Distillation

Date: Sep 16, 2026

• 🔴 (11:19) User supplied packages/core/src/session/recovery-id.ts: message(executionID, assistantMessageID?) returns the supplied assistantMessageID or deterministic SessionMessage.ID.make(\msg_recovery_${executionID}`); continuation(executionID)returnsSessionMessage.ID.make(`msg_recovery_continue_${executionID}`)`.
• 🔴 (11:19) User specified that recovery must never recover an execution superseded by a newer assistant.
• 🔴 (11:19) User specified that recovery must never promote a recovery superseded before scanner validation.
• 🔴 (11:19) User specified that a quarantined recovery continuation must never be promoted as an ordinary steer.
• 🔴 (11:19) User specified that an exact retry of an already promoted prompt must never wake execution.
• 🔴 (11:19) User specified that a conflicting execution retry must never arm durable wake.
• 🔴 (11:19) User supplied recovery tests in packages/core/test/session-projector.test.ts covering: atomic reconciliation of one abandoned execution into one continuation; recording a recovery continuation without forging an assistant identity; quarantining malformed and cross-session recovery continuations with recovery_error: "Invalid recovery continuation input"; quarantining a full poison batch before redriving valid work; stale recovery-wake handling; supersession before scanner validation, including assistant-less recovery; and preventing quarantined continuations from appearing as pending ordinary steer work.
• 🔴 (11:19) User supplied poison-batch test details in packages/core/test/session-projector.test.ts: poisonCount = SessionRecovery.BATCH_SIZE + 1; the malformed rows are changed with UPDATE session_input SET prompt = '{}'; all poisonCount rows must be quarantined with "Invalid recovery continuation input" before the valid session at sessions[poisonCount] is woken.
• 🔴 (11:19) User supplied prompt retry/admission tests in packages/core/test/session-prompt.test.ts: delegate execution continuation through SessionExecution; wake execution when an exact prompt retry recovers a committed message; replay historical prompt admission without execution intent as admit-only; turn an admit-only exact retry into durable execution intent; promote steers only through the captured inbox cutoff; and return an exact retry of a legacy projected prompt.
• 🔴 (11:19) User supplied an 8,215-byte patch at /tmp/opencode/session-recovery-final-v5.patch.
• 🔴 (11:19) User supplied packages/core/src/session/runner/index.ts, where SessionRunner.Interface.run(...) accepts { sessionID, force, executionID, ownerID, recoveryInputID? }; recoveryInputID is an optional SessionMessage.ID.
• 🔴 (11:19) User supplied SessionRunner.RunError as the union of LLMError, SessionRunnerModel.Error, MessageDecodeError, ContextSnapshotDecodeError, SystemContext.InitializationBlocked, and ToolOutputStore.Error.
• 🔴 (11:19) User supplied the runner contract that one local continuation runs from already-recorded Session history; durable eligible work is drained, while an explicit run performs one provider attempt even if no work is eligible.
• 🔴 (11:20) User supplied packages/opencode/src/server/routes/instance/httpapi/server.ts wiring where createRoutes(corsOptions?) provides AppNodeBuilderV1.build(LayerNode.group([SessionV2.node, SessionRecovery.node]), [[LocationServiceMap.node, locationServiceMapV2], [SessionExecution.node, SessionExecutionLocal.node]]).
• 🔴 (11:20) User supplied server-layer ordering constraints in packages/opencode/src/server/routes/instance/httpapi/server.ts: Observability.layer must remain last because later-provided layers build beneath earlier ones; otherwise eagerly forked fibers such as the ModelsDev background refresh capture Effect’s default stdout logger and corrupt the TUI (#34730).