Distillation

Date: Sep 15, 2026

• 🔴 (16:39) User specified the ownership-fencing invariant: “never dispatches a provider after ownership changes.”
• 🔴 (16:39) User specified the ownership-fencing invariant: “never publishes streamed output after ownership changes.”
• 🔴 (16:39) User specified the ownership-fencing invariant: “never starts a local tool after ownership changes.”
• 🔴 (16:39) User specified the ownership-fencing invariant: “never fails interrupted tools after ownership changes.”
• 🟡 (16:39) packages/core/test/session-runner.test.ts configures OpenAIChat.route with limits { context: 20_000, output: 1_000 }, tracks activeToolExecutions, and includes application tools echo and a tool described as "Fail unexpectedly".
• 🟡 (16:39) packages/core/test/session-runner.test.ts:838 changes execution ownership with .set({ id: "interrupted-tool-successor", owner_id: "successor" }) in the interrupted-tool fencing test.
• 🟡 (16:40) Compaction-related tests located in packages/core/test/session-runner.test.ts, in source order: 1. line 1336, "rebuilds the baseline directly after completed compaction"; 2. line 1375, "automatically compacts into a completed summary and retained recent turn"; 3. line 1448, "keeps automatic compaction unknown until Compaction.Ended commits"; 4. line 1495, "retains only complete serialized messages during compaction"; 5. line 1556, "forces one compaction and retries after provider context overflow"; 6. line 1585, "persists a second context overflow after one recovery"; 7. line 1608, "recovers once from a raw context overflow failure"; 8. line 1636, "publishes the original overflow when recovery summarization fails"; 9. line 1656, "interrupts overflow recovery while the summary provider is running"; 10. line 1681, "preserves effective System updates while compaction rebaseline is blocked"; 11. line 3557, "does not recover context overflow after durable assistant output".
• 🟡 (16:40) Automatic compaction tests use model compact; successful overflow recovery embeds "<summary>\n## Objective\n- Recover overflow\n</summary>" in the retried request and projects { type: "compaction", summary: "## Objective\n- Recover overflow" }.
• 🟡 (16:40) The raw-overflow recovery fixture uses LLMEvent.providerError({ message: "prompt too long", classification: "context-overflow" }) and summary text "## Objective\n- Recover raw overflow".
• 🟡 (16:40) The interrupted automatic-compaction fixture uses text fragment ID "text-uncommitted-summary" with summary "## Objective\n- Interrupted"; no compaction message is projected, and the successor execution remains { id: "compaction-successor", phase: "unknown" }.
• 🟡 (16:40) packages/core/src/session/recovery-id.ts defines message(executionID, assistantMessageID?) as assistantMessageID ?? SessionMessage.ID.make(\msg_recovery_${executionID}`)andcontinuation(executionID)asSessionMessage.ID.make(`msg_recovery_continue_${executionID}`)`.
• 🟡 (16:40) packages/schema/src/session-event.ts:207-217 defines Step.Recovered with event type "session.next.step.recovered"; it is included in the schema unions at lines 485 and 518.