Distillation

Date: Sep 15, 2026

• 🔴 (16:35) User directed that local execution ownership be replaced with durable multi-node ownership when clustered.
• 🔴 (16:35) User’s runner checklist requires honoring interruption and rejecting stale work after runtime attachment replacement.
• 🟡 (16:35) Runner checklist marks incremental persistence of assistant text and usage events as completed.
• 🟡 (16:35) Runner checklist marks continuation for durable user steering accepted during an active provider turn as completed.
• 🟡 (16:35) The runner loads V2 history, translates it, resolves a model through a core service, and persists one provider turn.
• 🟡 (16:35) Runner implementation imports SessionExecutionClaim from "../execution/claim", SystemContext from "../../system-context/index", and llmClient from "../../effect/app-node-platform".
• 🟡 (16:35) Runner execution tracks ownerID: string, recoveryInputID?: SessionMessage.ID, wake_pending, execution phase "safe", assistant_message_id, and promotion mode "steer".
• 🟡 (16:35) Runner context maps tool state with executed: tool.provider?.executed === true, manages tool fibers through FiberSet.make<void, ToolOutputStore.Error>(), and waits using Effect.raceFirst(FiberSet.join(fibers), FiberSet.awaitEmpty(fibers)).
• 🟡 (16:35) Runner continuation includes { readonly _tag: "ContinueAfterCompaction"; readonly step: number } and continueAfterOverflowCompaction(step); context-overflow handling checks isContextOverflowFailure(overflowFailure ?? failure).
• 🟡 (16:35) When steering is promoted, if (promoted > 0) currentStep = 1; recovery suppresses steering promotion via promotion = recovering ? undefined : "steer".
• 🟡 (16:35) Interrupted tool execution is persisted through publisher.failUnsettledTools("Tool execution interrupted").
• 🟡 (16:35) The provider-event publisher is documented to “Persist one provider turn without executing tools or starting a continuation turn.”
• 🟡 (16:35) Provider-event publishing maintains fragmented text in Map<string, string[]>, tracks tools by call ID, lazily establishes assistantMessageID, flushes fragments with the current assistant message ID, and records provider-created tool state with providerExecuted: false.
• 🟡 (16:35) Fragment handling treats a delta before its corresponding start as fatal with the exact pattern Effect.die(`${name} delta before start: ${id}`).
• 🟡 (16:35) SessionRunner.publishLLMEvent handles event cases including "reasoning-delta", "reasoning-end", "tool-result", "tool-error", and "step-finish"; reasoning starts through reasoning.start(event.id).
• 🟡 (16:35) Provider tool-event validation fatally rejects a renamed tool input using Effect.die(`Tool input name changed for ${event.id}: ${tool.name} -> ${event.name}`) and a renamed tool call using Effect.die(`Tool call name changed for ${event.id}: ${tool.name} -> ${event.name}`).
• 🟡 (16:35) Tool results are normalized through settledOutput(event.output, event.result); published tool records preserve callID: event.id and the tool’s assistantMessageID.
• 🟡 (16:35) The provider-event publisher exposes failUnsettledTools; assistant lifecycle handling sets assistantActive = false when the assistant turn ends or fails and calls failAssistant(event.message) on provider failure.