Distillation

Date: Sep 16, 2026

• 🔴 (15:32) User specified a session-runner requirement/TODO: “Replace local ownership with durable multi-node ownership when clustered,” superseding reliance on local ownership for clustered operation. Located in packages/core/src/session/runner/index.ts:55.
• 🔴 (15:32) User requires recovery redrive to “never wakes work settled between selection and redrive CAS.” The regression test is in packages/core/test/session-projector.test.ts:797; associated identifiers include session ses_ordinary_settled_at_cas, execution recovery-settled-at-cas, and message msg_ordinary_settled_at_cas.
• 🔴 (15:33) User requires the runner to “never marks a failed Step safe while a sibling tool finalizer is active.” This is covered by a live test in packages/core/test/session-runner.test.ts:3745.
• 🟡 (15:32) Recovery projector test coverage in packages/core/test/session-projector.test.ts includes quarantining a full poison batch before redriving valid work at line 495. It sets poisonCount = SessionRecovery.BATCH_SIZE + 1, creates poisonCount + 1 sessions named ses_recovery_poison_${index.toString().padStart(3, "0")}, uses execution IDs poison-${index.toString().padStart(3, "0")}, corrupts prompts via UPDATE session_input SET prompt = '{}' WHERE time_created < ${poisonCount}, expects only sessions[poisonCount] to wake, and expects poisonCount quarantined rows.
• 🟡 (15:33) Relevant durable runner tests found in packages/core/test/session-runner.test.ts, in order: line 688 clears durable execution intent when ordinary input is promoted; line 1085 reuses one durable baseline after the context producer changes; line 1119 includes the effective default agent system before durable context; line 2147 keeps provider failure unknown until unresolved tools are durable; line 2305 restores durable reasoning provider metadata in a second-turn request; line 2362 replays durable provider-executed tool results inline in a second-turn request; line 2695 preserves durable queued input for a later wake after interruption; line 2738 preserves durable steering input for a later resume after interruption; line 3745 never marks a failed Step safe while a sibling tool finalizer is active; line 4123 does not recover context overflow after durable assistant output.
• 🟡 (15:33) Tool-settlement tests found in packages/core/test/session-runner.test.ts include: line 2417 starts recorded local tools eagerly and awaits settlement before continuing; line 3941 durably fails blocked local tools when interrupted while awaiting settlement, using prompt text "Interrupt tool settlement"; line 4225 durably fails a hosted tool left unresolved at normal provider EOF, using prompt text "Fail hosted tool at EOF" and error { type: "unknown", message: "Provider Step ended without settlement" }.
• 🟡 (15:33) packages/core/test/session-runner-tool-events.test.ts defines a capture harness around createLLMEventPublisher; it records durable event names via EventV2.versionedType(definition.type, definition.durable.version). Publisher input uses session ses_tool_event_test, agent "build", model ID "model", provider ID "provider", and no-op startCommit/eventCommit.
• 🟡 (15:33) The local-tool media fixture in packages/core/test/session-runner-tool-events.test.ts uses call ID "call-image", tool name "read", input { path: "pixel.png" }, MIME "image/png", filename "pixel.png", text "Image read successfully", structured output { type: "media", mime: "image/png" }, and base64 value "iVBORw0KGgoAAAANSUhEUgAAAAEAAAAB".
• 🟡 (15:33) Test 1 in packages/core/test/session-runner-tool-events.test.ts:92, "local tool success serializes media base64 once and reconstructs from structured content", publishes the tool call and result, locates "session.next.tool.success.1", verifies the serialized event contains the base64 exactly once via serialized.split(base64).toHaveLength(2), verifies success data has no "result" property, and verifies reconstructed text/file content with the data URI and MIME.
• 🟡 (15:33) Test 2 in packages/core/test/session-runner-tool-events.test.ts:115, "provider-executed success retains its compatibility result", publishes the call and result with providerExecuted: true and verifies "session.next.tool.success.1" data retains the "result" property.
• 🟡 (15:33) Test 3 in packages/core/test/session-runner-tool-events.test.ts:123, "binary failure emits no success event", publishes a tool result { type: "error", value: "Cannot read binary file" }; it verifies no "session.next.tool.success.1" event is emitted and a "session.next.tool.failed.1" event is emitted.
• 🟡 (15:33) Test 4 in packages/core/test/session-runner-tool-events.test.ts:139, "old success event data containing result still decodes", decodes legacy SessionEvent.Tool.Success.data containing structured, content, compatibility result, and provider: { executed: false }; it verifies decoded.result remains { type: "content" }, preserving backward compatibility.
• 🟡 (15:33) Test 5 in packages/core/test/session-runner-tool-events.test.ts:168, "step finish records settlement without publishing step ended", publishes LLMEvent.stepStart({ index: 0 }) followed by LLMEvent.stepFinish({ index: 0, reason: "stop" }); it verifies no "session.next.step.ended.2" event is published and publisher.stepSettlement() matches { finish: "stop" }.