Distillation

Date: Sep 16, 2026

• 🟡 (13:32) The supplied test code parameterizes no-store behavior with test.each([true, false])("no-store=%s", ...); after a provider response, it asserts stateFor(alias).recallStore.size is 0 when noStore is true and 1 otherwise.
• 🟡 (13:32) The supplied test code includes protocol/provider routing through x-lore-provider, explicitly setting req.rawHeaders["x-lore-provider"] = "anthropic" in multiple cases and including the tuple ["openai-responses", true, "openai-responses"] in a test matrix.
• 🟡 (13:32) Provider-response fixtures emit OpenAI Responses lifecycle events including response.output_item.added and response.output_item.done, and construct completed responses with IDs in the form resp_${round} and status: "completed".
• 🟡 (13:32) The fixtures also generate Anthropic-style message_delta events and tool definitions with inputSchema: { type: "object", properties: {} }, exercising more than one upstream streaming protocol.
• 🟡 (13:32) Completion fixtures vary whether answer text is present: message.content becomes "Completed answer" when withText is true and null otherwise; OpenAI-style output text uses { type: "output_text", text: "Completed answer" }.
• 🟡 (13:32) The supplied tests exercise stream-consumption modes including "answer", "exhausted", "cancel", and "mixed"; cancellation calls await reader.cancel(), while answer-mode expectations distinguish "Completed answer" from recall output beginning with "lore-recall:".
• 🟡 (13:32) Abort and storage behavior are exercised with new AbortController(), a mocked temporal.store, and session tracking read through loadSessionTracking(state.sessionID)?.recallStore ?? null.
• 🟡 (13:32) Recall behavior is mocked through vi.spyOn(core, "runRecallWithMetadata").mockImplementationOnce(...); visible assertions also include an expected zero count, toEqual({ count: 0 }), and a provider fixture call providerResponse("anthropic", 1, "answer").