Distillation

Date: Sep 16, 2026

• 🔴 (15:59) User directed that local execution ownership be replaced with durable multi-node ownership when clustered.
• 🟡 (15:59) Session runner implementation documentation designates specs/v2/session.md as the canonical location for tracking V1 runtime-context parity.
• 🟡 (15:59) Session runner checklist marks authorization and execution of recorded local calls through a core-owned registry hook as completed.
• 🟡 (15:59) The session runner uses SessionCompaction.make(...), obtains session context through store.context(sessionID), and defines internal turn transitions including { readonly _tag: "ContinueAfterCompaction"; readonly step: number }, constructed by continueAfterCompaction(step).
• 🟡 (15:59) Session runner requests set OpenAI provider options to { openai: { promptCacheKey } }.
• 🟡 (15:59) When maximum agent steps disable tools, the runner calls publisher.failUnsettledTools("Tools are disabled after the maximum agent steps").
• 🟡 (15:59) After interruption, unresolved tool execution is failed with exact message Tool execution outcome unknown after interruption.
• 🟡 (15:59) Runner execution inputs include sessionID, executionID, and ownerID; runTurn(input.sessionID, input.executionID, promotion, step, input.ownerID) performs a turn after failInterruptedTools(input.sessionID, claim).
• 🟡 (15:59) The runner layer includes ToolRegistry.node, supporting the core-owned registry used for recorded local calls.
• 🟡 (16:00) Session runner publisher utility defines safe(value: number | undefined) as Math.max(0, Number.isFinite(value) ? (value ?? 0) : 0), clamping undefined, non-finite, and negative values to 0.
• 🟡 (16:00) The publisher’s value serialization uses JSON.stringify(value) ?? String(value).
• 🟡 (16:00) The publisher is documented to persist one provider turn without executing tools or starting a continuation turn.
• 🟡 (16:00) Publisher state tracks tools in a Map, lazily creates and retains an assistantMessageID, and exposes currentAssistantMessageID() for associating emitted events with the active assistant message.
• 🟡 (16:00) Streaming text/reasoning chunks are tracked by ID with start(id) and end(id) behavior; cleanup iterates chunks.keys() and ends every remaining chunk.
• 🟡 (16:00) SessionRunner.flush flushes accumulated publication state and marks assistantActive = false.
• 🟡 (16:00) SessionRunner.publishLLMEvent handles provider stream event types including reasoning-delta, reasoning-end, tool-result, tool-error, and step-finish.
• 🟡 (16:00) Tool-input consistency is enforced by terminating with exact error Tool input name changed for ${event.id}: ${tool.name} -> ${event.name} when an input event changes a tool’s recorded name.
• 🟡 (16:00) Tool-call consistency is enforced by terminating with exact error Tool call name changed for ${event.id}: ${tool.name} -> ${event.name} when a call event changes a tool’s recorded name.
• 🟡 (16:00) Tool results and errors are published with the recorded callID, tool-associated assistantMessageID, and provider metadata; provider error handling sets providerFailed = true.