Distillation

Date: Sep 14, 2026

• 🟡 (23:24) packages/core/test/session-runner.test.ts was updated to import and exercise EventV2, ContextSnapshotDecodeError, SessionRunnerLLM, and SkillGuidance; its runner layer is built with AppNodeBuilder.build(SessionRunnerLLM.node, [...]) and includes SessionExecution.Service.
• 🟡 (23:24) The SessionExecution test wiring in packages/core/test/session-runner.test.ts removed the direct wake: coordinator.wake assignment, reflecting the updated execution/coordinator integration.
• 🟡 (23:24) packages/core/test/session-runner.test.ts tracks tool concurrency with maxActiveToolExecutions, collects authorization contexts in authorizations: Tool.Context[], and includes synchronization via Deferred.succeed(toolExecutionsStarted, undefined).
• 🟡 (23:24) Added/retained the SessionRunnerLLM test "advertises and executes a globally attached application tool": registers application_context through ApplicationTools.Service, with input { query: Schema.String }, output { answer: Schema.String }, and execution that uppercases the query while capturing Tool.Context; the provider emits tool call ID call-application for query "hello".
• 🟡 (23:24) packages/core/test/session-runner.test.ts tests the built-in question tool using tool call ID call-question, name "question", and empty input {}.
• 🟡 (23:24) Added/retained the test "interrupts a blocked provider turn without local tool execution" using prompt text "Interrupt provider" and Deferred gates streamGate and streamStarted to hold and detect the provider stream.
• 🟡 (23:24) Maximum-step handling is tested with exactly 2 provider requests: the first has no toolChoice; the second has toolChoice: { type: "none" }, tools: [], and a final assistant message containing "MAXIMUM STEPS REACHED". Expected executions are ["done"], with context containing user text "Finish at the limit", completed tool call-terminal, and errored tool call-forbidden.
• 🟡 (23:24) packages/core/test/session-runner.test.ts includes interruption coverage titled "interrupts overflow recovery while the summary provider is running" at line 1307.
• 🟡 (23:24) packages/core/test/session-runner.test.ts includes interrupted-tool fixtures using local call ID call-interrupted and hosted/provider-executed call ID call-hosted-interrupted; the hosted fixture carries provider: { executed: true, metadata: { openai: { itemId: "call-hosted-interrupted" } } }.
• 🟡 (23:24) Unexpected local tool defects are expected to become the structured error { type: "unknown", message: "Tool execution failed: unexpected tool defect" }; the test tool implementation defects with Effect.die("unexpected tool defect").
• 🟡 (23:24) packages/core/test/session-runner.test.ts includes the test "durably fails blocked local tools when interrupted while awaiting settlement" at line 3024, exercising durable failure behavior during interruption.