Distillation

Date: September 16, 2026

• 🔴 (23:31) User requires transformed full-history replay with store:false; never rely on client state or previous_response_id.
• 🔴 (23:31) User requires recall-aware OpenAI Responses handling to validate protocol identities, roles, usage, terminal reasons, sequencing, limits, and OpenAI recovery envelopes before accepting output.
• 🔴 (23:31) User requires parseSSE before forwarding recall follow-ups and requires asserting that the follow-up is an actual SSE response.
• 🔴 (23:31) User stated recall-aware streaming must never leak recall data, ciphertext, diagnostics, references, synthetic lifecycle events, private provider state, or hidden recall/continuation chunks into client-visible output or GatewayResponse.
• 🔴 (23:31) User stated internal metadata needed for recall, source coverage, calibration, diagnostics, persistence, and post-response processing must remain private and must never be emitted to the client.
• 🔴 (23:31) User requires only validated, sanitized Responses events to be forwarded, with stable public error messages; the terminal response must match the identities, indices, content, and lifecycle events actually exposed to the client.
• 🔴 (23:31) User prefers true streaming over buffering while retaining an accurate internal accumulated response; finalized streamed content is authoritative over intermediate accumulated text.
• 🔴 (23:31) User requires comprehensive regression tests for recall-aware streaming changes across principal and continuation paths, including invalid private item_reference lifecycles, identity/index collisions, type changes, terminal and nested boundaries, suppression, retries, execution limits, retained-state/resource accounting, sparse or shifted output, terminal reconstruction, usage accounting, reasoning lifecycle cases, and private/public state separation.
• 🔴 [requested-tests] (23:31) User requires regression tests to assert exact lifecycle behavior, failure category, call counts, transaction outcomes, public error classification (PUBLIC_GATEWAY_ERROR versus PUBLIC_RECALL_ERROR), and absence of private provider state from client-visible output and GatewayResponse.
• 🔴 (23:32) User stated the real server path always supplies a fully formed GatewayRequest.
• 🔴 (23:32) User stated the malformed-invocation guard if (!req?.rawHeaders) exists for fuzzers and direct module calls and must return errorResponse(400, "Malformed request: missing headers") instead of allowing a deeper TypeError on req.rawHeaders.
• 🔴 (23:32) User stated removeOrphanedToolResults() should never fire during normal operation because resolveToolResults removes redundant tool_result parts and loreMessagesToGateway reconstructs them from the assistant’s completed tool parts; it is a belt-and-suspenders safeguard that removes orphaned references before they reach the API.
• 🟡 (23:31) Project recall used native sqlite-vec v0.1.9 from /tmp/lore-native/pid-739550/vec0.so; it found 90 results, showed the top 15, and reported a score range of 0.064–0.049.
• 🟡 (23:32) Project recall used native sqlite-vec v0.1.9 from /tmp/lore-native/pid-742172/vec0.so; it found 90 results, showed the top 15, and reported a score range of 0.064–0.052.
• 🟡 (23:33) Project recall used native sqlite-vec v0.1.9 from /tmp/lore-native/pid-743552/vec0.so; it found 90 results, showed the top 15, and reported a score range of 0.066–0.047.