: "hook"`). No spawn is raised for a revival of a parked or retained agent, a drive of a parked workflow agent, or an observer leg a definition starts beside its child. Known limit: the observed spawn's `e` does not name its observer, so neither hiding the observer's type nor a refusal keeps an observer from running beside a child the chain let start.
110
+ - **Now — `turn.complete` and `session.end`.** Observations. `turn.complete` once per `turn_end` the leg emits, in order, at the next await point — the boundary after it, or the leg's settle for turns no boundary followed (a recovery re-prompt, an interrupted last turn) — at the start of the boundary's own pass and before the leg's `classic.Stop`, `e = { answer, reason, turn, isAborted, usage?, agentId? }` (`usage` absent when the brain reported none; `reason: "aborted"` when the leg's cancel was in force or the reply's own stop reason says so, as a reply cut while streaming does — a turn whose tool batch an immediate steer cut reads `"answer"`); `session.end` once per leg as its terminal settles, `e = { reason, sessionId, agentId? }` with the terminal status word — like the delegation lifecycle's terminal, never for a leg that throws out of the runner. A handler's answer changes nothing and a failing handler is skipped. No handler on an event: nothing is raised and nothing is awaited on its path.
111
+ - **Failure columns, station by station** (the contract's fourth trigger): `agent.offer` skip · grace · pinned `agent`, `provider`; `agent.spawn` refuse_until_run · immediate · sixteen pinned members; `turn.complete` skip · grace · pinned `answer`, `agentId`, `turn`; `session.end` skip · grace · pinned `reason`, `sessionId`, `agentId`. None has a `call` column: `$.agent.spawn` is not on `$` in this version.
112
+ - **Hard-breaking, open to veto: a mod package written against contract 5 is refused (@server @cli @test @client-core).** What changed: `KERNEL_CONTRACT_VERSION` is 6, and `loadModPackage` loads only a package whose `hooks.json` states `"contract": 6` (`src/core/mods/package.ts:72-73`); 7.59.0 shipped 5. Who is hurt: every mod package written for 7.59.0 — a deployment loading a real package (server S-2388), the CLI's package leg (MODS155), test's real-package fixtures (C-R915), and a host comparing two kernel copies for equality (client-core's generated table gains four station rows); the server's own hard-coded `5` is in four places, none on its runtime path (its deployment loader reads the engine constant, `src/boot/deployment-mods.ts:33` and `:111`): the acceptance harness's fixture contract number (`acceptance/s2388-mods-deployment/run.mjs`), the version pin `test/mods-loader-pickup.test.ts`, the 7.59.0 pickup case in `test/mods-deployment-loader.test.ts`, and one sentence of `docs/ASSISTANT-WIRE-CONTRACT.md` (read from the server's tree before its 7.60.0 pickup); this repo's three fixtures (`test/fixtures/mods/{diff,sec-default,telemetry}/hooks/hooks.json`) were raised to 6. What to do instead: set `"contract": 6`; contract 6 adds these four stations and changes `tool.call`'s failure column (#2675, above: a failure once the body ran withholds the result), so a package whose handlers name none of the four, and none on `tool.call` that fails after `next(e)`, needs no other change.
113
+ - **Pins.** `test/backlog2457-2458-mods-wave3-stations.test.ts` — real Runner, a parent delegating once: a hidden type is off the listing and refused when named; a failing offer handler leaves it offered; spawn's `e`, `{ model }` coming up, a deny / a changed member / an own answer / an early throw refusing with no child started, a late throw or late deny reported beside a child that ran; the workflow lane's raise; `turn.complete` per turn before `classic.Stop` and `session.end` last; observations that throw change nothing; the rows and the contract number. `test/mods-kernel-296.test.ts` counts the four new stations and names `turn.step` as the station-less example.
114
+
115
+ ### What an execution environment can do is one required block on every environment — `ExecutionEnv.capabilities` (`ExecutionEnvCapabilities`, eight keys), each key read through its one reader, a key left out or of the wrong shape announced and read on its safe side (#2947, design/2891 S1, ruled C-R965 and C-R964; hard-breaking: the block is required on the base type and the two loose booleans are retired, three sentences each below, open to veto; closed sets: `config.execution_env_capability_invalid` `detail.key` +4 words and minted for any environment; exports: +1 root type, +2 root values) (@server @cli @client-core @test)
116
+
117
+ - **Before.** The capability block lived on `RemoteExecutionEnv` alone (`isolation`, `suspendable`, `canonicalPathAuthoritative?`, `sharesProcessTable`); every reader asked the remote interface first and gave any other environment a fixed answer, so an environment without the remote seam could declare nothing (KL-2348-b). Two more facts sat on the base type as loose optional booleans with absent readings of their own (`hostLocalPaths?` — absent fell back to the remote shape or `instanceof NodeExecutionEnv`; `externalContentTarget?` — absent read as trusted). The admission census looked at remote environments only, and absence of three of the four keys was a silent reading.
118
+ - **Now.** `ExecutionEnv.capabilities: ExecutionEnvCapabilities` is required on every environment: `isolation`, `suspendable`, `sharesProcessTable`, `canonicalPathAuthoritative`, `hostLocalPaths`, `externalContentTarget`, `io` (the filesystem and shell faces reach a target), `keepsReadFloor` (#2944, the next section). `RemoteExecutionEnv` inherits the block. Each key has one reader (`isIsolated`, `declaredSuspend` / `isSuspendable`, `processTableOf`, `canonicalPathIsAuthoritative`, `pathsAreHostLocal`, `targetIsExternal`, `performsIo`, `keepsReadFloor`), none of which asks the interface. Every key is required and has no default reading: absent, not a boolean, unreadable — or a block that is not an object or cannot be read — is one `config.execution_env_capability_invalid` per key per run at admission (`detail: { key, valueType }`, the sentence naming the reading) and the reader's safe side: isolation false, suspendable undeclared (never snapshot), process table undeclared (pattern kills refused), canonical authority true (silence refuses), host-local undeclared (no path presented as the host's; a worktree child refused — #2948 below), external undeclared (the content fold reads it as outside; a call that reaches it owes what reaching the target and what leaving it both owe on the cross-boundary lane), io undeclared (no shell face; the project marker still read through the environment — #2948 below), floor not kept. `NodeExecutionEnv` declares `{ isolation: false, suspendable: false, sharesProcessTable: true, canonicalPathAuthoritative: true, hostLocalPaths: true, externalContentTarget: false, io: true, keepsReadFloor: false }`; `StubExecutionEnv` the same but `canonicalPathAuthoritative: false`, `io: false`, `keepsReadFloor: true` (it refuses every start). The frame's `remote` and the file-history identity read `pathsAreHostLocal` where they read the loose `hostLocalPaths`; the run's content fold and the cross-boundary lane read `targetIsExternal` (`BoundaryLegsGiven.env` is `Pick<ExecutionEnv, "capabilities">`). An oracle grader is `out_of_process` only when it declares isolation AND has the remote interface whose sandbox identity the check compares.
119
+ - **Behavior on the lanes that exist.** None on `NodeExecutionEnv` (its new `canonicalPathAuthoritative: true` moves no read-boundary verdict: its resolver always answers `/`) or on a TypeScript adapter that states its block. A run without hands (the Stub) now has file-history identity `"host"` (it writes no file). An adapter that leaves a key out — a JS adapter, or one written against 7.59.0 — gets a notice per key and the safe side above: for an undeclared `hostLocalPaths` its frame reads `remote: true`, for an undeclared `externalContentTarget` its content is treated as from outside (the memory session mark) and its tree is armed, and a call reaching it owes both the on-target and the off-target obligations (a person's confirmation on a governed tree, the off-target read record).
120
+ - **Hard-breaking, open to veto — the block is required on every environment (@server).** *What changed:* `ExecutionEnv` gains the required member `readonly capabilities: ExecutionEnvCapabilities` (eight keys, `keepsReadFloor` and `io` among them); `RemoteExecutionEnv.capabilities` is that member, with `canonicalPathAuthoritative` required where it was optional. *Why there is no default:* each reader would have to guess one side for an environment that says nothing, and that side is wrong for some other environment (assuming isolation is a hole, assuming a shared process table refuses an ordinary `pkill` in a sandbox, assuming a kept floor hides bytes that reach the model). *Who is hurt:* every host that implements `ExecutionEnv` — the server's seven remote adapters and any adapter written against 7.59.0 — fails `tsc` until it states the block; a JS adapter that leaves a key out compiles, is announced at admission and is read on the key's safe side. *Migration:* each of the seven remote adapters adds `hostLocalPaths` (the host lane `true`, the others `false`), `externalContentTarget` (the device lane `true`, the others `false`), `io: true`, `keepsReadFloor: false` (until a lane keeps the floor or refuses the start, §11.1) and states `sharesProcessTable` and `canonicalPathAuthoritative` explicitly; the two subclasses of `StubExecutionEnv` inherit the Stub's block and compile unchanged (override it where an operation does real I/O); a JS adapter that leaves a key out is announced at admission and read on the key's safe side.
121
+ - **Hard-breaking, open to veto — the two loose base-type booleans are retired (@server).** *What changed:* `ExecutionEnv.hostLocalPaths?` and `ExecutionEnv.externalContentTarget?` (and `StubExecutionEnv.externalContentTarget?`) are removed; the facts are `capabilities.hostLocalPaths` / `capabilities.externalContentTarget`. *Why:* one fact in two places is two sources that can disagree. *Who is hurt:* an adapter or host that sets or reads `hostLocalPaths` / `externalContentTarget` on the environment itself (the device adapter's top-level `externalContentTarget` is one) fails `tsc` on the removed members. *Migration:* the device adapter moves its top-level `externalContentTarget` into `capabilities`; a host that read either member reads `pathsAreHostLocal(env)` (root export) or the block.
122
+ - **Consumers.** @server: the seven blocks and the device lane's move above; the execution-lane table's `hostFilePlane` can agree with `pathsAreHostLocal` (#2891 X-1 retires the duplicate bits when it lands). @client-core / @cli: `config.execution_env_capability_invalid` `detail.key` gains `hostLocalPaths`, `externalContentTarget`, `io`, `keepsReadFloor` (and `capabilities`) and arrives for local environments too — a table that pins the key set adds the rows. @test: a NON-remote environment that declares `isolation: true` is now read as isolated by every reader (the remote interface is no longer a precondition), except the oracle's out-of-process class.
123
+ - **Same release.** The structural readings that answered capability questions off the remote interface or the Stub's class retire to the readers in #2948 (design/2891 S2, the next-but-one section), with the host lane's five readings that move there.
124
+ - **Exports.** +1 root type `ExecutionEnvCapabilities`; +2 root values `pathsAreHostLocal`, `keepsReadFloor`. Removed members: `ExecutionEnv.hostLocalPaths`, `ExecutionEnv.externalContentTarget`, `StubExecutionEnv.externalContentTarget`.
125
+ - **Pins.** `test/backlog2947-capability-table.test.ts` (P1 the source scan: every remote-predicate call, class check and direct key read left in src/ is a row — the interface whitelist or a reading #2948 retires — and no loose member survives; P2 Node's eight answers, its read-boundary verdicts unchanged by `canonicalPathAuthoritative: true`, its frame and file-history readings; P5 no block / `"true"` / `1` / `null` / a throwing key or block ⇒ one notice per key and every reader on its safe side, a non-remote env declaring isolation, the oracle's interface conjunct; P7 the Stub's answers and no notice for a run with no env or a Stub; P7b the marker read through a limited host-disk env; P8 the census walks exactly the type's keys, a block missing a key does not compile); the remote doubles mint their blocks through `test/env-capability-block.ts`.
126
+
127
+ ### A run whose read-deny table holds rows on an environment that does not keep the byte floor is announced once — or refused at prepare when the deployment requires the floor (`readDenyFloorRequired`); the engine's own starts whose output reaches the model carry the floor (#2944, #2937 ①, design/2936, ruled C-R964; behavior change: one operator notice per run on `NodeExecutionEnv` with the default tiers; hard-breaking: none beyond the required `keepsReadFloor` key above; closed sets: `ENGINE_NOTICE_CODES` +1, `SWAPPABLE_DEP_SEATS` +1; exports: `keepsReadFloor` (above); members: `RunnerDeps.readDenyFloorRequired?`, `HandsToolkitOptions.readDenyFloorRequired?`) (@server @cli @client-core @test)
128
+
129
+ - **Before.** `NodeExecutionEnv` ignored the floor every shell start carries, in silence (INTEGRATION-CORE §11.1 said so in two sentences), so on the local lane a recursive read under the root carried a deny row's bytes into the conversation with nothing said (KL-r2-a), and nothing could declare that a lane does or does not keep the floor. The env facts probe, the git status frame and the Grep tool's ripgrep — starts the engine makes itself, whose output it relays to the model — carried no floor at all: a repository's `include.path` pointing at a deny-listed file put its fields into the git frame on any lane.
130
+ - **Now.** Judged per table owner, as each table is compiled: the Runner for every table a run compiles — the prepare's, and each one a person's persisted rows bring mid-run (with the requirement such a table fails to compile, so the call that would judge with it is refused; the notice is said once per run) — each delegated child on its own environment, and a library-direct band for a table it compiles itself (a band handed the Runner's face judges nothing). The table holds rows and the environment does not declare `keepsReadFloor: true` ⇒ with `RunnerDeps.readDenyFloorRequired: true` the prepare refuses with `config.read_deny_floor_unkept` before any command starts; otherwise one operator notice under the same code (`detail: { rowCount, sessionId, runId }`). An empty table is not judged. A run without hands runs on the Stub, which keeps the floor (it refuses every start), so it is never judged unkept. The requirement is the read-deny family's deployment seat: a boolean, absent = not required; any other value refuses (`config.read_face_seat`) at the prepare and at `swapDeps`, which now takes it (`config.read_deny_swapped` `detail.floorRequired`). The three relaying starts carry the run's floor (the Grep tool the call's own); an environment that keeps the floor and cannot keep it for one of them refuses it, and the probe's facts are missing / the frame degraded by the arms each already had.
131
+ - **Honest scope.** On `NodeExecutionEnv` a deny row's bytes still reach the conversation through a recursive read: this release makes that visible (the notice) and refusable (the requirement), not prevented. A lane that keeps the floor is a future one (the server's sandboxed lanes, S-2218; a local OS confinement, #2945 T3).
132
+ - **Consumers.** @server: each remote adapter declares `keepsReadFloor` (above); a deployment that must not run without the floor sets `readDenyFloorRequired: true` — on the local lane every run is then refused while the default tiers are on. @cli: every session on the local lane carries one operator notice `config.read_deny_floor_unkept` while the default tiers are on — how it is shown is the shell's. @client-core / @test: `ENGINE_NOTICE_CODES` +1 (audience `operator`); the same word is a prepare `errorCode`.
133
+ - **Deviations from the reference (registered as candidates).** `read-floor-rides-relaying-engine-starts` — ours: the env facts probe, the git status frame and the Grep tool's ripgrep carry the floor; the reference sandboxes the Bash tool's commands only. `read-floor-unkept-announced-when-table-nonempty` — ours: a non-empty deny table is the request, and an environment that does not keep it is announced; the reference warns only when its sandbox was asked for and cannot start.
134
+ - **Pins.** `test/backlog2944-read-floor-kept.test.ts` (J1 Node + default tiers: one notice, `rowCount ≥ 1`, and a recursive grep still reads a fake `.netrc` in a temp directory; J2 an empty table: no notice, the requirement refuses nothing; J3 the requirement on Node: the run fails with the code and nothing started, a self-compiling band refuses the same; J3b an empty table at prepare that gains a person's Read deny mid-run: no start carries a floor on Node; J4 a declared keeper is silent, `"true"` and a missing key are the census's notice and the floor's; J5 bad requirement values refuse at the prepare and the swap, a good one swaps and the next run reads it; J6 a library-direct band once per construction, the Runner's path once; J7 no env / an injected Stub: no notice, the requirement refuses nothing; J7b a delegated child judges its own environment, both ways; J7c rows with no floor for the starts refused as an engine defect; J8 the probe, the git frame and ripgrep carry the run's rows; J9 every environment start in src/ classified relayed or not; J10 a keeper refuses the git frame's start that would read a deny row through `include.path`, and the field is not shown).
135
+
136
+ ### Where an environment's actions land is read off its declaration alone — the remote interface and the Stub's class stop answering capability questions, and a host lane reads as the host it drives (#2948, design/2891 S2, ruled C-R965 and its addendum 1; behavior change: the host lane's five readings below, and smaller ones — the worktree exemption's narrowing open to veto; hard-breaking: HB-3 the host lane's readings flip, and `PersistenceFacts.remoteHands` is renamed `handsOffHost`, three sentences each below, open to veto; closed sets: none; exports: none) (@server @cli @client-core @test)
137
+
138
+ - **Before.** Four questions had more than one reading. *Is a path this environment names the host's?* — the frame's `remote` was `isRemoteExecutionEnv(env) || !pathsAreHostLocal(env)`; the file history asked the remote handle before the declaration; the memory write instruction, the capture carrier and the roster's persistence judgment asked the remote interface; a worktree child on the remote interface had to declare isolation (shape (d)) and one off it was compared by path; the delegated child's placement hint asked the interface. *Is there I/O?* — four `instanceof StubExecutionEnv` checks. *Can it snapshot or park?* — one strict reader, three truthy reads of `capabilities.suspendable`, and a park-only target of "remote and not snapshot-capable", which parked an environment that never said `false`.
139
+ - **Now.** One reader per question. Where paths land is `pathsAreHostLocal`: the frame `workspace_changed.remote` (and `ResolvedWorkspace.remote`) is `!pathsAreHostLocal(env)`; the file history is the host's one tree for a host-disk environment, else the remote handle's identity, else its own; the memory write instruction and the roster's persistence judgment drop the hand band's tools when the band is off the host's disk; a `capture: "off"` declaration needs a supplied carrier store when the environment is off the host's disk; a worktree child is compared by physical path whenever its environment is on the host's disk (isolated or not), exempt only when it declares itself off the host's disk (`hostLocalPaths: false`) and a sandbox (`isolation: true`), and refused otherwise — an environment that does not say where its paths land can be neither compared nor exempted (`isolation.unavailable`); the delegated child's placement hint follows. Whether a shell face mounts (the band's shell, the shell seat, the full shell's reachability) is a declared `io: true` (`performsIo`, now three states); the project marker is read through the environment unless it declares `io: false`. Whether a run can snapshot or park is `declaredSuspend` (three states): `true` (with the restore surface) is the snapshot target, `false` the park-only target, undeclared neither — such a run cannot suspend durably: a resource or platform suspend is INACTIVE, a durable ask is refused, a plan review is skipped, each naming the declaration; a snapshot-less handle not marked park-only is tolerated on resume only by an environment that declares `false`. The remote interface is read only where a remote method is called (the whitelist in backlog2947 P1).
140
+ - **Behavior on the lanes that exist.** None on `NodeExecutionEnv`, on the injected Stub, or on a sandbox lane that states its block. The host lane — the remote interface with `hostLocalPaths: true`, `isolation: false` (the server's default lane) — moves on five readings (HB-3 below). Three smaller moves: an environment off the host's disk WITHOUT the remote interface (an in-memory class) is no longer taught the memory write instruction, refuses `capture: "off"` without a carrier store, and refuses a worktree child (it was compared by path); a declared sandbox ON the host's disk (a bind-mounted container) has its worktree child compared by path (it was exempt); an environment that does not declare `io` mounts no shell face, one that does not declare `suspendable` as a boolean is not parked, and an isolated one that does not declare `hostLocalPaths` (a JS adapter) has its worktree child refused (it was exempt on the remote interface).
141
+ - **Hard-breaking, open to veto — a host lane is no longer read as remote (HB-3) (@server @cli @client-core).** *What changed:* an environment that declares `hostLocalPaths: true` on the remote interface (the server's host lane) now reads `workspace_changed.remote: false`, shares the host's file-history tree (`"host"`), is taught the memory write instruction, accepts `capture: "off"` without a supplied carrier store, and has a worktree child compared by physical path and accepted when it is its own directory (it was refused `isolation.unavailable`). *Why:* each of these asked the interface, and the host lane drives the engine's own machine and disk through it, so it got a sandbox's answers (no memory writes, no worktree children, a refused opt-out). *Who is hurt:* a projection that reads `workspace_changed.remote` (or `ResolvedWorkspace.remote`) to tell the host lane from a sandbox now reads `false` for the host lane — none downstream today — and a deployment that relied on the host lane being refused a worktree child or `capture: "off"` sees it accepted. *Migration:* none on the server side to get the new readings — the declaration already present (`hostLocalPaths: true`, #2947) is read; the explicit carrier store and `memoryPersistenceCapable` stay valid; the execution-lane table's `coreRemoteExecutionEnv` bit and the memory-boundary note that the file plane is a second axis can retire (#2891 X-1); the frame value change has no reader downstream today (the server projects `cwd` only; the CLI and client-core read `cwd`).
142
+ - **Behavior narrowing, open to veto — an isolated environment that does not say where its paths land gets no worktree exemption.** *What changed:* an environment declaring `isolation: true` whose `hostLocalPaths` is left out, not a boolean or unreadable has its worktree child refused (`isolation.unavailable`, a new sentence, no new code); on the remote interface it was exempt, off it it was compared by path. *Why:* the exemption stands on the environment's paths not being the host's, and an undeclared location is neither that nor something the engine can compare — `hostLocalPaths` has two consumers with opposite safe sides (no path presented as the host's; no exemption), read through one three-state reader. *Reach:* a JavaScript adapter only — a TypeScript adapter must state the key (the block is required, #2947); declare `hostLocalPaths: false` on a sandbox lane.
143
+ - **Behavior narrowing, open to veto — a plan review on a per-task environment in neither suspend target is skipped, not parked.** *What changed:* every durable suspend lane — the resource and platform lanes, a durable ask, a plan review — refuses on one fact, a per-task environment in neither the snapshot nor the park-only target, and that fact now names a per-task environment without the remote seam too (a factory-minted host env, a block that is absent, no object or unreadable); a plan review on one is skipped with a notice where it was parked with no workspace handle. *Why:* that checkpoint resumed onto a fresh environment, and the other lanes already refused it. *Reach:* a factory that mints environments without the remote seam; a static environment still suspends in process.
144
+ - **Hard-breaking, open to veto — `PersistenceFacts.remoteHands` is renamed `handsOffHost` (@server @cli @client-core).** *Who needs it:* the roster judgment's facts now say whether the hand band is off the host's disk, which is no longer "remote" — a host lane is remote and on the host's disk — so the old name would state the reading this release retires. *Who is hurt:* a host calling `rosterCanPersistMemory` or `persistenceCapableToolNames` with a facts literal naming `remoteHands` no longer compiles (no reader in the server, CLI or client-core sources today). *Compensation:* rename the member; the value is the same question asked of the declaration (`handsEnabled && !pathsAreHostLocal(env)`).
145
+ - **Consumers.** @server: the host lane gets the five readings above with no code change; the shell faces now mount only on a declared `io: true` — every adapter states it since #2947, and the two `StubExecutionEnv` subclasses inherit `io: false` (no shell, as before). @cli: a local session is unchanged; a session on the server's host lane sees the memory write instruction and worktree children it did not see. @client-core: the frame key is unchanged and its value is `false` on the host lane. @test: P3 / P3b / P4 / P6 / P7 below read the host lane against a sandbox lane item by item.
146
+ - **Pins.** `test/backlog2948-four-readings-one.test.ts` (P3 the host lane's frame, file history, write instruction, capture carrier and worktree child; P3b `isolation` × `hostLocalPaths` on and off the remote interface — a compared env refused at the shared root and accepted at its own, a sandbox off the host exempt, anything else refused, and an isolated env whose `hostLocalPaths` is left out, a string or a throwing getter refused; P4 the sandbox lane as before; P6 `suspendable: "false"` neither snapshot nor parked on the ask lane and the plan-review lane, each refusal naming the declaration, a declared `false` parking, and a per-task env without a usable remote seam (a throwing, absent or non-object block, a plain host env) refused on both lanes; P7 the injected Stub and an env without `io` keep their file tools without a shell; P7b an env without `io` has the marker read through it); `test/backlog2947-capability-table.test.ts` P1 (the interface whitelist is the only structural reading left; no class check, no direct key read).
147
+
148
+ ### An environment's capability block is received once — copied, frozen and the only thing any reader reads — and a member outside the eight keys is announced; a read face is a deployment statement on every lane (#2964, #3030; behavior narrowing, three sentences below, open to veto; hard-breaking: none; closed sets: `config.execution_env_capability_invalid` `detail.key` may name a member outside the eight keys; exports: none) (@server @cli @client-core)
149
+
150
+ - **Before.** One run read the adapter's `capabilities` getter several times (the census, each reader, the remote predicate's block conjunct), so what the census judged and what a reader acted on could differ for a block that changed; a member outside the eight keys (a misspelled `isolaton`) was ignored in silence, leaving the key it meant undeclared with no word about why. A read face's default on a multi-lane deployment, and how a governed deployment sets it, were not written down.
151
+ - **Now.** The block goes through the engine's one receipt of a caller's object (the reading a Model and an options bag get): read once per environment object — at its first read, a run's admission point — its own members each read once, copied and frozen; every reader, the census and the remote predicate read that copy. A member outside the eight keys is one `config.execution_env_capability_invalid` under its own name (`detail: { key, valueType }`), and no reader takes it. A read face is a deployment's trust statement, never an environment capability: no declaration of the environment (a sandbox's `isolation` included) enters the face's resolution order, so every lane's default is `roots`, and a lane that is to read open says so at a seat allowed to open — the deployment seat in force for its run, or the task seat where no organization governance is declared; under governance only the deployment seat opens.
152
+ - **Behavior narrowing, open to veto.** *What changed:* a host that changes an environment's block after the engine first read it is no longer read (the block is a constant of the adapter, as documented); a Proxy block is refused by the receipt before any trap runs (one notice under `capabilities`, every key on its safe side), and a member the receipt cannot copy (a function, a getter that throws) is `unreadable`; a member outside the eight keys is announced. *Why:* one reading — what was judged is what is read, as for a Model (#2846) and a store's options (#2847). *Migration:* none for a TypeScript adapter whose block is a plain data member (the server's seven); an adapter that rebuilt its block per read keeps the first one per environment object — mint a new environment object to change it.
153
+ - **Consumers.** @client-core / @cli: `detail.key` of `config.execution_env_capability_invalid` may now name any member the adapter wrote — a table that pins the key set reads an unknown word as a member that is no capability. @server: the seven adapters declare plain data blocks and see no change; a multi-lane engine under governance has one deployment read-face seat per Runner (the open question is #3035).
154
+ - **Pins.** `test/backlog2964-capability-receipt.test.ts` (R1 one getter read for a run with hands and a shell call, none for a second run, a member getter once; R2 `isolaton` announced under its own name beside the undeclared `isolation`; R3 a block mutated and replaced after the receipt changes no reader and the census repeats the received findings; R4 a Proxy block refused with no trap run); `test/backlog3030-read-face-lanes.test.ts` (the engine's own host, a host lane and an isolated sandbox × no seat / deployment open / task open / both, ungoverned and governed: roots by default on every lane, the deployment seat opens, the task seat opens only ungoverned and only tightens a deployment's open); `test/backlog787-capability-shape-census.test.ts` H6 (a lying Proxy block hides nothing: the receipt refuses it).
155
+
156
+ ### Org memory admission is asked by who asks — the resolver's input is discriminated by caller presence, read once the way the task door reads it, and the memory of an owner with no identity is written down as an opaque key, never a sentinel principal (#900 C0 / C1; hard-breaking: the seat's input type, open to veto; behavior narrowing, open to veto: a task whose principal is the empty string, and a resolver that rewrites its own input; closed sets: none; exports: +2 root types `MemoryScopeAdmissionInput`, `DeploymentMemoryScopeRequest`) (@server @cli)
157
+
158
+ - **Before.** `MemoryScopeAdmission` took `{ principal: string | undefined; requested; requestedWriteScope }`, any origin either way. The door refused a request-origin org scope without a caller only when the principal was `undefined`, while the task door reads the empty string as no principal too (`TaskSpec.principal: ""` passes it as absent): such a task put its request-origin org scopes to the resolver as if `""` were a caller, and a deployment-only request reached the resolver with `principal: ""`. Nothing in the type said the door had already refused the caller-less request-origin shape, so a resolver kept its own refusal arm for it (the server's does, with two cells that build that shape).
159
+ - **Now.** `MemoryScopeAdmissionInput` = `{ principal: string; requested: ReadonlyArray<MemoryScopeRequest>; requestedWriteScope: MemoryScopeRequest | null } | { principal: undefined; requested: ReadonlyArray<DeploymentMemoryScopeRequest>; requestedWriteScope: DeploymentMemoryScopeRequest | null }`, and `MemoryScopeAdmission` takes it. The door reads caller presence once, through `readPrincipal`: a principal name is the caller, handed as written; omitted or `""` is no caller — a request-origin read scope or write slot is refused `memory.admission_denied` before the resolver is asked (the sentence is unchanged), and a deployment-only request is put with `principal: undefined` (the key present, as before); a provided value that is no name is refused `memory.admission_denied` before the resolver (reachable only through the library entrance `admitMemoryScopes` — a task's prepare refuses it earlier as `config.invalid_argument`). The write slot's origin is read once, with the rest of `requested` (it used to be read a second time; a plain origin-stamp record answers the same either way), and the resolver is handed copies: a resolver that rewrote the origins of its input used to change the door's own later reading of them — a deployment scope it granted nothing for then stayed mounted — and now changes nothing the door decides. The resolver-absent leg reads no caller and is unchanged. Not changed here: the compliance posture resolver, consulted before admission, still receives the task's principal as given (`""` included).
160
+ - **The identity-absence contract** (`@contract memory.identity_absence.opaque_key`, on `isCentralAuthorityScope`; one pointer each in the scope-contract module header and on `RuleOwner`): the memory of an owner with no identity lives under an opaque, unprefixed key the deployment chooses (such as `local`), never a sentinel principal or a `user:` key (`user:local` is the key of a principal named `local`), so its authority is not central, and the engine mints no key for it; read as an owner it is `RuleOwner`'s `local-owner` arm; moving it to a principal is a re-keying of its entries (`migrateScope`), and the engine never reads the opaque name as a principal; admission authorizes by the caller, never by a layer's owner. Comments and the registry row only — no key is reserved, no value is refused for being `local`, and the sync client still syncs whatever scope a host hands it.
161
+ - **Hard-breaking (open to veto), three sentences.** *What breaks*: a call that builds a caller-less admission input holding a request-origin scope (read or write slot) no longer type-checks, and a task whose principal is `""` is refused `memory.admission_denied` where its request-origin org scopes used to reach the resolver (its deployment-only request now reaches the resolver with `principal: undefined` instead of `""`); absence is never spelled as a principal string on this seat, `local` included. *Who is hurt*: a host whose tests call its resolver directly with `principal: undefined` and request-origin items; a resolver that narrows on `input.principal === undefined` and then compares an item's origin with `"request"` (the comparison now has no overlap — that is the arm to delete); and a host that sends `principal: ""` with request-origin org scopes. A resolver typed against the wider earlier input is still a seat, and a destructuring one that filters by origin before it checks the principal keeps compiling, its caller-less arm now dead code. *Migration*: delete the resolver's own "principal absent" refusal arm and the cells that build the impossible shape — narrow on `input.principal === undefined`, whose arm holds deployment-origin scopes only — and send a principal name for any request that needs a tenant plane.
162
+ - **Consumers.** @server S-2394: `src/org-memory-admission.ts` keeps a "principal absent" arm (outcome `principal_missing`) that the type now proves unreachable, and `test/org-memory-admission.test.ts` builds the caller-less request-origin shape twice (read and write slot) — those two cells stop compiling and retire with the arm; the resolver's own source compiles unchanged (read in the server checkout and mirrored as a type cell, not compiled there); its `attributed` details lose `principal: ""` on a caller-less deployment-only request. @cli: no call of the seat in the cli source; no change. The memory-owner reading itself (an owner function that names the global layer's owner, the re-keying arc) is the server's and a later core slice; this section changes neither.
163
+ - **Pins.** `test/backlog900-admission-actor-seam.test.ts` (the real Runner over a File memory store, caller {name, omitted, `""`} × request {deployment only, request-origin read, request-origin write slot} × resolver {present, absent}: the caller is handed by name with every origin, a padded or non-ASCII name exactly as written; a provided value that is no name stops at the task door (`config.invalid_argument`, the resolver not asked); no caller ⇒ deployment-only reaches the resolver with `principal: undefined` and the key present, request-origin is refused before the resolver; no resolver ⇒ unchanged: deployment-only mounts and request-origin is `memory.admission_required`; through the library entrance a provided principal that is no name is refused before the resolver, and a resolver that rewrites its input's origins and grants nothing leaves the deployment scope dropped; the refusal sentence names each request-origin scope once; type cells: the caller-less arm cannot hold a request-origin scope, a narrowed resolver's "no caller but request-origin" comparison has no overlap, a wider and a destructuring resolver still compile). `test/export-surface.snapshot.json` +2 types; `docs/CONTRACT-SENTENCES.md` +1 row.
164
+
165
+ ### Retirements — fifteen faces on the retirement line are deleted; sixteen more are re-windowed to 7.61.0 with their reason on the row (RL-1336, #2100, #2909; hard-breaking: fifteen, three sentences each below, open to veto; closed sets: the `loadProjectMemory` phase union −1 (`post-clear`), the prepare refusal codes −1 (`rewind.rewind_files_retired`, now `config.spec_retired_key`); exports: −4 root values (`ARTIFACT_CARD`, `MEMORY_GUIDANCE`, `CODE_SYSTEM_PROMPT`, `createWebFetchTool`); members: −1 `ImportPreview.translated`, −1 `TaskStats.costBreakdown.memoryConsolidationMicroUsd`, −1 `stats.mechanisms.finalVerifyInjected` (and the same member on `task.end`'s mechanisms summary), `CheckpointStore.listByScope` optional → required, `RunnerDeps.loadProjectMemory` return narrowed to `ProjectMemoryLoad`) (@server @cli @client-core @test @sdk)
166
+
167
+ Every row here was on `docs/RETIREMENT-LINES.json` with `retireBy: 7.59.0`; the commit that deletes a face deletes its row, so the table now holds sixteen rows, none due before 7.61.0. `gate:retirement-lines` reads 16 rows online and 35 survival claims in `src/` (11 governed by a row, 24 exempted).
168
+
169
+ - **RL-1336-rewindfiles — `TaskSpec.rewindFiles` is refused by name in both senses.** *What changed:* the key joins `RETIRED_TASK_SPEC_KEYS`; any defined value refuses prepare with `config.spec_retired_key` (a resume whose re-supplied config carries it is refused before the claim as `checkpoint.config_refused` with `detail.code: "config.spec_retired_key"`) whose sentence names the replacement — the capture sense's tolerated no-op with its `config`-phase disclosure is gone, and the restore sense's own code `rewind.rewind_files_retired` is retired with it. *Who is hurt:* @server, whose request resolver forwards `rewindFiles: true` from the HTTP body into the spec (`src/boot/resolve-spec.ts:964`): every request that still carries it now fails prepare; @sdk's `TaskRequest.rewindFiles?` (`packages/sdk/src/types.ts:591`) keeps advertising it; any consumer that branches on `rewind.rewind_files_retired` (@server's `test/resume-anchor.test.ts:148` pins it). *How to migrate:* stop forwarding the key (refuse or strip it at admission, `src/http/admission.ts:700` already reads it); capture needs no request; restore with the conversation is `resumeAt` + `restoreFiles: true`, a code-only restore is `rewindFilesTo`; read `config.spec_retired_key` where you read the retired code.
170
+ - **RL-1336-import-translated — `ImportPreview` has no `translated` member.** *What changed:* the always-empty array is gone from the type and from every preview `prepareCcImport` returns; an entry imports as the file writes it. *Who is hurt:* @server, whose rules-consent face imports `prepareCcImport` and `ImportPreview` (`src/rules-consent.ts:51,60`), and @sdk, which declares `translated` required (`packages/sdk/src/resources/rules.ts:183`, `spec/openapi.yaml:22587`) and pins its presence (`packages/sdk/test/s177-server-767-wire.test.ts:258`); any client rendering a "rewritten spellings" column. *How to migrate:* drop the member from the wire type (or emit `[]` from your own layer if your contract must keep it for a release); render `candidates` and `skipped` only.
171
+ - **RL-1336-artifact-card — the root no longer exports `ARTIFACT_CARD`.** *What changed:* the constant is deleted; the Artifact tool's card id is the string `"artifact"` in `details.type`, as it always was. *Who is hurt:* a host that imports `ARTIFACT_CARD` (none found in @server / @cli / @client-core `src`). *How to migrate:* compare `details.type === "artifact"`.
172
+ - **RL-1336-toolsearch-matched — ToolSearch `details` carries one list of matched names, `matches`.** *What changed:* the second list `matched` (same content) is gone from both the match and the no-match result. *Who is hurt:* a renderer reading `details.matched` (none found in the four neighbours' `src`). *How to migrate:* read `details.matches`.
173
+ - **RL-1336-finalverify-bool — `stats.mechanisms.finalVerifyInjected` is gone; `finalVerifyInjections` is the reading.** *What changed:* the boolean is no longer minted on `TaskResult.stats.mechanisms` nor on `task.end`'s mechanisms summary; the count, present iff at least one injection fired, stays. *Who is hurt:* @server's run-terminal projection table, which lists the key (`src/observability/run-terminal.ts:1221`), and its scan script (`scripts/tb-1249-scan.py:215,230`); dashboards that read the boolean. *How to migrate:* drop the key from the table; read `(stats.mechanisms?.finalVerifyInjections ?? 0) > 0`.
174
+ - **RL-1336-memory-guidance — the root no longer exports `MEMORY_GUIDANCE`.** *What changed:* the concatenation constant is deleted; its two parts `MEMORY_SAFETY` and `MEMORY_HYGIENE` stay exported (core injects none of them). *Who is hurt:* a host that composed its prompt from the one constant. *How to migrate:* `${MEMORY_SAFETY}\n\n${MEMORY_HYGIENE}` is the same text.
175
+ - **RL-1336-code-system-prompt — the root no longer exports `CODE_SYSTEM_PROMPT`.** *What changed:* the alias is deleted; `CODE_AGENT_PROMPT` is the one name of the code prompt (the same catalog text). *Who is hurt:* a host that imports the alias — @server's golden pin imports it (`test/prompt-assembly-golden-7512.test.ts:26,111`); no `src` import found. *How to migrate:* import `CODE_AGENT_PROMPT`.
176
+ - **RL-1336-consolidation-cost — `TaskStats.costBreakdown` has no `memoryConsolidationMicroUsd`.** *What changed:* the permanently-0 line is gone from the type and from every assembled breakdown (and from the delegated child's cumulative breakdown); the breakdown is `{ llmRootMicroUsd, nestedSubagentMicroUsd, compactionMicroUsd }` plus the optional lines. *Who is hurt:* @server's bench arms that build and sum it (`src/bench/s1/arms.ts:155,184,224`, `src/bench/s1/run-firm.ts:208,365`, `test/cost-taxonomy.test.ts:35`) and its run wire, @sdk whose `types.ts:2418` and `openapi.yaml:9163-9167` require it, @client-core's projection (`src/adapter/downstream/terminalToSdkResult.ts:797`, optional there). *How to migrate:* drop the member from the objects you build; a wire that must keep it for a release emits `0` from its own layer, which is what core sent.
177
+ - **RL-1336-post-clear-public / RL-1336-post-clear-local — the `loadProjectMemory` phase union is `"fresh" | "resume" | "post-compact"`.** *What changed:* the reserved word `post-clear`, never emitted, leaves the public `ctx.phase` union and the engine's local one. *Who is hurt:* a loader whose `switch` handles `post-clear` (@server's `src/project-memory.ts:31` declares a wider union of its own, which still accepts the narrower one). *How to migrate:* delete the arm; a session clear starts a new session, which reads `fresh`.
178
+ - **RL-1336-git-mark-param — `stripGitStatusUnits(text)` takes no mark.** *What changed:* the ignored second parameter is gone (internal function; not on the package root). *Who is hurt:* nobody outside the engine. *How to migrate:* none.
179
+ - **RL-1336-journal-reexport — the journal cap has one import path.** *What changed:* `MAX_JOURNAL_RESULT_BYTES` and `oversizeJournalResult` are no longer re-exported by the file backend module `stores/file/workflow-journal-store`; the root keeps both names (now taken from the contract module `core/workflow-journal-store`). *Who is hurt:* a deep import of the file backend module for those two names. *How to migrate:* import them from the package root.
180
+ - **RL-1336-checkpoint-listbyscope — `CheckpointStore.listByScope` is required.** *What changed:* the member is no longer optional; the engine calls it without an optional probe, the store contract kit still names a JavaScript backend that lacks it, and the park self-check (`probeParkRoundTrip`) now reports such a store as `row_not_enumerable` where it used to pass with the step "NOT exercised". *Who is hurt:* a host's own `CheckpointStore` implementation or wrapper that omits it — it no longer type-checks (@server's SQL store implements it; a wrapper elsewhere in your tree may not). *How to migrate:* implement it (pending rows of the scope, the `summarizeCheckpoint` projection) or forward it to the store you wrap.
181
+ - **RL-1336-webfetch-factory — the root no longer exports `createWebFetchTool`.** *What changed:* the second construction path is deleted; `webFetchToolSpec` handed to `spec.tools` is the one path, and the Runner wraps it with its trusted context (host floor, reminder defuse) — a product built outside a Runner never had them. *Who is hurt:* a host that mounted the defined product directly (@server's consumption ledger names it; no `src` call found); tests that execute the tool outside a Runner. *How to migrate:* hand `webFetchToolSpec(config)` to `spec.tools`; to execute it outside a Runner, `defineTool(webFetchToolSpec(config))` (core's own tests do this through one helper).
182
+ - **RL-1336-project-memory-bare — `RunnerDeps.loadProjectMemory` answers a `ProjectMemoryLoad`.** *What changed:* the seam's return type is `ProjectMemoryLoad | Promise<ProjectMemoryLoad>`; a bare string, `null` or any non-object is not read — it is disclosed on `onError` (phase `config`, the sentence names the shape and `{ content: null }`) and the leg runs with no project layer, exactly as when the loader throws. *Who is hurt:* @server, whose `makeLoadProjectMemory` returns `string | null | ProjectMemoryLoad` (`src/project-memory.ts:126,236-263`): its bare-string and `null` answers stop injecting and say so once per prepare; any host loader returning a string. *How to migrate:* answer `{ content: text }`, and `{ content: null }` for no project layer.
183
+
184
+ **Not retired, re-windowed 7.59.0 → 7.61.0 with the reason written on each row** (the rows stay until their carrier lands): RL-1608 ×9 (the TaskOutput tool, its env-direct factory, two catalog rows, five retired-name hints) — deleting the tool before the replacement read seats of design/1056 §2.5 (R1–R7) exist leaves five lanes with no read path; carrier = the #1608 design + split car. RL-1654 ×4 (`MemoryStore`, `consolidateScope`, `runMemoryConsolidation`, the v1 ledger reading) — the three names are one face of a store family still exported under other names, and the v1 arm sits in the locked migration's retry machinery; carrier = the #1654 design section. RL-1103 (the anchor backfill arm) — with it gone a legacy repo partition adopted by `adoptLegacyRepoDirs` refuses at its first open, so the adoption must record the anchor first. RL-1333 (the v1–v9 checkpoint ladder) — a v≤9 row carries no requirement words, so the change needs the parked-row census of the backends holding checkpoints. RL-2164 (the kind-less batch member reading) — its own condition, the server's confirmation that no stored row carries the form, is open.
185
+
186
+ - **Consumers.** @server: `resolve-spec.ts:964` (stop forwarding `rewindFiles`), `project-memory.ts` (answer `{ content }`), `bench/s1` (drop the consolidation line), `observability/run-terminal.ts:1221` (drop the `finalVerifyInjected` key), any `CheckpointStore` wrapper (add `listByScope`; `SqlCheckpointStore` already has it), the rules-consent face's `translated`, and the tests that pin the retired faces (`prompt-assembly-golden-7512`, `resume-anchor`, `cost-taxonomy`); and two requests ride the 7.60.0 pickup sheet — the parked-checkpoint census for RL-1333 and the kind-less batch-member confirmation for RL-2164. @sdk: `TaskRequest.rewindFiles?`, `costBreakdown.memoryConsolidationMicroUsd`, `ImportPreview.translated` in the types, the OpenAPI spec and two tests (`runs-supervisor-cost`, `s177-server-767-wire`). @client-core: the optional consolidation member in the SDK-result projection. @cli: no `src` read of any retired face was found (its `rewindFiles` hits are the out-of-run verb and wire-manifest notes). @test: black-box readings that pin `rewind.rewind_files_retired`, `details.matched` or the bare-string loader are now red by design.
187
+ - **Pins.** `test/retirement-lines-gate.test.ts` (the shipped table: every row's face present, so a row kept for a deleted face reds as STALE; the dangling-citation discriminance now reads RL-2164 / RL-1654 rows), `test/rewind-files.test.ts` and `test/prepare-file-history-transcript.test.ts` (both senses refused by name, `true` and `false`, nothing restored), `test/backlog1050-file-history-capture-off.test.ts` (the refusal sentence names the capture switch), `test/prepare-project-context-transcript.test.ts` (a bare string, `null` and a number each disclosed once, no project layer), `test/tool-disclosure.test.ts` (no `matched` on either branch), `test/final-verification.test.ts` (no `finalVerifyInjected` on the result or on `task.end`, the count is 1), `test/design252-park-selfcheck.test.ts` (a store without `listByScope` is `row_not_enumerable`; with it, `round_trip_ok`), `test/budget-readcontext-and-taxonomy.test.ts` and `test/side-call-usage-account.test.ts` (no consolidation line), `test/full-body.test.ts` (no `CODE_SYSTEM_PROMPT` on the root), `test/backlog490-cc-wildcard-import.test.ts` / `test/permission-rule-consent.test.ts` (no `translated`), `test/export-surface.snapshot.json` −4.
188
+
189
+ ### A judge's refusal sentence never carries text the caller wrote — one shape for a caller's key or value, `[redacted] (N characters)` (#2888; behavior change: the wording of the declaration judges' sentences; codes unchanged; exports: none) (@server @cli)
190
+
191
+ - **Before.** The judges of a declaration quoted what the caller wrote: the `sessionMetaKey` judge its key, the request-meta judge a member's name, `readToolAxes` / `readToolFaces` the tool name and an unknown member's name, the face judge the word a closed table refused and the parameter names and list entries it could not place, the preparation door the `contentOrigin` value, an entry or a task list that was a string, a lock key outside the registry, a credential reference's kind, and every key on a refused member path. A credential pasted into the wrong field rode the sentence into logs, panels and probes; a host could only redact after the fact.
192
+ - **Now.** One rule for the table: a sentence says where (the declaration's own member names, a list index), why (the closed vocabulary, a kind, a count and its bound) and whose (the declarer's name — the entry's published identity, already minted into every tool name the engine mounts). Text the caller wrote — a key or a value — is shown as `callerText`: the redaction mark (`REDACTION_MARKS.whole`) and its length in code points (`[redacted] (12 characters)`), or its kind when it is no string. A closed-set refusal now lists the closed words instead of echoing the word it refused (`pathTarget.access [redacted] (5 characters) is outside the closed set (read | create | edit)`). A member path spells the declaration's member names and shows any other key in the same shape (`toolFaces.[redacted] (9 characters).pathTarget`). The mount's announcement of a face that does not fit reads the same judge, so it changes the same way.
193
+ - **Consumers.** @server S-2476: `src/task-mcp.ts` passes the engine's reason sentences through `redactSecrets` at `:499` only and the deployment leg's two warns (`web-admin-client/skills-mcp.ts:267/283`) not at all — with this release no engine sentence of the table carries the caller's text, so the special case can go; a pin that matched a quoted tool name, key or word in an engine sentence re-pins to the mark. @cli: a pin on one of these sentences re-pins the same way.
194
+ - **Pins.** `test/backlog2888-judge-sentences-withhold-caller-text.test.ts` — one cell per refusal of the table (the `sessionMetaKey` judge's four refusals and the preparation door's; the request meta's two; `toolAxes` unknown member / word / receipt fault; `toolFaces` eleven shapes; the face judge with a schema five; the door's content class, receipt path, string entry and list, lock key and credential kind): the sentinel never appears, nor a prefix of it, and the shape does. Existing sentence pins re-pinned in place.
195
+
196
+ ### A peer that echoes its connection's credential into what it says about itself is read without it — one eraser, one needle (#2946; behavior change: an echoed credential is erased in a peer's account of itself, a name holding one is refused; closed sets: `ENGINE_NOTICE_CODES` +1 `mcp.credential_echoed`; exports: none) (@server @cli @client-core)
197
+
198
+ - **Before.** The request-echo eraser covered only failure text. A server that echoed the bearer it was sent into `serverInfo.version`, its instructions or a tool description, or an A2A peer that echoed it into its card, handed the credential to the status record, the model's tool card and every panel and probe downstream (server S-2476, v7.137.0 the same).
199
+ - **Now.** One eraser (`eraseSentCredentials`), one needle: the credential values the connection's seat answered that the withheld table holds (`PeerCarrier.credentials` — a value that names a secret, `namesASecret`: never a blank one, a placeholder word or a key shorter than eight characters, which would otherwise erase ordinary words from every description). Every remote text a successful dial or listing hands the engine — an MCP server's `serverInfo`, instructions and tool listing; an A2A peer's card — is read through it before anything bounds or neutralizes it: free text holding the needle, in any spelling the request-echo derivation gives it, is erased in place (the redaction mark); a NAME holding it is never rewritten, and neither is a tool's input schema (its structure: the property names the model sends) — the tool is not mounted (dropped through the disclosed lane, the drop record carrying the name erased), an A2A skill is not minted, an interface URL the card names is not used, a `serverInfo` record or a card name is not projected (a transport label a refusal names is free text, erased). Nothing else is taken: an account with no needle in it reads byte for byte.
200
+ - **The notice, three sentences.** *Who needs it*: the operator of a deployment wired to an endpoint that repeats what it is sent — its own logs and panels may hold the credential. *Who is hurt*: a consumer that pins `ENGINE_NOTICE_CODES`' cardinality (+1, nothing removed) or the operator-notice audience table. *Compensation*: `mcp.credential_echoed`, audience `operator`, `detail: { protocol: "mcp" | "a2a", server, where, sessionId, runId }`, `where` the closed set `CREDENTIAL_ECHO_SITES` (`server_info` / `instructions` / `tool_listing` / `agent_card`), never a value; once per (materialization, server, site) for MCP, once per (materialization, peer) for A2A.
201
+ - **The same family, the text pass (#2953).** `redactSecrets` / `redactHostLeaks` took `api_key=…` out of a text but let `apiKey=…`, `{"accessToken":"…"}` and `{"clientSecret":"…"}` through on both tiers, and the escaped form of a JSON string inside a JSON string (`\"api_key\":\"…\"`) for every name: the keyword row now starts a keyword where a name's words start — a camel hump included, the one hump `splitCamelCase` cuts the record-key view's names at — and reads its quotes escaped as well, so a name the record-key view calls a credential is taken out of a text too; a word that only spells a keyword (`hockey`, `myKeyboard=`) is still none. Pins: `test/backlog2953-camel-credential-names.test.ts` (three camel names × bare / escaped / nested × two tiers; the snake-case names unchanged; the controls; one name, one reading) and seven fixtures in `test/fixtures/secret-detect-fixtures.json` (marked tightening).
202
+ - **Consumers.** @server S-2476: the panel and probe faces that publish `serverInfo` and the tool listing now receive the erased account; the notice joins the operator lane's allow list if it filters by code. @cli / @client-core: a renderer of the operator notice lane gains one code.
203
+ - **Pins.** `test/backlog2946-credential-echo-in-metadata.test.ts` — a real Streamable HTTP MCP server and a real A2A peer on loopback, the seat answering a sentinel bearer: `serverInfo.version`, instructions and a tool description erased in place, a tool named with it not mounted and its drop record erased, one notice per site; a `serverInfo.name` holding it drops the record; the positive control reads byte for byte with no notice, and so does an account echoing a short dev key (no needle); the A2A card's description erased before its 240-character bound could cut inside the value, a skill id and the card name holding it refused; the Runner path's notice (operator, the detail above, no value).
204
+
205
+ ### A seat answer the engine refuses is indexed before it is judged — the S-2463 shapes (#2868; behavior change: refused answers' credentials are withheld from every child; exports: none) (@server)
206
+
207
+ - **Before.** Since the peer seat (7.58.0), the seat's answer was judged first (`readPeerAnswer`) and only an accepted answer's credentials were taken into the withheld-value table; the release before took the credential headers in first and judged after. An answer refused for a value with a line break at its edge, a header named twice in two letter cases, or a line break inside a value therefore left its credential unindexed, and a variable of this host holding the same value reached every child (server S-2463).
208
+ - **Now.** One intake (`PeerCarrier`, C-R946 A): the answer's credential values are read from the answer as given — a header member by the credential vocabulary over its value as written and as the transport would send it, an environment member by the env entry reading, both lanes whichever the peer is — held, and only then judged. A blank value indexes nothing. Each value is held on its own, so one that cannot be withheld does not keep the others out (the answer is then refused after the rest were held). The header resolution (`resolveProtocolHttpHeaders`) no longer holds anything: the intake is the one place.
209
+ - **Pins.** `test/backlog2955-seat-answer-indexed-before-judged.test.ts` S1–S3 (one cell per S-2463 shape: refused, and the value held); `test/backlog1754-held-values.test.ts` re-pinned to the intake (the five call sites of `holdOwnCredentialValues(`; the protocol door cells drive a carrier).
210
+
211
+ ### A worker variable equal to a refused request credential no longer reaches the host shell (#2955; behavior change: the 7.56 order restored; exports: none) (@server)
212
+
213
+ - **Before.** The same order as above (judge first, hold after) let a worker variable whose value equals a credential the seat answered in a refused answer reach a `Bash printenv` on the host lane (server S-2535, 7.138.0 review H5: `{ Authorization: V, authorization: V }`).
214
+ - **Now.** The intake above holds the refused answer's credentials first, so the variable is withheld from the child, as it was before 7.58.0; a plain variable is handed on.
215
+ - **Consumers.** @server S-2535: the 🚧 cell (row 12 of the real-process group D) turns green on this release (PICK7600).
216
+ - **Pins.** `test/backlog2955-seat-answer-indexed-before-judged.test.ts` S4 (a real dial through the public probe, a real `NodeExecutionEnv` child under `"all"`: the refused credential's worker variables do not reach the child, the control does), S5 (a member of the other lane), S6 (blank values index nothing), S7 (an accepted answer held once, at the same intake).
217
+
218
+ ### A sync cursor refused on its store names two ways out and takes neither — rebind or clear (#2957; behavior narrowing: the documented answer to a store-axis refusal changes; open to veto, C-R(随发车帖); exports: none) (@server)
219
+
220
+ - **Before.** #2836 (7.59.0) refused a cursor bound to another store — or to none — and named one answer for every refusal: run the round from no cursor. A host that took it after a store moved re-converged from an empty baseline: an entry deleted on either side since the last completed round came back from the other side, and no conflict was reported.
221
+ - **Now.** The store-axis refusal (`config.memory_sync_cursor_mismatch`, the same judge and code) names two ways out and what each costs, and the engine takes neither: *rebind* — when this store holds the entries the baseline was taken over (the same store moved to another root), persist the cursor with `store` set to this store's identity (`memorySyncStoreOf`) and sync again; *clear* — delete the cursor and sync from none, accepting that an entry deleted on either side since the last completed round comes back. A cursor with no store binding (written before cursors carried one) reads the same sentence. A scope mismatch names the partition's own cursor instead (a partition key, not a binding); a peer mismatch names its own two ways out (#2968).
222
+ - **Behavior narrowing, three sentences (open to veto).** *Who needs it*: a deployment whose sync layer's store moved, which must not have its deletes resurrected without being told. *Who is hurt*: a host that answers every `config.memory_sync_cursor_mismatch` by running from no cursor — @server's layer executor (`src/memory-sync-client.ts`, the `memory_sync_cursor_rebound` arm) — which now has to pick a way out per layer, as it already does for a cursor written before the binding (its `memory_sync_cursor_predates_store` arm). *Compensation*: the refusal is raised before any read or transport call, so a layer that stops on it loses nothing; rebind is one member set from `memorySyncStoreOf` (root export, #2956), and the two arms of the server's executor read one sentence.
223
+ - **Pins.** `test/backlog2957-sync-store-move-refused.test.ts` — a store moved whole to another root with a local and a central delete since the last round: the old cursor is refused with zero calls on the wire and the sentence names rebind and clear; rebind sends the local delete and lands the central one (neither comes back); clear costs what it says (no delete sent, the local delete undone); the unbound cursor reads the same way out; a scope mismatch names the partition's own cursor. `test/backlog2836-sync-cursor-store-binding.test.ts` comments re-worded (the cells unchanged).
224
+
225
+ ### An MCP server can be mounted on a pipe the host supplies — `transport: { kind: "host" }` plus the leg's host seat (`RunInternals.mcpHostSeats`), through the one MCP pipeline every server uses; a seat that did not hand a request on is one failure word, `host_refused` (#2841, design/2825 §3 Q1-B, C-R932 + addendum 1; hard-breaking: two closed sets widen, open to veto)
226
+
227
+ - **Before.** A host that carries MCP over its own channel (a device bridge) had no mount: it could stand up a loopback HTTP MCP server per leg — where "the request never left" could only read as an HTTP status (`unknown`) or a JSON-RPC error (`yes`) — or pre-fill the exported connection pool under a stdio / http declaration that was not the connection's truth.
228
+ - **Now.** A declaration `{ name, transport: { kind: "host" } }` (no other transport member; no `elicitation`) dials through `RunInternals.mcpHostSeats.get(name)` — an `McpHostSeat` with one method, `request("tools/list" | "tools/call", params, signal)`. The engine's one adapter answers `initialize` itself (server name = the declaration's `name`, version `""` = not declared, capabilities `tools` only, no instructions), hands list / call to the seat with the request's own signal (aborted by the engine's abort, timeout or idle watchdog, and by the connection's close), drops an answer that arrives after its request was cancelled, and refuses any other method without handing it on. Listed `_meta`, the call's `_meta` (`sema/toolUseId`, the request meta, the session key), the intake, the manifest row and `tool_end` are the ones a stdio / http server gets. A seat that throws `McpHostSeatFailure({ delivered: "no", condition })` ⇒ `host_refused` / `delivered: "no"` / no liveness observation, the line "was not sent … sending it again is safe"; `{ delivered: "unknown", condition }` (or any other throw, in the engine's own words) ⇒ `unknown` / `unknown`, the outcome-unknown line and, for a write-capable tool, the verify-first warning; the remote's own `{ error }` ⇒ `protocol` / `yes`. A host seat's connection never enters the run tree's pool. A delegated child is handed the seats of the host servers it inherits; a retained child's snapshot keeps none and each wake hands the waker's (`createSubagentResume({ currentMcpHostSeats })`, filled by SendMessage and the steer-handle resume); a checkpoint resume reads the row from its config and the seat from its internals. No seat for a host declaration ⇒ that server fails its dial (`invalid_config` / `no`, `config.mcp_host_seat_missing`), siblings mount; a seat no host declaration names ⇒ one `config.mcp_host_seat_undeclared` operator line, nothing refused.
229
+ - **Hard-breaking (open to veto).** *What changed:* `McpServerSpec.transport` gains a third arm `{ kind: "host" }` (`MCP_TRANSPORT_KINDS` = `stdio` | `http` | `host`) and `McpFailureKind` gains `host_refused` (wire: `mcp[].errorCode`, `tool_end.errorCode`, a request error's `details.code`); `unknown` widens to "or the host seat cannot tell how the request ended". *Who is hurt:* a consumer that switches exhaustively over `transport.kind` or over the failure words fails to compile, and a runtime closed table sends the new words to its default arm. *How to migrate:* add the `host` arm (it names no command or url — render it as a host-supplied pipe) and a `host_refused` row (`delivered: "no"`, no liveness: not sent by the host, safe to retry); read `unknown` with its widened definition. @server (S-2426: implement the seat, hand it on every entry's internals, keep the hub's verify-first tail out of `condition` — the engine adds the tail once) @client-core @cli (projection / rendering of both words).
230
+ - **Closed sets.** `MCP_FAILURE_KINDS` +1 `host_refused` (delivery `"no"`, liveness none); new frozen list `MCP_TRANSPORT_KINDS` (lockstep with the union).
231
+ - **Exports.** +2 root values `McpHostSeatFailure`, `MCP_TRANSPORT_KINDS`; +4 root types `McpHostSeat`, `McpHostSeatMethod`, `McpHostSeatAnswer`, `McpTransportKind`; `RunInternals.mcpHostSeats`; `materializeMcpTools` +1 trailing parameter (`hostSeats`).
232
+ - **Pins.** `test/backlog2841-host-seat.test.ts` (real Runner: mount + list/call only + the call id in `_meta`; a refused listing ⇒ failed / host_refused / no, no liveness; an unknown call ⇒ unknown / unknown, the warning once for the write-class tool only; the remote's error ⇒ protocol / yes; a child dials the parent's seat itself; a retained wake uses the waker's seat, never the spawn's; a checkpoint resume dials the resume's seat; an undeclared seat ⇒ one line. Adapter under the SDK client: 1 000 failures leave no handler; an answer the protocol cannot carry (six shapes), an answer rewritten after it was returned, and a thrown value the engine cannot inspect each settle at once as `unknown` with no handler left and nothing unhandled; abort / timeout / close reach the seat's signal, a late answer is dropped; a method outside list / call is not handed on; a malformed declaration beside a seat of its name does not end the prepare. The door and the library entrance refuse a host transport with another member or with elicitation; no seat ⇒ that server fails, siblings mount; the probe reads it failed; a remote's error shaped like the verdict is `protocol`; the pre-filled pool path as it stands (#2843)).
233
+ - **Known limits.** Requests only (CC's SDK servers carry any message both ways) — no progress notifications, so a host call sits in the 5-minute idle bucket; the pre-filled pool path still leases a host-made connection and a re-dial dials its declaration (#2843 decides). C-R986 has ruled: from 7.61.0 a host-made connection entering the exported pool is refused loudly (`config.*` +1, with car-2929-fix1); do not rely on this path in 7.60.0.
234
+
235
+ ### A Model the engine is handed is read once into a frozen snapshot, judged there, and only the snapshot is read afterwards — the model door at every seat, the catalog door at `new Runner` / `swapDeps`, `modelOriginOf(snapshot)` for a host's own identity lookups (#2846, with #2781 / #2778; design 2772-2781, C-R933 and its addenda 1 / 2; behavior narrowing; hard-breaking: three sentences below, open to veto; closed sets: none; exports: +1 root value) (@server @client-core @cli)
236
+
237
+ - **Before.** One resolution read the caller's Model object two to four times — the role-seat test, the identity judgment, a `select`'s judgment of every candidate and of the pick, the tier default's re-judgment — and each read ran the caller's code on the live object. An object whose members came from its prototype and threw read as an empty role record and fell to the default without a word; an accessor that deleted `id` between two judgments passed both; a Proxy throwing on `cost` surfaced from a `select` as a bare `TypeError`; a getter inside `cost` threw past the judgment. A catalog entry was judged only when a task named it, a tier binding a catalog key shadowed was never judged, the classifier seat's tier-relative pick handed on an entry no judgment read (#2778), and a degrading brain's fallback was deduplicated by an `id` read before the one it was dispatched with.
238
+ - **Now.** A Model is received by the model door wherever it is written — `TaskSpec.model`, a role seat, a role record's `model`, a tier binding, a catalog entry, a degrading brain's fallback — through the engine's one receipt in the declaration mode: a plain record, every own member read once, every level copied and frozen; then judged on the copy: every member one of the `Model` contract's (a retired one keeps its own `config.deps_retired_key`), a non-empty string `id`, a string `api`. Every refusal is `config.unknown_role` / `not_a_model_ref` naming the seat (the code and reason unchanged; the sentence now names what could not be read). The code does change for a member value the receipt cannot copy (an object of another prototype, a function, a bigint, a symbol) that a member's own judge refused before: on a task's model, a `Model.extraBody` (`config.invalid_argument`) or `Model.promptTraits` (`config.prompt_trait_unknown`) of that kind is now refused at the seat by the model door's code first (a host brain's own request-time model spec, no seat of the door, still meets `extraBody`'s judge). The catalog door receives a catalog generation whole where a Runner takes one — `new Runner` and `swapDeps` (atomic, as before) — every entry and every tier binding (a shadowed one included), then expands the bindings over the received catalog and freezes it. Every read of a catalog entry hands its snapshot (an entry of a catalog a host built and hands an exported function is received at that read). A `select` reads frozen data. A role map stays the caller's between operations (a deployment may change it in place) and is read once per operation — a task's resolution, a role's model if set, the preset's role base, the classifier seat, the deployment's subagent tier — so a branch and the resolution it takes read the same copy.
239
+ - **`headers` is a read port.** An own accessor at `Model.headers` is never called by the receipt: the snapshot's `headers` calls the caller's getter at each read — each request — so a credential the deployment rotates without a restart is read at the next request, and a getter that throws throws at that request, not at construction. A data `headers` is received like any other member.
240
+ - **Identity.** A snapshot is a fixed point (handed back as itself), and snapshots are interned by origin and content: the same unchanged object received again (a swap that keeps it, a tool-local roster that shares it) is the same snapshot, so "this word names the very model the child was seated with" answers as before; an object changed in place and received again is a new snapshot. The engine hands back its snapshots, never the caller's objects: `modelOriginOf(snapshot)` (root export) is the object a snapshot was received from; a value the engine never received is its own origin.
241
+ - **Hard-breaking, open to veto (@server @client-core @cli).** ① A Model that is a Proxy, a class instance or an `Object.create(…)` value, or that holds a member the receipt cannot copy (a function, a getter that throws, an object of another prototype), ran as long as its members could be read; it is now refused at the seat it is written. Who needs it: every reader that judged a Model and then used it (what was judged is what is used). Who is hurt: a host building Models from classes, Proxies or prototypes, and test fixtures doing the same (server's Models are object literals). Compensation: build the Model as a plain record (`{ ...instance }`). ② A catalog entry or tier binding that is no Model was refused only when a task first named it; it is now refused when the catalog is received (`new Runner`, `swapDeps`), naming `RunnerDeps.models.<key>` / `swapDeps.models.<key>` / `tiers.<name>`. Who is hurt: a deployment carrying a broken entry nothing selects (it now fails at boot or at the swap); server's entry judge drops such entries before they reach the engine. Compensation: remove or fix the entry the refusal names. ③ A Model member outside the contract was ignored; it is now refused. Who is hurt: a JS host or fixture carrying a key of its own on the Model. Compensation: put a provider's own parameters in `Model.params`. And one reading: a host that compares a Model the engine returns with its own Model objects by `===` (server: the `{ select }` default's catalog name lookup, the consolidation seat's billing comparison) reads `modelOriginOf(model)` instead.
242
+ - **Not in this change.** `roles` is not a hot-swappable seat (#2845, its three preconditions are a separate cut); between operations the role map is the caller's live object, as before.
243
+ - **Pins.** `test/backlog2846-model-door.test.ts` (five hostile forms × nine seats through the real Runner where the seat is the Runner's; the accessor that deletes `id`; a never-named entry; the member set; fixed point, interning in four forms and `modelOriginOf`; the classifier role read once; the `headers` read port, its poison and a real degrading chain rotating per hop; the tier marks; the tier-relative pick; the workflow allowlist).
244
+
245
+ ### A store's constructor options and the retention door's input are read once, by the receipt law — no store keeps the caller's bag, and a TTL, sweep interval, eviction word, durability word or size bound that had no exit is refused (#2847, design 2772-2781 §1.3 / §1.4; behavior narrowing; hard-breaking: one reading, three sentences below, open to veto; closed sets: none; exports: none) (@server @client-core)
246
+
247
+ - **Before.** The bundled stores read their options off the caller's bag as they went: `InMemoryCheckpointStore` read its judge twice (once for the judgment, once for the `fence` declaration), `TtlSessionStore` read `repo` three times, re-read `opts.hooks` at every deletion and held `placements` as the caller's object, `InMemoryToolResultStore` held the caller's bag and read its bound at every write, `FileStorageBackend` asked a `retention` presence through the `in` trap, read three options twice and spread three nested records off the caller's objects. A getter that threw was a bare throw; a Proxy ran its traps; `defaultTtlDays: "abc"` never expired a session, a negative TTL expired every one, `evict: "nuke"` read as `forget`, `maxTotalChars: "abc"` read as unbounded. `assertRetentionCapability` / `assertRetentionWords` read `policy.maxAgeDays` two or three times and each store item twice.
248
+ - **Now.** Every bundled store — `InMemoryCheckpointStore`, `TtlSessionStore`, `InMemoryToolResultStore`, `FileStorageBackend` and each File store it builds (`FileCheckpointStore`, `FileSessionRepo`, `FileMemoryStore`, `FileToolResultStore`, `FileSessionPolicyStore`, `FileWorkflowJournalStore`, `FileUsageWindowStore`, `FileBackgroundAgentStore`, `FileMailboxStore`, `FileRosterStore`, `FileFileHistoryStore`) — receives its options through the options door (the receipt law: a plain record, every member read once; `config.invalid_argument`) and reads them only off that copy; a nested record is received where it is read; faces (a repo, an embedder, a hooks object, a judge) are kept as they are and called as they are. `TtlSessionStore` calls its deletion hook on the hooks object read at construction and holds `placements` as a frozen copy. The retention door receives its input bag, its policy (a receipt fault is `config.retention_policy`) and each store item once; a store is a face, its `retention` read once where it is judged (on its prototype too) — a getter that throws is `config.retention_capability`, as the executor's members already were. Refused, `config.invalid_argument`: `defaultTtlDays` not a finite non-negative number, `sweepIntervalMs` not a finite number, `evict` outside `delete` / `forget`, `durability` (and `placements.subagent.durability`) outside `durable` / `process-local`, `maxTotalChars` not a number.
249
+ - **Hard-breaking, open to veto (@server @client-core).** What changed: an options bag that is a Proxy, a class instance or an `Object.create(…)` value — and a nested options record of `FileStorageBackend` of those shapes — was read through; it is refused. Who is hurt: a host that builds store options from a class or wraps them in a Proxy, and one passing one of the bad values above (it ran with the silent reading). What to do: hand a plain record (server builds every store's options as object literals); write the value in its domain.
250
+ - **Pins.** `test/backlog2847-store-options-door.test.ts` (per store: a getter that throws, a Proxy that must not be asked, a class instance; an accessor that rewrites another option while read; the bound not held from the caller's bag; the judge read once with its `fence`; the hooks face read once; the placements copy; eleven bad values and their controls; the retention input: a Proxy list, an item getter, a store's prototype declaration admitted and one that throws refused, a policy read once, a Proxy policy).
251
+
252
+ ### A host that takes a peer seat's answer at its own door reads it with the engine's own reader — `readPeerAnswer` is on the package root (#2863; behavior narrowing: none; hard-breaking: none; closed sets: none; exports: +1 root value, +2 root types; members: none) (@server)
253
+
254
+ - **Before.** The reader of a seat's answer (the kind and members, the lane's own member only, header and variable names, no line break or NUL in a value, `expiresAt`, the `authorizeUrl` law) lived in the leaf module; the package root exported the five answer types but not the function, so a deployment's intake gate that wanted to refuse a bad answer before dialing had to write the rules a second time, and a bad name could only fail loudly at the dial.
255
+ - **Now.** `readPeerAnswer(raw, lane)` is a root export, with `PeerLane` and `CarriedPieces` (its parameter and result types). It is the very function the engine's send runs: `{ carry }` (frozen pieces and their digest), `{ unauthorized }` (the seat's validated URL, if any) or `{ problem }` (names a member, never a value). The root name is the leaf's function (`===`), and the problem it returns is the sentence a failed send carries.
256
+ - **Consumers.** exports +1 value, +2 types, nothing else changes. @server: the 7.138.0 acceptance gate reads an answer with this function instead of its own copy of the rules; delete the copy.
257
+ - **Pins.** `test/backlog2863-small-root-exports.test.ts` P1–P4 (identity; sixteen refusals each naming the member and none echoing the value; the pieces an accepted answer gives are the pieces a real `PeerCarrier` send carries, same digest; a refused answer's problem is the failed send's text). Smoke decision: `readPeerAnswer` in `SMOKE_CHECKED_EXPORTS`.
258
+
259
+ ### An A2A peer's failed row carries the failure record an MCP server's failed status carries — `errorCode` from the same closed set, `delivered`, `authorization` — read by one classifier and written by one row writer (#2865; behavior narrowing: none; hard-breaking: none; closed sets: none new — `MCP_FAILURE_KINDS` is now read by a second protocol; exports: +1 root type (`McpFailureFields`); members: +4 optional on `A2aPeerStatus` and on a failed `A2aRefreshResult`) (@server @cli @client-core)
260
+
261
+ - **Before.** A failed A2A peer row had `name`, `status` and `error` only, while the same seat refusal on an MCP server read `credential_unavailable` (or `unauthorized`) with `delivered: "no"` on the status, the manifest row and `tool_end`: one fact, two protocols, two readings — a consumer could promise "the same word on every face" for MCP only.
262
+ - **Now.** A peer whose materialization (or whose refresh) threw gets `errorCode`, `delivered` and, for `unauthorized`, `authorization` (`{ reason: "none", authorizeUrl }` for a seat that answered `unauthorized`), minted by `classifyMcpFailure` at the dial and spread by `failureFields` — the same two functions an MCP status row uses, so a member the record gains rides both. A seat that throws, times out or answers what the engine cannot carry is `credential_unavailable`; a card naming an endpoint outside the declared origins is `invalid_config`; a refused connection is `connect_refused`. **Stated reading:** an error the A2A client itself raises for a non-2xx status or a JSON-RPC error carries no structure the classifier reads, so it reads `unknown` (delivered `no` at the dial) — `error` stays the text to show, never the field to key on. A ready row, a successful refresh entry and the refresh entry for an undeclared peer name (nothing was attempted) carry none of the members.
263
+ - **Consumers.** members +4 optional on `A2aPeerStatus` / `A2aRefreshResult`, present iff the row records a thrown failure; `McpFailureFields` joins the root types. @server: the "same word on three faces" promise can cover A2A rows for the seat refusals; @cli @client-core: a roster that already keys on an MCP status's `errorCode` reads A2A rows the same way.
264
+ - **Pins.** `test/backlog2865-a2a-failure-code.test.ts` (three seat refusals on a real loopback card peer, each compared with the MCP status of the same seat — equal records; invalid_config, connect_refused, the stated `unknown`; a ready row has no member; every word is in `MCP_FAILURE_KINDS` and `authorization` rides `unauthorized` only; the refresh receipt's failed, refreshed and undeclared entries).
265
+
266
+ ### The redaction mark table is a root export and frozen — `REDACTION_MARKS` (#2887; behavior narrowing: the table is frozen (a write throws in strict code); hard-breaking: none; closed sets: none; exports: +1 root value; members: none) (@server)
267
+
268
+ - **Before.** The root carried only the bare mark (`REDACTION_MARK`); the three shape marks (`[redacted-credentials]`, `[redacted-url]`, `[redacted-path]`) were spelled by hosts that needed to tell which shape the engine withheld — a second copy of strings the engine's passes read from one table.
269
+ - **Now.** `REDACTION_MARKS` (`whole`, `credentials`, `url`, `path`) is exported from the root and is `Object.freeze`d: every pass reads it at call time, so a writable exported table would let a host change what the engine writes. `REDACTION_MARK` is the table's `whole`.
270
+ - **Consumers.** exports +1 value. @server: delete the hand-written marks in the workflow notify journal and read the table.
271
+ - **Pins.** `test/backlog2863-small-root-exports.test.ts` M1–M4 (identity with the leaf and four spellings, frozen; each shape's spelling is what `redactSecrets` / `redactHostLeaks` write for that shape; every write, delete and add on the table throws and the engine's output does not move). Smoke decision: a presence row and an identity row (`SMOKE_IDENTITY_EXPORTS`, which also asserts frozen).
272
+
273
+ ### A store twin outside the package folds a rule-store sync-join delta with the engine's own landing pipeline — `applySyncJoin` and `RuleSyncJoinBase` are root exports (#2831; behavior narrowing: none; hard-breaking: none; closed sets: none; exports: +1 root value, +1 root type; members: none) (@server)
274
+
275
+ - **Before.** `RuleWriteDelta` exported the `sync-join` arm "for out-of-repo store twins" and the declaration said `applySyncJoin` was for them, but only the function's own module exported it, and the package exposes no deep path: a SQL backend meeting the arm could only refuse it.
276
+ - **Now.** `applySyncJoin(current, delta, nowMs)` and the type of the current raw facts it takes and returns (`RuleSyncJoinBase`) are exported from the root — the function both bundled stores run in their write step, with its refusals and drops (a tombstone naming no observed add refuses the whole delta; an inbound record claiming this replica's actor above its counter is dropped) unchanged.
277
+ - **Consumers.** exports +1 value, +1 type. @server: the SQL rule store can implement the `sync-join` arm with it and stop refusing.
278
+ - **Pins.** `test/backlog2863-small-root-exports.test.ts` J1–J5 (identity with the leaf; an inbound rule lands beside the current one; the whole-delta refusal for a tombstone with no observed add reaches the twin; a delta of another kind is refused by name). Smoke decision: `applySyncJoin` in `SMOKE_CHECKED_EXPORTS`.
279
+
280
+ ### A session's offloaded tool results can be enumerated for a move — `ToolResultStore.listBySession` on the memory, file and example Postgres stores, with a contract-kit leg (#2894; behavior narrowing: none; hard-breaking: none; closed sets: the contract kit's optional-member report word `member` +1 (`"listBySession"`); exports: +2 root types; members: +1 optional on `ToolResultStore`; the file backend's `*.owner.json` record gains a `ref` member) (@server @client-core)
281
+
282
+ - **Before.** A store had `put`, `get`, `ownerOf` and the sweep `deleteBySession`, and nothing that named what a session owns: the results a transcript points at by reference could not leave the store they were offloaded to, so a session moved to another store read references it could not dereference.
283
+ - **Now.** `listBySession?(sessionId)` answers `{ entries: [{ ref, provenance }], unattributable, unaddressable }`, selected exactly as `deleteBySession` selects: by the owner recorded at the winning write, never by parsing the ref, `taskId` ignored in the match (the entry's `provenance` keeps the one recorded), an unowned entry never listed and counted in `unattributable` (store-wide, the figure a sweep reports). There is no second content path: the bytes are read with `get` (bounded slices) and written to the other store with `put(ref, content, entry.provenance)`, whose write-once election makes a repeated move a no-op and a ref already owned by a different owner a typed refusal. No order is promised. The scope wrapper does not forward the member, as it does not forward the sweep. The file backend's folded file names are a digest of the ref, so its owner record now states the ref (`{ sessionId, taskId?, ref }`; a reader that ignores members it does not know — this version's, unchanged since 5.28.0 — reads it as before).
284
+ - **Reading the two counts.** `unattributable` is rows no session can claim (store-wide); `unaddressable` is one thing only: rows this session owns that were written **before** the record carried the ref (file backend, versions before this one) — their ref cannot be recovered from a digest, so they are counted and not listed (never under a guessed ref); they are still deleted by the sweep, so `entries.length + unaddressable` is what a sweep of the session then deletes. A store whose rows were all written by this version or later reports `unaddressable === 0` for every ref form; it is not a second kind of entry and there is no migration for old records. A move that must be complete should check `unaddressable === 0`.
285
+ - **Consumers.** members +1 optional on `ToolResultStore` (implementers are not broken); `ToolResultSessionListing` / `ToolResultListedEntry` join the root types. The contract kit's `onOptionalMember` report word widens from `"deleteBySession"` to `"deleteBySession" | "listBySession"`: a caller that annotated its callback with the narrower literal type needs the wider one (an inferred callback is unaffected). @server: your `toolResultStoreContract` call sites, and the session-sync manifest can include the external results of a session (list, then `get`/`put`); @client-core: the orchestration of that move reads the two counts above. Out of scope here: a fork to a new session id does not carry offloaded results, and the orchestration.
286
+ - **Pins.** `test/backlog2894-tool-result-list-by-session.test.ts` (L: a session moved between two stores through the listing alone on memory and file — bytes, owners, idempotent repeat, typed refusal, other session's and unowned rows stay, listing = what the sweep deletes; F: the file layout — ref read from the record for folded names, a record without the member unaddressable and still swept, a record naming a ref stored elsewhere not believed, identity-named entries, damage and owner-less content unattributable, an owner record whose content an interrupted sweep removed listed as the sweep counts it, seven ref forms all listed with `unaddressable === 0`; W: the scope wrapper) and the contract kit's new optional leg run on all three bundled backends (`test/cross-backend-shared.ts` reports `verified`).
287
+
288
+ ### The file checkpoint ledger's event kinds are one exported closed set, and a data root's ledger reads with the store's own fold and no store — `CHECKPOINT_LEDGER_EVENT_KINDS`, `readCheckpointLedger` (#2832; behavior narrowing: none; hard-breaking: none; closed sets: one new, `CheckpointLedgerEventKind` (7 words); exports: +2 root values, +3 root types) (@server)
289
+
290
+ - **Before.** The seven event kinds the file checkpoint store folds out of `checkpoints.jsonl` (`put`, `resolve`, `reopen`, `execution`, `expire`, `delete`, `steer_append`) lived in a private registry, and the fold ran only inside the store, behind the ledger directory's writer fence. A program that wanted to know what a data root's ledger held without opening the store (the server's drain check) copied the kinds and wrote its own fold. Its copy had six kinds; `delete` has been a seventh since 7.56.0, so the first `delete` line in a ledger counted as an event nobody understood.
291
+ - **Now.** `CHECKPOINT_LEDGER_EVENT_KINDS` (frozen, seven words) is the one list: the store's discriminator fence reads it, and it is total over the event union both ways at compile time (`CheckpointLedgerKindsAreTheEventUnion`). `readCheckpointLedger(root, { onCorruptRead? })` replays snapshot and ledger through the store's own fold — `applyCheckpointEvent` behind the same fence, via `replayLedger`, the one function the authority's own bootstrap now runs after taking its fence — and answers `{ settled: true, rows: ReadonlyMap<token, Checkpoint>, events }`. It takes no writer fence (`checkpoints/LOCK` is neither created nor consulted: a live store keeps its directory, a dead owner's fence is not pruned), opens no append handle and creates nothing (a root with no ledger reads as no rows and the directory stays absent). A kind outside the list refuses the read as it refuses the store's replay (`checkpoint.unsupported_version`, `detail.reason` `ledger_event_unknown`); an adoption in flight refuses it (`adoption_in_flight`). Beside a compacting writer the pair is bracketed by the snapshot's version: a snapshot replaced between the two reads is redone, at most three times, and a directory that will not hold still answers `{ settled: false }` — never the last read.
292
+ - **Closed sets.** New `CHECKPOINT_LEDGER_EVENT_KINDS` / `CheckpointLedgerEventKind` (registered, `docs/CLOSED-SETS.md`); no existing set moves.
293
+ - **Consumers.** exports +2 values (`CHECKPOINT_LEDGER_EVENT_KINDS`, `readCheckpointLedger`), +3 types (`CheckpointLedgerEventKind`, `CheckpointLedgerReading`, `ReadCheckpointLedgerOptions`); nothing else changes. @server: in `scripts/drain-check-local.mjs` delete the hand-copied `KNOWN_LEDGER_EVENTS` and the fold; call `readCheckpointLedger(root, { onCorruptRead })`, count `pending` off `rows`, take `corrupt` from the sink, read `settled: false` as your existing "files kept changing" answer and a thrown refusal as rc 2.
294
+ - **Pins.** `test/backlog2832-checkpoint-ledger-reader.test.ts` (A: the seven words, frozen, in the root, each admitted by reader and store and the retired `steer` refused; B: a six-word ledger reads, a `delete` removes its row and a snapshot row too, the rows equal a real store's after real operations; C: a root with no ledger creates nothing, a live store's fence is neither contended with nor disturbed, a dead owner's fence is not pruned; D: unknown kind / adoption in flight refuse like the store, damage reaches the sink; E: a snapshot replaced under the read is redone and never served, three replacements yield `settled: false`).
295
+ - **Known limits.** The reading is a view at one instant, not a lease: a live store may resolve a row a moment later.
296
+
297
+ ### A task whose role chain supplies no model is refused with the roles family's code — `config.unknown_role` with `detail { field, reason: "no_model_for_role" }` — instead of a bare `Error` (#2833; hard-breaking, open to veto: type only; behavior narrowing: the refusal gains a code and a detail, its sentence is unchanged; closed sets: `CheckpointErrorReason` +1 (`no_model_for_role`, 31 → 32); exports: none) (@server @client-core)
298
+
299
+ - **Before.** A task that names no model, whose requested role and every role down its fallback chain supply none, ended `resolveTaskModel` in `throw new Error('No model for role "…": …')`: no `code`, no `detail`. Every other refusal of that family — a word the catalog does not hold, a value that is no model reference, a role word outside the vocabulary — was already `config.unknown_role` with `detail { field, reason }`, so a host telling a caller's configuration fault (400) from an engine fault (500) matched the sentence's opening words for this one arm, and the run's terminal carried the message alone.
300
+ - **Now.** The arm is the family's fourth: the one constructor (`modelBindingRefusal`) builds `config.unknown_role` with `detail { field: "roles.<requested role>", reason: "no_model_for_role" }` (`roles.default` for a task that names no role); the sentence is unchanged, so a text match still holds. A fresh run's terminal states it as every sibling arm does (`terminal.code`, `terminal.detail.detail.{ field, reason }`); a seat that names a word the catalog does not hold still wins with its own `unknown_model_ref`.
301
+ - **Hard-breaking (type only, open to veto).** *What changed:* the closed set `CheckpointErrorReason` gains `no_model_for_role`, and a refusal that had no `code` has one. *Who is hurt:* a host or a backend that keeps an exhaustive table over the type (`Record<CheckpointErrorReason, …>`, a `switch` ending in a `never` default) is tsc-red until it places the word; a consumer that treated a failed terminal with no `code` and this sentence as an engine fault now sees a caller's configuration fault. *Compensation:* place the word beside `unknown_model_ref` (same disposition: the caller's configuration, fix the roles or name a model); read `code` / `detail.reason`, not the sentence — the sentence is unchanged, so nothing is lost while you move.
302
+ - **Closed sets.** `CheckpointErrorReason` +1, `no_model_for_role` (31 → 32; `docs/CLOSED-SETS.md`, `docs/INTEGRATION-CORE.md`).
303
+ - **Consumers.** @server: `side-query.ts` classifies the two arms of this refusal by the sentence's English prefix; read `code` (`config.unknown_role`, with `detail.reason` where you need the arm) for both, and retire the prefix table. @client-core: a failed terminal with `code: "config.unknown_role"` now also arrives for a task that named no model, with the `coded` account its siblings carry.
304
+ - **Pins.** `test/backlog2833-no-model-for-role.test.ts` (A: a fresh run with no model — `failed` / `config.unknown_role` and the `coded` account on the result, the done event and the resolved result, the seat the requested role's; B: the direct throw carries code and pair and keeps its opening words; C: an unheld word still wins with `unknown_model_ref`, a bound role runs; D: the set and the registry name the word).
305
+
306
+ ### `delegation.transcript_integrity` is announced only where transcript durability is declared — the `full` subagent transcript tier — and is no longer minted under `rows` (#2919; behavior narrowing: one notice stops being minted on one tier; hard-breaking: none; closed sets: none; exports: none) (@server)
307
+
308
+ - **Before.** The notice says a durable agent row binds a transcript session the session store attests gone — "the declared transcript durability is being contradicted" — and it was minted on every deployment that kept agent rows, whatever the deployment declared. Under `rows` (rows and terminal snapshots survive a restart, transcripts follow a process-local store's fate) a gone transcript is the declared fate, not a contradiction: the notice was false there, and a deployment on that tier met it for each agent it read after a restart.
309
+ - **Now.** The three sites that meet a gone transcript — SendMessage's backlog-drain preflight, SendMessage's revival preflight and AgentTranscript's durable leg — hand the announcer the tier their mount derived (`resolveSubagentTranscriptTier`, the derivation the wiring manifest's `fleet.subagentTranscripts` rides), and the announcer asks one predicate, `declaresTranscriptDurability` (`full` only). `full`: unchanged (once per (scope, handle) per process). `rows` / `none`: no notice, and a `rows` call does not spend the once-per-handle latch. The per-call refusals the model reads (`resume.session_not_found`, `not_found`, the `queue_full` drain sentence) are unchanged on every tier. `createAgentTranscriptTool` now derives its tier at mount from its own seats, as the SendMessage mount already did: a `sessions.placements.subagent.durability` that is no `StoreDurability` is refused there (`config.store_durability_invalid`), as the prepare-time manifest refuses it.
310
+ - **Consumers.** @server: expect no `delegation.transcript_integrity` from a deployment whose session store declares `process-local` subagent placement (or declares none); one from a `full` deployment is now a true statement of a contradiction — alert on it.
311
+ - **Pins.** `test/backlog2919-transcript-integrity-tier.test.ts` (A: the predicate over the three tiers, the announcer per tier, the latch not spent under `rows`; B / C / D: each of the three sites at `full` (one notice, the refusal intact) and `rows` (no notice, the same refusal byte for byte); E: the first-party mount through `prepareDelegationSurface` with the real factories — the notice reaches `deps.onNotice` at `full`, neither mounted tool delivers it at `rows`).
312
+
313
+ ### `config.models_swapped` names the catalog generation it announces — `detail.generation`, a digest of the candidate's key material — so one swap announced by N Runners folds to one (#2920; wire +1 detail key, additive; behavior narrowing: none; hard-breaking: none; closed sets: none; exports: none) (@server @client-core)
314
+
315
+ - **Before.** A host that hot-swaps the model catalog on several Runners at once (the server's four boot Runners) announced one event as four lines: each line mints its own `occurrenceId` — the identity of THIS delivery — and the rest of the payload was two counts, so nothing in a line said which lines were the same swap.
316
+ - **Now.** `detail` is `{ models, tiers, generation, occurrenceId }`. `generation` is `cgv1:` + a sha-256 over the catalog's key material: the model names the candidate declares (sorted) and the tier bindings in force over them (each tier word with the catalog word it binds, sorted; a binding that is a `Model` object has no word and stands as `null`) — never the catalog's content. Every Runner swapped to the same model names under the same tier bindings announces the same `generation` (and its own `occurrenceId`, unchanged); a swap that ends on another model name or another tier binding announces another; the order the names were spelled in does not matter. Equal key material is one generation — two swaps that change only what a name points at read alike, because the line is an operator breadcrumb, not a content fingerprint. A swap with `tiers` omitted digests the Runner's current bindings.
317
+ - **Wire +1 (additive).** *What changed:* one detail key on the existing `config.models_swapped` line (an untyped `detail` record: no exported type moves). *Who needs it:* a consumer that deduplicates one swap announced by several Runners — fold by (`code`, `generation`). *Who is hurt:* none — a consumer that ignores unknown detail keys is untouched; one that pinned the line's key set (`models`, `tiers`, `occurrenceId`) must add `generation`. *Compensation:* none needed beyond the key.
318
+ - **Consumers.** @server: S-2497 — the notice allow-list reads `generation` and folds the four boot Runners' lines to one row by (`code`, `generation`); keep the per-delivery `occurrenceId` for exact-duplicate suppression. @client-core: forward `generation` with the line if you project the detail.
319
+ - **Pins.** `test/backlog2920-swap-generation.test.ts` (A: four Runners swapped to one candidate — one generation, four occurrence ids; Runners booted on different catalogs still agree; one Runner swapped twice; B: one model name more / different, a tier binding or tier word moved or dropped, none declared ⇒ another generation; tiers omitted ⇒ the current bindings' generation; spelling order and content do not move it; C: counts and `occurrenceId` unchanged, no catalog content in the line, the face lines carry no `generation`; D: the digest's shape and the object-binding `null`).
320
+
321
+ ### A background agent's task-list row names the session its child runs on — `TaskRegistry.list()` rows gain `sessionId?`, read from one anchor that every successful release of that session ends (#2924 / #2925, design/2861 r1 ①, ruled C-R958 ①; additive wire key; hard-breaking: one engine method's signature, open to veto; closed sets: none; exports: +1 root type `TaskListRow`) (@server @client-core @cli)
322
+
323
+ - **Before.** A list row carried no child session id, so a serving layer that wanted to open a background agent's conversation had no way to get there from the row. The id existed twice inside the engine — on the in-process handle and on the durable record — and they could drift: releasing the child's session cleared only the durable copy, and each place that released a bound child session wrote its own "release, then clear" (four of them; the stillborn background fork's never cleared, and the orphan-placement sweep's release cleared nothing).
324
+ - **Now.** A `background_agent` row carries `sessionId` while its anchor stands: absent in the startup window before the bind (the key is not there at all — not an `undefined` value), present from the bind (fork rows: the branch; other rows: the child's pre-minted session), still present on a terminal row whose session was kept (until its release or the terminal GC), absent once a release of that session succeeded; absent on every other task type (a workflow's originating session is not this key). One reading: `childSessionAnchorOf(handle)`. One law: a successful release of the child's session ends the anchor on the handle and the durable record together, matched on the pair (row, session) — a late clear for a session the row no longer names leaves the current anchor alone. Every release of a bound child session in the delegation lanes goes through one helper handed the release of the store that OWNS the session (a host-forked branch releases through the host store); the orphan-placement sweep's release ends a terminal handle's anchor too.
325
+ - **Also fixed (#2925, existing defect).** A background child whose retention degraded — the pin failed, or the retain ledger was disposed while the pin was in flight — had its pre-minted session released and its anchor cleared, yet still launched on that same id: the running child's durable record carried no transcript anchor (a store that recovered later kept it that way). The lane now binds the launch session again right before the launch (a no-op when the row already names it; `bindBackgroundAgentSession` leaves a row already anchored to that session as it is). And a disposal that dropped the entry but was still releasing it when the pin landed is now waited for: the entry's release runs once, and both degrade arms wait for it — before, the launch could overtake it, so the late release deleted the session the child had just created on that id and ended the running row's anchor.
326
+ - **Hard-breaking (open to veto).** *What changed:* `TaskRegistry.releaseDurableTranscriptAnchor(id)` is now `releaseDurableTranscriptAnchor(id, sessionId)` — the session id is required and the clear happens only while the row still names that session. *Who is hurt:* a caller of the old one-argument form; outside core there is none (`git grep releaseDurableTranscriptAnchor origin/main` in sema-server `772c2943`, sema-cli `7fb306ab`, sema-client-core `45c9279`, sema-sdk `65d8beb`, sema-connectors `fa0cbcb`, sema-desktop `32d66de`: 0 hits). *How to migrate:* pass the id of the session whose release just succeeded.
327
+ - **Consumers.** wire +1, additive and optional: `TaskRegistry.list()` / `liveTasks()` rows `sessionId?` (present-iff above). @server S-2450: `ROW_KEY_DISPOSITION` in `src/http/routes/background.ts` turns compile-red on the new key by design — state `keep` and project `rows[].sessionId?`; the server's `ReturnType<TaskRegistry["list"]>[number]` is unaffected, `TaskListRow` names the same type. @client-core CC-1054: the session row's navigable id. @cli: a row may now carry `sessionId`; nothing to change unless you render it. Only this process's handles are listed — a row known only to another process's durable store is not in `list()` (KL-2924-a).
328
+ - **Deviation candidate.** CC 2.1.309's SDK task rows (`background_tasks_changed`) carry no session id; its `local_agent` task id is the transcript key itself, and its `remote_agent` state carries `sessionId` — the nearest precedent. sema's child transcript key is a separately minted session id, so the row carries it; to be registered in the divergence ledger.
329
+ - **Exports.** +1 root type `TaskListRow` (the named row of `list()` / `liveTasks()`; zero behavior).
330
+ - **Pins.** `test/backlog2861-task-list-session-anchor.test.ts` (I1 source scan: the row mints the key only from `childSessionAnchorOf`; absent before the bind / bound after; store-less registration carries it; a release ends it on row, handle and durable record; a late clear for an earlier session leaves the current one; shell / monitor / workflow rows never carry it; the orphan sweep's release ends a terminal handle's anchor, a failed one keeps it; each bound release site — retained ledger disposal, resolve leg, reject leg, stillborn fork — ends it on success and keeps it on a throwing release; a host-forked branch is released through the host store; a fork row keeps its branch once terminal; #2925: after a failed pin and after a disposal mid-pin the running row names the session the child runs on, a disposal still releasing when the pin lands is waited for (nothing releases the child's session after its launch), and with the store down through the window the recovered store's record names that session; real load path: the root-exported `defaultTaskRegistry` under the session's own `{ owner: sid, sessionId: sid, scope }` lists the child's session and drops it with the release).
331
+ - **Known limits.** KL-2924-a — `list()` lists this process's handles only; a row carrying `sessionId` says the registry has not released the anchor, not that the session store still holds the session.
332
+
333
+ ### An ancestor's extensions judge a delegated child's call at the child's gate, on their own bus, as the child's — `createPreToolUseConstraintPolicy` is removed and a screening entry is `{ installation, hookEnv?, …regime }` (#2429, mods wave three S2b, design/1398 wave 3 §3.3 B, C-R777 / C-R940; HARD-BREAKING, open to veto: a host that rebuilds a screening entry with the factory, or assembles its own screening policy for `runToolGate`, no longer compiles; behavior change V1–V8 below)
334
+
335
+ - **Before.** An ancestor's `classic.PreToolUse` chain rode a child's delegation as a **policy layer** folded after the child's caller policy (`createPreToolUseConstraintPolicy`), while the same installation's `tool.check` was raised at the child's gate: one fact (an ancestor's installation judges a descendant's call), two mechanisms. The policy form carried no identity (`agentId` absent), dropped the screen's `additionalContext`, recorded its deny as `deniedBy: "policy"` with a `classic.PermissionDenied` `source: "policy"`, said nothing when it allowed, and judged only the bytes the child's policy had rewritten.
336
+ - **Now.** Each inherited installation is one judge at the child's gate, on its own bus (`InheritedInstallation`, `inheritedInstallations` — one list read by both judging events): `classic.PreToolUse` at phase 1 right after the child's own bus (and at every re-screen: the fold's second look, the approved-edit pass), `tool.check` at the verdict read as before. V1 the e is the child's (`agentId`, `identity`, `permissionMode`, `cwd`) with the environment the installation was installed against (`env`); V2 its `additionalContext` reaches the child's call result; V3 its deny is the hook's own refusal — `deniedBy: "hook"`, `extension` = the registration, no `classic.PermissionDenied`; V4 its allow is no objection (an installation answers only its own leg's questions) and is disclosed as the `hook` member's shadow on the child's card (`allowRuleShadowed`); V5 it judges the bytes it is given, before the child's policy, and the policy's rewrite of them once more (a deny or another form there refuses); V6 its ask is settled where it was frozen (the installation's approver, durable mandates, classifier) at phase 1 — refused ⇒ the call is refused, approved ⇒ that question is answered, approved with an edit ⇒ the gate's approved-edit pass takes the edit, floated ⇒ the call stays marked for the park; V7 the synthetic read / consolidation-write judges no longer see the screen as an inherited policy (it refused every probe as "nobody can be asked"); like the child's own PreToolUse, it is not consulted for synthetic probes. When the installations rewrite one after another, each earlier judge of the call sees the new bytes once; a refusal is recorded over the bytes its judge judged. V8 the screen's waits — every raise, the leg's own included, and every installation's settlement — end with the run or with the call (the per-call signal, as the gate's other synchronous waits): a turn that cancels a call no longer leaves a PreToolUse hook pending on the run's lifetime (the call is refused, fail-closed, "cancelled").
337
+ - **Hard-breaking (open to veto) — H2.** *What changed:* the inherited screen is no longer a policy layer: `createPreToolUseConstraintPolicy` is removed from the root, and `InheritedGate.parentConstraints[]` is `{ policy, …regime }` (a caller policy) or `{ installation, hookEnv?, …regime }` (a screening installation: the bus `admitModChain` returned, or a leg's own); an entry carrying both, neither, or an installation no engine admission mounted is refused (`config.mods_invalid` at prepare, `resume.parent_constraint_mismatch` before a resume's claim); the entry member `mods` is retired. *Who is hurt:* a host that rebuilds a screening entry across processes with the factory (@server `createParkedReviveInheritedGate`, if it rebuilds one), or that assembles a screening policy of its own for `runToolGate` — a compile error. *How to migrate:* rebuild with `{ installation: await admitModChain({ deployment, task }), hookEnv: <the env>, …the frozen regime }`; the checkpoint's constraint digest and the park format are unchanged (the entry still projects opaque).
338
+ - **Behavior change (open to veto) — H3.** *What changed:* an ancestor's screen judges a descendant's call in the gate's phase 1 (V1–V6). *Who is hurt:* a host reading the child's `classic.PermissionDenied` `source: "policy"` as "the ancestor's screen refused" (that observation is gone; the call's `tool_end.gate.disposition` reads `deniedBy: "hook"` with `extension`), and a screen written for post-policy bytes only (it now sees the original bytes first, then the rewrite). *How to migrate:* read `tool_end.gate.disposition.deniedBy` / `.extension`; make the screen idempotent over two judgments of one call.
339
+ - **Consumers.** @server: a parked revive that rebuilds an inherited chain rebuilds a screening entry as above (S2a found no factory call in `src`); trace projection of an ancestor's screen deny moves `policy` → `hook`. @cli: a child's card / refusal text for an ancestor's screen deny is the hook's sentence (via server). @client-core: no closed-set change; `ToolGateInput.inheritedToolChecks` is renamed `inheritedInstallations` (a host driving `runToolGate` directly). @test: the cells below are the black-box criteria.
340
+ - **Exports.** −1 root value `createPreToolUseConstraintPolicy`. `ToolGateInput.inheritedToolChecks` → `inheritedInstallations`; `ConstraintChainLayerView.policy` optional.
341
+ - **Pins.** `test/backlog2429-mods-w3-s2b.test.ts` (real Runner, the root delegating through a spawn tool: V1 the child's `agentId` / identity on the ancestor's e; V2 its context in the child's result; V3 `deniedBy: "hook"` + `extension`, no PermissionDenied on the child; V4 the child's own question stands, the card's hook shadow names the ancestor's registration; V5 original then rewritten bytes, a refusal of the rewrite refuses; V6 refused / approved at the frozen approver with the child's approver never asked, an approved edit judged and run; K2′ two judgments both the child's; stand-in one judgment; C1 a cancelled call ends a pending inherited screen; C2–C4 a refusal in phase 1, in the fold's second look and in the approved-edit round is recorded over the bytes its judge judged; S the entry fault and a leg refusing a malformed chain). The inherited-gate / mods families are re-pinned, each change marked intended (V1–V8, H2) in place; `gate-station-boundary-ticks` "screening → fold" counts 11 microtask ticks, was 9 — two awaits inside the screening face (one look per seat), none between station groups.
342
+ - **Known limits.** The ancestor's approver is asked at phase 1, before the child's caller policy runs — a call the child's policy then refuses has already put one card to the ancestor's person (before, the policy's refusal came first), and a policy rewrite of the call puts the installation's question again over the rewritten bytes. An ancestor's handlers now see the model's bytes before the child's policy rewrites them (as the child's own handlers always have): a child policy that redacts an argument no longer hides it from them. The resumed leg's re-screen of an approved call raises the leg's own bus only, as before.
343
+
344
+ ### A background task's output or full result has a reader that is not TaskOutput — a cut agent report and a workflow's full record are stored and named by a `ReadToolResult` ref, a Monitor watch writes its stdout and stderr to an output file, and every list of tasks names each task's read handle (#1608 R2 / R3 / R4 / R7 with the #469 placement of the new reader, design/1056 §2.5, ruled C-R675 / C-R976; behavior change: notification, receipt and table text; the TaskOutput tool itself is still mounted — its removal is a later step; hard-breaking: none; closed sets: none; exports: none; members: +1 optional `ToolExecuteContext.taskArtifacts`, +1 optional `SendMessageToolOptions.taskArtifacts`, +1 optional `RunWorkflowToolDeps.taskArtifacts`, +2 `TaskRegistry` methods `recordArtifactRef` / `readHandleOf`, `formatWorkflowRun` takes an optional bounds argument) (@server @cli @client-core @test)
345
+
346
+ - **Before.** A background agent's completion frame carried 2 000 characters and said "read it with TaskOutput where mounted"; the overflow reached a store only when a TaskOutput poll happened, and a report between 2 000 and 4 000 characters was never stored. A workflow's per-agent rows were only in TaskOutput (and capped at 60 rows / 5 partial outputs there). A Monitor watch had no output file — its stderr was read only through TaskOutput's spool. The post-compaction table said "check output with TaskOutput", and a task whose every pending notification was evicted was told TaskOutput "may still serve its spool".
347
+ - **Now.** A report longer than the frame carries is stored whole (redacted, content-addressed: task id + a digest of the text) in the run's offload store BEFORE the frame is announced, and the cut line names it: `read the full text with ReadToolResult (ref "…")`; with no store, or a write that failed, the cut names no ref and says so (and keeps the SendMessage / TaskOutput readings while TaskOutput is mounted). A workflow completion stores its full record — the result and every agent's row and finished output, with no display bound — the same way and names it in `diagnostics`, lists the agents that did not complete inline (up to five), and names the journal apart, with its own use (the step log `resumeFromRunId` replays); a rejection whose thrown value carries no run of its own reads the run store's record, and building the record can never cost the failure frame. A Monitor watch mirrors stdout and stderr into an output file minted by the background Bash's one mint; the start receipt and the terminal frame name it, its last writes land after TaskStop's kill and before the terminal frame, and Read's root-fence exemption covers it as it covers a background command's. Each row of the post-compaction table names the task's read handle (`output: Read <path>` / `result: ReadToolResult (ref "…")`; a revived agent drops its previous cycle's), and a dropped task's disclosure names its read handle or says its row is no longer readable. TaskStop's "already finished" receipt says the completion notification carries the final state. A page-back of a task's stored result is placed as a poll of that task is (`ReadToolResult`'s boundary legs are per call now: another tree's agent, a workflow run or a row gone read as an off-target read; the calling tree's own agent and any other ref travel none) — the new reader is no way around the cross-boundary table.
348
+ - **Behavior change (open to veto) — a retired name on an allow list.** *What changed:* a retired tool name on a `toolPolicy` allow list (the caller's policy or a session's stored `toolAllow`) no longer refuses the run with `config.legacy_tool_name`; it is reported by the `config.toolpolicy.unmatched_names` advisory with the name's retirement line (it grants nothing, so the policy is narrower than written). A retired name on a deny or ask list still refuses. *Who is hurt:* nobody loses a guard — a host that relied on the refusal to find a stale allow entry now finds it in the advisory. *How to migrate:* nothing; rewrite the entry to the current name when the advisory names it.
349
+ - **Consumers.** @server: the workflow completion notifier's `diagnostics` now carries the record's ref sentence, but the completion-inbox journal drops `diagnostics` (`workflow-notify-journal.ts`, on the premise "TaskOutput reads it") — that premise ends with the TaskOutput removal, so the bridge should carry `diagnostics` (or the ref) before then. @cli / @client-core: completion text changes only; no wire key changes; the wiring manifest's roster entry for `ReadToolResult` reads `boundaryLegs: "per_call"` (was `[]`). @test: the lanes below are the black-box shapes.
350
+ - **Pins.** `test/backlog1608-agent-result-artifact.test.ts` P1–P6 (a 3 000-character report stored whole; content-addressed refs; the frame after a slow put; failed and absent stores; the real Runner lane reading the ref with TaskOutput excluded); `test/run-workflow-tool.test.ts` "#1608 R4" R4a–R4f (rows inline, the record past the display bounds, a bare and a forged throw, the frame after a slow put); `test/backlog1608-monitor-output-file.test.ts` P1–P6 (stdout + stderr in the file, the exemption's access and env tests, no-file and failed-append disclosure, a stalled append never holding the kill, an in-flight append landing before the terminal frame, the real Runner lane reading the watch's stderr with TaskOutput excluded); `test/backlog1608-task-read-handles.test.ts` T0–T5 (T5: the page-back's placement against a poll's); `test/design469-r3-boundary-legs.test.ts` (the catalog row); `test/backlog1355b-unmintable-segment-judge.test.ts` "#1608 M9" M9a–M9c.
351
+ - **Behavior narrowing (open to veto) — a stored result's origin is in its ref (#3092).** *What changed:* the task whose result an artifact holds is a segment of the artifact's ref (`tr_<session>~fold.task~<task>~<digest>`, a segment the tool-output mint never produces), and the process-local ref → task table that remembered the newest 4 096 refs is gone; a page-back of a stored result whose task row is gone, belongs to another tree, or was stored by another process now reads as an off-target read — before, a ref that table had forgotten or never saw travelled no legs, so under an armed, governed tree the read went unrecorded and the next device call did not ask. *Who is hurt:* a governed run that pages back a stored task result after its row is gone, or after a restart, now owes the record and the confirmation an off-target read owes; offloaded tool output is unchanged (it never carries the segment). *How to migrate:* nothing; a ref stays an opaque handle — read it back verbatim, never parse it (the segment is the engine's own reading, not a format promise). Pins: `test/backlog3092-artifact-origin-in-ref.test.ts` A0–A6. The edge table that mapped a ref to its task only inside one process is gone with it, so a ref read in another process or after eviction is accounted against its task like any other (#3027).
352
+ - **Known limits.** A running background agent's progress is not readable before it ends (its completion frame is the reader). A missing or incomplete command output file is the next entry's (#3025). TaskStop on a watch waits for the file's last append after the kill (the bash lane's order) — an append that never settles holds the stop's receipt. A background agent's output spill (the overflow a TaskOutput poll stores under its own ref) is not a result artifact, and its page-back still travels no legs. The one-shot wait (R5) and the completion-sink resolution (R1) are the next entry's (#3025); the TaskOutput removal is a later step.
353
+
354
+ ### A one-shot submission waits for the background agents and workflow runs it started before it ends; a background command's missing or incomplete output file names a ref for what was retained; a whole-session notification loss carries the background-task table (#3025 = #1608 C1: R5 / R1 / R6 / R7 and the R2 gap arm of design/1056 §2.5.3; the CC 309 print-mode split, `cli309.js` :289218-289219; behavior change: the one-shot ending and receipt / notification text — open to veto; hard-breaking: four option members removed; closed sets: none; exports: none; members: −4 (below), +1 optional `taskArtifacts` on the background-command and Monitor registration inputs) (@server @cli @client-core)
355
+
356
+ - **Before.** A one-shot submission (`TaskSpec.oneShot`, e.g. headless `-p`) ended as soon as the model's turn did. Every loop-type launch receipt (background `Agent`, a forked background agent, `SendMessage` continuing an agent, `Workflow`, a still-running agent's poll) therefore told the model to block on `TaskOutput({ task_id, block: true })` itself, and a model that ended its turn lost the result; a background command's receipt said the same. Each background producer read its own completion sink (the call's run lane for an agent, the mount's sink for a command, a watch and a workflow). A command whose output file could not be created, or whose append failed, kept its retained output reachable only through TaskOutput. A whole-session notification loss said only that the lost tasks were no longer known. A read handle of `""` was named as one.
357
+ - **Now.** A loop-type background task (a background agent, a continued agent, a workflow run) admitted onto a run's lane owes that lane its terminal frame; where a one-shot run would end, it waits for the next such frame — already on the steer lane when the wait wakes — and hands it a turn. A cancel ends the wait (the run's walltime timer included: the answer the run gave stands and `task.interrupt_unconsumed` reports the stop), and so does a 600 s ceiling counted from the first wait — the reference's print-mode ceiling (`cli309.js` `mFt=600000`), a constant, not a knob. A background command or a Monitor watch is never waited for — its lifetime is its process's — and a one-shot command receipt says: "This is a one-shot submission — the engine does not wait for background commands; run it in the foreground (with a timeout) if you need its output before the submission ends." The loop-type receipts carry their ordinary notified wording in a one-shot run; none teaches a blocking wait. Every background producer resolves its completion sink through one function (the call's run lane, else the mount's sink). A command whose output file is missing or incomplete stores what the engine retained (the spool's head and tail, its gaps stated) and its terminal frame names the `ReadToolResult` ref — on every terminal face, the Monitor run-teardown face included. A whole-session loss line carries the run's background-task table — live and finished rows, each with its read handle (the post-compaction announce's renderer) — or says the run has none to read; a blank output path or ref names nothing. The wait owes per cycle (an agent continued while its earlier frame is on its way owes two; a parked frame a continuation voids releases its cycle), and any input reaching the run while it waits (a steer, a child's message) wakes it to take that turn first.
358
+ - **Hard-breaking (four option members).** *What changed:* `SendMessageToolOptions.oneShot`, `SendMessageToolOptions.retrievalToolMounted`, `TaskToolOptions.oneShot` (`createTaskOutputTool`) and `TaskPollOptions.oneShot` (`TaskRegistry.pollTask`) are removed — nothing reads them any more (a one-shot submission is no completion mode of its own; the retrieval-tool seat steered only the one-shot advice). *Who is affected:* a library caller that passes one of them in an object literal gets a TypeScript excess-property error; server, cli and client-core (origin/main) pass none of them. *What to do:* delete the argument — the behavior it chose is now the engine's own (the one-shot ending waits).
359
+ - **Behavior change (open to veto).** *What changed:* a one-shot run no longer ends while a background agent or workflow run it started is still owed to it — it waits for the completion (bounded by its own walltime, a cancel and the 600 s ceiling); a host that mounts Bash, Monitor or Workflow inside a run with its own notification sink now has completions delivered to the run's lane the call carries (the Agent lane's precedence). *Who is affected:* a `-p` caller whose submission used to return right after a background launch now returns after the background work finishes; a host relying on its own sink receiving a run-launched command's completion. *What to do:* bound a one-shot submission with `limits.maxWalltimeMs` if it must return by a deadline; a host that needs its own copy of completions observes them through `RunInternals.onTaskNotification` (the lane calls it first) rather than a per-tool sink.
360
+ - **Consumers.** @server: the one-shot contract text the SDK carries for `oneShot` ("guidance only … block-wait via TaskOutput") no longer describes the engine — the engine now waits for loop-type background completions before the submission ends; a request-level timeout on `-p` submissions meets the background work's duration. @client-core: `src/oneShotWireCaps.ts`'s header ("core 换发 block-wait 指引", "谁受害:无 —— 键只改模型指引文案") — the key now changes when a submission ends. @cli: a local port of the one-shot receipt wording, if kept, should drop the blocking-wait sentence.
361
+ - **Pins.** `test/backlog3025-oneshot-ending.test.ts` O1–O11 (real Runner: a child released from the parent's own Stop station — the one-shot ending waits for its completion and answers from it, an interactive run does not wait, a completion that never comes runs to the walltime, a background `sleep 60` is not waited for, a held workflow run is waited for, a steer during the wait gets its turn first; the obligation book: per-cycle debts, a voided cycle, its bound and the 600 s ceiling); `test/backlog3025-completion-sink.test.ts` S0–S3 (the one resolution; a Bash and a Monitor started by a call carrying the run lane complete into it, not into their mount's sink); `test/backlog3025-command-record-gap.test.ts` C1–C5; `test/backlog3025-session-loss-table.test.ts` E0–E5 (E4 / E5 real Runner: a running command's file, a finished agent's stored result); re-pinned: `run-workflow-mount` (the real mount now waits for the workflow's completion), `run-workflow-tool`, `rb288-295-agents-face`, `task-registry`, `bg-lifecycle-fidelity`, `hands-fs`, `design385-rescan-fixes` G, `up424-a-tier-ports`, `prepare-delegation-surface`.
362
+ - **Known limits.** A wait the walltime or the ceiling ends does not name the task ids still owed (`KL-3025-a`). TaskStop on a Monitor watch waits for the output file's last append after the kill with no bound (`KL-3025-b`). A loop-type task launched by an EARLIER run of the session completes into that run's lane — a one-shot run does not wait for it.
363
+ - **Fix (before release).** The obligation book names a task by its lane key (`taskNotificationLaneKey`, the pending store's own bucket key) at admission, discharge and release, so a frame of another lane that carries a background agent's `task_id` settles nothing; an operator resume of a retained child (`SubagentSteerHandle.resume`) completes into the mount's own sink, as it did before this entry — no call of a run launches it, so it has no run lane to resolve (design/122) — while a resume through `SendMessage` completes into the waking run's lane. No wire change. Pins: `defectscan-notify-budget-surface` F3 (the book is a per-task bucket, and a cross-lane cell), `rb201-handback-review` FO-3 retained-resume ×2, `subagent-resume` ⑦ (unchanged, green again).
364
+
365
+ ### `memorySyncStoreOf(backend)` is a root export — the one mint of a sync cursor's store identity, for a host binding a cursor taken before the binding existed (#2956)
366
+
367
+ - **Before.** The identity a sync cursor binds to (`file:<root of the store's directory>`) was minted only by `memorySyncStoreOf` in the memory-engine module, reachable by a deep path; a host that carries cursors persisted before `MemorySyncCursor.store` existed had to import that path or copy the formula to bind them.
368
+ - **Now.** The function is exported from the package root — the leaf's own function (one mint, pinned `===`), unchanged in behavior: a store without a directory still has no identity and is refused as a sync plane (`config.memory_sync_store`).
369
+ - **Exports.** +1 root value `memorySyncStoreOf` (a function; the consumer smoke list records it). @server S-2537: bind a legacy cursor before the round with `{ ...legacy, store: memorySyncStoreOf(backend) }` — `backend` is the store that layer's sync actually ran against when the cursor was written (the old executor's single `opts.backend`), not the store the layer resolves to today (S-2394 / #2818 move the global layer): binding to today's store reopens the #2836 deletion shape (C-R991 D-4 reading).
370
+ - **Pins.** `test/backlog2956-memory-sync-store-root-export.test.ts` — the root, the memory-engine barrel and the leaf hold one function; the smoke list decides it; the identity of a real File store and of a second instance over it; a legacy cursor is refused unbound and accepted bound through the root's function; a directory-less store refuses by code.
371
+
372
+ ### Two store options are read by one integer reading — `TtlSessionStoreOptions.sweepIntervalMs` is a whole number of milliseconds from 1 to 2147483647 and `InMemoryToolResultStore.maxTotalChars` a whole number of characters from 0 (#2959)
373
+
374
+ - **Before.** `sweepIntervalMs` was judged finite-and-positive only, so `1.5` (Node truncates it to 1 ms) and any value past 2147483647 (Node reads it as 1 ms) armed a background sweep that ran back to back, silently; `0` and a negative were read as "no sweep". `maxTotalChars` was judged "a number, not NaN" only, so `-1`, `1.5` and `Infinity` were taken as bounds (`Infinity` read as "no bound").
375
+ - **Now.** One reader (`readIntegerOption`, `src/core/integer-option.ts`) judges both: absent is the option's default (no sweep; an unbounded store), a whole number inside the range is itself, anything else refuses at construction with `config.invalid_argument`, naming the option and spelling a number (any other value is named by its type only, never repeated). The bound is read once at construction. `0` for `maxTotalChars` is still the tightest bound (everything but the just-written ref is evicted).
376
+ - **Behavior narrowing (open to veto).** *What changed*: a sweep interval of `0`, a negative, a fraction, a non-number or past 2147483647, and a result-store bound that is negative, fractional, infinite, past 2^53-1 or a non-number, now throw where they were accepted. *Who is hurt*: a deployment that passed `sweepIntervalMs: 0` to mean "off", `maxTotalChars: -1` to mean "tightest" or `maxTotalChars: Infinity` to mean "unbounded" (none in sema-server, whose result store passes 64000000 and whose own caching store has its own option; sema-cli's `sessionManager` has its own too — local checkouts read, inferred for the rest). *Compensation*: omit the option for "off" / "unbounded", and write `0` for the tightest bound; the refusal sentence says which.
377
+ - **Pins.** `test/backlog2959-integer-option-gates.test.ts` — eight bad sweep values and nine bad bound values, each refused with the code and with no timer armed (fake timers); the bundled and the Runner-shared store alike; the controls that stay green (1, 1000, 2147483647 arm exactly that delay; `0`, a real cap and the largest exact integer hold; omitted is unbounded); the bound is read once; the sentence spells numbers and never repeats a secret-shaped string; both constructors read through the one reader.
378
+
379
+ ### A peer seat's answer is held to the grammar of the send that carries it, read from its own members only, with `expiresAt` from 0 (#2961)
380
+
381
+ - **Before.** The engine's judge carried a header value the send then refused — U+2028, a character past U+00FF (Node's HTTP layer throws on it when the request is built), a control byte inside Latin-1, DEL — so "can carry" and "can send" disagreed; an `expiresAt` below zero was a finite number and was carried; an env value holding a lone surrogate was carried and the child read U+FFFD in its place; and the answer's members (`kind`, `headers`, `env`, `expiresAt`, `authorizeUrl`) were read through the prototype chain, so a member the answer's prototype (or a polluted `Object.prototype`) supplied was taken as the seat's own.
382
+ - **Now.** A header value is held to Node's own HTTP header-value grammar (tab, printable ASCII, U+0080–U+00FF; the empty value and U+0085 are carried as they are on the wire); an env value is held to what a spawn hands the child unchanged (no NUL, no line break, no lone surrogate); `expiresAt` is a finite number from 0 (a stamp past every clock, `Number.MAX_VALUE` included, still never expires, as an absent one); the answer's members are its own enumerable members, read once into an inert snapshot. A refusal is `config.peer_answer_invalid`, names the member and never repeats the value.
383
+ - **Behavior narrowing (open to veto).** *What changed*: an answer holding a header value outside the grammar, an env value holding a lone surrogate, a negative `expiresAt`, or a member supplied only by a prototype now refuses the send (`credential_unavailable`, delivered `no`, nothing sent) at the ask instead of failing at the send or being read as the seat's. *Who is hurt*: a seat whose vault holds a credential with a character past U+00FF (it already failed when the request was built), a seat that built its answer as `Object.create(base)` with members on `base`, and a seat that answered a negative `expiresAt`; sema-server's seat (`src/peer-seat.ts`) answers plain objects with own members (local checkout read). *Compensation*: put the members on the answer itself; encode a credential into the header grammar (ASCII, base64); omit `expiresAt` or answer 0 or a real stamp.
384
+ - **Pins.** `test/backlog2961-peer-answer-value-grammar.test.ts` — the judge against Node's `validateHeaderValue` for every character up to U+02FF and for the shapes beyond it on both header lanes (http, a2a); the exact boundary; prototype-supplied answer / headers / `expiresAt` / `env` / `authorizeUrl` members and a polluted `Object.prototype` not read; `expiresAt` from 0; the env lane (a lone surrogate, NUL, line breaks refused; every carried value reaches a real child byte for byte); and through the real MCP http entrance a value the send cannot carry is `credential_unavailable` with no call sent while an upper-Latin-1 value reaches the endpoint.
385
+
386
+ ### A background agent's session anchor changes by release and a fresh bind, never by being written over; an empty session id is no anchor (#2963)
387
+
388
+ - **Before.** `bindBackgroundAgentSession(id, sessionId)` on a row already anchored to another session overwrote the anchor without a release (the earlier session lost its owner row, nothing said so), and an empty session id was taken as an anchor (`sessionId: ""` on the row and its durable record).
389
+ - **Now.** A bind of another id onto an anchored row throws (the row keeps its anchor on the row, the handle and the durable record, and nothing is written); an id that is no non-empty string throws whatever the row holds; the same id again is still the no-op it was; a bind for a row that is gone or is not a background agent's is still silent. The sanctioned change is `releaseDurableTranscriptAnchor(id, sessionId)` then a bind. CC 2.1.309's `remote_agent` rows are built with their session id once (and rebuilt from the sidecar on restore); it has no re-anchor path, so there is nothing to follow.
390
+ - **Behavior narrowing (open to veto).** *What changed*: the two binds above throw where they were accepted. *Who is hurt*: a caller that re-bound a row to a different session without releasing the first (the engine's three bind sites bind the same id again or an unanchored row — none re-binds), and a caller that passed `""`. *Compensation*: release the first anchor, then bind; pass the real session id. Two existing pins are re-pinned to the new law (the late-clear pair-match pin and the tier-3 late old-lane write pin): the first now releases before the second bind, the second enqueues its late write as the release of the anchor.
391
+ - **Pins.** `test/backlog2963-bind-session-anchor-law.test.ts` — the second bind throws naming the row and no session id, with row / handle / durable record on the first and no durable write; repeated refusals; the same id is a no-op with no write; release then bind works; a store-less registration holds the law; five bad ids on a fresh and on an anchored row; a missing row and a workflow row stay silent.
392
+
393
+ ### A woken retained child is a running agent in its tree's delegation pool — a resume cycle is counted for exactly its own life, so the concurrency cap sees every child that runs (#2998, ruled C-R982 (a); security narrowing — runaway fan-out, open to veto; hard-breaking: none; closed sets: none; exports: none)
394
+
395
+ - **Before.** The delegation concurrency cap counted a running child by its registry row or, on a lane with no row, by the pool's local in-flight count — and only the spawn doors charged that count. A synchronous child retained for its steer handle and woken with `handle.resume(…)` ran with no row and no charge: invisible to the cap, so a full pool still admitted a fresh spawn (cap 1: two running). A child woken by SendMessage revived its row and was counted.
396
+ - **Now.** The gate's one reservation, `occupy(kind)`, is split from its verdict: `admit` is the two cap verdicts in front of `occupy`, and a resume cycle occupies without the verdicts — a cycle that revived its row is counted by the row, a cycle with no row is charged to the pool's local half (the handle's own spawn gate), given back where the retained entry's running claim is (the settle, and a launch that throws). A launched wake always reaches its settle: cancelling the previous cycle's parked completion frame is contained and disclosed on the resume's notifier (site `subagent.resume.park_cancel`), where a throwing registry unsubscribe used to leave the cycle undrained and its entry running for good. A woken child is still never refused by a full pool: CC 2.1.309's resume path takes a concurrency slot without reading the cap.
397
+ - **Behavior narrowing (open to veto).** *What changed*: while a handle-woken child runs, the tree's next Agent spawn is refused at a full pool (`delegation.concurrency_cap`), where it was admitted. *Who is hurt*: a deployment that wakes synchronous children through steer handles and spawns beside them at the cap (no sema-server or sema-cli reader of the in-flight count; the count is engine-internal). *Compensation*: none needed beyond waiting for a woken child to settle; the cap is the deployment's `RunnerDeps.delegationEntryCaps`.
398
+ - **Not done, put to clay.** Refusing a wake at a full pool (C-R982 (b)) diverges from CC 2.1.309 and is not done. A tier-3 revival of the same row is refused at a full pool today while a retained wake is not; which way the two are made one is the same question.
399
+ - **Consumers.** none to change (@server @cli: no reader of the in-flight count).
400
+ - **Pins.** `test/backlog2998-resume-occupies-pool.test.ts` (real Runner: the SendMessage wake at a full pool resumes and is counted — 3 and 5 running read 3 and 5; the handle wake at a full pool runs and the gate reads 2 for 2 running; a handle wake on an empty cap-1 pool refuses the next spawn and its settle gives the slot back; a launch that throws gives it back and the child stays resumable; a parked frame whose cancel throws after the launch still lets the cycle settle and give the slot back, the throw disclosed; the occupancy kind follows the row and is given back once).
401
+
402
+ ### A caller's write-capable tool that declares no permission seat is named at load under every asking mode, whatever policy layers are wired — the auto classifier never rules on it unless a rule or a policy asks (#2966; disclosure only, the routing is unchanged — C-R218 ②, registered divergence `caller-tools-self-seat`; hard-breaking: none; closed sets: none; exports: none) (@server @test)
403
+
404
+ - **Before.** The load-time ungated-write warning (`RunnerDeps.onError`, phase `config`) named a caller's tool (`TaskSpec.tools`) only when it declared `effect: "write"` and the leg wired no policy layer and no PreToolUse judge. A tool with no `permissionSeat` answers for itself (C-R218 ②): under `default`, `auto` and every other asking mode the mode's question never reaches its calls, so in `auto` the classifier never saw it — and once any policy layer was wired the operator was told nothing; a caller's tool answering to `Bash` / `Monitor` without a declared write was never named. A caller write tool the mode's question does reach was named UNGATED when no layer was wired — one declaring `permissionSeat: "passthrough"` (its `unstated` question; the classifier, in `auto`), one under `plan` (plan's own question), one declaring a write path target (the write gate) — and a request tool the run excluded was named too, spending the deployment's one warning for that shape.
405
+ - **Now.** One reading for a request tool (`spec.tools`, mounted under its own name — an excluded or displaced entry is not named): it is a surface when it is write-capable (declares `effect: "write"`, or answers under its name or an alias to a shell-face name) and carries no egress / irreversibility mark. A caller's own tool (one the engine does not define) under a mode that asks is named when no question of the mode reaches it (`modeQuestionReaches`, `src/core/mode-policy.ts`: the unstated question reaches a tool declaring the `passthrough` seat, the write gate a declared path write and, under `plan`, every tool that is not read-only), whatever policy layers are wired (a layer that does not name it covers nothing, the reading the server-tool arm already gives); every other request tool — and every one under `bypassPermissions` — is named when no effect-aware layer is wired, as before. The component reads `write tools under <mode> [...]`, and the warning adds that such a tool answers for itself, that in `auto` the classifier rules only on what a rule or a policy asks, and that `permissionSeat: "passthrough"` puts each call to the mode's question. When an effect-aware layer is wired, the warning's lead says no question of the mode reaches the named calls instead of telling the operator to wire a policy. Once per deployment and tool-set shape, at prepare; never per call.
406
+ - **Not changed.** The auto classifier answers every ask that survives the fold, as before; a seatless caller tool still runs unasked by the mode (the reference mounts a host's tool as an in-process MCP tool whose `passthrough` permission check asks and, in auto mode, is classified — sema keeps the self-answering seat, C-R218 ②). The engine's removal guard still asks a person for a protected removal on a call that answers to a shell-face name, a caller's tool included (`dangerous_delete`, every mode).
407
+ - **Consumers.** @server: the warning text gains the caller-tool component and sentence; a log reader keyed on `UNGATED` keeps matching. @test: black box — a seatless `effect: "write"` caller tool under `auto` with a policy layer runs, the classifier is called 0 times, one warning names it; the same tool declaring `permissionSeat: "passthrough"` is classified and not named.
408
+ - **Pins.** `test/backlog2966-caller-tools-unasked.test.ts` ①–⑨ (real Runner; base readings in the file head); `test/major1-ungated-write-tools.test.ts` unchanged.
409
+
410
+ ### The auto strip's protocol-namespace arm is deleted: a covering allow entry is judged like a path / name / host entry (#2967; rule −1, no behavior change on any call a prepared leg mints; hard-breaking: none; closed sets: none; exports: none) (@test)
411
+
412
+ - **Before.** `autoStripped` stood a covering entry (`mcp__srv__*`, `mcp__srv`) down in `auto` when the call answered to a shell-face name — an arm no call reaches: a covering allow entry reaches only names in a protocol namespace, a caller tool's name or alias carrying `__` is refused at prepare (`config.tool_name_invalid`), and a server's or a peer's tool is minted without aliases.
413
+ - **Now.** A covering entry stands down on a shell call only under `TaskSpec.classifyAllShell`, as every non-shell kind does. 7.59.0's criterion G9.d carries a correction: its last clause (the protocol-namespace entry reaching a shell alias) is retired, not observed — the package face cannot build it.
414
+ - **Pins.** `test/backlog2558-classify-all-shell.test.ts` A1 re-pinned (A1a the door refuses a caller name / alias in a protocol namespace, A1b a materialized server tool has no alias, A1c the covering entry's strip reads only the declaration).
415
+
416
+ ### An agent definition declares the MCP servers its leg connects to — `AgentDefinition.mcp`, in the `TaskSpec.mcp` form, judged by the door that judges a task's own list (#2595; additive; behavior: two choices stricter than the reference, open to veto; hard-breaking: none; closed sets: none; exports: none; members: +1 optional `AgentDefinition.mcp?`, +1 optional `servers?` on `ToolSpec.agentToolFaces` entries; internal, not index-exported: `snapshotAndValidateMcpEntries` (locked-config), `MCP_RESOURCE_FACES` (mounted-mcp), `withDefinitionServers` / `definitionAllowTools` (inherited-tool-pool), and `delegationCallIsExternal` takes an optional `inheritedServers`) (@server @cli @client-core)
417
+
418
+ - **Before.** A definition had no MCP seat: the leg spawned as an agent held only the servers its spawning run declared and its
419
+ lists admitted, so a host holding an agent's own servers (an agent file's `mcpServers`, a request's `agents[].mcpServers`)
420
+ could only refuse them, and an agent restricted to a few tools lost the servers it was meant to bring.
421
+ - **Now.** `AgentDefinition.mcp?: McpServerSpec[]` — the entries a task's `mcp` takes, copied and judged where the definition
422
+ is received (the delegation tool's assembly, the `TaskSpec.agents` door) by the same judge, code and sentence as the task's
423
+ own list (the list-level sentence names `AgentDefinition.mcp of "<agent>"`), and judged again at the leg's prepare. They are
424
+ declared on the leg spawned as the agent alone — its children inherit them as they inherit any spawner's servers, the
425
+ spawning run never mounts them — after the servers the leg inherits: a definition server whose `mcp__<server>__` namespace an
426
+ inherited one already holds is not declared (the spawning run's server holds the name); two of the definition's own under one
427
+ namespace refuse the leg's mount (`config.mcp_server_name_collision`), as a task's own pair does. The definition's own allow
428
+ list, when it has one, admits every tool of each of its own servers; its deny list still wins. Dialling, the shared connection
429
+ pool of the run tree, a host-transport server's seat (the spawning run's seat of that name, read off the same mount as the
430
+ pool, as a wake hands one; no seat ⇒ `config.mcp_host_seat_missing` on the leg, as for a task's own), `skipped`, the
431
+ `mcp_unavailable` notice on the leg's first request (#1551) and refresh / re-dial are the leg's as for any server it declares. The workflow lane's
432
+ `agent(…, {agentType})` judges and folds the definition's servers by the same rules. The memory content-origin judgment
433
+ counts a delegation to the agent as reaching, through the child's own valve readings over the face's lists, the namespace of
434
+ each of its own servers (classified by the server's declared `contentOrigin`; under a namespace the child also inherits a
435
+ server for, that server too, by its own class) and the cross-server resource readers those lists leave (external, as their
436
+ mint marks them). The judgment before the run reads `mcp__<server>…` entries naming a definition's own server as naming a
437
+ declared server, on the definition's lists and on the spawning run's own (a ceiling that admits a definition's server).
438
+ - **Stricter than the reference (open to veto).** *Ceiling*: the layers above the agent's leg (the spawning run's tool lists
439
+ and the ones it inherits) still bound its own servers — a server they leave the leg's valve nothing of is never dialled (the
440
+ reference appends an agent's server tools past its parent's tool pool). *Lock / compliance*: under an `mcp` configuration
441
+ lock or a compliance deny of MCP servers the leg's door refuses the spawn (`config.locked_key` / the compliance refusal; the
442
+ reference skips the servers, notifies and runs the agent). *Bad entries* are refused (the reference drops them with a debug
443
+ warning). *Name references* (the reference's string form) have no form: a definition that should hold a server its spawning
444
+ run declares names it in `allowTools` (`mcp__<server>`).
445
+ - **Today's form, not the contract (KL-2595-a; temporary, retires with #3007).** A spawning run that declares no MCP server
446
+ of its own has no mount: the host seats it holds reach no leg (a definition's own host-transport server fails its dial,
447
+ `config.mcp_host_seat_missing`), and two legs spawned as the same agent in one turn dial a connection each instead of sharing
448
+ one. With any server declared by the spawning run, both hold as above.
449
+ - **Today's form, not the contract (KL-2595-b; temporary, retires with #3017).** A per-task definition's MCP credential
450
+ references (`TaskSpec.agents[i].mcp`) are held by no one between the return of `runTaskStream` and the start of the leg
451
+ spawned as that agent: a host that releases its hold in that window makes the leg's dial fail loud with
452
+ `credential_unavailable`. The fix is the receipt taking `spec.agents[i].mcp` with `spec.mcp` and the door judging that copy.
453
+ - **Consumers.** @server: project the request face `agents[].mcpServers` onto `AgentDefinition.mcp` (one `McpServerSpec` per
454
+ record key; a name-only item maps to an `allowTools` entry `mcp__<name>`) and the agents wire projection lists the declared
455
+ server names — S-2182 is unblocked; the request's per-key judges for `mcpServers[i]` apply to `agents[j].mcpServers[i]`
456
+ unchanged, and a refused entry is the same refusal a task's own gets. @cli: an agent file's `mcpServers` reaches the engine
457
+ through that request face (CC-996); an item the reference would have skipped is now a refusal. @client-core: the agents wire's
458
+ server names are a projection position; no core call.
459
+ - **Pins.** `test/backlog2595-agent-mcp.test.ts` (real Runner + stdio servers: own server offered and called, allow list,
460
+ deny list, ceiling, namespace clash and an own duplicate pair, inheritance beside own, pool sharing, grandchild, per-task
461
+ definition, `mcp_unavailable`, a host-transport server through the spawning run's seat in both lanes and without one; one
462
+ judge at the definition, at the workflow fold and at the lock; the list judgment on the definition's lists and the spawning
463
+ run's own; the memory face, resource readers and an inherited server of the same namespace included).
464
+
465
+ ### The classifier seat is not asked again once its provider has rate limited it on three classifications in a row, and a status frame says whose model call it is about (#2876; behavior change, open to veto; hard-breaking: none; closed sets: +1 notice code `classifier.seat_rate_limited` (108 → 109), +1 knob refusal code `config.auto_mode_rate_limit_break_invalid`, +1 word set `BrainStatusSeat`; exports: none — the word set reads as `BrainStatus["seat"]`; wire +1 optional key `seat` on `status` events)
466
+
467
+ - **Before.** Every gated call walked the classifier ladder from its first rung. A classifier seat whose gateway answered HTTP 429 slept the brain's whole backoff ladder (`AUTO_MODE_CLASSIFIER_MAX_RETRIES`, 4 waits) on every call before the walk reached the main model behind it, and never recovered inside the run: a field run of 296 turns spent 5 133 s that way. The backoff's liveness frames (`status`, `phase: "rate_limited"`) carried nothing that said they were about the classifier's call, so on a client's screen the countdown read as the person's own turn's model being limited.
468
+ - **Now.** A rung whose dispatch ended on the provider's rate limit (an HTTP 429 that outlived the brain's own backoff, read from the reply's structured `apiErrorStatus`, not from prose) on `RunnerDeps.autoMode.rateLimitBreakAfter` classifications in a row (default 3) is not asked again for the rest of the run: the walk starts behind it, on the main model, and the record on each gate says the move it is (`GateDisposition.classifier`: requested the seat, used the main model, `fallback` `{ from, to, cause: "error" }`). One operator notice, `classifier.seat_rate_limited` (`detail: { model, to, rung, consecutive }`), is minted on the classification that tripped it. Any other outcome of the seat (a ruling, a timeout, a different failure) zeroes the count; an outcome that ends after the run's signal aborted neither counts nor zeroes (a 429 the seat answered while the run was live counts, whatever becomes of the classification after it). A retired rung is skipped wherever it sits on the ladder, and a walk that retires several rungs reports each once. Only a rung with a rung behind it retires: under `classifierFallback: "none"` (one rung) the seat is asked every time and the round stays fail-closed exactly as before. The count is the run's own, in memory — never persisted, never shared; a later run asks the seat afresh. A bad `rateLimitBreakAfter` (not a whole number of at least 1) refuses every leg's prepare with `config.auto_mode_rate_limit_break_invalid` — never clamped, never read as the default. Every `status` event for a call made for the classifier lane (the seat's model, or the main model standing in behind it) now carries `seat: "classifier"`; the frames of the run's own turn carry no `seat` key at all, byte for byte as before.
469
+ - **Who needs it / who is affected / compensation (open to veto).** *Needs it:* every deployment whose classifier seat shares a gateway or a quota that can answer 429 — it stops paying a full backoff per gated call for an answer that does not change, and its person stops reading the classifier's wait as their own turn's. *Affected:* a deployment that wants the cheaper seat to keep judging through a sustained rate limit — after three in a row the run's judgments come from the main model (dearer, and its judgment stands in for the seat's) until the run ends. *Compensation:* the count is a knob (a larger `rateLimitBreakAfter` asks the seat for longer), `classifierFallback: "none"` keeps asking the seat every time, the notice names the seat and the model now judging, and the next run asks the seat again.
470
+ - **Wire +1 (@cli @server @client-core).** `BrainStatus.seat?: "classifier"` on `status` events (a one-word closed union): present exactly on a frame about a classifier-lane call, absent for the run's own turn — a consumer reads absence as "the main loop", never "unknown". Additive and optional: a client that never reads it renders every frame as before; a client that does can show the classifier's backoff as the classifier's. The notice `classifier.seat_rate_limited` is a new operator-audience code (`ENGINE_NOTICE_CODES` 108 → 109): a host that pins the code count or renders operator notices from a closed table adds the row.
471
+ - **Consumers.** @cli: render a `status` frame with `seat: "classifier"` as the classifier's wait, not the turn's. @server: no change required; a deployment that maps its own auto-mode face from settings may pass `rateLimitBreakAfter` through. @client-core: the `status` frame gains the optional key. @test: construction — a gateway that answers 429 for the classifier model only; N+1 gated calls in one run dispatch the seat N times and the main model on every call; the notice appears once.
472
+ - **Known limits (stated).** (1) The memory is the run's: a session that spans several runs asks the seat once per run until it retires again. (2) A 429 whose `Retry-After` pushes the dispatch past `RunnerDeps.autoMode.timeoutMs` ends as a `timeout`, which is not counted. (3) The cross-process redemption's rebuilt decider (`rebuildAutoModeDecider`) has one rung and so never retires. (4) The frame names the lane, not the rung: the main model standing in behind the seat is also `seat: "classifier"`.
473
+ - **Pins.** `test/backlog2876-classifier-rate-limit-breaker.test.ts` (real Runner: two rate-limited classifications keep asking the seat, the third retires it and later calls never dispatch it; one notice; the record per call; a plain failure, a ruling and a timeout each zero the count; `none` and a one-rung ladder never retire; the last model is never retired; a cancel neither counts nor zeroes; the knob's default, values and refusals including through the Runner's prepare; the 429 marker comes from the structured status only; every classifier-lane frame carries the seat and the turn's own frames do not; `withStatusSeat` stamps through the active sink, a timer-fired frame included).
474
+
475
+ ### A retired `TaskSpec` key the caller's object carries by lookup is refused on every prepare path — the door reads the caller's spec before any rebind (#2914; behavior narrowing, open to veto; hard-breaking: none; closed sets: none; exports: none) (@server @cli @client-core)
476
+
477
+ - **Before.** The retired-key door (`config.spec_retired_key`) and the mode word's door ran on the spec after prepare had rebound it. Two of those rebinds — the per-task agents rebuild (`TaskSpec.agents` with a delegation tool mounted) and the tool model gate — copy the spec by spreading it, which keeps only its own enumerable members. On those paths a retired key the caller's object inherits (a class instance's prototype, a spec built with `Object.create`, a proxy answering by lookup) or holds as a non-enumerable member was dropped unseen — for every row of the retired table — and an inherited `permissionMode` read as absent, so the leg ran in the default mode.
478
+ - **Now.** The door runs once, on the caller's spec as handed (the roster snapshot's copy keeps the prototype and every own descriptor), before the config doors — the order the resume ladder's pre-claim screen already uses. On every path the door judges what lookup finds on the caller's object, and the leg's mode is the mode word that object answers. A spec holding a retired key and another fault now reports the retired key first, as a resume already did.
479
+ - **Behavior narrowing, three sentences (open to veto).** *Who needs it*: a host that builds specs as class instances or proxies and relies on a retired seat it still writes being refused rather than ignored. *Who is hurt*: a host whose prototype-backed spec still carries a retired key on a task that declares `agents` or trips the tool model gate — prepare now fails with `config.spec_retired_key` where the run used to go ahead without the seat — and a spec whose first fault was another door's now reads the retired-key refusal first. *Compensation*: the refusal names the key and its replacement (the sentence and code the plain path has always given), it comes before the session is acquired, and dropping the key fixes it.
480
+ - **Consumers.** @server @cli @client-core: a host that hands a Proxy or a class-instance spec — look for a retired key on its prototype; nothing else changes.
481
+ - **Not in this release.** The other members of a spec, read after those two rebinds, still lose an inherited or non-enumerable value on those paths (#3039).
482
+ - **One reader per retired row.** The retired `skills[].manifest` row moved out of the spec door's table into the skills seat's own (`RETIRED_SKILLS_SEAT_KEYS`, internal), judged only by the seat's one reader over its one read of `skills`. Moved ahead of the config doors, the spec door had become a second reader of the caller's `skills` (a `skills` accessor answering differently on its second read failed a run that the seat's reader would have prepared); the resume ladder's pre-claim screen had the same second read. Same sentence, same code; a retired skill member is reported where the skills seat is read, as before this release.
483
+ - **Pins.** `test/backlog2914-retired-keys-caller-spec.test.ts` ① an inherited, a non-enumerable and a dotted retired key on the agents path; ② an inherited mode word read, a bad one refused; ③ the clean spec prepares, the path with no rebind refuses as before. `test/backlog2003-seat-totality.test.ts`: a `skills` accessor is read once on a fresh run; the spec door reads no member of the skills seat; the two retired tables are disjoint and together the former table. `test/backlog2914-retired-keys-caller-spec.test.ts` ①: an inherited `skills` entry carrying `manifest` on the agents path is refused by name. `test/backlog1303b-loader-and-surface.test.ts` ③: the skills seat's reader and the resume ladder refuse `manifest` with the same code.
484
+
485
+ ### The Bash tool's description states one fact about the working directory — each call starts where the previous one ended (#2873; prompt text only; hard-breaking: none; closed sets: none; exports: none) (@cli)
486
+
487
+ - **Before.** "Working directory persists between calls … a `cd` carries over" stood beside "the shell is re-initialized each call, starting at the configured root". Only the first describes the tool: a call runs in the tracked per-task directory and the directory it ends in is read back. Models read the second sentence and put a `cd` into the project in front of every command (cli: 1 202 times in one session).
488
+ - **Now.** "Working directory persists between calls, but prefer absolute paths — each call starts in the directory the previous call ended in, so a `cd` carries over and moves the base for every later command. Shell state (env vars, functions) does not persist; the shell is re-initialized each call." This is the reference's sentence shape (`Working directory persists between calls, but prefer absolute paths — …`) with sema's reason kept. The tool listing's prompt-cache fingerprint changes once.
489
+ - **Pins.** `test/backlog2873-bash-cwd-sentence.test.ts` (the real tool's description, and a `cd` in one call setting where the next call starts on a real shell); `test/prompt-golden-1302.json` refrozen for this one sentence.
490
+
491
+ ### The `workspace_changed` contract says who receives the frame: the leg's own event stream (#2890; contract text only, no engine change; hard-breaking: none; closed sets: none; exports: none) (@server)
492
+
493
+ - **Now.** docs/INTEGRATION-CORE.md §1.8 adds one sentence to the frame's row: the frame rides the leg's own event stream, so `Runner.runTask` and the compositions that drain their legs' streams themselves (`runWithVerification` / `resumeWithVerification` / `runDeveloperTask`, `runCascade`) hand their caller no leg stream, and that caller receives no `workspace_changed` frame from those legs. This is within the contract (server [11986] ④). A deployment that needs where a verify or cascade leg stood would need a frame hook, which is a separate design request and is not in this release.
494
+ - **Pins.** `test/backlog2890-workspace-frame-contract-sentence.test.ts` (the row's sentence; the stream half: a consumer of a prepared leg's stream gets the leg-start frame).
495
+
496
+ ### A URL password holding `'` is taken out of a text, and the brain's request-error strip reads the same userinfo grammar (#3018; security narrowing — a credential leak, open to veto; hard-breaking: none; closed sets: none; exports: none) (@server @cli)
497
+
498
+ - **Before.** The vocabulary's userinfo row read a userinfo as characters that are no `/`, white space, `"`, `'` or `@`. `'` is one of RFC 3986's sub-delims, so `https://user:pa'ss@host` matched nothing, and the password left through every exit that reads the egress pass (`redactSecrets`, a sync refusal sentence, a response error, a tool result). On the human tier the whole-URL pass then stopped at the `'` and left `ss@host…` in the text. The brain's `stripUrlCredentials` (a request-build failure's text) kept its own copy of the class and stopped at the first `@`, so `https://u:p@ss@host` kept `ss@` of the password.
499
+ - **Now.** One grammar, `URL_USERINFO_SOURCE` (`credential-vocabulary.ts`): characters that are no `/`, white space, `"` or `@`, up to the last `@` before the host. It covers every RFC 3986 userinfo character (`'` included) and what an unencoded password carries (`#`, `?`, …), so it narrows nothing the row took before. Both readers match it and each renders its own way: the egress pass keeps the username, the brain's strip removes the whole userinfo — and `https://u:p@ss@host` is now removed whole. The two readers share the scheme grammar too (`URI_SCHEME_SOURCE`, as does the human tier's whole-URL pass), now of any length: it is read from the start of its run of scheme characters, so each run is tried once (linear), where the 64-character cap it replaces missed a longer scheme — a scheme glued behind digits keeps them in the text. The vendor-prefix rows keep their front boundary: `sk-…` glued behind a letter (`losk-…`, `xsk-…`) is still no credential.
500
+ - **Narrowing, three sentences (open to veto).** *Who needs it*: anyone whose URL password holds a `'`. *Who is hurt*: a text holding `scheme://` and then, with no white space, a `'` and an `@` that is no credential — a single-quoted list with no spaces (`'https://a','b@c'`) — now has the span between `://` and that `@` replaced; on the human tier a URL glued behind digits or marks (`1https://…`) is replaced together with them. *Compensation*: the span is replaced by the credentials mark, never refused, and double-quoted (JSON) text reads as before.
501
+ - **Pins.** `test/backlog3018-userinfo-grammar.test.ts` ① four passwords (`'`, `;`, `=`, a percent escape) replaced on the egress pass with the username kept, a password holding `@`, one written unencoded with `#` / `?`, a JSON string's closing quote; ② the human tier leaves no part of the password; ③ the brain's strip on the same passwords and on `@`; ③b a 65-character scheme in both readers, a scheme glued behind digits; ④ the front boundary kept.
502
+
503
+ ### A language server's root is a project directory, never the home directory — a file with no project marker below the home directory gets no server, and the LSP tool says why (#2872; behavior narrowing, open to veto; divergence from the reference D-2872, open to veto; hard-breaking: none; closed sets: none — the degrade reuses the tool's `LSP is unavailable` line; exports: none; members: `NodeLspManagerOptions.resolveRoot` may answer `undefined`) (@cli @server)
504
+
505
+ - **Before.** `NodeLspManager`'s default root was the environment's cwd whenever it had one (else the nearest marker above the file, else the file's own directory). A session started in the home directory therefore rooted every language server there — a Swift server indexing the whole home directory (cli IMG156-11).
506
+ - **Now.** The default root is read from the project markers the manager already keeps (`ROOT_MARKERS`: `.git`, `package.json`, `go.mod`, `Cargo.toml`, `pyproject.toml`, `Package.swift`, …): the cwd when it holds one (as before), else the nearest directory holding one above the file. The home directory (the environment's declared `homeDir`, else this process's) and every directory above it are never a root, marker or not — a dotfiles `.git` makes the home directory look like a project; directories are compared by their physical paths (the one physical-path mint, `physicalPathOf`), so a symlinked spelling of the home directory is the home directory. A file with no marker below them has no root: no server starts, the LSP tool answers `LSP is unavailable for "<file>" (no project root: …). Fall back to Grep/Read for this lookup.`, and the manager's `log` hook receives `lsp_root_refused`. A custom `resolveRoot` may answer `undefined` for the same outcome.
507
+ - **Behavior narrowing, three sentences (open to veto).** *Who needs it*: a person who starts a session in the home directory — no language server indexes everything under it. *Who is hurt*: a file outside every project (a loose script in a folder with no marker), which had a server rooted at its folder or the cwd and now has none; and a host whose `evict(root)` passes a cwd that holds no marker — that task's servers are rooted at the files' project directories and are reclaimed by the idle sweep instead. *Compensation*: the degrade line names the cause; any marker in the folder (`git init`, an empty `package.json`) brings the server back; a host that wants the old reading passes its own `resolveRoot`.
508
+ - **Divergence (D-2872, open to veto).** The reference roots a language server at its configured workspace folder or the session's cwd, home directory included (cli309.js:266959, `n.workspaceFolder||pe()`). sema refuses the home directory because a server rooted there indexes everything under it (cli IMG156-11): this is about what gets read and its privacy, not a feature trade.
509
+ - **Consumers.** @cli (the L-2122 family): a session started in the home directory gets the degrade line for a file outside a project, and a server rooted at the project for a file inside one. @server: a host lane that calls `evict(env.cwd)` at task end keeps working for a cwd that holds a marker.
510
+ - **Pins.** `test/backlog2872-lsp-root-not-home.test.ts` ①–⑥ (cwd = home with a file in it and under a markerless folder; home holding `.git`, also under a symlinked spelling of the cwd or of the declared home; a project below home; a project cwd with a nested package; a markerless cwd; the LSP tool over a real `NodeExecutionEnv` started in a declared home). `test/backlog1073-lsp-child-env.test.ts` and `test/backlog1753-lsp-withheld.test.ts` re-pinned: their temporary folders now hold a `package.json`.
511
+
512
+ ### A stop reaches the main model call whether or not the brain looks at the signal (#2882; behavior change: a halt, an interrupt or a cancel no longer waits for a brain that ignores the abort; hard-breaking: none; closed sets: none; exports: none) (@server @cli @test)
513
+
514
+ - **Before.** The loop read the brain's stream and waited for it. A halt (and an interrupt, a caller's abort, the walltime cut) aborted the turn's signal and left the rest to the brain. A brain that does not honour the signal — a third-party or wrapped brain, a transport that stalls — held the run until it answered. The compaction pass's summary call had the same gap until 7.59.0 (#2875).
515
+ - **Now.** One rule for both lanes: the model call is an await of the turn, settled by the turn's signal. The loop reads the brain's answer through one relay (`engine/loop/agent-loop.ts`); every await of the call is one — the stream's opening, each event and the final, which is read while the brain's iterator is still open, as before (the Runner's call guardrail stays armed until it is in). The brain's events pass as they come. When the signal fires and the brain's final is not in by the next macrotask turn, the turn ends with the loop's own aborted final — the shape the pre-dispatch gate already answers with: no content, `usageMissing` — and whatever the brain sends later is dropped. A brain that answers its own abort through the abort's own continuations keeps its answer, as the first-party brain does. A halt still reads as the person's stop (`haltedByUser`, completed); a caller's abort that the loop's final answered reads `run aborted`.
516
+ - **Known limit.** The ignored call keeps running: the engine cannot stop a call that does not look at its signal. Its turn's usage reads as unknown (`usageMissing`), and its late answer is never counted (KL-2875-a, now both lanes).
517
+ - **Consumers.** @server @cli: a stop on a provider that hangs now ends the run promptly; nothing to change. @test: black box — a brain that never answers and ignores the signal, halted while its call is open, ends the run within a second with `haltedByUser`.
518
+ - **Pins.** `test/backlog2882-halt-main-lane.test.ts` ①–⑦ (a silent brain, one that streamed part of an answer, one whose call never returns a stream; a caller's abort; the late answer reaching nothing; a brain that answers its own abort keeping its own words, and a halt on it as before; a brain whose final never comes after its terminal event, and one whose `result()` throws — the throw fails the run as it always did).
519
+ - **A delivery that throws lets the brain go (#3093).** When the loop's handling of a model event throws before the terminal event (a `message_start` or `message_update` sink), the relay stops reading the brain, calls its iterator's `return()` once without waiting on it, and drops its abort listener — before, the relay's producer kept reading the provider's stream with nothing consuming it until the stream ended on its own. A brain whose own read throws is not returned (its iteration is over, as a `for await` over it read it), and a call whose events were read to their end keeps its final cuttable by the turn's signal. Behavior fix; hard-breaking: none. Pins: `test/backlog3093-relay-consumer-exit.test.ts` R1–R5.
520
+ - **A guarded call left before its answer is cancelled (#3103).** Letting a brain's iterator go does not stop its request — a stream's own `return()` need not cancel it — and under the call guardrail (on by default) that `return()` also disarmed the gate, dropping the forward from the turn's signal into the signal the brain holds; so after a sink threw, the provider request ran on and nothing could cancel it any more. The guardrail's `return()` now aborts the signal the brain holds before it disarms, when the brain's terminal event has not passed yet; after the terminal event, or once the call ended (its iteration over, its result settled), it only disarms, so a call that ends on its own is never aborted. The run still fails with the sink's own error. With the guardrail off, the brain holds the turn's signal itself and the relay's exit does not cancel its request: the provider request runs to its own end — after the sink failure nothing holds a handle to cancel it (KL-3103-a). Behavior fix; hard-breaking: none. Pins: `test/backlog3103-guardrail-abandon-aborts.test.ts` G1–G6.
521
+ - **Wake cycles.** A wake cycle's first tool call can now land after the parent run's teardown abort (the relay adds one producer/consumer hand-off before dispatch); a parent's end aborting in-flight resume cycles is the existing D4 contract — hosts must not rely on the first tool racing ahead of the parent's end.
522
+
523
+ ### CronCreate and CronList report the job's next run when the scheduler backend knows it — `details.nextFireAt` and a sentence in the result (#2682; aligned with Claude Code; wire +1 optional key; hard-breaking: none; closed sets: none; exports: none)
524
+
525
+ - **Before.** The tools returned the schedule and the id; the backend's `ScheduledTaskSummary.nextRunMs` was read by nobody, so a model that scheduled a reminder could not say when it would fire.
526
+ - **Now.** CronList prints `, next run <ISO instant>` on a row and puts `nextFireAt` (epoch ms) on that row of `details.jobs`; CronCreate appends ` Next run: <ISO instant>.` to its receipt and puts `nextFireAt` on `details`, read with one more `list()` of the same leg inside the call's serialized section (the seam's `schedule()` returns only the handle, and the next-run fact has one reading: the listed summary). Only what the backend reported is shown: absent, non-finite or above `Date`'s upper bound ⇒ no clause and no key, never a guess; an instant already past — one below `Date`'s lower bound too — is clamped to now, as the reference clamps it. A listing that fails after the schedule landed leaves the clause out; the created job is never reported as an error. The call's own context is read for its cancellation only: whether a session-scoped job has a session to bind to is still read from the identity the tools were mounted with.
527
+ - **Consumers.** @cli @client-core: `details.nextFireAt?: number` on `cron-create` and on each `cron-list` job (the upstream output schema's own optional key); a card that ignores it is unchanged. A backend that never fills `nextRunMs` changes nothing for its users.
528
+ - **Pins.** `test/backlog2682-schedule-next-fire.test.ts` (the instant on both tools and in the text, absent / replaced / failing-listing / past / non-finite and out-of-range cells); `test/rb449-scatter-review-fixes.test.ts` (the serialization fixture does not count the read-back as a section opening).
529
+
530
+ ### An unparseable tool call is retried up to twice in a row, and a turn whose tool calls ran clears the count (#2684; aligned with Claude Code 2.1.307; behavior change, open to veto; hard-breaking: none; closed sets: none; exports: none)
531
+
532
+ - **Before.** With the malformed-tool-use recovery on (the Runner's default), the engine nudged once per RUN: the second turn whose every tool call was dropped as malformed ended the run as `malformed_tool_use_exhausted`, however many working turns lay between the two.
533
+ - **Now.** `recovery.malformedToolUse.maxRetries` defaults to 2 and counts consecutive malformed turns: any turn with a parseable tool call (the tools ran) resets it, so the terminal reason now says "this many in a row" and no longer "this many over the run".
534
+ - **Who needs it / who is affected / compensation (open to veto).** *Needs it:* a long task where a model emits a truncated call now and then no longer dies on the second one. *Affected:* a run that would have ended on its second malformed turn now spends one more turn (and tokens) before `malformed_tool_use_exhausted`; a host reading that reason as "twice in the run" reads it as "three in a row" now. *Compensation:* an explicit `maxRetries: 1` keeps the old budget (still counted in a row).
535
+ - **Pins.** `test/loop-transitions.test.ts` (the default budget is two nudges then exhausted; an explicit budget; recovery on the last of the default two; the streak clears across working turns and still ends after a fresh budget of two).
536
+
537
+ ### A missed `Skill` call says which listed skill it likely meant (#2688; aligned with Claude Code 2.1.299; hard-breaking: none; closed sets: none; exports: none)
538
+
539
+ - **Before.** `Unknown skill "<name>". Available skills: …` for every miss, including a model that asked for `anthropic-skills:pdf` where the listing says `pdf`.
540
+ - **Now.** The refusal adds `Did you mean <listed name>? Invoke it by that full name.` when a listed name is within two edits of the asked one, else when the asked name minus the other product's `anthropic-skills:` prefix is listed. The suggestion is a sentence: the one name resolution is unchanged, so nothing activates, binds or digests under an inexact name.
541
+ - **Pins.** `test/backlog2688-skill-name-suggestion.test.ts` (prefix, typo and transposition, no suggestion for a far or blank name, the nearest name before the prefix reading, nothing activated, an exact name served).
542
+
543
+ ### A deny / ask rule on an MCP name reads it as the reference does — a server whose own name carries `__` is covered — and an allow entry on an ambiguous spelling is refused with the reference's sentence (#2980 ①; security; behavior narrowing of grants, widening of restrictions; open to veto; hard-breaking: the session rule grammar is 7, open to veto; closed sets: none; exports: none)
544
+
545
+ - **Before.** The mint folds a server segment's run of `_` (`a__b` is spelled `a_b`), and every rule name was split once at its first `__`. A rule written the reference's way, `mcp__a__b` for the server `a__b`, read as tool `b` of server `a` and matched none of that server's tools: a deny / ask that looked protective and was inert.
546
+ - **Now.** A deny / ask MCP name with `__` in its body reads every way it can be read — a cut at each `__` (server = before, tool = after, the server spelled as the mint spells it) and the whole body as a server — so `mcp__a__b` covers tool `b` of server `a` AND every tool of server `a__b`, and `mcp__a__b__t` covers tool `t` of server `a__b` and tool `b__t` of server `a`. The same reading serves the policy rule lanes, the covering name lists, a sub-agent's denied list and a stored whole-tool row. A granting list keeps the one exact name: `mcp__a` never covers `a__b`'s tools, and an allow entry the run's roster shows to read as another server's spelling (no tool of that exact name, a tool on another reading) is read as nothing in the reference's words — "the name reads as more than one MCP server's. To allow it, use the tool's full name." — with the spelling the roster carries; a toolPolicy allow list reports it as an unmatched name instead of refusing the run. The mint is unchanged. A stored whole-tool rule still names one tool (the stored-rule grammar asks `namesASetOfTools`, not the covering predicate).
547
+ - **Behavior change, three sentences (open to veto).** *Who needs it:* anyone who carries rules written for the reference's server spelling. *Who is affected:* a restricting list now also matches tools of a server named like the rule's whole body (more deny / ask, never less); an allow entry that was silently inert next to a roster that shows the other reading is now named. *Compensation:* write the minted spelling (`mcp__a_b__<tool>`) — the sentence names it.
548
+ - **Consumers.** @server @cli @client-core: no wire key; `RuleToolNameVerdict`'s `unmintable_segment` gains an optional `ambiguous?: true`, and `isCoveringRuleName(name, "deny" | "ask")` is true for a name with two readings (it was false for an exact tool name); a host that stored the old reading of `isCoveringRuleName` for exact MCP names re-pins.
549
+ - **The rule grammar moved with the meaning: `SESSION_RULES_GRAMMAR` is 7.** A session's stored `toolDeny` entry `mcp__a__b` meant the one exact tool in every earlier grammar, so a row stamped 6 or below keeps that meaning and is announced once per layer (`config.session_rules_earlier_grammar`, reader `grammar_revised`: what its writer read, what this grammar reads, the stored meaning stands); a row this engine writes is stamped 7 and covers both spellings. Globs, protocol peers and allow entries are not revised. The persisted rule store and the rule-sync wire carry the same stamp; that store keeps no earlier-stamp reading, so a whole-tool deny / ask `mcp__a__b` in a bucket stamped 6 reads the grammar-7 way (wider; the earlier reading it is owed is #3085). So does a stored session row handed back to a write (an export → import replay): the write stamps 7, and the entry takes the wider reading unannounced — the same as a row kept by an earlier revision (a grammar-4 `toolDeny: ["Re*d"]` replayed reads the pattern); #3085 too.
550
+ - **Hard-breaking, three sentences (open to veto).** *What changed:* session rule rows, persisted rule buckets and rule-sync states this engine writes carry grammar 7. *Who is hurt:* a mixed-version fleet — a 7.59.0 engine (grammar 6) reading a session row or a persisted bucket a 7.60.0 engine wrote refuses it loudly (the row refuses every call; the bucket loads zero rules with `permission_rules.later_grammar`; a sync round is refused whole) instead of reading the deny as the narrower exact tool; @server (a shared session store or rule bucket across engine versions; the rule-sync twin), @cli, @test (fixtures that pin the number 6). *Migration:* upgrade the engines that share a store in one step; the refusal sentence names both grammars ("written in rule grammar 7, newer than this engine reads (grammar 6)").
551
+ - **Known limits (stated).** The allow-side refusal is as true as the roster the entry is read against: a server that connects after the roster is read leaves an ambiguous entry as it was (inert for the other reading). `peerReadingsOf` reads MCP names only; other protocols' peers are named as the roster spells them.
552
+ - **Pins.** `test/backlog2980-mcp-server-spelling.test.ts` (the readings; deny and ask each on both readings; the tool-level spelling; stored rule and sub-agent list; `mcp__a` does not cover `a__b`'s tools; the allow verdict, entry and sentence); `test/backlog1879-rule-name-one-reader.test.ts` F4 (a grammar-6 row's exact MCP deny keeps the one tool, announced; a grammar-7 row covers both spellings; the write mints 7; the later-stamp sentence); `test/backlog1385-entry-grammar-one-render.test.ts` / `test/backlog7480-smallfix.test.ts` (a deny list compiles the exact MCP name to a covering row, an allow list to a tool row).
553
+
554
+ ### A file `writeFile` replaces is staged owner-only until the target's own mode is applied — the new content of a 0600 file is never readable by another user while it is written (#3120; security fix; behavior change: a replace only, a create is unchanged; hard-breaking: none; closed sets: none; exports: none) (@server)
555
+
556
+ - **Before.** `NodeExecutionEnv.writeFile` stages a whole-file replace in a same-directory `.tmp.<pid>.<hex12>` sibling and renames it over the target. The sibling was opened with no mode — 0o666 under the umask, 0644 on a typical host — and the whole new content was written before the fchmod to the target's mode, so replacing an owner-only (0600) file left its new content readable by every other user of the machine, in the same directory, for the length of the write (seconds for a large file). A host lane that writes through this primitive inherited the window; the in-place `O_TRUNC` write it replaced kept the file's inode and mode.
557
+ - **Now.** A replace opens the sibling owner-only (0o600) and applies the target's mode before the rename, as before: the sibling is never more readable than the file it replaces, and a 0755 script is already 0755 when it is renamed. A create still opens the sibling with 0o666 and leaves the umask (or the directory's default ACL) to the kernel, so a new file's mode is byte-for-byte what it was; it has no older content to protect, and its final mode is what it carries while written. A process that dies mid-write now leaves its staging residue 0600 instead of a half-written 0644 file holding the new content.
558
+ - **Behavior change, three sentences.** *What changed:* replacing an existing regular file stages it 0o600 instead of the umask mode. *Who is affected:* a filesystem that refuses the fchmod with the support errors the replace already tolerated (EINVAL / ENOTSUP / EPERM / ENOSYS) yet honours the open mode now ends a replaced file 0600 instead of the umask mode — fail closed (KL-3120-a); directory watchers and indexers can still list the staged sibling, and what they see is now owner-only. *What to do:* nothing on the engine path; @server, in the server's words: server 7.139.0: the five remote lanes stage a sibling that is never more readable than the target from creation (FIXSTAGEMODE moves staging into a writer-private 0700 directory and brings the adb / docker host-side `os.tmpdir` relay files under the same rule); the host lane inherits this fix through core 7.60.0 #3120.
559
+ - **Divergence (open to veto).** The staging mode is a 0o600 constant where CC 309 uses `existingMode & 0o777 | 0o200`; 7.61.0 #3121 moves to the CC form.
560
+ - **Pins.** `test/backlog3120-staging-tmp-private-mode.test.ts` — the staged sibling's mode observed right after open, right after the content write and right before rename: P1 replacing a 0600 file (4 MiB) and P1b through a one-hop symlink stay owner-only throughout; P2 a 0755 script is staged owner-only and is 0755 at rename; P3 a create ends at the umask create mode (022 → 0644, 002 → 0664); P4 a create under umask 077 ends 0600; P5 an fchmod the mount refuses (EPERM) ends the replaced 0644 file 0600, owner-only throughout; a multi-hop symlink chain keeps the in-place write; the Write tool reaches the same rule.
561
+
562
+ no-notes: #454(repo tool scripts/probe-external-go.mjs and its package.json script `probe:external-go`: the SDK v2 external-condition probe, exit 0 GO / 1 NO-GO / 2 could not judge; not in the package, no gate-all roster row), #2916(repo release tooling: the car reader of the ship lint and the re-window reads a fast-forwarded car as in the release; scripts/release/lib.mjs, not in the package), #2921(repo tool post-ship: the PLANS row it writes opens the next window; not in the package), #2922(new repo gate gate:changelog-released-blocks with a frozen baseline docs/changelog-released-blocks.baseline.json; not in the package), #2923(repo release tooling: the chain's r-commit gate table asks gate:debts:rc at the release's own version; not in the package), #2932(repo gate verify-debts: a hex string stuck to digest: / sha256: / sha1: / integrity: / md5: is a content digest, not a commit coordinate; not in the package), #954(the credential-vocabulary parity pin's ledger text names the new keyword row; a test-file text change only, no engine behavior), #2979(the module-size ratchet rows of three oversize files cite the 7.61.0 split design ticket; test-file bookkeeping only; not in the package), #2861(the flaky-fix car only re-pins three merged tests; test files only, no behavior), #1549(the fix-2846 car re-pins the class-instance extraBody arm against the model door snapshot; test files only; no behavior beyond #2846), #1663(the fix-2846 car adds cells to the prompt-traits pin for the model door snapshot; one test file only; no behavior beyond #2846), #2966(the chain3-fix car re-pins four merged tests whose fixtures mount a seatless caller write tool or a caller Bash under default: each cell now sets aside exactly the load-time ungated-write notice this section documents; test files only, no behavior beyond #2966), #2994(docs/INTEGRATION-CORE.md Steer paragraph names which judgments read a steer, plus three pins; no engine change, the verb has read the live mode since 7.34.0), #2899(repo gate gate:changelog-coverage: a merge commit's subject line and a ledger row RESOLVED-PENDING-TEST-VERIFY at a commit of the range now count as tickets the range delivered; scripts only, not in the package), #2971(repo gate gate:backlog-ledger: a passed MM-DD window is a warning row naming the row, never red; red stays for format, closed sets and a second source of one fact; scripts only, not in the package), #2974(repo gate verify-debts: the pending-window category set drops the state word 候裁, seven words to six; scripts only, not in the package), #2494(repo gates and release tooling: the ledger gate caps the cars in flight (eleven, FIX cars not counted) and no longer the rows a window holds, a leftover window-cap declaration is red by line, rewindow writes no capacity and reports the next window's count; scripts and docs only, not in the package), #3037(the 2944 read-floor pin reaches git through the scripts/git-cli.mjs seam; one test file only, no behavior), #697(a #3025 commit re-pinned the parent-spec-seat pin's launch-note line — the one-shot note arm it read is the #3025 entry's; one test line only), #3041(repo tools only: the ledger gate accepts the terminal status word CLOSED-DESIGN-LANDED as a landed state and the release re-window no longer refuses a DESIGN-DONE row for lacking a car coordinate and lists per design what its implementation tickets say; scripts and pins; not in the package), #3045(repo tool scripts/wire-changes.mjs: the manifest generator reports a moved version literal — an exported VERSION integer constant or a version-named integer-literal member — as a version literal changed line of breakingItems and a versionLiterals row; WIRE: none cannot cover one; scripts and pins; not in the package), #3080(the chain-8 fix car's changes are told in the #2682 entry — the call context is read for its cancellation only — and the #2980 entry — the rule grammar 7 and its hard-breaking sentences; no entry of its own), #7(not a ticket: the number is a release-chain run label in the subject of the #2987 fix commit that follows chain 7; that fix is told in the #2987 entry — the tool-result seat resolves the run's image processor only once a block claims media; no entry of its own), #3096(a comment-only sentence on `SubagentToolOptions.background.notify`: an operator resume completes into that sink whether or not a run drives the mount; no entry of its own)
563
+ no-notes: #3065(a fix inside this window's own #3025 code: the delivery-obligation ledger keyed on the bare task_id instead of the notification lane key and routed an operator's resume completion into the spawn run lane; both corrected before release; no released behavior changed), #3066(a fix inside this window's own #2914 move: the retired-key table is split by reader `RETIRED_SKILLS_SEAT_KEYS` and five continuation pins re-pinned; no released behavior changed), #3067(a stale pin re-pinned after #2948; zero src), #3069(repo tool scripts/verify-debts.mjs: two new ledger states read; not in the package)
564
+
565
+ ## 7.59.0 — 2026-10-27
566
+
567
+ ### A model call waiting on a gateway gives up within 599 s however large the request, and a host that slept through a wait reconnects at the first second after waking — `sleptMs` on the reconnect frame, `waitSegment` on the wait frame (#2870; behavior narrowing: the first-byte window's upload allowance is capped; hard-breaking: none; closed sets: one new, `BrainStatus.waitSegment`; exports: none; members: +2 optional on `BrainStatus` / the `status` event) (@cli @server @client-core)
568
+
569
+ - **Before.** The first-byte window was `knob + 1 s per 32 KiB of body` with no bound but the timer ceiling: a 40 MB request (a context full of screenshots) armed 300 s + ~1 280 s and its retry 600 s + ~1 280 s — about 26 minutes of silence before the first retry. And no wait looked at the wall clock: a host that slept (a laptop closed over lunch) woke up still waiting on the connection the sleep took, until a TCP error happened to arrive or the window armed before the sleep ran out.
570
+ - **Now.** The upload allowance grows a window up to `FIRST_BYTE_WINDOW_CAP_MS` = 599 000 and no further (40 MB: 599 000, retry 600 000; 2 KB unchanged at 301 000; a knob already at or above the cap is the window as set). While any of the three engine waits is armed — response headers, first token, the silence between deltas — the wall clock is read once a second; a gap over five seconds is counted as slept, and at the first reading where a wait's sleep exceeds half its window the attempt ends and is re-sent on a fresh connection (shared retry budget; in the stream, the mid-stream tiers), announced as `reconnecting` / `errClass: "stall"` with `detail: "the host was suspended for <s>s during the <…> wait, reconnecting on a fresh connection"`.
571
+ - **Consumers.** wire +2, additive and optional: `sleptMs` — present exactly on a retry-wait frame for an attempt a host suspend ended (absent on a deadline's stall); `waitSegment` — present on every `waiting_first_token` frame (`response_headers` | `first_token`), absent on every other phase. @cli: render the wake line off `sleptMs` and the wait row's segment off `waitSegment`; @server / @client-core: project both members on the `status` face.
572
+ - **Pins.** `test/backlog2870-first-byte-sleep.test.ts` (the cap pure and at the call: 40 MB cut at 599 s, 2 KB at 301 s; a 15-minute suspend in each of the three waits ⇒ the frame within 1 s of waking and the new request within 1 s of the frame, text already streamed kept as incomplete; a deadline that runs before the next reading names the suspend; half-window threshold from both sides (200 s / 100 s of 301 s); five-second gaps never add up to a suspend; both members on the task stream's `status` event). KL-1121e retired; KL-2870-a (a forward wall-clock step, or an event loop blocked for more than five seconds at a time, reads as a suspend once it adds up past half the window: one reconnect).
573
+
574
+ ### An MCP server the user must authorize is one failure word, `unauthorized`, with one authorization fact that every face reads — the status record, the wiring manifest, a refused call's error and `tool_end`, a `mcp.authorization_required` notice, and the model's needs-authorization line (#2295, slice S2 of the MCP credential seat, design/1494; hard-breaking, open to veto; closed sets: `MCP_FAILURE_KINDS` +1 (`unauthorized`), `ENGINE_NOTICE_CODES` +1 (`mcp.authorization_required`), new `MCP_AUTHORIZATION_REASONS`; exports: +1 root value, +2 root types) (@server @cli @client-core @connectors)
575
+
576
+ - **Before.** A 401 from an MCP server's HTTP endpoint was `errorCode: "http_status"` with `httpStatus: 401` — read like a gateway outage, with the endpoint's `WWW-Authenticate` challenge dropped — and a refused call told the model that the endpoint "answered 401 instead of an MCP response" and that the request "may or may not have executed". A 403 whose challenge said `insufficient_scope` was the same status word. A deployment seat answering `unauthorized` (#2294) was `credential_unavailable`, the deployment's-fault word. The model was never told a server needed authorization: the needs-authorization arm of the reference's tool-delta notice had no state to read.
577
+ - **Now.** One word, `unauthorized`, carries `authorization: { reason, authorizeUrl?, resourceMetadataUrl? }` (`McpAuthorization`; `reason` ∈ `MCP_AUTHORIZATION_REASONS`): `rejected` — the endpoint answered 401 and the seat's re-ask did not answer a credential that passed (`delivered: "unknown"`); `scope` — a 403 with `Bearer error="insufficient_scope"` (the seat is not asked again, nothing is sent again; `"unknown"`); `none` — the seat answered `unauthorized` (nothing sent; `"no"`, no liveness observation). A 403 without that challenge stays `http_status`. `authorizeUrl` comes only from the seat — its answer to the send or to the re-ask after a 401 — and must be an absolute http(s) URL (≤ 2048, no control character) or the answer is refused (`credential_unavailable`); `resourceMetadataUrl` comes only from the endpoint's challenge, kept under the same URL law as the parser serializes it, and is never rendered to the model. The engine's one HTTP fetch reads the challenge off the 401 / scope-403 response and raises the refusal itself (the SDK's own error keeps no header). The faces: the status record (`MaterializedMcp.statuses`, a re-dial's `status`) and the manifest's `mcp[]` row carry `authorization` (no `httpStatus`), through the record's one projection; a refused call's `details.authorization` and its `tool_end.errorCode`; the notice `mcp.authorization_required` (user audience; `detail: { server, reason, authorizeUrl?, principal?, sessionId, runId }`) once per (leg, server); the `mcp_unavailable` frame on the leg's first request lists a server whose dial failed `unauthorized` in its own needs-authorization arm — the reference's `needsAuthMcpServers` sentence (CC 2.1.309, minted `cli309.js:265582`, read and worded `cli309.js:267398-267401`), always its cannot-start-the-flow form, `<where>` = the seat's URL or the deployment's connector settings, "Do not ask the user for authorization codes, tokens, or callback URLs" — and never in the failed-to-connect arm; a refused call's line says the user needs to authorize the server (and where), with the delivery tail its verdict owes. A refusal is announced when it is the request's answer: a repeat-safe call refused 401 and sent once more on a renewed credential announces only if that second send is refused too; a write-class 401 after which the seat answered another credential is announced, and its line says a new call carries that credential and that only its refusal is the user's (a seat that answers a fresh assertion on every ask must still reach the user). Nothing parks on it and a resume replays nothing (C-R725 加注 4: a deployment that waits keys on the notice).
578
+ - **Hard-breaking (open to veto) — sentence 3 of design/1494 §8.3.** *What changed:* an MCP 401, and a 403 with an `insufficient_scope` challenge, are no longer `errorCode: "http_status"` + `httpStatus`; they are `errorCode: "unauthorized"` + `authorization` on every face (status, manifest, error `details`, `tool_end`), and a seat's `unauthorized` answer is no longer `credential_unavailable`; a write-class call answered 401 is still sent again by no layer (#2294). *Who is hurt:* every consumer that branches on `errorCode === "http_status" && httpStatus === 401` (or 403) — @server's MCP status face, @client-core's `engineErrorCodes.ts` 401 classification, @cli's rendering — and every `switch` over `MCP_FAILURE_KINDS` (one more word; the "previously only" 401 reading moved to it). *How to migrate:* read `errorCode === "unauthorized"` and `authorization.reason`; send the user to `authorization.authorizeUrl` when present (your seat's own URL), else to your connector settings; treat `resourceMetadataUrl` as the remote's untrusted claim; key a wait-for-authorization flow on `mcp.authorization_required`.
579
+ - **Closed sets.** `MCP_FAILURE_KINDS` +1 `unauthorized` (its delivery and liveness rows read the status class's: `"unknown"`, kept at the dial); a refusal by the engine's own seat is `"no"` / no observation by one rule over both seat words. `ENGINE_NOTICE_CODES` +1 `mcp.authorization_required` (audience `user`, minter `engine`). New `MCP_AUTHORIZATION_REASONS` (`rejected` / `scope` / `none`). Consumers that pin these sets closed: @server (notice whitelist, MCP status projection), @client-core, @cli.
580
+ - **Consumers.** @server: project `unauthorized` + `authorization` on the MCP status face; forward `mcp.authorization_required` to the user (route by `sessionId`; `authorizeUrl` is your seat's) and, for a managed wait, key the authorization callback on it — no park gate exists for it; your `peerSeat` answering `{ kind: "unauthorized", authorizeUrl }` (to a send or to the `rejected` re-ask) is how the user is sent to your authorization page. @client-core: `engineErrorCodes.ts` reads `unauthorized`; your local seat's `unauthorized` answer is the same. @cli: render `unauthorized` and the notice. @connectors: answer 401 with `WWW-Authenticate: Bearer resource_metadata="…"` per RFC 9728 and it rides `resourceMetadataUrl`; a scope shortfall as 403 `error="insufficient_scope"`.
581
+ - **Exports.** +1 root value `MCP_AUTHORIZATION_REASONS`, +2 root types `McpAuthorization` / `McpAuthorizationReason`; `McpServerStatus` / `WiringManifestMcpEntry` +`authorization`; `McpFailure` +`authorization`; `materializeMcpTools` +1 trailing parameter (`onAuthorizationRequired`).
582
+ - **Pins.** `test/backlog2295-unauthorized.test.ts` (W: 401 not renewed ⇒ rejected / unknown with the challenge's resource_metadata and the user's line; 403 scope ⇒ scope, no re-ask, not sent again; 403 without the challenge stays a status; the seat's `unauthorized` ⇒ none / no, nothing sent; the re-ask's URL on a 401's record; a renewed write 401 ⇒ the same word, the "a new call carries it; if refused too, the user" line, announced; a re-send that passes announces nothing; a seat renewing on every ask still announces. T: the challenge never supplies `authorizeUrl`, its resource_metadata is percent-encoded, a non-http(s) one dropped; a seat URL off the URL law refused loudly; a Bearer challenge after another scheme is read, `not_error=` and an `error=` inside a quoted description are not its error; markup in a seat URL percent-encoded, user-info refused; a control character in resource_metadata dropped. F: the status record and the announcer one fact; the seat's dial refusal no observation; through the Runner the manifest row, the needs-authorization arm with the server in it only, one notice; two refused calls ⇒ two frames, one notice; a listing refresh refused for scope announces; a changed authorize URL is told again on the session's next run, the same one is not; the receipt files the told pair only; the closed sets. P: a durable leg that parked after a refused call does not send it again on resume — a control, green before and after). Re-pinned to the new word: `backlog2294-peer-seat` (R2 / R3 / R4; F's `unauthorized` row moved here), `backlog562-manifest-mcp-face`, `backlog610-connect-delivery`, `backlog874-mcp-liveness`, `mcp-failure-record`.
583
+ - **Known limits.** A write-class call refused 401 once with a credential the seat then renewed still tells the user to authorize (it is the call's answer; a seat that renews on every ask must still reach the user) — KL-2295-a; a 403 scope refusal never re-asks the seat, so it carries no `authorizeUrl` — KL-2295-b; an unauthorized dial still raises the operator's `mcp.server_unavailable`, and A2A peers have no `unauthorized` word — KL-2295-c.
584
+
585
+ ### A frame's delivery record advances on the receipt of the message that carried it, never at collection — every turn-boundary attachment and the limit-approach notice (#565, design/565 slice S1, ruled C-R882; behavior narrowing, open to veto; hard-breaking: none; closed sets: none; exports: none; members: none) (@server @cli @test)
586
+
587
+ - **Before.** The turn boundary collects its attachment bundle, then awaits the postToolBatch dispatch and the instructions probe, then issues one steer that the next request drains. Several frames recorded "delivered" at collection: the `mcp_dropped_tools` batch (its entries left the lane and their once-per-session digests joined the session's announced record), the post-compaction `background_tasks` latch, the `tools_delta` pending lists, the `instructions_change` / `workflow_size_guideline_change` announced values, the `changed_files` surfaced marks, the todo / task / tool-search reminder windows and the limit-approach threshold. Every bundled frame's `steering_injected` echo and `stats.mechanisms.attachmentsInjected` count went out at collection too. A stop landing after the collection (`TaskStream.halt()` or a tool's stop-after-turn, inside the postToolBatch dispatch or the probe) still let the steer be issued, and the loop exited before draining it: the session never heard the dropped MCP tool again (its digest was recorded), and the background-task restatement vanished without its settle-time `compaction.announce_dropped` disclosure.
588
+ - **Now.** One rule: collection records nothing; each frame registers a receipt on the steer that carries it, and the run loop's `message_end` walk settles it when the drained steer is in the transcript — the frame's echo, its count and its record step run there (the rule the listings, the date and the leg's first-request frames already followed). A frame on its way is not collected again. A steer that never drains, or is refused, records nothing: the frame is said again at the next quiet boundary, or the next run on the session (`mcp_dropped_tools`), or — the background-task latch — it is still owed when the run settles and the loss is disclosed. A frame the bundle cap drops is due again at the next quiet boundary (it used to spend its reminder window or its latch). An empty background-task registry still consumes the latch at collection (nothing to deliver).
589
+ - **Open to veto — the echo.** `steering_injected` for a turn-boundary frame (and for `limit_approach`) is now emitted when the drained steer's message is committed — after that message's `message_committed`, at the next request — not at collection; a frame that never reached the transcript emits none, and `stats.mechanisms.attachmentsInjected` counts delivered frames only. Who is hurt: a consumer that rendered the echo as "injected" at the boundary, or a test that expected an echo for an undelivered frame. Migrate: read the echo as delivered; order it after the message it rode.
590
+ - **Not in this slice (stated).** The `mcp_instructions` frame still records at collection (its told-ledger is written by asynchronous MCP lifecycle events and needs an in-flight state); a background-task restatement lost to an undrained steer is disclosed but not re-armed on the session's next run (the re-arm seed reads "an assistant message follows the compaction" as consumed); the LSP diagnostics drain and the memory announcement queue (drained at prepare into the system prompt) record before delivery. Each has its own design ticket.
591
+ - **Consumers.** @cli: the boundary lane's `steering_injected` rows arrive with the next request, not at the boundary; none for a frame a stop swept. @server: `announced_listing.advisories` (the dropped-tool digest arm) is written later — at the leg's exit after a delivered batch — and not at all for an undelivered one; the key shape is unchanged. @test: a halted boundary's frames are announced again by the next run.
592
+ - **Pins.** `test/backlog565-commit-on-receipt.test.ts` (real Runner: a halt inside the postToolBatch dispatch ⇒ the dropped-tool batch is not recorded and the next run announces it; a drained batch is recorded; the post-compaction restatement swept by the same halt is disclosed; a drained one is seen once; unit: the producers' record steps ride their frames, a cap-dropped frame hands nothing over, a clipped one keeps its step).
593
+
594
+ ### A parked approval covers the tool its roster row was, read whole: one classification of every roster member (identity / face / declaration / unbound), the declaration bound as one digest, a row that does not carry a whole tool member refused before the claim (#639 / #2345, C-R883; BREAKING — behavior narrowing, open to veto)
595
+
596
+ - **Before.** The tool member of a parked approval was a hand-picked subset of its roster row, and each member added later had its own reading of absence: a row with none of the four identity members bound the tool's NAME alone (a replacement wearing the name ran on resume); one with some of them reopened as a replacement after the claim; a path target without its resolution members was compared as `cwd` / `none` by a second statement of those defaults (a row parked before #635 by a tool that resolved against the root was read as a face it never had, and vetoed with "the face changed" after its approval was spent); the recorded mount site (`engineMounted`) was never compared; and the tool's declaration — its effect, egress, irreversibility, content origin, boundary legs, rule face, model gate (an MCP server's `destructiveHint` folds into `irreversibility`) — was not part of the approval: a tool re-declared between the park and the resume ran under the new declaration.
597
+ - **Now.** `ROSTER_APPROVAL_BINDING` classifies every member of `ToolRosterEntry` once (a new roster member is a compile error until it is classified; the judge's identity and face lists are fenced to it at compile time). Unbound, each with its reason: aliases, the server's declaring origin, the two value provenances no decision reads (`egressProvenance`, `contentOriginProvenance`), the card id, the wire schema's prose and the input keys (the schema's shape is identity), result storage and transcript handling, the preview / validator flags, loading, render hints. The park writes the identity (the four members and the mount site) and the face by value, and the declaration members as one digest, `pendingAction.declarationDigest` (`toolDeclarationDigestOf`, requirement word `tool_declaration`). `judgeApprovalObject` reads the tool member whole or not at all: at the claim station (now `["tool", "grants"]`) a row missing any member the park always writes (the four identity members, `family`, `declarationDigest`) is refused — `checkpoint.unsupported_version`, `detail.reason: "tool_identity_missing"`, the row pending, a deny is not judged by it (a member present in another shape — `engineMounted` not `true`, a slot that is not a record naming its `param` — is read as damage the same way; past the claim the same reading vetoes); with a whole member, another identity (the mount site included) reopens (`resume.tool_contract_mismatch`) and another face or declaration vetoes (`resume.target_moved`). A face is compared as it resolves, through the one reader of a slot's silences (`pathTargetResolutionOf`): a definition that spells its default base matches one that states none.
598
+ - **What changes / who is hurt / how to migrate.** (1) Every tool approval parked before 7.59.0 carries no declaration digest: an allow of it is refused before the claim — deny it so the model can re-issue the call and its new card is asked (a row that also predates the ask's origin, 7.5.x, refuses a deny too), finish it on a worker of the release that parked it, or expire it. (2) A deployment that writes its own `tool_approval` rows must mint the tool member as the park does (`ROSTER_APPROVAL_BINDING` / `toolDeclarationDigestOf`), or its rows bind no tool. (3) A tool re-declared after the approval (a caller axis; an MCP annotation the engine folds into an axis — `destructiveHint`, `openWorldHint`) no longer runs on that approval: the call is vetoed and the model may re-issue it. (4) Consumers listing `CheckpointErrorReason` gain `tool_identity_missing`; those listing the requirement words gain `tool_declaration` (a 7.58.x worker refuses a 7.59.0 approval row `version_newer`, pre-CAS, retryable on a newer worker).
599
+ - **Consumers.** @server: the row gains `pendingAction.declarationDigest` and the word `tool_declaration` (a store that drops the member fails `probeParkRoundTrip`, whose representative row now carries the whole tool member, and every such row reads `requirements_damaged`); pending approvals across the upgrade must be re-asked (three exits above); the reason closed set +1. @dgx: the same through the server's store. @cli / @client-core: the new veto sentence rides the existing `resume.target_moved`; the reason word +1 where the error-code tables list it.
600
+ - **Exports.** +3: `ROSTER_APPROVAL_BINDING`, `type RosterApprovalBinding`, `toolDeclarationDigestOf`. `APPROVAL_STATION_MEMBERS.claim` is `["tool", "grants"]`; `ApprovalObject.tool` is the whole member or `{ missing }`; `ApprovalTool`'s always-written members are required.
601
+ - **Pins.** `test/backlog639-parked-identity-face.test.ts` (real Runner: base-only face move vetoed; a row before the resolution members, a row with no identity and a replacement, a row with a lost member — each refused before the claim, pending, nothing ran; a deny settles; a caller tool re-declared irreversible and an MCP tool re-listed `destructiveHint` — vetoed, never run; the controls; the classification covers every roster member, the digest moves with each declaration member and with no unbound one, the mount site is identity, the face resolves its silences once, a member present in another shape is damage at every station).
602
+
603
+ ### A revived background agent runs under the restrictions it was born with — the durable row records the clamp table's projection of the child's spawn spec (`policyFloor`), and a cross-process revival folds it with the waker's own clamps; a row that states no floor is not revived (#887, design/887; behavior narrowing and security fix; hard-breaking, open to veto; closed sets: none; exports: none — one optional member `BackgroundAgentRecord.policyFloor`, one on `RegisterBackgroundAgentInput`) (@server @dgx)
604
+
605
+ - **Before.** A background child woken across processes (SendMessage's durable rung, or a host's parked-approval drive through `ctx.reviveClaim.parkedResume`) had its spec rebuilt from the WAKER's mount: every tighten-only seat of the clamp table — `handsReadOnly`, `interactiveTools`, `approvalReach`, both containment faces, the three add-only deny lists, `oneShot`, `memoryPersistenceCapable`, `memory.enabled`, `fileHistory`, `projectMemory`, `promptForm` — came from whoever woke it, and the row carried none of the spawning run's. A child spawned by a read-only parent and woken after a restart by its unclamped root wrote files. The in-process wake already folded the spawn snapshot with the waker's clamps; the two wakes followed two rules.
606
+ - **Now.** One rule for both wakes: a woken cycle runs under the table's fold of the floor it was born under and the waker's clamps — a restriction either side carries stands, add-only lists union, the waker can narrow and never widen. The registration that creates a row writes `policyFloor` (`{}` when nothing was clamped) and nothing rewrites it. `readClampFloor` (internal) is the one reader: a floor that is absent, not plain data, names a seat this build's table lacks, states a non-restricting value, or carries an entry its seat's law refuses is unreadable. A revived cycle's interaction posture is read under the folded clamps (a floor that brings the hard-headless clamp onto a cycle woken by an interactive caller, or by a host's bare drive that states no posture, runs it headless instead of tripping the posture door). A parked-resume refusal's receipt names its cause (`configuration`). "A clamp leaves nobody to ask" is one predicate (`clampsReachNoPerson`, internal) that `postureUnderClamps` and the revival read; the two other lanes that read `postureUnderClamps` (the in-process wake, the workflow lane) are unchanged — an absent posture under such a clamp still falls through to the deployment default there (a known gap, #2748: the workflow lane's spec is a journal key, so its fix is its own design).
607
+ - **Hard-breaking (open to veto).** *What changed:* a revival of a row without a readable `policyFloor` — rows written by 7.58.x and earlier, or by a host store that does not persist the member — is refused `revive.policy_floor_unreadable` instead of running under the waker's clamps: SendMessage refuses before its claim (the row is untouched, nothing is charged), the Agent tool's revive arm refuses before any allocation, and a parked-resume drive refuses before the consume and rolls its claim back (the approval stays pending). *Who is hurt:* a deployment with its own `BackgroundAgentStore` that maps columns by hand (server's SQL twin) — without a column for the member every revival is refused; rows in flight across the upgrade cannot be revived by the new engine. *How to migrate:* persist `policyFloor` verbatim (JSON; `{}` must read back as `{}`, absent as absent) and run the contract kit's new case; for in-flight rows, let them finish, start a new agent, or continue them on an older worker during a rolling deploy; a parked old row is expired.
608
+ - **Consumers.** @server: add the column to the SQL `BackgroundAgentStore` and its local twin; the contract kit's new case names a store that drops it; `revive.policy_floor_unreadable` is a new `details.error` on the SendMessage receipt and the Agent tool's not-started result. @dgx: a deployment store persists the member.
609
+ - **Not in this release.** The inherited gate chain (permission-mode ceiling, `shellGate`, ancestor rules and constraints) and the tool-selector ceiling of a cross-process revival still come from the waker (#2739).
610
+ - **Pins.** `test/backlog887-revive-policy-floor.test.ts` (the floor recorded at spawn, `{}` for an unclamped spawn; six tier-3 shapes over a real Runner — write hands with a real `Write`, `oneShot`, both faces, memory off, the add-only union, posture — and a second wake still held to the birth floor; the waker's own clamp lands and an unclamped floor tightens nothing (positive control: the probe observes a write); an old-shape row refused before the claim with the row untouched, the Agent arm's own refusal of a directly handed claim, a parked-resume drive's refusal with the claim rolled back; the reader's accepted shape and sixteen refusals, one read per accessor; `clampsReachNoPerson` row by row; a waker that states no posture on an interactive-default deployment); the store contract kit's round-trip case (each CAS on a floor-bearing row must land and keep the floor). Re-pinned: `agent-revive-tier3` (a hand-built row states its floor), `flaky-guard` (the new file's bounded effect poll declared). `gate:error-surface` `src/agents` 92 → 93 (the floor refusal names the agent like its sibling receipts).
611
+
612
+ ### A call an extension refused names that extension on the gate's record — `tool_end.gate.disposition.extension` (#2810), one reading for every hook-layer refusal
613
+
614
+ - **Before.** An extension's refusal at `tool.check` — and its failure, which the engine fails closed — told the model "denied by extension "<n>": <reason>" and recorded `{ kind: "denied", deniedBy: "hook" }`: the registration name lived in the sentence alone, so a card or an audit trail stripped it back out of prose. A `classic.PreToolUse` deny and a `tool.call` refusal had the same gap (the latter only on `answeredBy`).
615
+ - **Now.** The denied arm of `GateDisposition` gains `extension?: string`: the registration whose handler refused the call at the hook layer — its `deny` answer at `tool.check`, `classic.PreToolUse` or `tool.call`, or its failure. Present exactly when `deniedBy` is `hook` and the refusing handler's registration is known; absent beside every other layer word, beside an ask a person refused, and beside a `hook` refusal that names no registration (the engine's own failed raise, a cancel, the engine's refusal of a form an extension rewrote). The record is the gate's one mint: the `tool_end` frame, the `classic.PermissionDenied` observer and the durable row hold the same value. Invariant I12 screens it against one table (`DENIED_BY_MAY_NAME_EXTENSION`) the mint and the archive reading both read.
616
+ - **Consumers.** wire +1 (additive, present-iff the hook layer refused through a named registration): @cli the "denied by" line reads the member instead of the sentence; @server projects `gate` whole (no key to add); @client-core reads it where it renders a refusal. `answeredBy` (and `gate.disposition.extension`) names the handler whose run produced the answer — when a `.catch` hands back the very answer the run below it already produced, the producer below keeps the name (`src/core/mods/kernel/dispatch.ts:830`; before, the `.catch`'s handler was named); otherwise the `.catch`'s own handler is named, as before. A `tool.call` refusal carries both, the same name.
617
+ - **Exports.** none (a member of an exported type).
618
+ - **Pins.** `test/backlog2810-hook-layer-extension.test.ts` (a `tool.check` deny, a deny beside an allowing bystander, a throwing handler, a `classic.PreToolUse` deny and throw, a `tool.call` refusal, the approved edit's re-screen, a resumed leg's re-screen, the observer's record equal to the frame's; policy / ask-refused / allow carry none; I12 in both readings).
619
+
620
+ ### A call the loop refuses after its gate let it go is `notExecuted` — and keeps its `rewrite` (#2817; hard-breaking: two wire invariants reversed, open to veto; closed sets: none; exports: none) (@cli @client-core @server)
621
+
622
+ - **Before.** When an extension (or the policy, or an approver) replaced the arguments and the loop then refused the form the gate cleared — its re-validation, the tool's own input check, its re-reading of the bytes, a run aborted before the body — the frame was `isError` with no `notExecuted`: the body had not run and the frame did not say so.
623
+ - **Now.** The loop marks the end of such a call `notExecuted` on the same loop-authored channel a batch-halted sibling uses, and the runner projects it (frame and persisted row). The same mark covers a call the host's execution station (`tool.call`) settled without running the body. No member and no status word is added. `tool_end.rewrite` now stays beside the mark: it is dropped only on the engine's own evidence that the call was never let go or was answered in place. A call refused before any gate saw it (an unknown tool, arguments the model's own call got wrong) carries no mark.
624
+ - **Consumers.** `notExecuted` is now present on frames whose `gate.disposition` is `allowed` and `isError` is true (it was always present only beside a gate refusal or an answer in place): @cli @client-core read it as "the body did not run", never as "the gate refused"; the persisted row repeats it as `details.notExecuted`, so a reloaded transcript agrees. Such a call no longer arms the final-verification gate (`wroteThisRun`) and no longer enters the post-tool batch.
625
+ - **Hard-breaking, open to veto (@cli @client-core @server).** What changed: two invariants of the 7.58.0 wire are reversed — `tool_end.notExecuted` was present only beside a gate refusal, a halted batch, a reconcile close or an answer in place, and is now also present on a frame whose `gate.disposition` is `allowed` with `isError`; and `tool_end.rewrite` was present only on a frame that carried no `notExecuted` (the contract sentence 7.58.0 shipped with C-R922, `src/core/task-event.ts`), and now stands beside `notExecuted` on a call the loop refused after the gate let it go with other arguments. Who is hurt: a consumer that reads `notExecuted` as "the gate refused" (a card or audit line worded as a refusal, a replay guard that treats the mark as a gate verdict) or that infers whether `rewrite` is present from the mark's absence (a "the arguments were changed" line dropped whenever `notExecuted` is on). What to read instead: the two keys independently — `notExecuted` as "the body did not run", `gate.disposition` for who decided, `rewrite` for whether the gate replaced the arguments — and never infer one key from the other's absence.
626
+ - **Pins.** `test/backlog2817-loop-refusal-not-executed.test.ts` (a rewrite the schema refuses, a rewrite the tool's input check refuses — both with `rewrite` beside the mark; the row; controls: a body that runs and fails, a gate refusal, a pre-gate refusal, an answer in place; the loop's own end event for a gate that throws / blocks / is absent, a station that runs or does not run the body, a run aborted before the body).
627
+
628
+ ### A parked entry has one identity, the one its producer minted — a mailbox record carries an `id` and a store's `append` is a compare-and-set on it, the session park seat re-asks a store whose reply was lost instead of keeping a second copy, and a wake message under a parked steer's `inputId` is judged by the parked queue's own rule (#988, #2705; design/988, ruled C-R886; behavior change; hard-breaking, open to veto: the store contract gains the id obligation; closed sets: +2 mailbox store codes; exports: +8 root values, +1 type) (@server @dgx @cli @test)
629
+
630
+ - **Before.** When the session park seat's durable append was rejected, it read the rejection as "nothing was written" and parked the frame in process too. A store can commit and lose its reply, so the box then held the record AND the process held a copy under a different key (`task_notification:<child>:event:<seq>` beside `session.<id>`/box seq): the session's next run read the message twice — once as `<cross-session-message>`, once as an `<agent-message>` twin — and the worker's receipt said "parked in process … a restart of this process loses it" for a message the box held. A second seat (another process, or this one restarted) parking the same frame wrote a second record. A wake message sent under the `inputId` of a steer the row already held was delivered as a second instruction (identical) or delivered beside it (different), where `setPendingSteer` replays the first and refuses the second.
631
+ - **Now — one rule, two seats, one judge each.** An identical write under an identity the seat holds is a replay (one entry, one delivery); a different record under it is refused by name; it is never a second entry.
632
+ - `MailboxAppendMessage.id` / `MailboxMessage.id` (optional): the record's producer-minted identity, persisted and handed back on `claimLease`. While a box holds a record under that id, `append` answers through `judgeMailboxAppendById`: identical on `MAILBOX_REPLAY_FACETS` (`from`, `content`, `hopChain`, `peerMeta` — what the record says; `sentAt` is when a write happened) ⇒ the held `seq`, nothing written; any facet different ⇒ `mailbox.duplicate_record_id` naming the facets. A malformed id (not a 1–1024 character string) ⇒ `mailbox.invalid_record_id`, before any side effect. After the record is acked its id is free again: a box answers for what it holds. Both bundled backends run the judgment inside their write critical section; an absent id is a new record each time (unchanged).
633
+ - The park seat's two record forms (the peer-words `peer_message` record and the `engine_frame` record) carry the frame's key as `id`. A rejected append is disclosed once and asked again with the same record; only a second fault parks the frame in process (`parked`).
634
+ - The in-process receipt says what the seat knows: "this copy does not survive a restart of this process" (it no longer asserts that nothing else holds the message).
635
+ - The wake entrance reads `judgeParkedSteerRepeat`, the parked queue's identity judgment now shared with `appendPendingSteer` (which calls it; no change there): an identical repeat of a parked entry is that entry, delivered once; a different one is `steering.duplicate_input_id` (`detail.differs`) before the claim, the row pending and byte-identical.
636
+ - **Limits, stated.** A box answers for the records it HOLDS. Two paths still deliver one message twice (KL-969j, narrowed): the re-ask faults too while the first append had committed (the box and the process each hold it); and a consumer drains and acks the record between the first ask's commit and the re-ask (the re-ask then writes a new record) — a store whose rejection arrives a whole turn after its commit. A record the producer re-parks after it was consumed is a new record (at-least-once, unchanged). A different record under a held id is kept, not folded: the seat parks it in process (it is not lost, and the receipt says in process).
637
+ - **A replayed wake message is screened as the entry it replays.** When the wake message folds into the parked entry, the deployment's prompt screen runs once, on the parked entry's redelivery (post-claim), not on the wake message (pre-claim) — exactly what `setPendingSteer` of the same steer followed by a message-less wake does.
638
+ - **Hard-breaking (open to veto).**
639
+ - **The store contract gains the record-id obligation** — changed: every `MailboxStore` persists `id` and, inside the append's critical section, answers an id it holds through `judgeMailboxAppendById` (held seq / `mailbox.duplicate_record_id`) and refuses a malformed id with `mailbox.invalid_record_id`; who is hurt: a backend outside this package (@server's SQL store, @dgx if it ships one) — `mailboxStoreContract` has a new T1 case that is red until it does; migration: read the caller's record once with `readMailboxAppend` (it validates `peerMeta` and `id`), store the `id` column, constrain `(scope, handle, id)` among held rows, read the held rows in the append's transaction and call `judgeMailboxAppendById`.
640
+ - **The park seat re-asks** — changed: a rejected session-box append is appended again with the same record; who is hurt: a backend that has not met the obligation above writes a second record where today a second copy waited in process (the delivery count is the same, the twin moves into the box); migration: meet the obligation first.
641
+ - **A wake message under a parked `inputId`** — changed: identical ⇒ delivered once; different ⇒ `steering.duplicate_input_id` from `resume` before the claim; who is hurt: a host that resends a parked steer's id on a wake to deliver it again, or that branches on the wake's refusal codes; migration: a new instruction takes a new `inputId` (the `setPendingSteer` rule).
642
+ - **Exports.** +8 root values: `MAILBOX_INVALID_RECORD_ID_CODE`, `MAILBOX_DUPLICATE_RECORD_ID_CODE`, `MAILBOX_RECORD_ID_MAX_CHARS`, `MAILBOX_REPLAY_FACETS`, `readMailboxRecordId`, `readMailboxAppend` (one read of a caller's record: validated, detached — what a backend judges and stores), `judgeMailboxAppendById`, `judgeParkedSteerRepeat`; +1 type `MailboxReplayFacet`; `MailboxStoreErrorCode` +2 members. None removed.
643
+ - **Consumers.** @server: the SQL mailbox store implements the obligation (the kit case names it); the daemon's wake handler gets `steering.duplicate_input_id` from `resume` for a colliding wake message. @dgx: a store you ship runs the kit. @cli / @test: the `parked` receipt sentence changed; a wake that replays a parked steer delivers it once.
644
+ - **Pins.** `test/backlog988-park-idempotent.test.ts` — both bundled backends behind a store that loses the reply of a committed append: the seat answers `parked_durably` with one record under the frame's key and nothing in process (peer-words and engine-frame forms); a genuine fault parks in process with one disclosure; a second seat parking the same frame writes no second record; two frames keep two ids; the store's id judgment facet by facet, the malformed ids, the round trip and the acked-id-is-free cell; end to end over a real Runner per backend: the worker's receipt is durable and the leader's next run reads one cross-session frame and no `<agent-message>` twin; #2705 — replay delivered once, a different record refused pre-claim with the row untouched, a fresh id delivers both in order. `test/backlog969-main-park-seat.test.ts` re-pinned on the receipt sentence.
645
+
646
+ ### What a literal printer spells for a shell is read for what it refuses — `echo 'cat ~/.ssh/id_rsa' | bash` is refused by the deny table again; nothing that text reads clears the produced-text question (#1539 S1, design/1539, ruled C-R872; behavior narrowing, open to veto; hard-breaking: none; closed sets: none; exports: no new name — the exported type `ShellReadSegment` gains the optional member `printed`, and each entry of its `stdin` the optional member `printed`) (@server @cli @test)
647
+
648
+ - **Before.** Since 7.49.0 (#1505) a text a program produces for a shell — what a pipe carries, what a `<(…)` prints, a `< <(…)`
649
+ input — is not read at all: the call is one `unverifiable_shell_text` card (bypass-immune; a person may confirm it), whatever
650
+ the text says. 7.47.0 – 7.48.0 read a literal printer's text and refused `echo 'cat ~/.ssh/id_rsa' | bash` by the deny table;
651
+ that refusal became a card a person can approve. The literal reading was removed because it was used to EXEMPT harmless text,
652
+ and a redefined printer, a second writer on another descriptor or an inherited DEBUG trap made the stream differ from what the
653
+ printer's words say.
654
+ - **Now.** The read station also reads what a LITERAL PRINTER spells for the shell — one simple `echo` / `printf` of literal
655
+ words with no redirection (no backslash; a `printf` format with no `%` and no option), as the producer before a pipe, the body
656
+ of a `<(…)` script file or of a `< <(…)` input — and only a deny hit in it counts: that refuses the call; whatever else the
657
+ text finds, or fails to find, counts for nothing. The text is never known to be what the shell receives, so it is planned
658
+ beside the texts that run (`ShellReadSegment.printed`), never among them (`nested`), and judged on its spelling alone (no
659
+ environment is asked about it: a lookup it cannot finish cannot make the command's answer silence). The removal guard, the
660
+ standard-device marks, the kill guard, the write guards and the card do not read it, and the produced-text question stands
661
+ whatever it finds. The reading is monotone: it can only refuse — making the text look more dangerous only gets the call
662
+ refused. (The command's own re-entered texts — a `-c` string, a here-document, a here-string — keep their reading: their deny
663
+ relation, a hit or an unanswered lookup, counts.)
664
+ - **Behavior narrowing (open to veto).** *What changed*: `echo 'cat ~/.ssh/id_rsa' | bash`, `printf 'cat ~/.ssh/id_rsa' | sh`,
665
+ `bash <(echo 'cat ~/.ssh/id_rsa')`, `bash < <(printf 'cat ~/.ssh/id_rsa')` — a literal printer's text naming a deny-listed
666
+ path, on either face — are refused by the deny table on every arm (no card; the 7.48.0 refusal) where they were one
667
+ approvable card. *Who is hurt*: a command whose printer is redefined or whose stream is something else than the printer's words
668
+ (a function or an alias of `echo`, a trap), or whose shell does not run what it reads (`… | bash -n`, `… | bash 0<&-`), and
669
+ that would not read the path is refused too — the reading is the printer's words, not the stream; a path the text spells
670
+ through a link is not followed (the spelling is judged). *Compensation*: a text that does not read the path need not spell
671
+ it (drop the path from what the printer prints); a check that runs nothing (`bash -n`) can be run on a file; the deny table
672
+ is the deployment's to edit.
673
+ - **Unchanged (pinned).** A harmless printed text keeps its card (`echo 'echo hi' | bash`); every form the literal exemption
674
+ could not close (a function of the printer's name, `enable -n`, `BASH_ENV`, a second writer on another descriptor, an
675
+ inherited DEBUG trap, a here-string in the substitution body, a named `coproc`) keeps exactly one card, bypass included;
676
+ `echo 'rm -rf ~' | sh` stays `unverifiable_shell_text` (the removal guard reads what runs); a format the reader does not read
677
+ (`printf 'cat %s' ~/.ssh/id_rsa | bash`) keeps the card; a printed deny path no shell runs (`… | cat`) is no refusal.
678
+ - **Limits (stated).** Only the three producers above are read; `bash -c "$(echo …)"` and `… | xargs -I{} bash -c '{}'` are
679
+ candidate slices. What a stream carries beyond the printer's words (a trap's output, a second writer, a file's content) is not
680
+ read, and the printed text is judged on its spelling (a link in it is not followed: `echo 'cat ./link' | bash` stays the card
681
+ where `bash -c 'cat ./link'` is refused; a pattern is judged by its spelling, as the rows judge any pattern — `echo 'cat
682
+ ~/.s*/id_rsa' | bash` is refused): such a call stays the approvable card (KL-1417-c). The printed
683
+ text rides where the shell's input goes, so it is also read where another reader drains the stream first (`… | { read l;
684
+ bash; }`), where a here-document stands beside the pipe (`… | bash <<'EOF'`), and in a `-c` child that reads it (`… | bash
685
+ -c 'bash'`) — refusals in the safe direction.
686
+ - **Consumers.** @server: the exported type `ShellReadSegment` gains the optional member `printed` (additive: a value built
687
+ without it is the same plan; no engine seat hands a plan to a host). @cli: a refusal now arrives
688
+ where a card did for the forms above — the deny refusal's sentence and `read_path_denied` details, as for `bash -c '…'`. @test:
689
+ `core-1505-streamfed-shell-recheck` keeps its must-judge cell (bypass × an unread produced text is still a card); add
690
+ `echo 'cat ~/.ssh/id_rsa' | bash` ⇒ refused, no card.
691
+ - **Pins.** `test/backlog1539-printed-text.test.ts` (① the refusal back, five forms × three arms × two faces; ② seven hole forms
692
+ keep one card on every arm; ③ controls) and `test/backlog1539-printed-plan.test.ts` (④ the plan: `printed`, never `nested`; the
693
+ producers not read; `printedText`). Re-pinned: `backlog1417` ⑧ (`echo 'cat ~/.ssh/id_rsa' | bash` refused by the deny table,
694
+ as through 7.48.0).
695
+
696
+ ### A delegated child opens on `classic.SubagentStart`, not `classic.UserPromptSubmit`: the prompt a parent's model wrote for its child is no longer screened as a person's prompt (#1500, design/1500, C-R888; behavior narrowing; hard-breaking: two readings, three sentences each below (the prompt screen; the mod-package contract), open to veto; closed sets: one station row; exports: +2 types)
697
+
698
+ - **Before.** A delegated child's objective — the prompt the parent's model handed the Agent tool (and a workflow, fork or retained child's next prompt) — passed the child leg's `classic.UserPromptSubmit` screen as `source: "objective"`. A deployment's intercepting prompt filter therefore refused every child whose prompt it did not like, and its context was injected into children; the reference runs that hook only on what a person submits (CC 2.1.309: the prompt-submit path and cross-session / poll intake; the subagent runner raises no UserPromptSubmit). `classic.SubagentStart` was named in the event table but had no station.
699
+ - **Now.** Both ends of a leg read one table (`legEventsOf`; `stopEventOf` is retired into it). The delegation fact picks the lifecycle: a delegated child opens on `classic.SubagentStart` and ends on `classic.SubagentStop`, a top-level leg ends on `classic.Stop`. Who wrote the objective picks the screen: a person's prompt passes `classic.UserPromptSubmit` — a top-level leg's objective, and a retained child's resume through its steer handle (the deployment routes an operator's words there; a trusted run fact, `RunInternals.operatorObjective`, says so; that leg is screened first and then opens on SubagentStart); a delegated objective its delegator wrote (a parent model's Agent prompt, a workflow script's, a peer's SendMessage resume) is not screened as a person's. `classic.SubagentStart` is raised on the child's own leg as it opens on its objective, before its first request — never on a top-level leg, never on a checkpoint resume. Its `e` carries `agentId` and `agentType` (both pinned; `agentType` is absent on a workflow-lane agent) and the base; no prompt. Its answer is `{ additionalContext? }`, delivered ahead of the child's prompt as a `<system-reminder>` (the shape UserPromptSubmit's context had). It never keeps a child from starting (CC: a SubagentStart hook "can't block subagent creation"): failure column `skip`, abort `immediate`; a handler that throws or overruns is skipped and reported on `onError` (`phase: "hook"`); an answer carrying `block` is an answer the station cannot take — skipped and reported, the sentence pointing at `classic.PreToolUse`; a cancel during it ends the child as any cancel does (no `prompt_blocked_by_hook`).
700
+ - **A resume's origin is its caller's.** `createSubagentResume` hands out one face per origin: the face it returns speaks for its creator (a SendMessage: a peer; otherwise an operator), and the steer handle a resumed cycle re-emits is always the operator's face. Before, a handle re-emitted by a SendMessage-woken cycle reused the peer's face, so an operator's words sent through it were framed as a teammate's (`teammate resume`) — and, with this change, would have skipped the person's-prompt screen; now they are framed `operator resume` and screened.
701
+ - **Where a child's prompt is refused.** The parent's `classic.PreToolUse` on the delegating `Agent` call reads `input.prompt`; its deny keeps the child from starting and answers the parent, and a throwing handler refuses the call (fail-closed) — the reference's seat, unchanged here, now the documented one. A workflow agent's individual prompt has no such seat until `agent.spawn` (#2457).
702
+ - **Hard-breaking, open to veto (@server @cli @client-core).** What changed: `classic.UserPromptSubmit` no longer sees a delegated objective its delegator wrote (an operator's resume of a retained child is still screened). Who is hurt: a deployment that screened or refused child prompts, or injected context into children, through UserPromptSubmit (a settings-hook bridge mapping the reference's `UserPromptSubmit` command hooks onto `classic.UserPromptSubmit` now matches the reference's own reach; a probe pinning `source: "objective"` on a child leg turns red). What to read instead: refuse at `classic.PreToolUse` with `tool === "Agent"` (`input.prompt`); inject at `classic.SubagentStart` (`agentType` is the reference's `agent_type` matcher word) — a bridge carrying the reference's `SubagentStart` hooks has a station to land on from this release.
703
+ - **Closed sets / contract.** `MOD_EVENT_TABLE["classic.SubagentStart"]` gains its station (`skip` · `immediate` · pinned `agentId`, `agentType`); `ClassicEventMap` gains the key; `KERNEL_CONTRACT_VERSION` 4 → 5 (`@sema-agent/mods-kernel`: a host comparing two kernel copies for equality sees the raise; 7.58.0 shipped 4). A consumer holding the set of raised events (client-core's generated table) gains one row.
704
+ - **Hard-breaking, open to veto: a mod package written against contract 4 is refused (@server @cli @test @client-core).** What changed: `loadModPackage` loads only a package whose `hooks.json` states `"contract": 5` — any other value is refused with `config.mods_invalid` naming `hooks.json` and the contract (`src/core/mods/package.ts:72-73`); the 7.58.0 engine's kernel was contract 4, so every package written for it is now a refused one. Who is hurt: every mod package written for 7.58.0 on 7.59.0 — a deployment loading a real package (server S-2388), the CLI's package leg (MODS155) and test's real-package L1 fixtures (C-R915); this repo's own three fixtures (`test/fixtures/mods/{diff,sec-default,telemetry}/hooks/hooks.json`) were raised to 5 for the same reason. What to do instead: set `"contract": 5` in the package's `hooks.json`; contract 5 adds only the `classic.SubagentStart` station (`src/core/mods/kernel/events.ts:28-29`), so a package whose handlers do not name that event needs no other change.
705
+ - **Exports.** +2 types: `SubagentStartEvent`, `SubagentStartResult`.
706
+ - **Known limit.** KL-1500-a: a retained child's every new leg re-raises SubagentStart and prepends its context again (the reference injects once while the copy is still in the child's context) — the same count UserPromptSubmit gave those legs before.
707
+ - **Pins.** `test/backlog1500-subagent-prompt-hooks.test.ts` — real Runner, a parent delegating once: an intercepting UserPromptSubmit filter screens the root only and the child starts (base: it screened and refused `CHILD go`); the root's screen still fails closed; SubagentStart raised once for the child with its id, type and identity, never for the root; its context ahead of the child's prompt and never in the root's; `agentType` pinned; `block`, a throw and an overrun each skipped and reported with the child started; a cancel ends the child as an aborted run, never as a held prompt; PreToolUse on the Agent call reads the prompt and its deny or throw keeps the child from starting; a fork child opens on SubagentStart unscreened; an operator's resume through the steer handle is screened (its block keeps the words from the child and the leg never opens) and, let through, opens the leg on SubagentStart again — screen first, start second, the start's context leading; a SendMessage-woken background child opens on SubagentStart unscreened, and the handle that cycle re-emits screens an operator's words; the table's ends; the station's columns and the contract number. `test/mods-classic-redo.test.ts` S3 reads the stop end off the same table; `test/delegation-lifecycle-seat.test.ts` makes its non-completed child fail at the model instead of blocking it with a UserPromptSubmit filter.
708
+
709
+ ### A stream frame is read once, by the wire's shape, at the parser's door — a member in no shape is absent and said, never coerced, dereferenced or run (#1529, design/1529; behavior change and security narrowing, open to veto; hard-breaking: none; closed sets: none; exports: none — the reader lives in `src/brain/stream-shared.ts`, not index-exported) (@server @cli @test)
710
+
711
+ - **Before.** Each read site of an OpenAI-family chunk and an Anthropic event guarded (or did not guard) its own member: a `data: null` frame, a `content_block_delta` with no `delta`, a `content_block_start` with no block or `tool_calls: [null]` threw inside the parser and the turn failed with its streamed text and usage lost — after the stream's end too, losing a finished turn; a `choices` object, a `tool_calls` object, a `usage` number were dropped in silence; `delta.content` as an array became `"[object Object]"` in the answer, a `text_delta` number became digits; a usage count that was a string reached the usage seats as a string (`totalTokens` `"1002"`); an argument fragment that was a number or an object was coerced into the call's arguments — a numeric fragment between two string fragments ran the call with arguments the model never sent (both wires); an Anthropic `stop_reason` that was not a string ended the turn as a silent clean stop.
712
+ - **Now.** The OpenAI-family parser reads each chunk once (`readChunk`), the Anthropic parser each event once (`readEvent`), through one member reader: an object, an array, a string, a count — `null` and absent state nothing; any other value reads as absent and its loss is said where that member's loss is already said. A content-bearing member in no shape (or a frame that is not an object) counts the frame with the unparseable ones — the one dropped-frame note to the model (its sentence now reads "were not valid JSON or not in the stream's frame shape; what could not be read was dropped") and the person's `task.tool_turn_incomplete` (`bad_frame`). A call-stream member in no shape — or a container of the call stream (`choices`, the choice, its `delta`; an Anthropic block frame not placed on a text / thinking block, an event object with no `type` word) — leaves none of the response's still-open calls runnable: a call that closed in full before keeps its cached call, every call still open lands in the invalid-arguments drop. An Anthropic `error` event's envelope members are read as words (a member in no shape no longer throws while the error is worded). A usage in no shape makes the turn's usage unknown (`usageMissing`). An ending seat holding a non-word is an ending in no word (`?`) on both wires — the Anthropic wire now carries `finishReasonUnknown: "?"` and the run announces `route.finish_reason_unknown`. After the stream's first ending a frame that offered nothing readable and held a member in no shape is refused like any late frame (`route.stream_end_ignored`, content `?`); a late offer is still named by what it offered. A falsy `finish_reason` still states no ending, and a provider that spells its stop word `?` is a word, not a non-word; a stream whose every frame is in its shape reads as before (pinned on both wires, including the nulls real gateways send).
713
+ - **Behavior change (open to veto).** More turns carry the dropped-frame note and `task.tool_turn_incomplete` (`bad_frame`); more turns carry `usageMissing`; the Anthropic wire can now raise `route.finish_reason_unknown` (`?` only — an unknown `stop_reason` WORD still folds to a stop, a named follow-up). **Security narrowing (open to veto).** A tool call whose argument stream held a fragment in no shape no longer runs.
714
+ - **Consumers.** No new code, key or detail member: `task.tool_turn_incomplete`, `route.finish_reason_unknown`, `route.stream_end_ignored` and `turn_end.usageMissing` carry their existing shapes on more occasions. @server @cli: a pin or renderer that matched the old note words in full (`"were not valid JSON and were dropped"`) reads the new sentence; the first words (`"N provider stream frame(s) were not valid JSON"`) are unchanged. @test: the fake-gateway cells below are the black-box forms.
715
+ - **Pins.** `test/backlog1529-chunk-decoder.test.ts` — a local fake gateway fed frame by frame through the real `fetch`: O1–O14 (OpenAI-family: `data: null` before / after the end, the r4 / r7 controls, `choices` / `tool_calls` / `usage` / `delta.content` / a usage count / an argument fragment in no shape, one note for many frames and one person notice per run, an all-in-shape control, the gateway nulls control) and A1–A12 (Anthropic: `data: null`, a missing `delta` / `content_block`, a numeric `text`, a numeric `stop_reason`, a numeric usage, a numeric `partial_json`, a typeless event carrying a fragment, an `error` envelope in no shape, a late `message_delta` in no shape, a late `data: null`, a reference-endpoint control, a control); S1 the reader, the ending word and the one sentence. Re-pinned: `test/backlog7490-brain.test.ts` (a `tool_calls: {}` frame — the turn still goes on and nothing runs; it now carries the note).
716
+
717
+ ### A credential-origin refusal names the no-credential door — the request's fix hint and the run's announcement carry one clause: an endpoint that takes no credential is handed none (#1853; no behaviour change — the wording of two refusal texts; hard-breaking: none; closed sets: none; exports: none in the package root)
718
+
719
+ - **Before.** `route.credential_origin_mismatch` told an operator two ways out — change the entry's credential reference together with its `baseUrl`, or point the entry back at the bound origin — and, for a provider that only moved domains, a restart. A local multi-node catalog that hands every node the placeholder word `EMPTY` met the refusal on its second node and neither way applied: the nodes need no credential, and the way out the law serves (hand the node an empty key; the integration guide says "none where the endpoint needs none", the refusal said nothing) was in neither text. The run's announcement was worse: it said the deployment "has to give the model a credential for" the new host.
720
+ - **Now.** One clause, `KEYLESS_EXIT` (`src/brain/route-adjudicator.ts`), is in both texts: the verdict's `fixHint` — the text of the request's own failure (`[invalid_request] route.credential_origin_mismatch: …`) and the `fixHint` (and the sentence that ends the message) of a derived seat's fallback notice — reads "…or point the entry back at <origin>; if the endpoint takes no credential, hand it none — answer an empty apiKey and no credential header (this works under any brain; a brain configured with neither a baseUrl nor a credential of its own also accepts no apiKey at all), because a placeholder word handed in its place is still a credential and binds like any other; a binding lasts as long as the engine process, …"; the run's announcement (`src/core/runner/credential-origin-seat.ts`) ends its advice with "If <origin> takes no credential, <the same clause>." The three older suggestions are unchanged word for word.
721
+ - **What it does not say.** The clause is the same for every refusal of this code: the engine cannot tell a placeholder from a real key by its spelling and does not try (a short real key and `EMPTY` bind alike), so the text is advice, not a diagnosis. The door it names is read off the law: an empty `apiKey` is a per-model answer with no secret, so it binds nothing, never falls back to the brain's own key or root, and none of the three lanes writes a key header for it; leaving the answer out reaches the same door only under a brain with no credential of its own whose `baseUrl` is absent or is the endpoint itself — under a brain with a root on another host it is refused `route.credential_missing`, under one with a key or credential header meant for another host `route.credential_mismatch`, and under one whose key or credential header is paired with this very endpoint it is the brain's own credential that goes out. The entry and the hook's answer must carry no credential header either: a placeholder carried in a header binds like one carried in the key. The other pairing refusals (`route.credential_mismatch`, `route.credential_missing`, `route.credential_answer_conflict`) keep their hints.
722
+ - **Known limits.** KL-1821-c stays: a value handed over on two origins still refuses the one sent second, `EMPTY` included — that is the law, unchanged; the row now says the refusal names the door.
723
+ - **Consumers.** No one is asked to change. @server @cli @web-admin: a program that matches `fixHint` or the announcement's `message` as text matches each older sentence still (kept word for word and in order; one clause is inserted, so a program that compares the whole text sees it); `detail` is unchanged (`{ modelId, boundOrigin, requestOrigin, sessionId, runId }`).
724
+ - **Pins.** `test/backlog1821-credential-origin-binding.test.ts` K1 – K5 (the second `EMPTY` node's request fails naming the door, and both doors the text names serve it under a bare brain with no credential header on the wire; the hint holds the clause beside the three older suggestions verbatim and the other three pairing hints do not hold it; the real Runner's failure and its one user-audience notice carry the same clause, the notice's `detail` keys unchanged; K4 — under seven brain configurations (bare; a root on another host, alone / with a key / with a credential header; the endpoint itself as the root, alone / with a key / with a credential header), with the placeholder carried in the hook's `apiKey`, in its `Authorization` header and in the entry's own — the second node is refused naming the door, the empty key serves it under every brain with no credential header and none of the brain's own, and leaving the answer out does what the law says for that brain; K5 — the same seven configurations on the openai, anthropic and open-responses lanes against a stubbed transport).
725
+
726
+ ### A blank credential is the keyless door, one reading for every source — a key or a credential header with nothing but white space in it is no credential: its source keeps its rank and nothing of it goes out, and the route a refused request states is what it carried (#1882, ruled C-R894 and its note 1; behavior change, open to veto; hard-breaking: none — one exported function's output shape narrows, open to veto; closed sets: none; exports: +1 root value `isBlankCredential`) (@server @dgx @web-admin)
727
+
728
+ - **Before.** Three readings of "an empty credential". The empty `apiKey` (`""`) was the keyless door: the hook answer or `config.apiKey` that held it kept its rank, no key header went out, nothing was bound. A key of white space only (`" "`, `"\t\n"`) was a key: the lanes' truthy lock wrote it (`Authorization: Bearer <spaces>`, which `fetch` sends as `Authorization: Bearer`; `x-api-key: <spaces>`, sent empty; a non-Latin-1 space made `fetch` throw and the request failed), and beside the lane's own key header it was a conflict (`route.credential_answer_conflict`). A credential header whose value was blank rode as an empty header. The withheld table's entry rule already read white space only as no credential, and the memory distiller's chat seat refused `""` but took a white-space key.
729
+ - **Now.** One predicate, `isBlankCredential` (a string `String.prototype.trim` empties — the empty string, white space only), read everywhere the question is asked. `resolveRouteCredential` reads what was handed ONCE: the rank is still the source's presence (a key handed, a credential header named — no fall-through), and the credential it returns holds only what rides — a blank key is no `apiKey`, a bag whose credential headers are all blank is no `headersBorne` (`{ source: "per-model" }` carries nothing). So every first-party lane's key lock writes no header for a blank key, no identity is bound, a blank key beside the lane's own key header conflicts with nothing (the header is the credential and rides), the header strip keeps a blank credential header off the wire whichever source it came from, and the route a request states (`requestRouteOf` → `AssistantMessage.apiErrorRoute`, `terminal.detail.route`) is `{ credentialSource: "absent" }` when nothing went out. The withheld table's entry rule, the credential intake (a blank value is not said: no request carries it) and the distiller's chat seat (a blank key is no key) read the same predicate. A key that is not blank rides as written, padding included. The published conformance corpus is unchanged except the `per-model-empty-string-key` intent sentence; its expectations are the same.
730
+ - **Who sees a difference.** A request handed a white-space key (per-call, from the hook, or as `config.apiKey`) now goes out with no key header instead of a junk one; a white-space key beside the key header now rides the header instead of being refused; a blank credential header no longer goes out empty; a provider refusal after a blank key or a blank credential header states the absent route (it said `carrier: "apiKey"` / `"header"`); the distiller's chat seat refuses a white-space key as it refuses `""`; a white-space `config.apiKey` no longer prints the `config.credential_not_withheld` line. `resolveRouteCredential({ optionsApiKey: "" })` returns `{ source: "per-model" }` (it returned `{ source: "per-model", apiKey: "" }`), and a blank `configApiKey` likewise loses its `apiKey` member (open to veto).
731
+ - **Consumers.** @server: `src/hooks/hook-llm.ts` writes `x-api-key` from `credential.apiKey ?? ""` for a per-model credential — the same empty value as before for a blank key; `model-route-gap.ts` reads `""` as unset but white space as set (`declaredValue`), and reading the same predicate (`isBlankCredential` from the root) makes its env reading the engine's. @dgx: the keyless door for several local nodes is unchanged — hand no key, or a blank one; white space now reads as `""` does. @web-admin: no action; a catalog field saved as white space is keyless, as an empty one is.
732
+ - **Pins.** `test/backlog1882-blank-credential.test.ts` — real Runner over real local listeners: a blank hook key (three spellings), a blank `config.apiKey`, a blank hook credential header and a blank `config.headers` credential header are served with no credential header and no deployment key; a blank key beside the key header rides the header; a blank key and a blank credential header bind nothing (the entry moves origin and is served keyless twice); controls: no key ⇒ the deployment key rides, `""` reads as white space does, a padded key rides; every lane's lock and every lane's stated route after a 401 (OpenAI, Anthropic, Open Responses); the predicate from the package root, `keyCarriedBy`, resolution's rank and carried shape, `requestRouteOf` on a hand-built credential, the withheld table's refusal and the distiller seat's. Re-pinned: `route-adjudication` (a blank per-call key resolves to `{ source: "per-model" }`), `backlog1754-held-values` (a blank construction key is not said). `backlog1399-terminal-detail-projection`'s "a route that is not absent leaked a credential" cell now holds for blank keys too.
733
+
734
+ ### A host moves a per-model credential's origin binding without a restart — `rebindCredentialOrigin(model, perModelAuth, config, onNotice?)`, announced to the operator as `route.credential_rebound` (#1844; additive verb; hard-breaking: one notice word, two sentences reworded; open to veto)
735
+
736
+ - **Before.** (7.52.0, #1821) A per-model credential (the per-model hook's key, a credential header on the entry) is bound to the origin of the first request the engine sends carrying it; a catalog that moves the entry to another host with the same credential is refused there (`route.credential_origin_mismatch`). Inside a process nothing could move a binding: a provider that only changed domains and accepts the same key on the new one needed an engine restart — also after the control plane's audited "this key is valid for the new address" (web-admin `rebindCredential: true`), which the engine never saw.
737
+ - **Now.** `rebindCredentialOrigin(model, perModelAuth, config, onNotice?)` (package root) is the one way a binding moves inside a process. The host hands the entry as it is in force now (`id`, `baseUrl`, `headers`), the per-model hook's answer the request will get for it (`{ apiKey?, headers? }`, or `undefined` when the entry's own credential header is the credential) and the brain's pairing config (`{ baseUrl?, laneDefault? }`, as for `admitRouteRequest`). The engine resolves the credential itself, by the request's own resolution (the answer outranks the entry's headers; the brain's deployment credential is never the verb's), and binds every secret it carries — the key and each credential header of the winning bag — to the origin the entry's request goes to: the root judged by the brains' own root judge (`resolveRequestRootOf`), the origin folded as the request gate folds it. It returns `{ origin, previous, changed }` (display spellings). A call that changed the table announces `route.credential_rebound` once — audience `operator`, minter `engine`, `detail: { modelId, origin, previous, occurrenceId }`, through `onNotice` (else as the notice's line on `console.warn`), never a credential or a digest of one; a call that finds every secret already there writes and announces nothing. Refused with `config.invalid_argument`, nothing written: an entry with no per-model credential, an answer that carries no secret (the keyless door), an entry whose request has no servable root (none at all, a non-http(s) root, a fragment, userinfo — the lane default the host hands judged like the entry's and the brain's roots when it is the root chosen). A non-string id, URL half or key is a `TypeError`. The notice is delivered just before the write, in the same synchronous step: a console that throws in the no-sink arm leaves the table as it was and the call throws (a retry rebinds). The rebind moves the credential, not the entry: another entry carrying the same value and still targeting the old host is refused from then on (one credential, one root, KL-1821-c). A request the gate judged before the rebind and commits after it is judged again at its commit and refused there.
738
+ - **Unchanged.** Nothing a request, a catalog swap or an entry's own members carry moves a binding: re-publishing the moved catalog stays refused (the #1821 W2 shape, kept as this change's control). `adjudicateModelRoute` and `routePairingStatus` stay reads. `__resetCredentialOriginBindings()` stays the test seam (it empties the whole table; the rebind moves one credential). A restart still empties the table.
739
+ - **Hard-breaking.** (1) *What changed:* `EngineNoticeCode` gains `route.credential_rebound`. *Who is hurt:* a consumer that keys a record by `EngineNoticeCode` — a notice allowlist or router, an audience table, a renderer — gets a compile error until it gives the word a row. *Migration:* add the row (operator; minted outside any run, so it carries no `sessionId` / `runId`). (2) *What changed:* two sentences say the binding's other source and the new way out: the refusal's fix (`route.credential_origin_mismatch`'s `fixHint`, carried in the request's failure text) now reads "bound to X, the origin it was first sent to or the host rebound it to … needs the host's explicit rebind (rebindCredentialOrigin) where the host offers one, otherwise an engine restart", and the run's `route.credential_origin_mismatch` notice sentence says "(the host it was first sent to, or the one the deployment rebound it to)". *Who is hurt:* a test or reader that pins those sentences' words. *Migration:* re-pin; codes, audiences and every `detail` member are unchanged.
740
+ - **Consumers.** @server: when applying a catalog whose control plane recorded an audited rebind for an entry, call `rebindCredentialOrigin(entry, keyResolver(entry), { baseUrl: declaredRoot })` with the entry of the catalog generation in force (dedupe by the control plane's own event; the verb's target is the entry as handed, so a stale copy rebinds to a stale host, KL-1844-a), and give `route.credential_rebound` its notice-wire row (operator, logs only). @web-admin: publish the audited `rebindCredential` act as a signal the server can read beside the catalog (which cell is the control plane's to choose); until the server consumes it, the console's "confirm rebind" is still met by the engine's refusal until restart. @dgx: a host that runs the engine directly can call the verb; `EngineNoticeCode` +1.
741
+ - **Exports.** +2 root names: `rebindCredentialOrigin` (function, tier advanced), `CredentialRebind` (type). Nothing removed.
742
+ - **Limits (stated).** KL-1844-a (ordering and replay are the host's), KL-1844-b (the notice is the table's account, not an audit of attempts; the call's return value is the result), KL-1844-c (the rebind shares the table's one write, so it refreshes and, at the 16 384 bound, evicts like a send). KL-1821-a (no in-process rebind) is closed.
743
+ - **Pins.** `test/backlog1844-credential-rebind.test.ts` (real Runner, real OpenAI brain, real local listeners: R1 moved → refused → rebound → served at Y with the key, X untouched, one operator notice; R2 a rebind moves, pointing back is refused naming Y; R3 control: no explicit act, three re-publishes stay refused, nothing announced; R4 a second rebind changes and announces nothing; R5 rebind ahead of the first send; R6 the refusal arms and the `TypeError`s write nothing; R7 a shared key moves for both entries; R8 the fix names the rebind and the restart; R9 key + gateway header both move; R10 judged before, committed after ⇒ refused at commit; R11 an entry with no root rebinds to the brain's declared root; R12 the `console.warn` arm, the audience and minter rows; R13 an entry's own credential header; R14 the hook's answer outranks the entry's header; R15 a console that throws in the no-sink arm leaves the table unmoved and a retry rebinds; R6 also refuses a lane default that is not servable). Fourteen of the fifteen are red on the base tree (the verb did not exist); R3 is green there by design — it is the positive control for the rule this change must not move (a mutation that lets a request move its binding turns it red).
744
+
745
+ ### Where the shell releases end a substitution body is read off their own parsers: the 4.0 – 5.1 counting pass is ported, every `$( )` / `<( )` / `>( )` / `${ list; }` body is compared, and an arithmetic command reads no here-document (#2792, #2787, #2788, #2789; behavior narrowing and security narrowing, open to veto; hard-breaking: none; closed sets: none; exports: none) (@server @cli @test)
746
+
747
+ - **Before.** The previous section's comparison read the counting pass of bash 4.0 – 5.1 through four approximations, and a reading that ran off the end of the text counted as agreement — so where an approximation was wrong the text was read as data although a release closes the body elsewhere and runs what follows (a document line ending in a backslash, a comment line right after a comment that followed a word, a comment right after a nested substitution, a quote in a here-document header's comment — that last one introduced by the previous section). A body holding no here-document was not compared at all, and the same comment rule leaves such bodies open to the same split. Bash 3.2 – 5.2 read `${ list; }` as a parameter expansion ending at its first `}` and run the lines after it; a `$((` that is not arithmetic was delimited with comments read, where 4.2 and later count none; a `<( )` was not compared with the scanner 3.2 – 4.2 re-read it with as the word expands; `((…))` was read as holding here-documents. Separately, the previous section refused texts every release delivers alike (a `#` comment after a word holding a `)`, a pair of quotes after a commented `)` in double quotes).
748
+ - **Now.** One rule, every body: a body is read only where each release's reading closes it at the same place, or never closes it (that release refuses the whole command). Bash 4.0 – 5.1's counting pass is ported state for state from their parsers (`src/shell/counting-pass.ts`, one flag row per release; a state it does not follow claims no close), 3.2's count and its expansion-time scanner, the count with no comment where a release reads so (`$((`, `<( )` / `>( )`), and the releases before 5.3 for `${ list; }`; bash 5.2 and later are the lexer's own reading. Two places the releases rewrite a body before it runs are refused outright: a backslash-newline in a here-document line of the body, and a comment ending in one. An arithmetic command — bash's own test, its matched pair followed by `)` — reads no here-document and no comment (two subshell parentheses still do). In double quotes, 3.2's early close is harmless only while its double-quoted word runs on to the same close. A refusal where the readings split a body the full reading reads a here-document in carries `undelimited_substitution_heredoc` (now also for `${ list; }`), whose sentence says the way out; a comment ending in a backslash-newline keeps `undelimited_substitution`.
749
+ - **Measured.** Ten binaries (3.2.57, 4.0 – 4.4, 5.0 – 5.2 built from the GNU releases with every published patch, 5.3.15) over the review's shapes, this change's table (`test/fixtures/backlog2792-measured.json`) and three random generators: no text the lexer reads has a release run what the reading holds as data (counts in the car receipt); the baseline reads such texts on every generator. Limits: KL-2792-a (the two rewrite places are refused even where every release agrees), KL-2792-b (the model's boundary). KL-1547-a is closed.
750
+ - **Who is hurt.** A command whose substitution body the releases split differently is refused where it was read; a commit message with a line ending in a backslash inside `$(cat <<'EOF' … )` is refused (way out: `git commit -F -`). In the loosening direction, the previous section's refusals of a comment after a word and of quote pairs after a commented `)` in double quotes read again; `((x<<2))` and `for ((…<<…))` read as arithmetic. @server @cli: the `unverifiable_shell_text` card text (the heredoc sentence is reworded, still naming `git commit -F -`). @test: the table is the differential to re-run on another host's bash.
751
+ - **Pins.** `test/backlog2792-lexer-shapes.test.ts` (the measured table, the host's bashes live, what every release delivers read, the release table and its merging, structure, a 1 MiB body in linear time, a seeded random differential); `test/backlog1547-heredoc-prefix.test.ts` re-pinned where the previous section's approximations pinned the wrong reading (its `${ list; }` block, a double-quoted row) and its structure cell; `test/backlog994-s2-one-lexer.test.ts` FIX4 re-pinned (the `${ list; }` forms the releases before 5.3 end at an earlier `}` are refused; 5.3's own reading of them is unchanged); `test/backlog1504-shell-words.test.ts` +1 role of the one word table (`pass_keeps_reserved`, the four-letter words the 4.4 – 5.1 pass keeps its reserved-word state after; internal, not on the package root); the #1102 equivalence ledgers' one entry re-hashed for the reworded sentence; the per-mutation table is in the car receipt.
752
+ - **Linear time (#2867).** The comparison first asked about the rest of the text once per adjacent `((` (where the pair it opens closes) and once per body (which releases read the rest alike, whether a `#` follows): a run of 9 999 `(` took 2 s where it took 4 ms. Each text is now read once for all of them, for as long as the lexer reads it (a compound value or a here-document opens its own; nothing is kept afterwards); no reading changes (the lexer's output is byte-identical to the previous head on 3 182 texts). Pins: `test/backlog2792-lexer-shapes.test.ts` Z (a run of `(`, a `((` that opens no arithmetic command, many bodies, and bodies between compound values: each under a ratio bound for a text four times longer and under three times the earlier tree's time) and R (the per-text answers equal a fresh read, three places where something is carried over, a trigger a backslash-newline splits).
753
+
754
+ ### A here-document in a substitution body is read by every bash version it can run under — a line that merely begins with the delimiter is data unless a `)` follows it, the counting passes of bash 4.0 – 5.1 are compared with the full reading, and the refusal says the way out (`git commit -F -`) (#1547, #2777; behavior narrowing and security narrowing, open to veto; hard-breaking: none; closed sets: +1 `ParseFailureReason` member `undelimited_substitution_heredoc` (not on the package root); exports: none) (@server @cli @test)
755
+
756
+ - **Before.** Any line in a here-document body inside `$( )` / `<( )` / `>( )` that began with the document's delimiter without being it made the whole command text unreadable, whatever the line said (`EOF marker handling changed`), and the card for it (`unverifiable_shell_text`) did not say how to write the text so it is read. Separately, a body whose here-documents share one header line was read as data by the lexer where bash 4.1 – 5.1 execute part of it.
757
+ - **Now.** The rule is measured, not assumed: ten real binaries (3.2.57, 4.0 – 4.4, 5.0 – 5.2 built from the GNU releases with every published patch, 5.3.15) were run over 61 body shapes and the table is the pin fixture (`test/fixtures/backlog1547-heredoc-measured.json`). A substitution body that holds a here-document closes only where every reading closes it at the same `)` — bash 5.x's full reading, bash 4.0 – 5.1's counting passes (as ported by the section above, which replaces this section's approximations of them), and bash 3.2's count — or where one of them never closes it (that bash refuses the text). In a body in double quotes the bash 3.2 parser reads no comment, and that count is compared too (its early close is harmless only while its double-quoted word runs on to the same close; the section above). A line that begins with the delimiter and holds a `)` anywhere after it stays refused (bash 5.2 and 5.3 end the document there and run the rest of the line); a line that begins with the delimiter and holds none is data and is read as data. In a `${ list; }` body (bash 5.3) a line that begins with the delimiter and is followed, after blanks, by a `}` is refused too (reason `undelimited_substitution_heredoc`, since the section above). A text no reading can place is refused, never read as data. The security narrowing holds as measured with the section above (its generators and shapes: no text the lexer reads runs on any release); what is left is KL-2792-a / b.
758
+ - **The refusal's own reason.** Every refusal of a here-document in a substitution body now carries `undelimited_substitution_heredoc` (the other undelimited forms — a `case` pattern, a body nested past the bound — keep `undelimited_substitution`), and its sentence — the card's account — says the way out: feed the text as stdin outside the substitution (`git commit -F -`), or reword the line / change the delimiter. It fits the card's 200-character cut.
759
+ - **Consumers.** @server @cli: the `unverifiable_shell_text` card text for these texts changes (it now names `git commit -F -`); a commit written `git commit -m "$(cat <<'EOF' … EOF)"` whose message has a line beginning with the delimiter and no `)` after it is read and no longer draws that card. @test: the pin fixture's table is the differential to re-run on another bash if your host has one the table lacks.
760
+ - **Pins.** `test/backlog1547-heredoc-prefix.test.ts` — the measured table (every shape × every version, one cell per shape: the lexer reads a text only where no measured version runs its marker), the same differential live on every bash the host has (versions it lacks skip with their column recorded), the lexer's reading of the shapes, the way out (reads, and runs as data on every host bash), the card through the real Runner lane in three arms, and a seeded random differential; the 1542 and 753 pins read the new reason.
761
+
762
+ ### The store that carries a memory scope has one reader, and the root exports it — a host's sync leg and a session now read and write a typed personal scope in the same store (#2818, #900 C2b; additive export; exports: +1 `resolveMemoryScopeStore`, +2 types)
763
+
764
+ - **Before.** On a deployment whose backend is a directory-pinned File store (`<root>/memory`), a session under the v2 scope contract runs its personal-plane scopes (`user:` / `userproj:` / `org:`) on a SECOND File store, `<root>/personal/memory` (the in-repo projection guard: personal memory must never be projected under a repo root). That choice was a closure inside the session assembly. A host leg working outside a session (the single-user sync executor, the operator faces) had no way to ask it and read "the deployment's backend" instead: with `MEMORY_SCOPE=user:<P>` and sync on, a round pushed nothing the sessions wrote into the global layer (or an older copy, if the pinned store held one), and pulled entries landed in the store no session reads. There is no workaround: sync on requires a typed `MEMORY_SCOPE`.
765
+ - **Now.** `resolveMemoryScopeStore({ scope, scopeContract, backend, memoryEngineDir, onIncident? })` → `{ plane: "project" | "personal", backend }` answers "which store holds this scope's entries", and the session assembly uses the same reading: `src/core/memory-engine/dual-root.ts` holds the plane reading (`scopePlanesOf`) and the plane's store (`memoryPlaneStore`), and `prepare-memory.ts` builds no store of its own any more. The session face and the sync leg resolve a scope to one store from now on. `backend` and `memoryEngineDir` are the values given as `RunnerDeps.memoryBackend` / `RunnerDeps.memoryEngineDir`; `scopeContract` is a required key (`undefined` written out), because it is what moves a typed key to the personal plane. On a directory-pinned deployment a personal-plane answer is a freshly constructed File store over the personal mount (resolve once per scope and keep it); every other answer is the deployment's backend itself.
766
+ - **Bad input is refused, never folded to the deployment's backend** (`config.memory_scope_key`): a blank or non-string key, a contract other than `"v2"` / `undefined`, a left-out `scopeContract` or `memoryEngineDir` key (`undefined` written out is the statement of absence), an engine dir that is not a non-empty string, and under v2 a typed key the scope contract cannot read (its own refusal kept as `cause`). The key is read through the Runner's own normalization, so a spelling a session trims and runs with resolves to the store that session uses. Without the opt-in a typed-looking key stays opaque and on the project plane — the same no-sniffing rule the session applies. Two values the Runner itself still reads silently — a `scopeContract` other than `"v2"` (read as no opt-in) and an empty `memoryEngineDir` (read as the working directory) — are refused here and recorded for the Runner's own door.
767
+ - **Not changed (stated).** Where a session's planes live (`<root>/personal/memory` + `<root>/personal/.engine`), which scopes route where, and every byte a session writes; the untyped default scope (`local`) keeps living in the deployment's backend.
768
+ - **Consumers.** @server S-2394: the sync executor (`createMemorySyncRunner`, today `backend: memoryEngine.backend`) and the three operator faces read the store per layer from `resolveMemoryScopeStore` — the global layer `user:<P>` resolves to the personal store, a host-minted project layer (`userrepo:…`, opaque under v2) to the deployment's backend; the layer enumeration of project layers keeps reading the deployment's backend (the plane they live on). Switching a layer's store relies on #2836 (next section, same release): the layer's old cursor is refused, and the round runs from no cursor. Until S-2394 lands, typed-scope sync stays broken on a pinned File deployment.
769
+ - **Pins.** `test/backlog2818-typed-scope-sync-store.test.ts` (the real Runner over a pinned File store laid out as a single-user host does: a typed session's write lands in the store the reader names, its ledger is that store's, a sync round over it pushes the entry and a pulled entry is served by the next session; the deployment's backend pushes nothing; untyped and both-planes layouts, the personal plane as the write face of a two-plane session; a store without a directory, through a real session; fifteen bad inputs refused with nothing constructed, padded spellings read as the Runner reads them; the session assembly holds no store construction of its own).
770
+
771
+ ### BREAKING — a memory sync cursor is bound to the local store it was taken against, and a round refuses a cursor from any other store (#2836; `MemorySyncCursor.store` required; behavior narrowing; hard-breaking: yes)
772
+
773
+ - **Before.** `MemorySyncCursor` was `{ scope, peer, baseRevs, updatedAtMs }`, and `syncMemoryScope` checked only scope and peer. A host that moved a layer to another local store (the typed global layer moving from the deployment's pinned store to the personal store, #2818) and kept the layer's persisted cursor ran the new store against the old store's baseline: every baselined id the new store lacks is "deleted here" to the push half (`src/core/memory-engine/sync-client.ts`, the baseline-stub reconcile), so the round sent DELETEs for entries the user still has on other devices.
774
+ - **Now.** A cursor carries `store` — the local File store's physical root (`file:<path>`, the path its root anchor records; minted once, by `memorySyncStoreOf`). `syncMemoryScope` mints the store identity of the backend it is given, binds every cursor it returns to it, and refuses a cursor for another scope, peer or store — or one with no store binding — with `config.memory_sync_cursor_mismatch`, before any read or transport call. The scope and peer checks are the same judge and now carry the same code (their sentences keep their opening words). A store without a directory has no root to bind a baseline to and is refused as a sync plane (`config.memory_sync_store`); the sync client's documented plane was always the File store. The store identity never goes on the wire. The pull half commits pinned to the root the cursor was judged against (`applyPatches(…, { anchoredRoot })`, the session engine's own discipline): a store re-pointed and re-anchored during the transport takes no entry (`memory.target_outside_store`). Known limit (stated): the identity is the root's physical path, so a different store later created at the very same path — the old one moved away with its cursor left behind — reads as the same store.
775
+ - **The host's answer to a refusal is a round from no cursor** (then persist the cursor that round returns). It pushes what the store holds and no deletes. Its one cost: a delete made locally since the last completed round has no baseline to travel on, so that entry comes back from the peer.
776
+ - **Hard-breaking, three sentences.** *What breaks*: `MemorySyncCursor` gains a required `store` member; a cursor persisted by an earlier release (no `store`) is refused once; a round over a backend without a directory is refused. *Who is hurt*: a host that rebuilds cursors from a fixed key list (it drops `store`, so every later round is refused) and any host that syncs a non-File backend. *Migration*: persist the returned cursor verbatim, `store` included; on `config.memory_sync_cursor_mismatch` run the round with no cursor and replace the persisted file with what it returns.
777
+ - **Consumers.** @server S-2394: `cursorOf` in `src/memory-sync-client.ts` rebuilds `{ scope, peer, baseRevs, updatedAtMs }` and must carry `store` through (shape check: a string; the comparison stays in core); the layer round answers the code by running from no cursor. After the upgrade each layer's first round is refused once (its cursor predates the binding), and the global layer's switch to the personal store (#2818) is refused for the same reason — both are the same one-time fresh round.
778
+ - **Pins.** `test/backlog2836-sync-cursor-store-binding.test.ts` (a minted cursor names its store and is accepted over that store, also through a linked spelling; the old cursor over the personal store is refused with zero transport calls and a fresh round pushes no deletes; an unbound or garbage `store` is refused; a refusal reads nothing from the store; a pull that arrives after the store was re-pointed and re-anchored lands nowhere; scope / peer mismatches carry the same code; a store without a directory is refused). `test/memory-sync-client.test.ts` re-pinned: its cursors are minted over the test's own store.
779
+
780
+ ### The store boundary is written down once — INTEGRATION-CORE §5.11–§5.13 say what core ships and what a deployment ships, list what a store-owning deployment reads from core at 7.58.0, and read the four packages against each other; four claims of a shipped database store are corrected (#2770; documentation only, no behaviour change; hard-breaking: none; closed sets: none; exports: none) (@server @dgx)
781
+
782
+ - **Before.** The line was implied across INTEGRATION-CORE §5.2–§5.7 and the CHANGELOG sections of #1468 S4, S4b and T4; no sentence said what core ships. Four places said more than was true: `docs/sdk/10-extension-points.md` §3 named `src/stores/pg.ts` (no such file) as ready-made PostgreSQL implementations, the root README offered Postgres as a `CheckpointStore` and `MemoryBackend` backing and counted two runtime dependencies (the package declares four), and `docs/ARCHITECTURE.md` drew `src/stores/pg` in its diagram. INTEGRATION-CORE carried two sections numbered 5.7.
783
+ - **Now.** §5.11 draws the line once: core owns the seat contracts, the facts a backend must not restate and the kits that accept a backend; it ships the in-memory defaults and the file family (`FileStorageBackend`) and **no database store** — the PostgreSQL / TiDB adapters under `src/examples/adapters/` are repository examples (not a root export, not in `dist/`), and `@sema-agent/core-stores-sql` carries the neutral factories without stores. §5.12 is the one list of what a deployment that owns session and checkpoint stores reads from core — each export with the reading it retires — followed by the five readings that changed in 7.58.0 (the owner-only import, the head an import answers and the conditional write, the two kit report words, the required `executor`, the retention word as a construction argument) and what a deployment that implements `SessionStore` itself, a local lane and the startup call each do. §5.13 reads the four packages against each other: one boundary sentence and one install form each, the exact-version rule between them, and that core does not depend on `@sema-agent/mods-kernel`. The second §5.7 heading is §5.7a. The four stale claims above are corrected, and the `@sema-agent/core-stores-sql` and `@sema-agent/mods-kernel` READMEs each gain the one sentence that places them against core.
784
+ - **Consumers.** @server: §5.12 is the list to read the 7.58.0 store faces from; its hard-breaking readings are stated, with migrations, in the #1468 S4 / S4b / T4 sections in the 7.58.0 notes below, which this does not restate; 7.59.0's own store-face changes are in the #2818 / #2836 sections. @dgx: the File lane's two readings (`FileSessionRepo.importEntries` as the row's owner, `FileStorageBackendOptions.sessions.beforeSessionDelete`) are in §5.12 and §5.5. No code, export or behaviour moves.
785
+ - **Pins.** None added (documentation). The existing document pins (`test/docs-contract-pins.test.ts`, `test/backlog1009-store-read-back-shape.test.ts`) read headings and anchors this change leaves in place; the package README pins (`test/backlog2104-mods-kernel-publish.test.ts`, `test/backlog2673-core-stores-sql-publish.test.ts`) read their READMEs' install lines, forbidden words and Exports tables, none of which moved; the front-door export tiers (`test/export-surface.test.ts` reads the root README) are unchanged — no root export name enters or leaves that file.
786
+
787
+ ### Seven gaps the 7.58.0 release found in its own chain are closed — `scripts/release/` and `scripts/verify-tmp-hygiene.mjs` (#2852, #2853, #2856, #2857, #2858, #2859, #2860; release tooling only, `src/` untouched; no behaviour change; hard-breaking: none; closed sets: none; exports: none)
788
+
789
+ - **#2852.** A live stamp and `RC_GREEN_SHA` now vouch for a commit only when its build inputs are the same too — every `package.json` with its dependencies, the lockfile, `tsconfig*.json`, `.npmrc` and the scripts the build runs — with the version cells a bump writes as the only thing allowed to move; before, only the source trees were compared.
790
+ - **#2853.** A `tag-push` whose push exits non-zero keeps the local tag and reads the destination's `main` and tag through the push URL: both there is a lost acknowledgement and the step goes on, neither is a refusal whose rerun pushes the same tag object, anything else is refused by name.
791
+ - **#2856.** `r-commit --dry-run` now runs the re-window half (nothing written) and prints its moved / refused / kept counts, on the CHANGELOG as the release commit would leave it.
792
+ - **#2857.** `gate:tmp-hygiene` prints every failed test file of a red run with its failing tests (it kept the last lines of the output before) and `--full` takes `--workers N`.
793
+ - **#2858.** The re-window keeps a car row only when the row's own car is in the release — the first car name of its status head, judged by the car's own head named in its `DONE` clause (before or after the word) when it has one — a sha that is the stage commit the row waits to be merged onto, or not the car's work, is passed over; a neighbour car named later in the head no longer counts.
794
+ - **#2859.** `gate:criteria` is one of `r-commit`'s heading-sensitive gates, and each gate's exit is kept in the step's chain-state record.
795
+ - **#2860.** The chain gains `post-publish` after `ship-post-lint`: the ship post, the criteria post and the pickup list reach the board only after the lint passed, in the body it passed; `ship-number` only claims the number.
796
+ - **Pins.** `test/release-chain-gaps.test.ts` and `test/backlog2857-tmp-hygiene-failed-files.test.ts`; `test/release-chain-steps.test.ts` re-pinned (a failed push keeps the tag; the step order).
797
+ - **Who is hurt.** The R operator: `post-publish` is a new step between `ship-post-lint` and `post-ship` (`--with-posts` names the extra posts), `r-commit` runs `gate:criteria`, and a stamp taken before a dependency change no longer carries over it.
798
+
799
+ ### An inherited screening entry carries its installation on the minted policy; one frozen regime per installing leg (#2428; structure only, no behaviour change; hard-breaking: none; closed sets: none; exports: none) (@server @cli @client-core)
800
+
801
+ - **Before.** The bus a screening constraint was minted over (`createPreToolUseConstraintPolicy`) lived in a module-level side table keyed by the minted policy object. The mint wrote it, and the installation identity (`sameInstallation`), the carrier check (`screeningInstallationsOf` / `unraisableGuardsOf`) and the inherited `tool.check` raise (`inheritedToolChecks`) looked it up there. A leg's frozen regime (its approver, its two durable mandates, its classifier) was spread twice when the leg handed its chain down: once onto its screening entry and once onto its caller-policy entry.
802
+ - **Now.** The minted policy carries its installation as a member of its own. The member is fixed and not enumerable, the mint is its only writer and one reader (`installationOf`) reads it, so the side table is retired. A copy, a wrapper, a prototype heir or a Proxy of the minted policy carries none, which is what the side table recognised before. That includes a clone that copies every own descriptor, and a deep clone as well, because the member holds the object it was minted on in a closure that no clone can redirect. The regime is minted once each time a leg hands its chain down, and both of the leg's entries carry that same regime. As a result the two entries' `autoMode` containers are now one object; the engine never mutates them, and the `decider` and the denial tracker inside were already shared.
803
+ - **Unchanged.** What runs, what asks and what is refused. The entry's keys and the policy's enumerable face (`Object.keys` is `["check"]`) are the same. Hosts rebuild the entry the same way: `createPreToolUseConstraintPolicy(chain, env)` over the same registrations. What a park records about the chain is unchanged too: `constraintChain` and `constraintDigest` are byte-equal to what the previous tree writes (pin K1).
804
+ - **Consumers.** @server (S-2388; `createParkedReviveInheritedGate`), @cli (MODS155) and @client-core (CC-1042) have nothing to change for this entry. **Announced, not in this entry (#2429):** inherited `classic.PreToolUse` will be raised in the gate's first phase on each installation's own bus, and the screening entry's folded-policy form will retire. The constraint channel stays, as the carrier of each installation's obligations. That change is a hard-break candidate for hosts that rebuild the screening entry with `createPreToolUseConstraintPolicy`, and its three sentences ship with it.
805
+ - Pins: `test/backlog2428-mods-w3-s2a.test.ts` (M / R / C / W / U / S / K).
806
+
807
+ ### Wording only: a recursive shell read's contract sentences say what 7.52.0 made of it, the program-text deny reading is written down (KL-1102-n), and four notes of 7.56.0 are corrected (#2499, #2504, #1714; no behaviour change; hard-breaking: none; closed sets: none; exports: none; members: none)
808
+
809
+ - **Recursive reads.** Five statements still described the tree walk that 7.52.0 retired (its #1714 section, which says it right, is unchanged). The `shell.read_boundary.friction_floor` contract sentence (`src/tools/fs/shell-read-judge.ts`, mirrored in `docs/CONTRACT-SENTENCES.md`) said "traversal roots enumerated, each within a bound"; the `traversal_root` line of `docs/CLOSED-SETS.md` said "enumerated within a bound"; the header of `src/tools/fs/read-deny.ts` called the active rows "the floor of EVERY read face … and their traversals"; and two JSDoc lines (`ReadEstablishment` in the judge, the account of what `env` answers in `shell-read-check.ts`) spoke of a traversal that is enumerated or cannot be. They now say what the code does and has done since 7.52.0: a recursive read's root is judged by what the command names, the root itself and the placed rows it contains, never by a tree walked below it; a credential row that is only spelled (`.netrc`, `.ssh/`) and lies below a root is met where a command names it, not on the walk (KL-r2-a); only a glob root's own expansion keeps its bound. Comments and documents only, the reading is unchanged. #2499 reported a client fence written for the refusal before 7.52.0; whether the walk should be judged again is a separate question, and #2505 is the filed direction (a credential floor in a sandboxed `ExecutionEnv`).
810
+ - **Program text (KL-1102-n, #2504).** The text of an inline program (`python3 -c "…"`, `perl -e '…'`, an `awk` program) meets the read deny rows as ONE spelling, not as the paths written inside it: it is one deny-only candidate (`src/tools/fs/shell-read-argv.ts`, `programTextReads`) resolved like an operand (`src/tools/fs/shell-read-establish.ts`, `lexicalReadsOf`; `src/tools/fs/bash-readonly-classifier.ts`, `resolveOperandSpellings`), so a text that starts with `/` is that absolute path (the working directory plays no part), one that starts with `~/` is judged both at its literal spelling under the working directory and expanded against the declared home, and any other text is placed under the working directory and judged together with the directory's own segments (a program run inside a guarded directory is refused whatever its text: under a `.ssh` directory even `open('.netrc')` is). The judge reads those spellings' segments, splitting at `/` and `\`, and a row is a run of whole segments (`segmentRowMeets`, `src/tools/fs/read-deny.ts`; the ends of the text are boundaries, and a quote, parenthesis or letter touching a segment makes it another segment). So (1) an unanchored spelled row (a built-in `.netrc` / `.ssh/` / `.aws/` / `.git-credentials`, a deployment's `secrets/**`) meets the text only where one of its segments stands whole: `open('<home>/.ssh/id_rsa')`, `open('<home>/.netrc/')` and `perl -e '/.netrc'` are refused; `open('<home>/.netrc')` (the closing `')` touches the last segment) and `open('.netrc')` are not, under an ordinary working directory and the default rows. (2) An anchored row (a home-anchored built-in such as `~/.npmrc`, a rooted deployment row `/abs/x`, a `Read(/…)` rule) meets it only where the resolved spelling really lies under the anchor: a text placed under the working directory only when the directory itself lies under it, a text that starts with `/` only when it lies under it itself; a path written inside the text does not. A home-anchored row compiled with no known home is read as unanchored. (3) When no row meets it, only a text that may open a file (any text of a language the station does not name, such as `python3`, `perl`; a token language's text, `awk` or `jq`, only with an opening token, so `awk 'BEGIN {print 1}'` is not) is `unestablishable`, and only where the read or the write face guards something; with both faces open and no deny row at all it is let through (`foldShellReadFindings`, `src/tools/fs/shell-read-judge.ts`; `programTextMayOpen`, `src/tools/fs/program-read-model.ts`). The question is mandated on the full shell's roots read face and ordinary on its open read face (a deny row wired, or a guarded write face); the read-only and the re-judging seats refuse it. This is how it has read; nothing changed. Whether program text should meet the operand judge, or the scan should be retired, is #2504.
811
+ - **Four 7.56.0 sentences, corrected.** The released notes stand as published; these replace the sentences named.
812
+ - #2302, *Before*: "the default mode handed the call to the classifier and `bypassPermissions` ran it". The classifier is auto mode's (`src/core/classifier-unavailable-arm.ts`). A misspelled deny read as "no rule speaks", so the call went on as if no rule had spoken: in `default` it took its ordinary path (a question a policy had raised reached the approver as an ordinary question), and `bypassPermissions` ran it.
813
+ - #1329, *Who is hurt*: "A policy that caches an ASK answer across calls and returns it for another call id has that call refused." Only an ask that carries the engine's record of another call (a guard policy's or a fold's record, bound to its `toolCallId` and `toolName`, `src/core/engine-attestation.ts`) is refused. A cached ask with no such record is asked or released as it was before. The *Behavior narrowing* paragraph of that section states it right.
814
+ - #971, *Now*: "So a socket is dialed and a pid probed for a local process and nothing else." That holds for a row of plain data members. `readPeerSessionRecord` does not snapshot an accessor member, so a getter that answers differently on each read can pass the check without it (the JSDoc of `PEER_SESSION_WHERE_VALUES` and `docs/sdk/07-multi-agent.md` already say so).
815
+ - #1551, *How to migrate*: "the switch rides the parent spec seat, so the request's engine-built child legs are switched off with it". True because an engine-built child authors no `attachments`. A delegation path that authors its own `attachments` keeps exactly what it authored (`src/core/parent-spec-seat.ts`; the *Behavior change* bullet of the same 7.56.0 section says it), so the parent's switch does not reach its legs: a lane its declaration leaves out takes the default, `mcp_unavailable` on (`mcpUnavailableOn`, `src/core/runner/turn-attachments.ts`), and an explicit `mcpUnavailable: false` of its own stays off.
816
+
817
+ ### A path segment is compared by canonical equivalence, not by code points: a deny row meets the other spelling of its own name (#2577; behavior narrowing, open to veto; hard-breaking: none; closed sets: none; exports: none) (@server @cli @client-core)
818
+
819
+ - **Before.** The segment engine behind the read deny rows and the sensitive-path write list folded case and separators but
820
+ composed nothing. A filesystem that compares names by canonical equivalence (APFS, HFS+) serves `ü` written as one code
821
+ point and `u` plus a combining diaeresis as one file, so a credential row `schlüssel/**` judged the composed spelling of a
822
+ path and answered `null` for the decomposed one — the same file, read or written past the row. The same held for a
823
+ home-anchored or rooted row whose home or run carried such a name, and for a configuration root with a non-ASCII name.
824
+ - **Now.** Both sides of every comparison are composed (NFC) before the fold of the row applies — the pattern at compile,
825
+ the anchors a placed row carries (a home, a rooted location, a configuration root) and the path's segments — under every
826
+ fold (`caseSensitive: true` rows and the write list's host-keyed fold included), by the one fold table
827
+ (`SEGMENT_FOLDS`) every consumer already read. Nothing else changes: ASCII case still folds, non-ASCII case does not
828
+ (the deny set's documented contract), a compatibility form (a ligature) is another name, and a canonical singleton is the
829
+ letter it decomposes to (the Kelvin sign is `K`, which the ASCII fold then lowers).
830
+ - **Behavior narrowing (open to veto).** *What changed*: a deny row or write-list entry with a non-ASCII name now also
831
+ refuses the other composition of that name; the write list and the read set share the compiler, so both widen the same way.
832
+ *Who is hurt*: a deployment whose directory holds two distinct files that differ only in composition (possible on a
833
+ filesystem that does not fold them, such as ext4) and guards one of them — the twin is now guarded too; one extra
834
+ refusal, never a missed one. *How to migrate*: nothing; list the name once, in either spelling.
835
+ - **Limits (stated).** ripgrep compares bytes: its exclusion globs stay spelled as the row is written, and a path it returns
836
+ under the other spelling is caught by the output tripwire (the judge now names it), which hands the search to the JS
837
+ scanner that prunes with the same judge. The data-root exemption's comparison stays byte-exact on purpose (an exemption
838
+ widened by a fold is a guard bypassed). Composition costs what the platform's NFC costs on the spelled segment: a name a
839
+ filesystem can hold (at most 255 units) takes well under a millisecond, but a hostile spelling of tens of thousands of
840
+ combining marks takes seconds (it grows with the square of the segment's length) — the same cost the tool gate's own path
841
+ normalization already pays; the judgment memoizes it per spelling and does not bound it.
842
+ - **Consumers.** @server @cli @client-core: no API or wire change; a deployment's non-ASCII deny entries and write patterns
843
+ now cover both compositions of a name.
844
+ - **Pins.** `test/backlog2577-segment-names-canonical-equivalence.test.ts` (N1 a row against the other spelling of its own
845
+ name under the ASCII and the case-sensitive fold, both directions, deeper paths, a wildcard row, and the names that only
846
+ start like the row; N2 what does not change: ASCII case, separators, dot segments, non-ASCII case, a ligature; N3 a
847
+ home-anchored and a rooted row, any spelling of the home; N4 the target judgment, a recursive read's root, a row at or
848
+ below a path; N5 the write list under all three folds and through `createSensitivePathPolicy`; N6 a configuration root with
849
+ a non-ASCII name still guards its own settings file). `backlog2117-read-deny-star-fold` F5 re-pinned: the `unicode` fold's
850
+ reference is the engine's own `i` flag over the composed name.
851
+
852
+ ### The read-deny table travels with every shell start — an environment that confines processes keeps those bytes from every program the command runs (#2505; behavior change only for a library-direct band: its shell starts wait for the band's deny placement; new field for environment adapters) (@server)
853
+
854
+ - **Before.** The shell read station judges what a command NAMES — its operands, its expansions, a recursive read's root (since 7.52.0 never the tree below a root: `grep -r x .` reads a `.netrc` the root holds, KL-r2-a ①). The documented hard floor under it — "a sandboxed ExecutionEnv" — had no way to learn which paths to keep: nothing handed the environment the table.
855
+ - **Now.** Every command start the shell faces make (Bash foreground and background, the read-only Bash, Monitor) carries `readFloor` in its exec options: the read-deny table in force for that call (built-in tiers, deployment and task rows, a person's `Read(...)` deny rules — the call's own pinned table), as `rows` (the strings the engine's refusals cite) and `denies(absolutePath)` (the same judgment the read station asks). One table, two readers: the station judges names, the environment judges bytes; the station does not loosen when an environment keeps the floor. A library-direct band (no Runner) whose own deny placement is still settling now holds its shell starts until it settles, as its read tools always did. New root type export `ReadFloor`.
856
+ - **Contract for an adapter that confines processes** (sandbox / container / microVM / remote agent): every program the command runs must be unable to read a path `denies` answers non-null for, judged where the environment resolves it; an environment that cannot keep the whole table for a start refuses the start (`spawn_error`, naming the row) rather than run with fewer rows — a runtime configured from the static `rows` that cannot place a row (one with no anchored directory) refuses too. An environment that cannot confine a process (the host, ssh, a device) ignores the field — no byte floor there, KL-r2-a unchanged. `NodeExecutionEnv` ignores it.
857
+ - **Consumers.** @server: the sandboxed lanes (e2b / k8s / local-docker) can now implement the deny floor from `ExecutionEnvExecOptions.readFloor` (and `BackgroundSpawnOptions.readFloor`); no change is needed to keep compiling — the field is optional.
858
+ - **Pins.** `test/backlog2505-read-floor-lane.test.ts` (real Runner; a confining environment's test double: `grep -r` / `find -exec cat` / `tar` keep the credential file's bytes out, a command naming the file is still refused by the station and never starts, deployment rows travel and an excluded built-in row does not, background and Monitor starts carry the floor, the host env reads nothing new), `test/backlog2505-read-floor-seat.test.ts` (the seat: rows = the refusals' strings, the table's own judgment, fail-closed on an unreadable path, the call's pinned table).
859
+
860
+ ### A role seat that names a model the catalog does not hold is read as an unset seat and said once, where it refused every turn (#2571; behavior change, open to veto; follows the reference implementation's split — the main model's seat refuses, helper and subagent seats retreat and say so; hard-breaking: the notice catalog, open to veto; closed sets: `ENGINE_NOTICE_CODES` 102 → 103, +`model.role_defaulted` (audience `operator`, minted by the engine); exports: none; members: none) (@cli @server @client-core)
861
+
862
+ - **Before.** A deployment that named a model its catalog does not hold at a role seat — the field shape is a host that maps
863
+ `MODEL_CHEAP_ID` onto `roles.summarize` / `roles.subagent` while the id it names is not a catalog key — had every run refused:
864
+ the compaction seat (`roles.summarize`) is read at every prepare, so each turn ended `failed` with `config.unknown_role`
865
+ (`detail { field: "roles.summarize", reason: "unknown_model_ref" }`) and no answer; a `roles.subagent` word failed each
866
+ delegation the same way. *Erratum:* this is not a 7.53.0 change. The refusal is older — in 7.52.0 and before the same read threw
867
+ `Unknown model ref "<id>". Provide it in RunnerDeps.models or pass a Model object.` (uncoded) from the same seat at every
868
+ prepare; 7.53.0 made it coded and gave it `detail`, and 7.55.0 (#2152) built it in one place. What a shell saw change between
869
+ its builds was the code and the text, not whether the run was refused.
870
+ - **Now.** One reading, `seatRead` in `src/core/roles.ts`: a seat other than `default` and `classifier` whose word the catalog
871
+ does not hold supplies no model and is read as the same seat without its `model` member — a `select` beside it still answers,
872
+ then the role's tier default, then its fallback chain (`summarize` → `default`, `subagent` → the parent run's model,
873
+ `verifier` → `default`). The walk (`resolveTaskModel`), the compaction seat (`roleModelIfSet`), the subagent seat
874
+ (`deploymentStatesSubagentModel` / `deploymentSubagentTierModel`, which now take the catalog) and the consolidation chain read
875
+ it through the one test `unheldSeatWord`. The person is told: at the root leg's prepare the engine says `model.role_defaulted`
876
+ for each such seat of the deployment's own map and of the merged deployment-and-task map, **once per session per (seat, word)** (the session's announced
877
+ record, as `model.capability_downgraded` is), with the seat, the word and what stands in. Still refused, by the seat's own
878
+ `config.unknown_role`: `roles.default` (the main model's seat), `TaskSpec.model`, `TaskSpec.compactionModel`, a value that is
879
+ no model reference (`not_a_model_ref`), and a chain that ends with no model (the first unheld seat's refusal, not "no model for
880
+ role"; the memory-consolidation chain `consolidate` → `summarize` reads the same way: a written-but-unheld `consolidate` now falls to
881
+ a held `summarize`, and with neither the refusal is still `config.unknown_role` naming the first). The `classifier` seat keeps its existing stand-in and its line,
882
+ `classifier.role_defaulted`; this line is not minted for it.
883
+ - **Behavior change (open to veto).** *What changed*: a run, a compaction and a delegation whose role seat names an unknown word
884
+ now go on (on the fallback model) and say it once, where they were refused at every prepare. *Who is hurt*: a deployment that
885
+ relied on the refusal to find a typo in a role seat — it now sees one operator notice per conversation instead of a failed run
886
+ (the cost of the fallback model is the deployment's to watch; `summarize` falls to the run's own model, `subagent` to the
887
+ parent's). *How to migrate*: fix the seat or declare the word in `RunnerDeps.models`; a host that wants the old refusal reads
888
+ `model.role_defaulted` and fails the session itself.
889
+ - **Limits (stated).** The line is said at a leg's prepare, from the catalog that prepare reads. A word that becomes unheld afterwards — a
890
+ catalog swap during a run, then a delegation that reads the live catalog — takes the retreat at once (the child runs on the parent's
891
+ model) and is said at the session's next prepare, not before; the old behaviour there was the spawn's refusal. A `subagent` seat is
892
+ read, at a spawn, from the deployment's own role map, so a task's own `roles.subagent` does not change which model its children use
893
+ (it never did); the line names the deployment's seat for that reason.
894
+ - **Hard-breaking (open to veto) — the notice catalog.** What changed: `ENGINE_NOTICE_CODES` gains `model.role_defaulted` (audience
895
+ `operator`, minted by the engine; it always carries `sessionId`). Who is hurt: a consumer switching exhaustively over the catalog
896
+ or holding a whitelist of notice codes. How to migrate: add the row; show the message by code (`detail.seat`, `detail.model`).
897
+ - **Consumers.** @cli (L-1984): read `model.role_defaulted` for the start-up line about a role seat that names a model outside the
898
+ catalog (`MODEL_CHEAP_ID` and the like) — the engine now falls back and says so once, so the shell's own pre-flight can drop its
899
+ copy of the refusal. @server: no change; `roles.summarize` / `roles.subagent` written from `MODEL_CHEAP_ID` now fall back with a
900
+ line instead of failing every task. @client-core: show the notice by code like any operator row.
901
+ - **Pins.** `test/backlog2571-role-seat-unheld-word.test.ts` (R1 the cheap seat unheld: the run completes and says it once; R2 once
902
+ per session per (seat, word), a re-pointed seat again, a task's own seat on the merged map; R3 the subagent seat: the child runs on
903
+ the parent's model and one line is said; R4 held words and a bare deployment say nothing; R5 `roles.default` still refuses, also with a tier table bound; R6 the
904
+ walk, the tier stand-in and the compaction seat read, and the end of a chain; R7 the listing and the one test; R8 the classifier seat keeps its stand-in; R9 the consolidation
905
+ chain; R10 what still refuses; R11 the closed sets). Re-pinned to the new reading: `backlog641-child-model-seat` (a dead tier
906
+ word is an unset seat), `backlog1279-child-undrained-to-parent` (the catalog count).
907
+
908
+ ### `TaskSpec.classifyAllShell` — in auto mode no shell allow entry clears a call, whichever source holds it: the classifier judges (#2558; behavior narrowing, opt-in)
909
+
910
+ - **Before.** The engine had no counterpart of the reference implementation's `autoMode.classifyAllShell` (KL-1327-c, retired by this entry). In `auto`, the allow layer's strip removed only the shell entries that hand over arbitrary code (the whole shell, an interpreter or launcher prefix); a narrow rule such as `Bash(git:*)` kept clearing its calls with no classifier — from a person's stored rules, a skill's `allowed-tools`, the request's `allowedTools` and the deployment's `RunnerDeps.allowedTools` alike. A host that honours the setting had no engine seat to say so: its shell allow lists reached the engine and cleared calls the person had asked the classifier to judge.
911
+ - **Now.** `TaskSpec.classifyAllShell?: boolean`. `true` ⇒ while the leg's mode is `auto`, no shell allow entry clears a call: the whole shell tool, a tool name that reaches the shell and every `Bash(…)` command rule stand down whatever they name, in all four allow members (one predicate, `strippedInAuto`; the mode is read on every call, so a steer into or out of `auto` applies on the next call — a steer that lands while a call's stored rules are still being read applies from the call after it). Under the key no allow entry of any kind clears a call on the shell face: a path, name or host entry — a list's or a person's stored row (stored rows are now read as the entry kind of their grammar) — that a shell-face tool's declared path target reaches stands down too. The Skill card no longer names shell entries in `auto` under it. A command the read-only reader waves through still runs without the classifier, as before (it never reaches the allow layer). Outside `auto` nothing changes; deny and ask rules, a hook's allow, non-shell entries and every question an allow rule may not clear read exactly as before. A tighten-only seat (one more row of the clamp table): `tightenTaskSpec` keeps a baseline's `true`, an engine-built child inherits `true` and its own `false` cannot lift it, a governed workflow script may declare it for its child. A present value that is not a boolean refuses the run before it starts (`config.invalid_argument`). Absent or `false` ⇒ byte-identical to 7.58.0, but for one shape of the existing strip: a protocol-namespace entry (`mcp__srv` / `mcp__srv__*`) that reaches a shell call through an alias the deployment gave the shell tool now stands down in `auto` like the whole shell does under every other name (it used to clear the call unclassified; narrowing, with or without the key).
912
+ - **Behavior narrowing (opt-in).** Only a caller that sends `true` sees a change: its shell calls in `auto` are classified (more classifier calls; a classifier block where an allow rule used to clear; the denial limit is reached sooner). Measured on a real Runner, 7.56.0 vs this head, `auto` with a blocking classifier and `Bash(git:*)` + `git add -A`: request / stored / deployment / person-invoked skill rows each cleared the call unclassified and now each is classified (and blocked); the same with the key absent, the same under `default`, and a `Write` entry under the key read the same in both trees.
913
+ - **Consumers.** @client-core (CC-997): carry `classifyAllShell: true` when the user, flag or managed settings plane says `autoMode.classifyAllShell: true` — the reference reads it from those three only, never from a repository's project or local settings — and keep projecting the allow buckets (settings `permissions.allow`, `--allowedTools`, a skill's `allowed-tools`) unchanged; do not build an arm that drops shell entries from the request (it cannot reach the engine's own rule store or the deployment list, and cannot follow a mid-run mode change). @server: pass the request field through to `TaskSpec.classifyAllShell`; a managed policy that forces it = a `tightenTaskSpec` baseline carrying `true`. @cli: D-081 / L-1961's remaining difference retires once the bundled engine and client-core carry the bit; no shell code change. A test or a table that pins the clamp table's rows (`CLAMP_SEATS`) or `ParentSpecSeat`'s members sees one more.
914
+ - **Pins.** `test/backlog2558-classify-all-shell.test.ts` (real Runner quadrants Q1–Q4, the four members S1–S3 / S1′, a mid-run steer into `auto` M1, the namespace entry through an alias A1, the path / name / host arms and the stored row's grammar P1, the bad-value door D1, the Skill card K1–K3, the seat's travel U1–U4); the resume door table row in `test/backlog1950-resume-config-doors-preclaim.test.ts`; the strip's unit calls in `backlog1327-allow-entry-unit` / `backlog1517-b-bash-pattern-rules` / `backlog1303-skill-grant-unit` re-spelled for the new parameter (same readings).
915
+
916
+ ### An image the provider cannot process no longer poisons its session: admission checks each format's header, and a provider's image refusal removes the image from what the model sees and re-sends the turn (#2874; hard-breaking: one notice word (`ENGINE_NOTICE_CODES` +1 `context.image_removed`, the #1844 / #2571 reading); closed sets: three +1 (`ENGINE_NOTICE_CODES` `context.image_removed`, `AssistantMessage.errorKind` `image_unprocessable`, `LoopContinueReason` `image_refusal_retry`); behavior narrowing, open to veto — an Apple-optimized CgBI PNG (a PNG whose first chunk is not a 13-byte `IHDR` of non-zero size), a JPEG whose fourth byte is below 0xC0, a GIF of zero width or height and a WebP without a `VP8 ` / `VP8L` / `VP8X` chunk were all admitted before and are refused now) (@cli @server @client-core)
917
+
918
+ - **Now.** An image's bytes must open with their format's header (one table: PNG = signature + `IHDR` first chunk, so an Apple-optimized `CgBI` PNG is refused; JPEG = SOI + marker; GIF = version + screen descriptor; WebP = `RIFF`/`WEBP` + `VP8 `/`VP8L`/`VP8X`), same refusal word `magic_mismatch`. A provider 400/422 whose envelope says an image could not be processed (OpenAI- and Anthropic-form) types the failed turn `AssistantMessage.errorKind: "image_unprocessable"` (+1 member); that failure and its persisted record make every request view replace the images of the newest message still carrying any with `[image N removed: rejected by provider: image_unprocessable]` (the transcript keeps them), the turn is re-sent in-turn (≤ 3), and later turns / resumed sessions read the same removal. Closed sets: `ENGINE_NOTICE_CODES` +1 `context.image_removed` (audience `user`, `detail: { images, sessionId, runId }`, once per run per refusal); `LoopContinueReason` +1 `image_refusal_retry`. Consumers: @cli @server @client-core — a closed-set pin on notice codes or `errorKind` goes red; show `context.image_removed` to the person. Pins: `test/backlog2874-image-reject.test.ts`.
919
+
920
+ ### Erasure takes the import's scope authorization seat — `eraseMemoryEntries({ …, expectedScopes })` is judged after resolution, inside the store's lock, and a refusal names nothing (#2304; security fix; behavior narrowing, open to veto; hard-breaking: one type-only edge)
921
+
922
+ - **Before.** `importMemoryBundle(bundle, { expectedScopes })` let a service endpoint pin its principal's grant set; `eraseMemoryEntries` had no such seat, so the question "which scopes may this caller erase" had two answerers — the import seat in core and an envelope a service built for itself, which could only read a `{ scope }` selector's literal and had to refuse `ids` / `sessionId` selectors outright (it cannot see what they resolve to).
923
+ - **Now.** One seat, one shape rule (`expectedScopesInvalid`: an array of non-empty scope names) and one membership predicate (`firstOutsideExpectedScopes`), both in `src/core/memory-engine/export-bundle.ts`, read by both verbs. On erasure, present ⇒ the selector's own scope (a place selector, judged before anything is read) and the **binding scope of every resolved id** must be members; an id the store cannot attribute to a scope — no account row, an unbound row, a non-entry-id ledger key, an unledgered file on disk, an id that never existed — is not a member. Any violation refuses the whole request with the new code `config.memory_erasure_scope`: nothing deleted, nothing recorded, and the refusal carries no count, no id and no scope (an out-of-set id and a never-existing id answer the same text, so the refusal is no existence oracle). The evidence lane judges in the lock, records the canonical set (sorted, de-duplicated) on the request's anchor row, refuses a replay under another set first (absent and present differ), and judges what a replay deletes where it is now: an id still (or again) holding an account row is judged by its current binding — an erased id the host re-added in another scope is not deleted by a replay of the old request — while an id the request already erased (no row, no bytes) is judged by the scope its first execution recorded on the request's own projection rows (one assignment under the anchor's selector hash; a conflicting record attributes nothing), so a commit-then-crash replay converges without a live row. The recorded set is held to the shape rule by the row validator and judged only where it is used — a record no first execution could have written authorizes nothing at replay. The degraded lane (no evidence face, `allowUnevidenced`) runs the same predicate over the read faces' answer before its CAS'd deletes. The seat authorizes erasure only; it never changes a session's memory scopes or write scope.
924
+ - **A declaration, never duck-typing.** An evidence face honours the seat only when its backend declares `judgesErasureScopes: true` (the File backend does); the engine refuses a grant-bearing erasure over an evidence face that does not declare it (`config.memory_erasure_scope`, the face is never called) — a face that predates the seat would otherwise ignore the set and erase outside it. Without `expectedScopes` every path is byte-identical (the anchor row carries no new key).
925
+ - **Other behavior changes (every caller).** The `memory.erasure_selector_mismatch` text no longer prints the two selector hashes (the first request's hash stays on its anchor row). The degraded lane deletes only the ids it was asked about: a read face that returns extra entries no longer widens the delete set.
926
+ - **Behavior narrowing (open to veto).** *Import*: a malformed `expectedScopes` — a `Set` or other non-array iterable (accepted before), a string (read character by character before), an empty or non-string member — is refused as `config.memory_import_request` before the package is read; the pure `memoryBundleInvalid` answers the same violation for a direct caller. *Erasure*: under a grant, a `{ scope }` erase of a scope holding an unledgered file refuses until a read adopts the file (the host, with no set, erases it as before).
927
+ - **Hard-breaking (type-only).** *What changed*: `MemoryBackendContractHooks.onOptionalMember`'s `member` union gains `"judgesErasureScopes"` (the suite reports a backend without the declaration as `absent`, a declaring one as `verified` after its new case). *Who is hurt*: a TypeScript hook that switches exhaustively over `member` with a `never` default stops compiling. *How to migrate*: add the arm.
928
+ - **Downgrade, stated.** The anchor's `expectedScopes` key is tolerated by an older binary's row validator (it refuses only the keys it names), so an older engine replaying an authorized request does not judge the set.
929
+ - **What still stands (named, not closed; KL-2304-a).** The degraded lane holds no lock between its read and its delete: an entry moved to another scope with unchanged content in that window is not seen. Two principals that present the same grant set share the requestId namespace (the engine holds no principal; the set is the identity it can compare). Under a grant, a `{ scope }` erase of a scope holding an unledgered file refuses without saying which (the behavior narrowing above).
930
+ - **Consumers.** @server: `POST /v1/memory/erase` can pass the principal's set as `expectedScopes` and retire `erasureEnvelopeRefusal` (`memory-operator-faces.ts`), which today refuses every `ids` / `sessionId` selector under an envelope; add `config.memory_erasure_scope` to the compliance failure code table (unclassified, it would answer 500). The two SQL twins have no evidence face — they ride the degraded lane, judged by the engine, and need no declaration; a twin that gains `eraseWithEvidence` must declare `judgesErasureScopes: true` and pass the contract's new case. @dgx: a self-hosted backend certified through `memoryBackendContract` that offers `eraseWithEvidence` declares the seat and passes the same case.
931
+ - **Pins.** `test/backlog2304-erasure-scope-seat.test.ts`: the three selector forms out of set / unattributable / place selector including an empty scope, the ghost-file cost, the canonical anchor record, replay under same / other / no set with the set check ahead of the selector check, two commit-then-crash forms converging without a live row, forged authorization records authorizing nothing at replay and a malformed recorded set failing the row validator, the shape rule on both verbs, the undeclared evidence face, the direct File face call, the degraded lane, a host re-add in another scope surviving a replay, conflicting projection records, the hash-free selector-mismatch text, an over-returning degraded read face, and the three actor counterexamples (no memory path reads an actor; lineage keyed by session; the seat leaves no standing state). `memoryBackendContract` gains the seat case (File verified; it judges the refusal by code — its text is pinned for core's own lanes, which mint one refusal).
932
+
933
+ ### The erasure attestation is version 2 — `complete` means every carrier of an erased entry in this store is gone, each carrier class answers what it removed and what remains, and no slug reaches the attestation or the evidence chain (#1917 S1b-1; hard-breaking, open to veto; behavior change, open to veto; closed sets: +1 `ERASURE_CARRIER_CLASSES` (eight words); exports: +3 type names `ErasureCarrierClass`, `ErasureCarrierAnswer`, `ErasureCountAnswer`; `MemoryErasureAttestation` and `ErasedBinding` reshaped) (@server @dgx @test)
934
+
935
+ - **Before.** `complete` judged only the selected ids' account rows: a quarantined copy of an erased entry stayed in the quarantine directory ("preserve-not-delete") and was listed by file name; the consolidation archive and plan workbook kept the text untouched and unmentioned; the attestation (`erased[].binding.slug`, `erasedPreviously[].from`) and every delete row on the append-only evidence chain kept the entry's slug — a model-chosen file name that can restate what was erased; and the degraded lane (no evidence face, the two SQL twins) answered `complete` with an empty, unenumerated residual set.
936
+ - **Now.** `v: 2`. `erased[].binding` is `{ state: "bound", scope }` or `{ state: "unbound" }`; `erasedPreviously` rows are `{ id, ev, at }`. `residuals` is the carrier table — `carriers` answers each of `projection`, `shadow`, `indexLine`, `quarantineCopy`, `planArchive`, `planWorkbook`, `evidenceBinding`, `backendDerived` with `{ removed, remaining }` or `{ unverified: true }` — plus `quarantineUnattributed` and `derivedEntries` (store-level counts that do not decide status), `replay` (`"converges"` on the evidence lane, `"unsupported"` on the degraded lane) and `propagation`. `status` is one law for every lane (`erasureStatusOf`): the v1 conditions, and every carrier class answering `remaining: 0`. A quarantine file whose frontmatter id is erased is deleted under the request's lock; a file no id can be read from is counted and left. A new delete row lands on the chain with its `scope` and no slug (the transaction journal keeps the full binding for recovery and is deleted at close). The paths the derived-index sweep owes live only in a transient control-plane record (`erasure-sweeps.json`, by requestId; each path relative to the store root and read at the root the store is opened at now), written from the request's journal right after its commit point — and by recovery from any erasure journal it rolls forward, including one an older binary left; `MemoryEngine.eraseMemoryEntries` sweeps every MEMORY.md line naming one of those paths — whoever holds the path now, an occupant's line returns in mechanical form at the next materialize — fsyncs each rewritten index, and drops the record row, all under the store's transaction lock and the record's lock (two sweeps of one index never interleave, and a commit, a recovery or a re-anchor never lands inside a sweep). The engine answers the index class on its own copy of the face's attestation (a frozen third-party object can no longer keep a failed sweep `complete`). The degraded lane answers every store class `unverified` and is never `complete`.
937
+ - **Contract sentences (the attestation's JSDoc).** The evidence chain is append-only and a delete row never expires with any retention period; on the evidence lane a replay of the same requestId converges with no time limit (a ≥ 24 h replay window holds); the degraded lane promises no replay convergence and says so.
938
+ - **Behavior change, open to veto.** *Until S1b-2 lands (same window)*: a File store with any file under `distiller-plans/` or `consolidation-plans/` answers `planArchive` / `planWorkbook` `unverified`, so **every erasure on a store with consolidation history is `partial`** — true: the archive keeps the prompts and answers verbatim and the workbook keeps product text, and this lane does not read them yet. *Permanently*: an erased entry that was ever renamed or moved, or whose delete was recorded before this version, keeps `evidenceBinding.remaining > 0` (the chain keeps those rows — KL-1917-a) and stays `partial`, and a replay cannot clear it. The store face called directly (`eraseWithEvidence`) answers `indexLine: { unverified: true }` — `partial` — while its record waits for the engine's sweep. A failed index sweep keeps the request `partial` (v1 said `complete` and listed the file on `indexUncleared`).
939
+ - **Hard-breaking.** *What changed*: `MemoryErasureAttestation.v` is `2`; `ErasedBinding` loses `slug`; `erasedPreviously[].from` is gone; `residuals.quarantineHits`, `residuals.quarantineOpaque` and `residuals.indexUncleared` are replaced by `residuals.carriers`, `quarantineUnattributed`, `derivedEntries` and `replay`; quarantine copies of an erased entry are deleted; a delete row on the chain carries `scope` instead of `from`. *Who is hurt*: a host that reads the attestation's v1 fields — @server `memory-operator-faces.ts` / `memory-compliance.ts` read `erased[].binding` and `residuals.quarantineHits` — and a reader of `transfers.jsonl` that takes a delete row's projection path from `from`; an operator who relied on quarantine keeping a copy of an erased entry for forensics. *How to migrate*: dispatch on `v` (a reader of `v > 2` refuses); read the carrier table; locate an erased entry by `requestId` in your own task state, never by slug; read a delete row's scope as `scope ?? from.scope`; export a quarantined copy before erasing if you need it.
940
+ - **Downgrade, stated.** An older binary reads a chain-form delete row (it does not refuse the unknown `scope` key), and its own recovery of a journal this version wrote is unchanged (the journal keeps `from`). An older binary that appends a journal's rows writes the slug-bearing form again, and its scoped export reads a chain-form delete row as unbound (counted as unsliceable; a full-store export carries it).
941
+ - **What still stands (KL-1917-a, KL-1917-b).** Chain rows minted before this version and every move / migration-bind row keep their slugs (exit: the candidate ticket that gives move forms a slug-free chain form). The record is keyed by requestId alone: when two executions of one requestId overlap (a host retry over a still-running first call), the first sweep settles the whole row, and an index line an outside actor restores between that sweep and the second execution's own sweep is not seen by the second — which then answers `complete`. A record that cannot be read or written, or a row whose path is absolute or leaves the store root, refuses like every strict control-plane file (the erasure, or the recovery leg). A non-File evidence face keeps no record, so the engine sweeps no index line for it — its own `indexLine` answer stands.
942
+ - **A store move while a sweep is owed.** A record row binds no root, so a store moved and re-anchored with a sweep still owed (its erasure answered `partial` on `indexLine`) replays that erasure at the new location: the sweep removes the line there and the row goes only then. `reanchorMemoryStoreRoot` refuses what it refused in 7.58.0 — a committed transaction pending recovery — and nothing more; 7.58.0 kept no sweep record, so a re-anchor is not narrowed. The erasure's settlement does narrow one thing: it now waits for the store's transaction lock (15 s, extended to a held lock's 30 s stale line), and one that does not get it in time answers `partial` (7.58.0 did not wait). The row goes only if the store root is still the anchored one after the sweep (read at that act): an engine still open at the old root when another holder moves and re-anchors the store — between the store's commit and the engine's sweep, or while the sweep reads — keeps the row and answers `partial` (`indexLine: unverified`), and a replay at the new root settles it. The settlement proves the store's transaction lock is still its own before it sweeps and again before the row goes, the same check a commit and a re-anchor make: an engine that held the lock past its stale line and had it taken keeps the row and answers `partial` the same way. A row path is split into segments the way the platform splits paths, so on POSIX a file name holding `\..\` is one segment and its row is recorded; the row is derived before the commit point, so a target it cannot name refuses with nothing committed. Pins: `test/backlog1068-sweep-reanchor.test.ts` (move, damaged anchor in place, move behind a link; the absolute row form refused; the stale holder after a re-anchor, a transaction in progress, a move during the sweep, the lock taken during the sweep; a `\..\` file name). The degraded lane (a sweep with no request id) answers `indexLine: unverified` when its holder loses the transaction lock mid-sweep — it cannot vouch for `remaining: 0` — where it answered `{ removed: 0, remaining: 0 }` before.
943
+ - **Consumers.** @server: the two SQL twins ride the degraded lane — their attestations are now always `partial` with `replay: "unsupported"`; read `erased[].binding.scope` and the carrier table, retire the `quarantineHits` / slug reads, and take erased-entry location from `requestId`. @dgx: the core counterpart of ha's `source_tombstones (root, source_event, …)` is the attestation plus the chain's anchor / request / delete rows — request (`requestId`), subject (`expectedScopes`, `erased[].binding.scope`, `sessions`) and entry ids, never text or a hash; a self-hosted evidence face answers the eight classes and `replay: "converges"` (`memoryBackendContract` checks the shape and that no slug or text appears). @test: the S1b-1 cells below are the black-box handles.
944
+ - **Pins.** `test/backlog1917-erasure-attestation-v2.test.ts` (14 cells; baseline 13 red; 17 mutations, all killed): an older binary's pending erasure journal rolled forward with its record and swept on replay; every sweep under the record's lock; the clean table; projection copies; a quarantined copy erased and an unattributable file counted; the consolidation plane answered `unverified`; a rename keeping `evidenceBinding` and the request `partial` across a replay; `derivedEntries` counted; slug-free attestation and new chain rows on the engine and organic lanes (byte search); commit-then-crash with the path reused by a new entry, the replay removing every line naming the path and the next materialize restoring the occupant's; a record surviving a reopen until the engine settles it; recovery accepting a journal row the chain already holds in its older full form; the degraded lane never `complete`. `test/design178-v2b-erasure-evidence.test.ts` re-pinned to v2 (quarantine, index sweep, frozen third-party attestation); the neighbour cells `backlog1094` ⑥ (an erasure through a linked root of an entry renamed earlier in the cell: `evidenceBinding` 1, partial) and ⑲d (the index class, not a file name) and `backlog1103` ⑪ (the source scan's row for the sweep's new method) read v2. `memoryBackendContract`'s erasure case asserts v2: every class answered, `replay: "converges"`, no slug or text in the deliverable.
945
+
946
+ ### Memory notices are filed per scope — a session reads the notices of the scopes it mounts and no other (#1896 / #2423; hard-breaking, open to veto; security)
947
+
948
+ - **Before.** The memory notice queue was one queue per control plane. Every enqueue (harvest tail, read-side adoption, consolidation, erasure, challenges) appended to it with no scope, and a session's first injection drained the whole queue: a session that mounted only a shared `home` scope read another member's private-scope notices — the rejected file names (`write rejected [secret] <path>`) and the counts (`N added, M updated`) — and, because the drain is destructive, the session that owned those notices never saw them. The bound (20) and its fold count were shared too, so one scope flooding the queue pushed every other scope's notices out and the fold count told every reader how busy the others were.
949
+ - **Now.** One bounded queue per scope inside the same `announcements.json` (each with its own bound, fold count and standing-condition sweep). `MemoryAnnouncement.scope` is required: a notice is filed under the one scope it reports on. `drainMemoryAnnouncements(controlDir, scopes)` / `peekMemoryAnnouncements(controlDir, scopes)` take a required scope list and read only those queues; the engine's injection drains the scopes the session mounts plus its write scope, and every other queue stays for the session that mounts it. A harvest that reports on its write home and on a read-only layer at once files each line under the scope that owns the path (the write face's own facts — patch counts, a refused harvest, a conflict, a degraded run — under the write scope; a held entry's warnings under the hold's scope; a link is judged by the directory it sits in, never by what it points at); a count over several layers is split per layer, and a count over the whole control plane (expired settlement rows, unsettled lineage transactions) is the host's. A fact of the whole store that names no scope, path or count (a ledger reset, a torn evidence tail, a v1-compat window) is filed under every registered scope; a notice that names no single scope — a count or ids over the whole store (the v2 ledger upgrade line, the unbound-row disclosure, an ambiguous projection across scopes) — reaches no model face and goes to the host's incident sink as `memory.announcement_withheld`; a session-level event with no scope in hand (an external delegation settlement, a failed settlement write) reaches no model face either and is written only to the process log (`console.warn`), not to the incident sink. A challenge notice is filed under the challenged entry's scope, read from the committed store after the synchronous ledger append; each filing is bounded on its own (5 s — past it the lines go to the host and a late answer files nothing) and the next mount waits for the filings at most as long (a wait that runs out, and a filing that fails, are said as `memory.announce_failed`). A store-read fault that relays its error text is the host's (the text can name any scope's rows). A cross-scope move or duplicate-projection notice no longer names the other scope's path.
950
+ - **Hard-breaking (open to veto).** *What changed*: `MemoryAnnouncement.scope` is required; `drainMemoryAnnouncements` and `peekMemoryAnnouncements` take a required `scopes` array (no every-scope spelling) and answer `{ folded, queue, unscopedDiscarded }`; `enqueueMemoryAnnouncement` throws a `TypeError` on a line without a scope; a queue file written before this version is discarded unread at its first locked rewrite and its line count is said once to the incident sink (`memory.announcement_withheld`); the unbound-row disclosure, the v2 ledger upgrade line and the two delegation-settlement lines are no longer model notices. *Who is hurt*: a host that calls the package-root `drainMemoryAnnouncements` / `peekMemoryAnnouncements` / `enqueueMemoryAnnouncement` directly, and test fixtures that write queue lines or the queue file by hand. *How to migrate*: pass the session's scopes to the readers (`[...memory.scopes, memory.writeScope]`), add `scope` to every hand-built line, and read the host-only notices from the engine's `onIncident` sink (code `memory.announcement_withheld`; the two delegation-settlement lines are in the process log only).
951
+ - **Consumers.** @server: no call of the three queue functions was found in the server source read for this change (a local checkout, not the release head) — a host that does call them passes the session's scopes and adds `scope` to its fixtures; the incident sink gains `memory.announcement_withheld` (a posture code — not a governance verdict). Whether a multi-tenant deployment shares one control plane across tenants was not read here — with this version it no longer decides who reads a notice. @cli: no call site found either (same reading); no change.
952
+ - **Pins.** `test/backlog1896-announcement-scope.test.ts` (per-scope drain / flood / sweep / required scope / upgrade discard / transfer; HOME-only session reads none of the private scope's notices while their owner still does; every model face of a HOME-only session — injection index, `memory_search`, `memory_get`, `memory_index`, the notice block, the read-only notice — free of the private scope after a privileged read proves the private rows exist; one harvest's write-home and read-only-layer lines each in their own queue; a rejected link in the private home pointing at a HOME file stays the private scope's line; a transfer whose destination cannot be written leaves the source whole; challenge notices in the entry's scope, the unknown id to the host, a hung filing bounded on its own and not holding the next one, the mount's wait bounded however many are queued, a failed filing said); re-pinned to scopes: `memory-engine-s2-announce`, `rb495-t1-occurrence-identity` (the unbound-row disclosure is now the operator's: said once per reading, A→B→A included, never queued), `memory-ledger-v2`, `design383-slice1-security`, `design178-v2b-erasure-evidence`, `memory-dual-root` and the other queue readers (through `test/helpers/memory-announcements.ts`, a test-only reader of every registered scope).
953
+
954
+ ### The A2A lock key and the A2A capability are words of the closed sets — `LockedKey` gains `a2a`, `ComplianceCapability` gains `a2a_peers`, so a deployment's A2A veto can be written and a task that declares a peer is refused under it (#2500; additive; behavior narrowing for a deployment whose posture resolver can fault, open to veto; hard-breaking: none; closed sets: `LockedKey` / `LOCKED_KEY_REGISTRY` +1 word `a2a`, `ComplianceCapability` / `COMPLIANCE_CAPABILITIES` +1 word `a2a_peers`; exports: no root name added or removed) (@server)
955
+
956
+ - **Before.** The two words did not exist, so neither veto could be written. A lock declaration listing `a2a` was an unknown key: `config.locked_registry` on every prepare of that deployment (measured on a real Runner at the previous release: a task with no peer at all failed), not a lock on peers. A posture naming `a2a_peers` was an unparseable posture, which prepare treats as a resolver fault; that fault denies every capability the engine manages, and none of them was the peers, so a task declaring a peer completed under it (measured: a posture denying `a2a_peers` and a thrown resolver both ended `completed`, the recording listener hit twice). A task that added `a2a_peers` to its own `complianceAdditionalDenies` was refused as `config.invalid_argument` (an unknown word), not as a denial.
957
+ - **Now.** One registry row and one capability word, each read by the code that already reads its set. `LOCKED_KEY_REGISTRY.a2a = { specFields: ["a2a"] }` — the `mcp` row's twin: a task that carries `TaskSpec.a2a` under the lock is refused wholesale (`config.locked_key`) before any peer is contacted, and the preflight reads the peers seat once — the lock check's read — and resolves nothing from it under the lock (a seat that answers differently to a second read reaches nothing). `a2a_peers` is a member of `COMPLIANCE_CAPABILITIES`: a posture's `additionalDenies`, a task's `complianceAdditionalDenies` and a governed script's list accept it through the one array they already read; prepare's conflict check beside `mcp_servers` refuses a task whose preflight snapshot declares a peer while the capability is denied (`config.compliance_denied`; `config.compliance_required` when the denial is the resolver fault's) before any peer is dialed (the check sits in task preparation, which every leg — fresh run, resume, delegated child — enters). No builtin profile floor names it (`BUILTIN_COMPLIANCE_DENIES` follows the policy-limits table, which has no A2A row), so a deployment that asks for nothing sees nothing.
958
+ - **Behavior narrowing (open to veto).** A resolver that throws, or returns a posture the engine cannot parse, now also denies `a2a_peers` (the fault arm denies the whole closed set), so a task that declares an A2A peer is refused `config.compliance_required` where it used to run and dial the peer. Measured on a real Runner against a recording loopback listener, previous release vs this one: thrown resolver + `a2a: [peer]` — `completed`, listener hit twice (the card probes) → refused, listener not hit. The narrowing reaches only a task that declares a peer under a faulting resolver (the newly valid words are additive: a lock declaration listing `a2a` no longer fails every prepare, and a task or posture may now name `a2a_peers`). Hard-breaking: none; a consumer that switches or builds a `Record` exhaustively over `ComplianceCapability` or `LockedKey` sees one more word to handle at compile time.
959
+ - **Reach, stated.** The veto rides the task's preparation: the public library entrance `materializeA2aTools` (like `materializeMcpTools`) passes no prepare, so no posture and no lock reaches a host that calls it directly — the module header of `compliance.ts` now says so; a refresh re-dials the peers prepare admitted and admits no new one. `TaskSpec.a2a` is the only channel that declares peers (`RunnerDeps.a2a` carries only `isolatePeers`).
960
+ - **Consumers.** @server: `task-a2a.ts` (`a2aInjectionHonored`) reads both words as spelled strings (`lockedConfigKeys?.includes("a2a")`, `denies.has("a2a_peers")`) and says in its header that the env channel could not write them; with this release it can (`lockedConfigKeys` is read against `LOCKED_KEY_REGISTRY`, `complianceAdditionalDenies` against `COMPLIANCE_CAPABILITIES`) — read the typed members (`LockedKey`, `ComplianceCapability`) instead of spelling the words, and retire the "not in core yet" notes there; a request body that adds `a2a_peers` through the tighten-only seat is now accepted by the same array. Read in the server checkout, not run by this car. @test: a black-box cell for a deployment that locks `a2a` (a task carrying `a2a` is refused `config.locked_key`, a task without one runs) and for a posture that denies `a2a_peers` (refused `config.compliance_denied`; resolver fault refused `config.compliance_required`; no peer contacted).
961
+ - **Pins.** `test/backlog2500-a2a-closed-set-words.test.ts` (12 cells; 8 red on the previous release's tree, 4 controls green on both); `test/prepare-caps-and-workflow.test.ts` (the degraded deny set gains the word; the conflict check on the preflight snapshot, with the empty-list and no-denial controls). `docs/CLOSED-SETS.md` registers both sets (rows and notes); the governed-agent README, the `RunnerDeps.lockedConfig` JSDoc and KL-469-k list the registry by hand and now say `a2a`; `test/backlog1473-task-tighten-seat-family.test.ts` quotes the closed-capability sentence and now lists `a2a_peers`.
962
+
963
+ ### A parked workflow call is a call that has not been answered yet — a decision delivered to the still-running run drives the parked child on in place (`redeemWorkflowParks`), and the run suspends only when nothing but a person can move it (#2561; hard-breaking, open to veto)
964
+
965
+ - **Before.** The first workflow child to park at an approval gate threw `WorkflowAgentParkedError` into the script at once, every later spawn was refused (a call queued behind the concurrency slot included), and the run ended on that park after its in-flight siblings drained. The only way to apply a decision was a second run resuming from the first — which `startWorkflow` refuses while the source is still active in the process. So a decision taken while a sibling leg was still running could not be delivered: the card kept reading `parked: true` until the run ended.
966
+ - **Now.** The parked call waits. The run keeps running its other calls (siblings, chains, calls queued behind the slot). `redeemWorkflowParks(runId, scope, decisions)` (package root) hands decisions — the same `WorkflowParkedResume` a resume takes — to a run that is live in this process: each decision whose token names a park the run is waiting on is disposed exactly as a resume disposes a parked ordinal (the checkpoint store's word, then one disposition): a pending checkpoint with an outcome drives the parked child on the same record and ordinal and the call returns its result; an acknowledgement without an outcome keeps the park; a checkpoint the store no longer calls pending settles the call failed (`workflow.park_not_pending`). The answer says what the run did per decision (`driving` / `settled` / `kept_parked` / `already_taken` / `not_parked_here` / `unreadable`), or that it cannot take decisions: `suspending` (with `ended`, settling at the run's end), `not_active` (no run of that id and scope runs here), `not_armed` (no checkpoint store to prove a park against). The run suspends when every call it still has unanswered is a parked one — checked one macrotask after any change, so the script's own continuations run first — or at once when it is cancelled or its deadline fires; then each waiting call is answered with `WorkflowAgentParkedError`, later spawns are refused with it, and the park is the run's terminal, as before. A body that returns while a call of its run is parked or being driven from a park waits for the run to suspend or for every call to be answered. The parked row reads `running` again (its token cleared, a fresh `agent_start`) the moment the drive starts, so `WorkflowRunSummary.parked` goes absent while the run is still running. The fork guard on `resumeFromRunId` of an active run is unchanged.
967
+ - **One journal rule.** The entries of one call reach the journal store in the order they were issued (an append for a call waits for the store's answer to that call's previous append; calls are not ordered against each other) — a park entry whose bounded wait gave up but which commits late can no longer replace the result the same call was driven to.
968
+ - **Hard-breaking (open to veto).** *What changed*: `ctx.agent` / `ctx.agentStream().result()` of a call whose child parks no longer reject at the park; they wait for a decision delivered to the live run or for the run to suspend. Spawns made after a park and before the suspension run instead of being refused. *Who is affected*: scripts that use the park error as control flow (a `catch` that does something else right after the park — it now runs only at suspension); tests and criteria that pin "the sibling queued behind the slot is refused after a park". *Migration*: none for a script that lets the park error propagate (same error, same terminal); write work that should happen while a call is parked in parallel with it rather than in its `catch`.
969
+ - **What still stands (named, not closed).** Calls that chain off concurrent answers (`A → A2` beside `B → B2`) get ordinals in completion order — the existing reason `pipeline` forces divergence; a run that keeps spawning after a park makes it likelier that a later resume of that run finds a live call on another call's parked ordinal and refuses loudly (`workflow.parked_call_changed`) — nothing runs twice. A park-capable run that crashed while running is still not resumable (`record_not_terminal`). A run whose claim grant was superseded drives with its own holder and re-parks on the lost CAS. A revived wa* row restarts its `seq` at 1.
970
+ - **Consumers.** @server (S-2092): the decide route's workflow arm calls `redeemWorkflowParks(runId, scope, [{ token, outcome }])` first — `live` + `driving` / `settled` ⇒ accepted; `kept_parked` / `unreadable` ⇒ the card stays pending; `already_taken` ⇒ an idempotent repeat; `suspending` ⇒ await `ended`, then the existing resume lane; `not_active` ⇒ the existing resume lane. No new bit: `parked` is still `runParkedOf`. @test: G23 / TF3 ("absent after the decision") read as above.
971
+ - **Pins.** `test/backlog2561-in-run-park-redemption.test.ts` (real Runner, R1–R12: both lanes, acknowledgement, delivery answers, replay of a redeemed ordinal, late park append, seat before publish, body return, a resume's re-parked ordinal, chain continuations, an expired checkpoint, a cancel while a body-returned run waits on a driven park); re-pinned to the new law: `backlog642` (a sibling queued behind a park runs; a hung failed-entry append still does not hold a suspending run) and `backlog1514` C13 (a refused drive's slot goes back before its park record is waited on — the queued sibling's leg runs at once).
972
+
973
+ ### The card-edit face reads a command rule's content through the validator's one spacing fold — a stray blank inside the parentheses no longer lands another row than the clean spelling (#2564; behavior change, open to veto; closed sets: none; exports: none) (@server)
974
+
975
+ - **Before.** The edit face's muscle-memory rewrite (one trailing ` *` becomes the canonical `:*`) looked for the star in the text as typed, while the validator reads a command rule's content with its spacing folded — two readings of one spelling. So a stray blank inside the parentheses changed the outcome: `Bash(adb *)`, `Bash(adb *)` and `Bash( adb *)` persisted the prefix row `Bash(adb:*)`; `Bash(adb * )`, `Bash(adb<TAB>*)`, `Bash(adb *<TAB>)` and `Bash( adb * )` reached the validator unrewritten and persisted the wildcard row `Bash(adb *)`; on a chain card `Bash(cd /tmp && adb pull *)` was accepted as the compound prefix while `Bash(cd /tmp && adb pull * )` was refused `unsupported.wildcard`. Blanks AROUND the rule or before its `(` (` Bash(adb *)`, `Bash(adb *) `, `Bash (adb *)`) are not part of this: since 7.56.0 (padded rule names) they are refused `invalid.grammar` with the spelling to copy (`respelling`), on the edit face as everywhere the rule syntax is read.
976
+ - **Now.** One reading: the normalizer folds a command rule's content with the validator's own `foldSpacing`, then looks for the trailing ` *` — one sentence, one canonical row (`Bash(adb:*)`, `Bash(cd /tmp && adb pull:*)`). A blank inside quotes stays data (the rule admits only the command spelled that way); unbalanced quoting passes through to the validator's own refusal; a non-command tool's content is not folded (a path's blank is part of the path). Blanks around the rule stay refused as written — the reference's editor trims what a person typed before it validates, so a surface trims before it submits; `precheckEditedRuleText` and the confirmation answer the same. The settled-replay arm reads a retry through the same normalizer (a retry spelled with a stray inner blank is the same choice). Consequence of the form space probed (the pin states it by corpus, not proof): the card-edit face lands no `wildcard` row for a command text — `exact`, `prefix` and `pattern` remain; the store, import and sync doors read `wildcard` as before.
977
+ - **Behavior change (open to veto).** (1) *What changed:* the four inner-blank forms above that persisted the wildcard row `Bash(adb *)` persist the prefix row `Bash(adb:*)`; the chain form with a blank before the closing paren is accepted as the compound prefix instead of refused. (2) *Who is hurt:* a host whose test pins the wildcard row as reachable through the edit gate (server's `permission-rules-lane` G2w pins the 7.55.0 form, ` Bash(adb *)` → wildcard: on 7.56.0 that text is refused, and here the inner-blank forms land prefix); a wildcard row an earlier version stored from such an edit is not rewritten (it still reads as it did). A settled card edit minted before this version keeps its receipt on a byte-for-byte retry; a retry spelled differently (the same sentence with another blank) no longer recovers it from a stored wildcard row — it reads as the prefix row now, a different row (`selection_mismatch`). (3) *How to migrate:* read the edit gate's reachable matches as `exact` / `prefix` / `pattern`; trim what a person typed before submitting.
978
+ - **Consumers.** @server: the S-2071 question is answered — blanks around a rule: not accepted, since 7.56.0 (loud, `invalid.grammar` + `respelling`); the edit gate's `wildcard` reachability ends with this version — flip the G2w pin to the refusal and narrow the contract's reachable words (§968) to `exact` / `prefix` / `pattern`.
979
+ - **Pins.** `test/backlog2564-edit-spelling-one-fold.test.ts` (E1 seven spellings of `adb *`, each a confirmation on a real card and the redeemed row, one canonical spelling; E2 no wildcard row, the validator itself still reading it; E3 chain card, three spellings; E4 a blank inside quotes is data; E5 unbalanced quoting passes through; E6 a star the body carries is not rewritten; E7 a path rule is not folded; O1 five spellings with a blank around the rule, refused on both faces with the card left open; O2 the respelling is accepted; R1 a settled retry through the fold; R2 a record settled before the fold keeps its receipt on a byte-for-byte retry, and another spelling of its sentence is `selection_mismatch`) — nine cells red on the base, and each guard family red under its own mutation.
980
+
981
+ ### `thinkingOffSpelled` is the one answer to "can this model's thinking be turned off" — derived from the wire table, no catalog key for it; a completions model whose `thinkingFormat` names no dialect now answers `false`
982
+
983
+ - **What changed.** `thinkingOffSpelled(model)` (already on the package root) is the one reader of whether the wire carries a disable
984
+ key for a model, and the one answer a host that lists models serves. It is derived from the entry's wire table (API family,
985
+ `compat`, `thinkingLevelMap.off`, `reasoning`), so `Model` grows no key for it, the way `reasoningEffortLevelsOf` derives the
986
+ supported effort tiers (Claude Code derives its catalog row's `supportsThinkingOff` the same way — it declares nothing per row).
987
+ The reader's JSDoc states the contract, and `docs/sdk/03-brains-and-models.md` and `INTEGRATION-CORE.md` §1.10f say how a host
988
+ wraps it. The chat-completions dialect names are now ONE closed table (`completionsFormatOf`, `brain/reasoning.ts`) that both the
989
+ reader and the wire applier's dialect lookup read, so what the reader claims and what the wire carries are one judgement of the
990
+ name.
991
+ - **Behaviour change (a malformed `compat` corrected; open to veto).** What changed: a completions-family model whose
992
+ `compat.thinkingFormat` names no dialect — a string outside the closed set (`openai`, `openrouter`, `deepseek`, `together`,
993
+ `zai`, `qwen`, `qwen-chat-template`), or a value that is not a string (an array, an object, a boxed string) — answered `true`
994
+ ("the wire spells off") while its wire carried no thinking key of any tier; it now answers `false`. Who is hurt: a host that
995
+ handed the engine such a model object and relied on the `true` — the classifier's answer-only seat then takes the `low` tier with
996
+ the thinking allowance in its output cap where it took `off` before, the `classifier.role_defaulted` line says "cannot turn
997
+ thinking off" for it, and a listing that wraps the reader serves `false`. A server's catalog never reached this: its compat
998
+ reader drops a bad format before the engine sees the model. How to migrate: declare `compat.thinkingFormat` as one of the seven
999
+ names (or leave it undeclared for the stock `openai` rules); nothing else changes — an undeclared format and every well-formed
1000
+ one answer as before.
1001
+ - **Consumers.** @server: a model row served by `GET /v1/models` can carry `supportsThinkingOff` as `thinkingOffSpelled(model)`
1002
+ (the same wrapping as `supportedEffortLevels` from `reasoningEffortLevelsOf`), next to the row's `reasoning`. @client-core: read
1003
+ that key where it is served instead of probing, and a probe's `cannot_disable` verdict need not delete the entry's
1004
+ `thinkingFormat` to be recorded. @test: the reader and the applier's `off` arm answer alike for every model of the reasoning wire
1005
+ matrix.
1006
+ - **Pins.** `test/backlog2566-thinking-off-reader.test.ts` (R1 the root carries the reader and a `supportsThinkingOff` on the entry
1007
+ changes nothing; R2 reader ⇔ the applier writes a disable key, every matrix model; R3 the same through the real brains; R4 a
1008
+ malformed `thinkingFormat` answers `false` from the shared judge; KL a malformed format is still accepted and writes nothing on
1009
+ every tier).
1010
+ - **Known limit.** A bad `thinkingFormat` is read as "no dialect" and accepted: nothing refuses it where the model enters the
1011
+ catalog and nothing announces that no thinking key will go out for it.
1012
+
1013
+ ### A case twin stored in a session's rules can be taken out by its owner: the write judge and the tighten comparison read one roster (#2573; behavior change (one refusal retired), open to veto; hard-breaking: none; closed sets: none; exports: none) (@client-core @server)
1014
+
1015
+ - **Before.** A session's stored rules holding a case twin of an engine tool (`toolDeny: ["BASH"]` beside the catalog's `Bash`,
1016
+ written past the write gate by an older release) could be neither kept nor removed by the session's owner. The write judge
1017
+ read an entry's tool name against the engine's catalog ∪ the names the store declares (`mountedToolNames`) — `BASH` reads as
1018
+ nothing there, so a write that kept it was refused `inert_entry` — while the tighten comparison (`loosenReasons`) read it
1019
+ against the declared names alone, where `BASH` is a tool's name: taking it out was a loosening, so the owner's (non-operator)
1020
+ write was refused `loosen_forbidden`. Every run of that session was refused at the policy audit (`config.legacy_tool_name`)
1021
+ in the meantime; only an operator write cleared the row. Two readers answered one question two ways.
1022
+ - **Now.** One roster, defined once (`storeRoster`: the engine's catalog ∪ the store's declared names) and read by both: the
1023
+ write judge's refusal of an entry that reads as nothing and the comparison's effective set. A case twin of an engine tool is
1024
+ nothing on both seats, like a retired name already was (#1156): the owner's write that drops it is accepted (the deny list
1025
+ guarded nothing), and adding one to an allowlist widens nothing. A write that KEEPS the twin is still refused `inert_entry`
1026
+ (the entry names nothing; the way out is to drop it). A twin the store declares is a real tool's name on both seats and
1027
+ guards as before.
1028
+ - **Behavior change (open to veto).** *What changed*: removing a case twin of an engine tool's name from a stored session deny
1029
+ list is no longer a loosening (the exported `loosenReasons` and the write judge it feeds, both bundled stores). *Who is
1030
+ hurt*: a deployment that mounts its own tool under a case twin of an engine tool's name (its own `BASH` beside `Bash`, or a
1031
+ tool alias spelled so) and did not declare it in the store — the write judge already refused writing such a deny (it read the
1032
+ name as nothing); a stored one (an operator's, written before the write gate read catalog twins) can now be removed by a
1033
+ non-operator write, as a retired name's already could (KL-2573-a). The store has no run's roster, so only its declaration can
1034
+ tell a live twin from a dead one. *How to migrate*: declare the name in the store's `mountedToolNames` (the store then guards it
1035
+ as before).
1036
+ - **Consumers.** @client-core: a removal write for a stored case twin (the `nothing_to_remove` class) is accepted now, in the
1037
+ same write that may tighten; a write that keeps the twin is still refused. @server: no change — a durable backend calling
1038
+ `judgeSessionRulesWrite` / `loosenReasons` reads the one roster with the new engine; a backend that re-implemented the
1039
+ comparison keeps its own reading.
1040
+ - **Pins.** `test/backlog2573-store-roster-one-reader.test.ts` (A the twin is removable through the write judge, the file
1041
+ store on disk and the pure predicate, the allow side too; B a write that keeps the twin, a real loosening beside it, a
1042
+ declared twin and a retired name keep their answers; C the one-answer law over a 28-entry corpus × five declarations: refused
1043
+ as written ⟺ removal is no loosening / adding widens nothing). `test/backlog1930-near-miss-refuses.test.ts` D3 re-pinned to
1044
+ the surviving reading (the owner takes the twin out and the session runs again).
1045
+
1046
+ ### The policy-name refusal carries its problems as data: `config.legacy_tool_name` `detail.problems` with the layer (#2574; additive; hard-breaking: none; closed sets: none; exports: none) (@client-core @server)
1047
+
1048
+ - **Before.** The prepare-time policy-name audit refused a run at the first entry it met (`config.legacy_tool_name`) with a
1049
+ sentence and no `detail`; a shell that wanted to say which entries of which layer were refused had to parse the sentence,
1050
+ and a layer with several refused entries showed one per rejected attempt.
1051
+ - **Now.** The audit reads one layer's whole pass before it refuses. The sentence and the code are the first entry's, unchanged;
1052
+ `detail` is `{ problems, total }`: one problem per refused entry in scan order (list by list, entry by entry), each
1053
+ `{ code: "config.legacy_tool_name", field, entry, reason, layer }` — the four keys of an entry problem of a rejection body
1054
+ plus the layer: `field` the policy list (`deny` / `ask` / `allow`), `reason` the name judgment's verdict word (`retired` ·
1055
+ `unprefixed_namespace` · `unmintable_segment`, and `near_miss` for another spelling of a carried name), `layer` one of
1056
+ `caller_policy` · `session_rules` · `ancestor_session_rules` (an ancestor's rules are not asked for case twins, as before).
1057
+ At most 32 problems are listed and `total` is the real count; an entry longer than 128 characters (a sentence, not a name) is
1058
+ counted in `total` and not listed — never shown cut (the failed terminal's `coded` snapshot drops a string past its bounds and
1059
+ a list that does not fit whole, so the list is bounded where it is minted). A single refused entry is a list of one.
1060
+ - **Limits (stated).** The layers are read in order (the run's own policy, then the session's rules, then each ancestor's):
1061
+ the layer that refuses first is the one listed, and a later layer's entries show once it is fixed (KL-2574-a).
1062
+ - **Consumers.** @client-core: read `detail.problems[].layer` instead of the sentence. @server: the same shape for the body
1063
+ it mints (`detail.problems` of the rejection body, A.1f — `layer` is core's one addition: the layers are core's).
1064
+ - **Pins.** `test/backlog2574-legacy-name-problems.test.ts` (A one pass lists every refused entry with its verdict word, the
1065
+ ask list a list of its own, a list of one is the same shape; B the layer word on each of the three layers; C the sentence and
1066
+ code are the first entry's; D the first refusing layer is the one listed; E through a real Runner the failed terminal carries
1067
+ the detail, cut at 32 with `total` 40, thirty-two names of 128 characters all kept, an entry past 128 counted and not listed).
1068
+
1069
+ ### A halt reaches a compaction pass's summary call whatever the brain does with the signal (#2875; behavior change: a summary call that ignores the abort is no longer waited out)
1070
+
1071
+ - **Before.** Since 7.52.0 (#1828) every stop verb cuts every compaction pass, and the first-party brains end the summary call on the abort in every wait (connect, first byte, stream, a 429 / 503 `Retry-After` backoff). But the summary call itself trusted the brain to look at `options.signal`: a brain or gateway wrapper that did not held the pass — and `done` — until it answered (KL-1828-d).
1072
+ - **Now.** The summary call is an await of the pass like the summary-provider consult and the credential callback: the pass signal firing first answers the call's abort outcome at once (a signal already fired dispatches nothing), the late answer is dropped, and the pass reads `compaction_outcome{outcome: "mooted", reason: "cancelled"}` with its `trigger` and the `pass` its `compaction_started` carried, as every cut pass does. One site (`completeSummarization`), every lane (turn boundary, prompt-too-long recovery, guard chain, end of task, `compactSession`). A halted run reads `haltedByUser: true`, `terminal.kind: "completed"`. KL-1828-d is closed.
1073
+ - **Consumers.** @server @cli: no wire change, no new word. A halt landing while auto-compaction is summarizing ends the run within a second on every brain; where it landed is the open pass (`compaction_started` without its concluding frame until the `mooted` / `cancelled` one).
1074
+ - **Pins.** `test/backlog2875-halt-compaction.test.ts` (① the turn-boundary auto pass: the summary request held open, a brain that ignores the signal; ② the recovery pass, both forms; ③ the summary call in a 429 backoff, a brain that sleeps a 429 out without the signal, the main call in a 429 backoff with its `gave_up` / "cancelled while retrying" status; ④ controls; ⑤ the call site: an already-fired signal dispatches nothing, a held call is cut, a stop inside the call's setup wins over the setup's throw, a late rejection goes nowhere, no listener per call); `test/backlog1828-stop-during-pass.test.ts` W1 re-pinned; `test/manual-compact-cancel.test.ts` ②c re-pinned to its own case — a real failure whose answer arrives before the cancel stays a failure (a cancel that fires before the answer is the call's abort outcome).
1075
+ - **Limits.** A brain that ignores the signal keeps its call running after the cut, and its usage may be booked after the result is assembled (KL-2875-a). The main model call still trusts the brain to look at the signal (#2882).
1076
+
1077
+ ### Every refusal of an approval decide names where the fault is — a defect of the caller's payload carries `detail.field`, a damaged row carries `detail.reason`, and `detail.reason` on `checkpoint.invalid_outcome` is a row-integrity word and only that (#2592; additive wire words; hard-breaking for a host's total table over `CheckpointErrorField`, open to veto; closed sets: `CheckpointErrorField` 10 → 14 (+`outcome`, `updatedInput`, `decision`, `reason`), `CheckpointErrorReason` 29 → 30 (+`bound_input_hash_missing`); exports: none — two exported types gain members) (@client-core @server @cli @sdk)
1078
+
1079
+ - **Before.** Seven refusals on the approval decide's path carried no word at all: the resume's outcome that is not an object, a person's edit (`updatedInput`) that is not plain JSON data, a decision word outside `allow` / `deny`, a deny `reason` that is not a string, a deny `reason` carrying the engine's `<system-reminder>` tag — and a binding-version row whose pending call has no persisted `boundInputHash`, the one refusal of the seven that is a damaged row (the seventh, the admission's "no lane captured it" belt, is an engine invariant). A host that told a dead row from a fixable decide by "no `field` ⇒ row integrity" — a reading from absence — took the five payload defects for dead rows and hid cards a person could have fixed.
1080
+ - **Now.** One rule. A refusal of an approval decide carries at most one of two words: `detail.field` when the decision, or the live state it was bound to, is what is at fault (`outcome`, `updatedInput`, `decision`, `reason` join `boundCallId` / `boundInputHash` / `boundGrantDigest` / `answer` / `hostDecision` / `approver` / `feedback` — each word's entry in the registry gives its recovery: correct the cell, re-fetch the pending list, deny and re-activate), or `detail.reason` when a facet of the stored row is damaged (`bound_input_hash_missing` joins the ten row-integrity words the resume screen already spelled — no re-decide of that row helps). `detail.reason` on `checkpoint.invalid_outcome` is therefore a row-integrity word and only that, which is what lets a host read "`reason` present, `field` absent" as positive evidence of a damaged row. A refusal with neither word is never a damaged row: it is a payload of a lane the set does not word yet (the review lanes' text payloads and the resource-limit lane's decision word) or an engine-invariant belt (the admission's lane-capture belt and the two post-claim belts of the resumed leg), and `docs/CLOSED-SETS.md` says so. The sentences are unchanged and a refused decide still leaves the row pending and redeemable.
1081
+ - **Hard-breaking (open to veto).** *What changed:* `CheckpointErrorField` gains four words (`outcome`, `updatedInput`, `decision`, `reason`), 10 → 14; `CheckpointErrorReason` gains `bound_input_hash_missing`, 29 → 30. *Who is hurt:* a deployment that holds a `Record<CheckpointErrorField, …>` (the server's `CHECKPOINT_ERROR_FIELD_ROLE`) is tsc-red on the four new words until it places them, and a host that switches over `detail.field` without a default meets an unplaced word at run time; a host that read the missing `reason` of the damaged-row refusal as "not a row" now reads it as a row. *Migration:* add four rows — all four are cells of the decision whose recovery is the one `feedback` / `hostDecision` already have (re-decide the same token with the cell corrected); treat `reason: "bound_input_hash_missing"` like the other row-integrity words (no re-decide of that row helps, and no worker version redeems a row that lost its hash: expire it and start the task again).
1082
+ - **Consumers.** @client-core: `refusedParkFromError` can delete its "no `field` ⇒ row integrity" inference and judge on the positive evidence — `field` present = an intact row whose decision is what to act on (by the word's own recovery), `reason` present and `field` absent = a damaged row, neither = not a row; the five payload refusals above stop reading as dead rows once the engine ships. @server: the HTTP 409 for `checkpoint.invalid_outcome` passes `detail.field` / `detail.reason` through unchanged (add the four field rows to the role table). @cli, @sdk: nothing to place; the two words are the ones the decision card already branches on.
1083
+ - **Docs made true.** `docs/CLOSED-SETS.md`: the `CheckpointErrorReason` bullet said twenty-eight words while the union held twenty-nine (7.56.0's `skill_scopes_damaged` was not counted) and now says thirty and names every word, the three `installation_record_*` words by their full names; the one rule and the line of row-integrity words are registered there, and the pin reads them against the source.
1084
+ - **Pins.** `test/backlog2592-policy-ask-refusal-words.test.ts` (F1–F5 each payload refusal on a real Runner over a durable park — `detail` exactly `{ field }`, the sentence as before, the row pending and then redeemed; R1 / R2 the damaged-row refusal `{ reason: "bound_input_hash_missing" }` against its control; C the words that already existed; T1–T4 the registry's counts and member names, the line of row-integrity words equal to what the source's `invalid_outcome` constructions spell, no construction spelling both words; Z1 / Z2 the census of constructions with neither word — the review lanes and the belts — exactly as registered); re-pinned: `backlog2240-feedback-refusal-field` T (the total table, 10 → 14 rows).
1085
+
1086
+ ### One composition for the environment a child spawned on this host starts with — the tool shell's children and a settings command hook's children read the same `hostChildEnv(policy, ...grants)`, so a hook starts from what the run's shell starts from (#2607; behavior narrowing, open to veto; the hook-runner widening it enables is a security widening, open to veto; hard-breaking: none; closed sets: none; exports: +3 `hostChildEnv`, `hostChildEnvProblem`, type `HostChildEnv`) (@server @cli @test)
1087
+
1088
+ - **Before.** The same settings hook saw two environments. A hook the engine host runs (the nine engine-owned events) started from a fixed list of twelve names (`PATH`, `HOME`, `SHELL`, `USER`, `LOGNAME`, `TERM`, `TZ`, `LANG`, `LC_ALL`, `LC_CTYPE`, `LC_MESSAGES`, `TMPDIR`) plus the request's `settings.env` and the hook's own variables — a list kept beside the server's hook runner, not by the engine; a hook the terminal client runs itself started from its whole process environment. The client documents the split as undecided (its KL0157). The engine composed its tool shell's child in `NodeExecutionEnv` alone (`inheritedHostEnv(policy)`, then `shellEnv`, then a per-exec `env`), so a hook runner that wanted the shell's reading had no function to call and wrote its own. A `shellEnv` / per-exec `env` entry whose value was not a string reached the child as Node's string of it (`[object Object]`, `8080`), and a name holding `=` became another name / value pair in the child.
1089
+ - **Now.** `hostChildEnv(policy, ...grants)` (`src/core/secret-env.ts`, root export) is the one composition: `inheritedHostEnv(policy)` (a deployment's own credential is never inherited, under any policy, by name or by value), then each grant in order, each read once and after the inherited part (a later grant's value wins, names compared as the platform compares them — case-blind on Windows, the value landing under the spelling the name first had; an `undefined` value unsets; a grant is the host's explicit act and is not read against the declaration, as before), and `withheld` = the inheritance's findings the composed environment does not set. `NodeExecutionEnv` composes every shell and background-shell child through it, and its `withheldEnv` / `homeDir` / `cdPath` reads read it. The reading for a command hook (INTEGRATION-CORE §11, "A hook's child"): `hostChildEnv(<the run's inheritEnv>, <the run's settings-env grant>, <the hook's own variables>)` — the same policy the run's tool shell inherits under (absent ⇒ `"scrub"`; where no tool shell runs on the hook's host, the run's word applied there), never a hook-specific list; the one place that configures it is the run's `inheritEnv`; no hook-specific knob. The reference's ordinary command hook builds its child on the same base as its Bash tool (`cli309.js:266734`, `:266386`; base `:253905`: the process environment with the trusted settings env, its own credentials removed); its exceptions — a first-party plugin allowlist handed the reference's own token as an explicit grant (`:266738`), organization hooks composed over the pre-plugin settings env (`:266737`), a hook run inside a served sandbox — have no counterpart here. `hostChildEnvProblem(policy, ...grants)` answers the same judgment without throwing. A grant that is not an object or throws while read, a name that is empty or cannot keep its spelling in a child's environment (`=`, NUL, an unpaired surrogate), and a value that is not a string or cannot keep its spelling (NUL, an unpaired surrogate) are refused with a sentence that names the grant and the variable (never a value, nor a bad name).
1090
+ - **Behavior narrowing (open to veto).** *What changed:* `NodeExecutionEnv` refuses a `shellEnv` that is not a map of variable names to strings at construction (`NodeExecutionEnv: env grant 1: X must be a string (got number)`), and a per-exec `env` or a `shellEnv` made bad later makes that spawn answer `spawn_error` with the same sentence (the reads answer "none"). *Who is hurt:* a JavaScript host passing a number, a boolean or an object as an environment value (it reached the child as Node's string of it), a name holding `=` (the child read another pair), or a name or value holding an unpaired surrogate (the child read U+FFFD); a `shellEnv` whose read throws. *How to migrate:* pass strings (`String(port)`) and names the platform keeps. Smaller changes, same direction (the account and the child now agree): a withheld name the child does not receive is counted withheld even when a grant mentions it — a per-exec `env` entry set to `undefined` over a `shellEnv` re-supply, or a grant property a spread never copies (non-enumerable, inherited), used to count as re-supplied; on Windows a later grant spelling a name in another case now sets that name's value (Node kept whichever spelling sorted first, so the earlier value could win); the inherited part is read before any grant, so a grant's getter cannot change what is inherited.
1091
+ - **Security widening the hook runner adopts (open to veto).** *What changes when a hook runner reads it:* an engine-run hook sees what the run's tool shell sees — on the terminal client's interactive lane (`inheritEnv: "all"`) the host's whole environment less the deployment's own credentials, under the default `"scrub"` every variable whose name and value are not credential-shaped — instead of twelve names. A server runs settings hooks on its worker in every single-user deployment, whatever the tool lane, so the widening reaches a single-user sandbox or in-process lane too (the run's word there; `"all"` is refused outside the host lane, so `"scrub"`). *Who is hurt:* a deployment that relied on the short list to keep a non-credential variable of the worker from a request-supplied hook command — a capability or routing variable such as `DOCKER_HOST` pointing at an administrative daemon is neither a declared own credential nor credential-shaped, and reaches the hook as it reaches the shell (a hook is a command run on that host with the host's user; nothing here changes what such a command can otherwise reach). *How to migrate:* declare such a variable with `declareWithheldEnv` (it is never inherited by any child, the shell's included), or narrow the run's `inheritEnv` to a list.
1092
+ - **Consumers.** @server: the hook runner (`src/hooks/hook-runner.ts`, `HOOK_ENV_ALLOWLIST` and its two readers) composes the command leg and the HTTP interpolation with the word the same request carries for its shell (bound when the request's hook registration is minted, not read later from the mutable session map) and the request's `settings.env` grant, the hook's own variables last — its `CLAUDE_PROJECT_DIR` rule (the request's value when it sets one, else the hook's directory) becomes how that last grant's value is computed, not an order; the twelve-name list retires; a `settings.env` name the composition refuses (`=`, NUL, an unpaired surrogate) is best refused at intake, where the request can be told; the host-lane shell's own composition (`composeHostSubprocessEnv` with `collapseWin32EnvKeys`) can read `hostChildEnv` so the shell and the hooks share one. @cli: the local executor (`localHookChildEnv`) and the HTTP hook's interpolation (it reads the raw process environment) compose with `hostChildEnv(<the word the session hands the engine>, <the settings env grant>, <the hook's variables>)`; the engine needs the same settings env grant for one reading (the default rules projection does not send `env`); the client process declares its own credentials (`declareWithheldEnv` — none is declared in the client today), which takes the place of the conditional subprocess list for hook children; on the interactive lane (`"all"`) the local hooks keep the whole environment less those, and where the client sends no word (non-interactive, missing capability evidence) the local hooks narrow to the scrub — the client's to rule on under C-R284 ④ (send the word, or accept); KL0157's environment half retires when both runners land (its compaction and run-placement halves stay). @test: one hook configuration run in the shell and in the engine reads the same terminal-exported variable and the same settings-env variable, and neither reads a declared own credential.
1093
+ - **Pins.** `test/backlog2607-host-child-env.test.ts` (C1 the root export; C2–C4 the three policies with a declared own name and a value copy; C5 grant order, re-supply, unset and a `__proto__` name; C5b Windows name comparison, platform stubbed; C6 the hook composition = the shell's plus the hook's variables, the withheld account the composition's (a hook variable re-supplying a withheld name); C7 / C7b thirteen bad shapes, a throwing grant, a revoked Proxy, a throwing policy element and the problem reader; C8 a grant read once, after the inherited part, a non-enumerable property not composed; C9 / C9b a real shell child and a real background child print what `hostChildEnv` composed; C10 the shell's refusal at construction and at the spawn, the reads composing once; C11 the shell's composition site calls the one reading).
1094
+
1095
+ ### Every model call the engine prices has one account line, and a host reads its lines through one fold: `Runner.sideQuery`'s result carries its engine-priced `account` (origin `side_query`), `accountRowOfResult` derives a finished run's line, `foldAccount` is the only reading (#2608; additive; hard-breaking: none; closed sets: +1 `AccountOrigin` = `run` | `side_query`; exports: +7) (@server @cli @client-core)
1096
+
1097
+ - **Before.** A one-shot side query (`Runner.sideQuery`: an auto title, a recap) belongs to no run, so its spend never reached a run's `stats`, and the result carried only the provider's own `usage.cost` (an adapter amount: higher than the engine's figure whenever a cache hit is priced at the full rate). A host that kept a session total had nothing to add for it, and every host that summed run results wrote its own sum, with its own reading of an unpriced run.
1098
+ - **Now.** One pricing function, `priceReply`, prices a reply for both readers: the run's side-call lanes (`recordSideCallUsage`: compaction, classifier, mods, advisor — same arithmetic, telemetry marks and trace row as before) and the side query. `SideQueryResult.account` is the query's line — `{ origin: "side_query", tokens, costMicroUsd?, usageMissing? }` — priced from the deployment's table (`RunnerDeps.pricing`, else the model's own `cost`) on the raw provider usage and the model that ANSWERED (a degraded reply is priced at the fallback). `costMicroUsd` is absent when that model has no price table (unknown, not a fabricated 0 — the run's law), and a reply that reported no usage gives `tokens: 0` with `usageMissing`. `accountRowOfResult(result)` is a finished run's line (`origin: "run"`): its own usage plus its delegated sub-agents' — the reconciled spend. `foldAccount(rows)` is the one reading: a total and one total per origin; an unknown cost makes every total containing it unknown; a malformed line (origin outside the closed set, a negative or non-finite number, a cost spelled as anything but an absent key) is refused with `TypeError` naming the line. The engine does not keep the lines: the host does.
1099
+ - **Limits (stated, not exceptions).** A run's line is exactly what the result's `stats` says, and `stats` is not one interval across a durable resume chain, so the line is exact for a run that did not durably resume and that no background delegate outlived. (1) The lines of ONE durably resumed chain are not additive: a leg's own usage is per leg (a `resource_limit` chain folds the earlier slices into its result), its delegated usage is cumulative over the chain, so one line per result double-counts the delegates and one line per chain loses the earlier legs' own spend. (2) An unknown cost does not survive a `resource_limit` chain (the resource ledger folds a slice's cost as a number) or the delegation rollup (a delegate's `usageMissing` mark has no seat in the nested usage), so a line can read complete when it is a lower bound. (3) Spend a background delegate books after the parent's result settled is in no line. (4) The prompt-suggestion pass and memory consolidation book after the result on their own budget-excluded faces and are not in it. (5) A malformed price table prices at its arithmetic on a side query, as it does for a run whose deployment governs no money. The structural fix for (1)–(3), an account the engine keeps and mints at each interval, with persistence and restore on resume, is its own design ticket (#2620).
1100
+ - **Consumers.** @server: a session-level aggregate is `foldAccount` over per-run `accountRowOfResult` lines plus the side-query lines you receive from your side-query verb (carry `account` through; a layer that rebuilds `SideQueryResult` must now supply it). @cli: the auto-title side query keeps the returned `account` line beside the run lines; the result frame field is the second section below. @client-core: the history-lines read port supplies rows, not transcript `usage.cost.total` (the adapter amount).
1101
+ - **Pins.** `test/backlog2608-session-account.test.ts` (the line's price, the table read from the Runner, the model's own cost with a cache subset, parity with the run's side-call pricing, usageMissing, the degraded reply priced at the fallback, the run line before and after a side query, the fold's totals / unknown propagation / loud refusals, the import graph: one pricing function with two readers, one mint per origin).
1102
+
1103
+ ### A session's total is the fold of the lines the host kept: a result frame's `total_cost_usd` is the run lines including a resumed session's history, and a side query's line is in the session's cumulative figure but not in that field (#2610; behavior change for a host that adopts the reading, open to veto; hard-breaking: none; engine behavior unchanged) (@cli @server @client-core)
1104
+
1105
+ - **Before.** A `-p` result frame's `total_cost_usd` was this run's cost only: resuming a session showed the new run's number as if the session had spent nothing before, and a side query's spend was in no number.
1106
+ - **Now.** The reading is fixed once, in `foldAccount`'s output, and every host maps its fields from it: a result frame's `total_cost_usd` = `byOrigin.run` over the history lines the host supplies plus this run's line (a new session = this run; a resumed session = history + this run; a fork's seed lines under the cli's fork ruling are ordinary `run` lines), `byOrigin.side_query` is kept apart, and a session's cumulative figure = the whole fold. This is the Claude Code 2.1.309 split: one cost ledger whose result-frame total is the main-loop total including the restored history, with a separate side-query line that enters only the cumulative metadata.
1107
+ - **Behavior change (open to veto).** A host that adopts the reading shows a larger `total_cost_usd` on a resumed session than before (history is now in it). One erratum against the ticket wording "a side query's cost enters the session account": the line IS in the account; what differs from the wording is only which host field reads it (frame total = `byOrigin.run`, cumulative = the whole fold). The engine itself changes no behavior in this section: it supplies the lines and the fold.
1108
+ - **Consumers.** @cli: map the frame field from `byOrigin.run`, the `/cost` cumulative from the whole fold, and supply fork / resume history as `run` lines; the numbers on resumed sessions grow. @server: the session-level aggregate states which total it serves ("includes fork inheritance" per the cli ruling). @client-core: the read port returns the lines.
1109
+
1110
+ ### The cwd-alignment observation asks the env's shell whether its directory is the task root instead of comparing two path strings — no misreport on a Windows Git Bash shell, and the sentence states only what was read (#1734; behavior change, open to veto; hard-breaking: none; closed sets: none; exports: none) (@test)
1111
+
1112
+ - **Before.** The env probe's `pwd -P` line and the task root were each passed through the env's `canonicalPath` and compared
1113
+ as strings. On Windows the task root is a native path (`C:\Users\x\proj`) while a Git Bash shell reports the same directory
1114
+ in its own POSIX spelling (`/c/Users/x/proj`, `/tmp/…`), which no engine-side canonicalizer maps — so every session (an
1115
+ ordinary `C:\` root as much as a TEMP-mapped one, and a host-lane remote env) reported `execution-env cwd misalignment …`
1116
+ through `onError` (`phase: "config"`) while the shell sat in the task root. The sentence also said an exec or an artifact
1117
+ "can land outside the workspace root", which was untrue on every such report. A POSIX host has one spelling and never saw it.
1118
+ - **Now.** The probe's plain `pwd -P` line becomes one self-framed line, `cwd-alignment:<word>:<the shell's pwd -P>`, found
1119
+ by its marker (a banner a login shell prints first moves nothing, and a reading that holds the separator stays one
1120
+ reading). The shell itself asks whether its working directory and the task root are one directory — `test . -ef <root>`,
1121
+ the same filesystem object, links followed, spelling and case irrelevant — and the word is `same`, `differ` or `unknown`.
1122
+ The engine reads only that word; no path mapping and no second canonicalizer exist on the engine side. `differ` reports;
1123
+ `same` is silent (a symlinked root, a differently cased spelling on a case-insensitive volume and an msys-spelled pwd are
1124
+ one directory to the shell); `unknown` — the root is not a directory in the shell's view, or its `test` has no `-ef` — and
1125
+ an absent framed line are no judgment and report nothing. The sentence now states the two readings and the shell's
1126
+ finding and nothing about where an exec or an artifact would land.
1127
+ - **Behavior change (open to veto).** *What changed*: (1) the observation no longer fires on a spelling difference — only on
1128
+ the shell's own `differ`; (2) a shell that cannot see the task root as a directory answers `unknown` and is silent (before:
1129
+ any string difference reported); (3) the sentence lost its cause and landing clauses; (4) the guard no longer calls
1130
+ `canonicalPath` (one round trip less per prepare on a remote lane); (5) the probe's ninth output line is the framed line, so
1131
+ a scripted `exec` double that answers the probe with a bare `pwd` line yields no observation (before: it was compared).
1132
+ *Who is hurt*: a deployment that matched the old sentence text beyond `cwd misalignment`, a test double of that shape,
1133
+ and a lane whose real fall-through also happens to hide the root from its shell (the observation is advisory — the
1134
+ sandbox fence is unchanged). *How to migrate*: match on `cwd misalignment` only; a double that wants the observation
1135
+ answers the framed line.
1136
+ - **Consumers.** @test (the Windows machine): run two shapes with Git Bash as the env shell — an ordinary `C:\` task root and
1137
+ a TEMP-mapped one, expecting zero `cwd misalignment` observations; and a lane that starts the shell elsewhere, expecting
1138
+ exactly one with the fact-only sentence. @server: the host-lane remote env needs no change — the question runs in the
1139
+ host's own shell, in the same POSIX-`sh` family as the rest of the probe.
1140
+ - **Limits (stated).** Pinned on a real bash with an msys-shaped `pwd` / `test` stand-in; the real Windows reading is the test
1141
+ seat's. A shell that cannot parse a native `C:\…` root at all (a WSL bash) answers `unknown` — silent, not a false report.
1142
+ The probe's earlier lines are still read by position, so a login banner shifts those facts as before; only the
1143
+ alignment line is immune.
1144
+ - **Pins.** `test/backlog1734-cwd-alignment-same-directory.test.ts` (Runner: one directory in two spellings is silent; the
1145
+ shell's word is read, not two strings — `same` over unequal strings silent, `differ` over equal strings reports, `unknown`
1146
+ and an absent framed line silent; a really different directory reports once with a fact-only sentence, also behind a login
1147
+ banner; a symlinked root silent; the same two shapes behind the remote seam; the fragment on a real bash: Windows-style
1148
+ root × msys-style shell `same`, different directories `differ`, unenterable root `unknown`, symlink `same`, a case alias
1149
+ under an inherited `PWD` `same`, hostile spelling stays data, a reading holding the separator or a newline; the reader's
1150
+ closed set); re-pinned: `env-git-probe` ([1451] drop-cwd sentence).
1151
+
1152
+ ### A store that copies a fork's rows as bytes picks them with the engine's own parent-chain walk: `selectForkRows` is exported from the package root (#2565; additive exports; behavior change: none; hard-breaking: none; closed sets: none; exports: +3 (1 runtime: `selectForkRows`; 2 types: `ForkPoint`, `ForkRowShape`); members: none; signatures: none) (@server)
1153
+
1154
+ - **Before.** A session store with no `SessionRepo.fork` to hand `settledForkOptions`' answer to (the two SQL stores move rows as bytes) had to pick the rows itself: the path from the root to the fork point, the loop refusal and the missing-parent refusal. Server wrote that walk a second time (`forkCopyOf`, copied from `getPathToRoot` and its `ParentChainGuard`, which is not a root export).
1155
+ - **Now.** `selectForkRows(rows, sourceLeafId, point, { idOf, parentOf })` answers `{ rows, leafId }`: `{}` copies every row and keeps the source leaf; `{ entryId, position: "at" }` copies the rows on the path from the root to that entry (the path, not a prefix of the log) in the log's order, and the leaf is that entry. It runs the very walk `getPathToRoot` now runs (one `pathToRoot` in `src/engine/session/path-to-root.ts`; the storage keeps no parent-chain loop of its own), so a loop is refused `invalid_session` with the shared sentence, a parent the log does not hold is `invalid_session` naming it, and a fork point the log does not hold is `invalid_fork_target`. A fork point that names an entry with a `position` other than `at` is refused `invalid_fork_target` (a point with no entry id is the whole-log form, as `getEntriesToFork` reads it). `ForkPoint` (the answer `settledForkOptions` gives, now its declared return type) and `ForkRowShape` ride with it. No behavior change on any existing call: `getPathToRoot` answers as before, including an empty parent id reading as "no parent" there (a separate ticket makes that a refusal).
1156
+ - **Who is hurt.** Nobody: one added runtime export and two types, no existing name or shape touched.
1157
+ - **How to migrate.** *server*: delete the hand-written walk in `forkCopyOf` (`src/plugins/session-fork.ts:34-61` at `e1ea5bd2`, line numbers move) and answer it with `selectForkRows(rows, sourceLeafId, point, { idOf: (r) => r.entry_id, parentOf: (r) => r.parent_id })`; the two call sites (`tidb-session-store.ts`, `pg-session-storage.ts`) stay. The refusals are byte-identical (loop, missing parent, missing fork point, an empty `parent_id` refused as a missing parent); the one sentence that changes is a non-`at` position's (same code `invalid_fork_target`, worded for the reader rather than for a store), and one corner tightens: a row whose id is the empty string is no longer followed as the parent of a row whose `parent_id` is the empty string (the copy followed it; the reader refuses, as it does when no such row exists). `type ForkPoint = Awaited<ReturnType<typeof settledForkOptions>>` may become the root's `ForkPoint` (the same shape). Server deletes its copy on arrival.
1158
+ - **Pins.** `test/backlog2565-fork-row-selection.test.ts` (R1 the root exports; R2 `{}`; R3 `at` copies the path not a prefix, log order kept, rows not cloned; R4 the sound / loop / broken shapes and a missing point answer as `getEntriesToFork` does over a storage holding the same log; R5 the refusals in full; R6 an empty parent id; R7 one walk — the storage takes it from the module and keeps no `ParentChainGuard` loop, a floor still ends its walk); `test/export-surface.snapshot.json` (+3).
1159
+
1160
+ ### A refused MCP call says what happened to this call and how a later one is judged, not that every later call is refused the same way (#2589; behavior change in the text the model reads, open to veto; hard-breaking: none; closed sets: none; exports: none; members: none) (@server @cli @test)
1161
+
1162
+ - **Before.** A call the connection's listing refuses before it is sent (`mcp.tool_declaration_changed`, `mcp.tool_listing_unreached`,
1163
+ `delivered: "no"`) ended: "Every call to this tool is refused the same way until a listing of the connection reaches the tool, so
1164
+ don't retry it." A host's re-dial or refresh lands between two calls, and the call after it runs: the sentence foretold the
1165
+ opposite of what happened, and its advice made an agent give up a tool that worked again (reported by the black-box acceptance:
1166
+ the next call after `reconnectMcpServer` ran).
1167
+ - **Now.** The same two refusals end: "The server did not execute it. This refusal is for this call only: each later call to this
1168
+ tool is checked again against the connection's latest listing." — the fact about this call, and the rule a later one meets (the
1169
+ rule `reach` asks on every call), never its outcome and no advice. One mint (`listingRefusal`, `src/core/mcp-server-link.ts`) for
1170
+ both refusals; the codes, `details`, `delivered` and `isError` are unchanged.
1171
+ - **Behavior change (open to veto).** *What changed*: the closing sentence of the two refusals. *Who is hurt*: a host or a test that
1172
+ matches the old sentence by its text. *How to migrate*: read `details.code`, as the refusals have always carried it.
1173
+ - **The same shape elsewhere.** The four revoked-server refusals and the task-required refusal said the same thing and are closed
1174
+ by the next section (#2616); the closing sentence is now one writer, `refusalScope`, for all six.
1175
+ - **Consumers.** @server @cli: nothing to do; a surface that shows the tool result's text shows the new sentence. @test: the
1176
+ refusal text now holds neither "refused the same way", "don't retry" nor "until a listing" (pin A).
1177
+ - **Pins.** `test/backlog2589-2591-one-sentence-one-reader.test.ts` A1 (a re-listing that stopped before the tool: the fact, the rule,
1178
+ none of the foretelling words; after the host's refresh the same object runs), A2 (a changed declaration: the same sentence; the
1179
+ tool the refresh minted runs while the one minted before is still refused), A3 (the foretelling words are minted nowhere in `src`,
1180
+ the new sentence once).
1181
+
1182
+ ### A directory fault is worded once: the fault states what is wrong with the directory, and the reader of the fault states what it does about it (#2590; behavior change in the text the model reads, open to veto; hard-breaking: none; closed sets: none; exports: none — one internal function, `peerRowFault` in `peer-directory`, not index-exported) (@server @cli @test)
1183
+
1184
+ - **Before.** The row reader's fault ended "; no peer row is used until its host fixes it", and ListAgents added "— no peer row is
1185
+ listed until its host fixes it": one fact said twice in one line ("…(s-odd: name_invalid); no peer row is used until its host fixes
1186
+ it — no peer row is listed until its host fixes it)"). The reference registry's vetting fault ended "— nothing was listed" under the
1187
+ same suffix.
1188
+ - **Now.** A fault states the fault: "the directory listed a row this engine cannot read (<id>: <reason>)" (one function,
1189
+ `peerRowFault`, for a host's rows and for the reference registry's) and "the reference agent's session registry is not a
1190
+ directory this engine can trust (<code>: <path>)". Whoever reads the fault states its own consequence once: ListAgents "— no peer
1191
+ row is listed until its host fixes it", SendMessage its refusal frame "Message not sent:" (the arm no longer repeats it, #2617
1192
+ below). The consequence tail of the row reader's fault and of the registry's vetting fault is gone.
1193
+ - **Behavior change (open to veto).** *What changed*: the text of a directory fault, as thrown and as shown by ListAgents and SendMessage.
1194
+ *Who is hurt*: a host that matches the thrown message for "no peer row is used" or "nothing was listed" (`CcPeerDirectoryError.code`
1195
+ and `.path` are unchanged). *How to migrate*: match the code, or the first clause.
1196
+ - **Pins.** `test/backlog2589-2591-one-sentence-one-reader.test.ts` B1 (ListAgents over a host's directory: the one fault sentence and the
1197
+ tool's own line, "until its host fixes it" once), B2 (the same over the reference registry), B3 (a registry that fails its vetting: the
1198
+ thrown fault, read whole, carries no consequence, and ListAgents states it once), B4 (SendMessage: the whole text is the frame and the
1199
+ fault sentence, the fact that nothing was sent said once).
1200
+
1201
+ ### The reference registry's directory reads its rows through the engine's one row reader — a row the reader refuses faults the whole listing, as a host's rows do, instead of being repaired or skipped (#2591; behavior narrowing, open to veto; hard-breaking: none; closed sets: none; exports: none; members: none) (@server @cli @test)
1202
+
1203
+ - **Before.** One question, two answers. A host's directory hands its rows to the engine as written and the engine reads each through
1204
+ `readPeerSessionRecord`: a row it refuses stops the listing (ListAgents lists no peer row and names the refusal, SendMessage sends
1205
+ nothing). `createCcPeerDirectory` read the reference's registry file first and repaired what that reader judges: a name that reads as
1206
+ nothing (only invisible characters, only blanks) became "(unnamed session)", an `agent` label over 120 characters was dropped, and
1207
+ a row the reader still refused (a session id outside the address grammar) was skipped (`invalid_row`). The same row was listed by
1208
+ one seat and refused by the other.
1209
+ - **Now.** A string name and a string `agent` are carried as the row wrote them and the reader decides; a row it refuses throws the
1210
+ same fault a host's row does (`peerRowFault`) out of `listPeerSessions`, so no seat holds half a set (refs are minted and names
1211
+ resolved over the full set). `invalid_row` is left for a registry file that is not an object. What the registry directory keeps is the
1212
+ reference's own reading of its own file: a row that states no name, or a name that is not a string, is the reference's unnamed
1213
+ session, listed and addressed by "(unnamed session)" (the reference's rows for non-interactive sessions state none); a member of
1214
+ another type is not carried; capability bits are read by the reference's grammar (`/^[a-z0-9_]{1,32}$/`, at most 16). Liveness
1215
+ comes before validity: a row of a dead process, a file with no socket, a file that fails the ownership / mode / size check is skipped as before.
1216
+ - **Behavior narrowing (open to veto).** *What changed*: a registry row whose name is only invisible characters or blanks, whose `agent`
1217
+ label is over 120 characters, or whose session id is outside `^[A-Za-z0-9_-]{1,80}$` stops the listing of the whole registry
1218
+ (ListAgents names the row, SendMessage sends nothing) until that row is fixed or its file removed; it used to be listed repaired or
1219
+ skipped. *Who is hurt*: a deployment that reads the reference's registry beside a peer whose row states such a value (a row the
1220
+ reference writes states a name only from a flag or a derived name, so the name cases come from another writer; the label and
1221
+ session-id cases were not checked against the reference's writer) — for as long as the row stands, which the owning process's
1222
+ exit ends. *How to migrate*: nothing in the engine; fix the writer of the row.
1223
+ Divergence from the reference's tolerant reading of the same file (every ill-typed member read as nothing): candidate, registered under #2862 before the next release.
1224
+ - **Consumers.** @server @cli: a host that composes `createCcPeerDirectory` beside its own directory now sees a thrown fault for these
1225
+ rows where it saw a listing; its `onSkip` no longer reports `invalid_row` for a row the reader refuses. Which seat a given cli or
1226
+ server mount uses is not checked here (inferred: both seats exist).
1227
+ - **Pins.** `test/backlog2589-2591-one-sentence-one-reader.test.ts` C1 (five rows the reader refuses — a zero-width title, a
1228
+ direction-mark title, a blank title, a 121-character label, and a session id outside the grammar — read from a host's directory and
1229
+ from the reference registry: the same fault sentence, the whole listing, the directory itself throwing, ListAgents showing it and no
1230
+ good sibling), C2 (the two seats' fault texts equal across every fixture), C3 (what is not a row's fault stays a skip: a non-object
1231
+ file, a row with no socket, a dead process), C4 (what the reference reads is kept: no name / a non-string name is the unnamed session;
1232
+ a 120-character label is the cap and reads). `design385-s3-cc-uds-channel` D1 no longer carries a bad-id row (it faults now, C1).
1233
+
1234
+ ### A revoked server's four refusals and the task-required refusal end as the listing's do — what happened to this call and what judges the next one, never "every call to it is refused the same way, so don't retry it" (#2616; behavior change in the text the model reads, open to veto; hard-breaking: none; closed sets: none; exports: none — one internal function, `refusalScope` in `mcp-server-link`, not index-exported; members: none) (@server @cli @test)
1235
+
1236
+ - **Before.** The four outlets of a server the operator revoked (the tool call, the resource read, the resource directory listing, the
1237
+ aggregate resource listing) ended "This server stays on this run's tool roster — every call to it is refused the same way, so
1238
+ don't retry it", and a tool that requires task-based execution ended "Every call to this tool is refused the same way, so don't
1239
+ retry it." Both foretold the next call. The revocation ledger is asked on every call, and a server the operator un-revokes runs on
1240
+ the very next call (`test/mcp-revocation.test.ts`, the first case); a tool's task-required declaration is the listing's as
1241
+ mounted, and the tool a host's refresh mints from a listing that no longer requires tasks runs. The advice made an agent give up
1242
+ a server that worked again.
1243
+ - **Now.** Six call sites end with the one sentence `refusalScope` writes — the listing's refusal (#2589; its two codes), a revoked
1244
+ server's four outlets and the task-required refusal: "This refusal is for this call only: <each later call to this tool is
1245
+ checked again against the connection's latest listing | each later call to this server is checked again against the operator's
1246
+ current revocation list | each later call to this tool is judged by the declaration of the listing it was mounted from, until a
1247
+ refresh or reconnect of the server mints the tool anew>." A revoked server's sentence keeps the standing fact it carried, "This
1248
+ server stays on this run's tool roster." (the roster is materialized once per run; a refresh of a revoked server does not unmount
1249
+ its tools). The task-required rule is a re-read of a mount-time fact, not a re-evaluation, and the sentence says so: the same
1250
+ mounted object is refused until something mints the tool anew. No site words any of it: `mcp.ts` calls the writer five times,
1251
+ `listingRefusal` once; the codes, `details`, `delivered` and `isError` are unchanged.
1252
+ - **Not changed here — the same shape, one more text.** The refusal of a call to a server whose transport is gone
1253
+ (`throwDeadServer`, `src/core/mcp.ts`; the tool call, the resource read and the directory listing) and the HTTP "request was not
1254
+ delivered" refusal end "will keep failing until the server is available again — do not retry them". For the tool call, a host's
1255
+ re-dial replaces the live connection and the same tool object runs on it (`reach` reads the live connection on every call), so
1256
+ that sentence is overturned the same way (the resource outlets were not checked here); it is a thrown error with its own pins and
1257
+ is left for its own change.
1258
+ - **Behavior change (open to veto).** *What changed*: the closing sentence of five refusals. *Who is hurt*: a host or a test that
1259
+ matches the old sentence by its text. *How to migrate*: read `details.code` (`mcp.server_revoked`, `mcp.task_required`), as the
1260
+ refusals have always carried it.
1261
+ - **Consumers.** @server @cli: nothing to do; a surface that shows the tool result's text shows the new sentence. @test: the refusal
1262
+ text now holds none of "refused the same way", "don't retry" or "every call to it is" (pin A).
1263
+ - **Pins.** `test/backlog2616-2617-refusal-scope-one-writer.test.ts` A1 (each of the four revoked outlets: the sentence, the standing
1264
+ roster, no foretelling words, no request of any kind reached the server; after the operator's un-revoke the same object runs and
1265
+ reaches it), A2 (the task-required
1266
+ tool: the sentence, the call never reaches the server), A2b (a host's refresh that takes in a listing without the requirement mints a
1267
+ tool that runs), A3 (a run: revoked before the model's first call, un-revoked before its second — the first answer is the sentence,
1268
+ the second call runs), A4 (the scope sentence is minted in one file and the foretelling words nowhere in the engine);
1269
+ `test/up424-a-tier-ports.test.ts` A-5 now pins the roster clause in the writer, and the four outlets calling it.
1270
+
1271
+ ### A SendMessage refusal says once that the message was not sent — the frame says it, the arm states why (#2617; behavior change in the text the model reads, open to veto; hard-breaking: none; closed sets: none; exports: none; members: none) (@server @cli @test)
1272
+
1273
+ - **Before.** A refusal opens with the tool's frame "Message not sent:" and eleven arms closed with the same fact again: "Nothing was
1274
+ sent." (the rebound-name refusal, the three socket-reached peer arms, the four `notify_when_idle` and cross-session addressing arms),
1275
+ "Nothing was delivered." (a socket that refused the connection), "— nothing was sent." (the directory fault) and "— the message was
1276
+ not delivered." (an agent that is still starting up). One fact, said twice in one sentence.
1277
+ - **Now.** The frame says it and the arm does not: the arms state the reason alone. What stays is another fact — the state of the
1278
+ mailbox ("nothing was parked", "the message was NOT queued") and the warning on a send that may have arrived ("It may have arrived:
1279
+ do not resend the same content blindly") — because a refusal that leaves a message queued ("Message not sent now: … Your message
1280
+ is queued") is a different answer from one that leaves nothing behind. One exception stays as it was: a socket send of uncertain
1281
+ outcome (`peer_send.send_uncertain`) still opens with the frame "Message not sent:" and then says it may have arrived — an older
1282
+ contradiction this change does not touch.
1283
+ - **Behavior change (open to veto).** *What changed*: the text of those arms, and with it `details.verdict.message` of the peer arms
1284
+ (the same text as the arm, minus the frame). *Who is hurt*: a host or a test that matches "Nothing was sent." in a SendMessage
1285
+ result or in that field. *How to migrate*: read `details.code` / `details.error`, which the arms have always carried, or match the
1286
+ frame.
1287
+ - **Consumers.** @server @cli: nothing to do. @test: the reference's own rebound-name arm reads "Nothing was sent." (its text has no
1288
+ frame); here the frame carries it.
1289
+ - **Pins.** `test/backlog2616-2617-refusal-scope-one-writer.test.ts` B1 – B4 (a `notify_when_idle` with no directory mounted, a socket-
1290
+ reached peer with no channel wired, a `notify_when_idle` to a socket-reached peer, a socket that refuses: each text is the frame and the
1291
+ reason, the fact once), B5 (the mailbox fact stays, once, under the same frame), B6 (no arm of the tool says it again);
1292
+ `send-message-name-rebind-guard`, `backlog971-session-where` W6 and `backlog2589-2591-one-sentence-one-reader` B4 read the arms they
1293
+ covered without the tail.
1294
+
1295
+ ### The request a Runner hands a policy carries the call's grant-binding words — the members its cards carry, from the same factory, over the request's own args (#1955, design/1955, C-R896; no behaviour change; hard-breaking: none; closed sets: none; exports: none; members: `ToolCallRequest` +3 optional) (@server)
1296
+
1297
+ - **Before.** Every card of a call carried the engine's binding words — `boundGrantDigest` + `boundGrants` on an activation of the engine's
1298
+ `Skill` tool, `boundGrantAbsence: "not_bindable"` on every other call — but the `ToolCallRequest` a deployment policy layer is handed carried
1299
+ none, so a layer deciding before it asks whether an answer kept by tool name may answer the call had to infer the binding from the tool name.
1300
+ - **Now.** `ToolCallRequest.boundGrantDigest?` / `.boundGrants?` / `.boundGrantAbsence?` — the `AskRequest` members by name and type, exactly one
1301
+ arm. **Present** on every request a Runner hands a policy: the gate's adjudication (every gate request — the first look, the post-screen
1302
+ re-mint, the approved-edit re-check), the resume belt, the park's stored-shape re-check, the pre-claim edit re-check (minted with the skills
1303
+ that resume serves), the synthetic read / write probes (`not_bindable`), and every copy a composed policy, the inherited edit chain or the
1304
+ inherited classifier makes over moved args — the copy carries the answer for ITS args. **Absent** on a bare request (a host driving a policy
1305
+ outside a Runner). On a request `boundGrants` is a frozen copy. A copy whose digest the engine did not mint for that call (a clone, a
1306
+ host-built request) keeps the digest over the same bytes and reads `boundGrantAbsence: "unreadable"` over moved ones.
1307
+ - **Behaviour.** No change: no engine judgment reads the request's members (a parked approval is judged by its row's digest). The cards, the
1308
+ parked row and the inbox projection are unchanged; `tool_start` / `tool_end` carry no binding (no consumer reads one there).
1309
+ - **Consumers.** @server: the deployment-list gate and the workflow question gate can read the request through the card reader
1310
+ (`grantBindingOf`) and retire `callGrantBindingOf` and the carrier choice in `questionGrantBindingOf`; a request you build yourself outside a
1311
+ Runner carries no word — copy the row's members when you re-judge a parked call. The exemption list filter reads store rows that keep only a
1312
+ tool name: this change does not reach it. @cli @client-core: nothing to render.
1313
+ - **Pins.** `test/backlog1955-call-grant-binding.test.ts` (C1 card form, C2 no card, C3 park → edited resume, C4 a composed policy's rewrite,
1314
+ C5 the probes, C6 the mover, C7 the inherited classifier's copies).
1315
+
1316
+ ### A plan-review pause names the call that presented the plan — `requestedByCallId` on the parked row, the paused cause and the inbox summary (#2593; additive; no behavior change for a consumer that ignores the key; hard-breaking: none; closed sets: none; exports: none) (@client-core @server)
1317
+
1318
+ - **Before.** A plan-review pause gated no call and named none: the parked row was `{ kind: "plan_review" }`, and the paused cause carried the gate, the token and the checkpoint id. A card that had to know which transcript call carried the plan under review joined the pause to the `ExitPlanMode` call by adjacency or by tool name.
1319
+ - **Now.** The park mint writes `requestedByCallId` — the id of the call that asked for the review — onto the `plan_review` pending action, once. For the first-party plan tool that is the `ExitPlanMode` call's own id (the one its `tool_start` / `tool_end` frames carry); a caller tool's `ctx.requestReview()` is rebuilt per call, so the id is the calling call's and a tool cannot name another call's (a revive context's id is minted for the revival and names no transcript call, so it is not attributed). The paused cause (`terminal.requestedByCallId` on a `paused` cause) and the inbox summary (`CheckpointSummary.requestedByCallId`, also on `listByScope`) echo it through one reader (`planReviewRequesterOf`). First request in a batch wins, and so does its id; a plan re-presented after a reject parks a new row carrying the new call's id. The key is OMITTED, never `null`, on a row whose requester carried no call id (a row a deployment minted itself, a row from an earlier engine, a revive context) and on every other kind of pause.
1320
+ - **What it is not.** Provenance, not a binding: the review still gates no call, so `toolCallId` (on the pending action and on the summary) and `gatedCallId` stay absent on a plan review, and nothing on the resume path reads the id — it is classified SHOWN in the reader-requirements table, so a lost or changed id damages no row. The distinct name is deliberate: `toolCallId` means "the call this pause gates", and a reader that tests its presence reads "this is a tool approval".
1321
+ - **Limits (stated).** A store that whitelists the keys of a `plan_review` pending action drops the member on read-back (the in-memory and file stores keep it — pinned; the server's SQL store keeps the checkpoint row as one JSON column and reads `pendingAction` members out of it — read in that checkout, not exercised here); the live run's paused cause carries it regardless. The id names the FIRST requester of a batch only.
1322
+ - **Consumers.** @client-core: join the pause to the plan card by `terminal.requestedByCallId` (live frame) or `CheckpointSummary.requestedByCallId` (inbox row) against the card's own `toolCallId` — equality, no adjacency. @server: the paused cause rides your seal unchanged (the token is the only key stripped); no store change; your `parkToolCallId` still answers absence for a plan review, by design.
1323
+ - **Pins.** `test/backlog2593-plan-review-requester.test.ts` (① row / cause / summary / `listByScope` carry the `ExitPlanMode` call's id, the id on its own frames, `toolCallId` and `gatedCallId` absent; ①b the file store round-trips it; ② two gates, two ids; ③ a caller tool's id is the calling call's, a forged one is not honored; ④ first request wins; ⑤ the reader — absent / empty / non-string / other kinds; ⑥ an approval park carries none; ⑦ SHOWN: the row judges ok with the id lost or changed); `test/prepare-inherited-gate.test.ts` (the per-call ctx wrapper: own id bound, a forged option ignored, a revive context's and an empty id attribute nothing). Re-pinned: `prepare-boundary-parks-transcript` (the committed plan review's row now names its requester), `backlog1320-reader-requirements` (the SHOWN read-site registry gains the member).
1324
+
1325
+ ### The `leg_end` rows of the undrained-input notices name WHICH inputs they count — `detail.inputIds` (#2594; additive; no behavior change for a consumer that ignores the key; hard-breaking: none; closed sets: none; exports: none) (@client-core @server)
1326
+
1327
+ - **Before.** `task.user_steer_undrained` / `task.user_followup_undrained` said how many inputs were never consumed (`detail.steer` / `detail.followUp`) and not which. A client that tracks each input by the id its receipt carried could say only "something I sent was lost", and had to fold every unconfirmed input into "could not confirm".
1328
+ - **Now.** The engine's own (`cause: "leg_end"`) rows carry `detail.inputIds`: the id of EACH counted input, in the order the run would have served them (queue order: an `immediate`-tier input leads the ones accepted before it) — the same value the live `human_input` frame carries as `inputId` and a parked entry as `inputId`, taken from the same queue frames the count is. It follows the account's own arithmetic: the live queue's inputs less the ones a durable park carried onto the checkpoint, plus a resumed leg's continuation frames the reopen did not bring back and the park did not carry (those lead: they were accepted first). A park that carries one entry and is refused another names only the refused one. The list rides only where it names exactly the counted inputs; a lane with an input that carried no id keeps its count and loses the list, so a reader never mistakes a partial list for the whole. A child leg's copy on the parent's session carries the same list. An `approval_feedback` row (the words an approving person wrote) never carries one — it is no input; the `crashed` arm is the serving layer's and is not minted here.
1329
+ - **Behavior change.** None for a consumer that ignores the key; a notice's `detail` gains one array key on a real stranded input (an exact-`detail` comparison in a consumer's test would see it).
1330
+ - **Consumers.** @client-core: key a stranded input by `detail.inputIds` against the receipt's `messageId` (the `inputId` you sent, or the one the `human_input` frame returned); where the key is absent the count is the whole account, as before. @server: the `crashed` arm you mint is unchanged and carries none; you may spell the same key for it when you can name the inputs.
1331
+ - **Pins.** `test/backlog2594-undrained-input-ids.test.ts` (① real lane: two steers — one caller id, one minted — and a `later` steer cut by `interrupt()`, each family lists the ids the `human_input` frames carried; ①b the harness seam — the sink's park records are the same predicate's frames, person inputs only, an input with no record leaves its lane's list short, an `immediate`-tier input leads in service order; ② a park that carries one and is refused one names the refused one, a park that carries nothing names all, a full carry announces nothing; ③ the account arithmetic as ids — held less migrated, continuation first, a reopened row's frames dropped, a short list is the count alone, no ids recorded is the count alone, the approval row carries none; ④ the minter — complete lists only, follow-up lane keeps its own ids, the parent's copy carries the list and nothing aliases the caller's array or the other notice's; ⑤ every row with ids is `leg_end`).
1332
+
1333
+
1334
+ ### Corrections
1335
+ - The 7.58.0 notes carry the #2670 section (checkpoint-store contract kit judges the run origin) inside their `## 7.57.0` block; #2670 shipped in 7.58.0 (merged 10-25 21:15, after the v7.57.0 tag) — the released blocks are left byte-identical to their tags, so the misplacement stays where it is and is stated here instead.
1336
+
1337
+ - #902: the 7.58.0 note said the divergence was registered; it was registered 2026-10-26, after the release, as `fs-throat-dotdot-resolved-by-filesystem`.
1338
+
1339
+ no-notes: #2613(repo tool verify-debts.mjs closed-row sha existence; not in the package), #1068(a two-line src deletion only: the memory engine's two awaited rescue timers stay ref'd and are cleared when the race settles, the RB-466 form; its two functions are new in this window, so there is no behaviour relative to 7.58.0), #2352(repo tooling only: package.json test:affected + scripts/affected-tests.*; no package face)
1340
+
1
1341
  ## 7.58.0 — 2026-10-26
2
1342
 
3
1343
  ### The release re-window refuses the rows it cannot decide and keeps each row's first promised window — `scripts/release/rewindow.mjs` (#2835; release tooling only, `src/` untouched; no behaviour change; hard-breaking: none; closed sets: none; exports: none)