Distillation

Date: Sep 16, 2026

• 🟡 (15:21) SHA-256 tool output returned 4 hashes: 4b247d3850983b60a249cd0be3a6c005509d5fb26b41da259c8ed1c6e9ede154, 39593a5f72dec344a670f89125245f3cc4ad75337a1b9a3611aa1a99256bea34, 504b5448846f43333bccba80f9a262171a3b6e0fde6146252ca6334297d3ecaa, and a83e976242db98545e26f5c53618f8b5713ab6b5d60cba449ac9ae2996d31333.
• 🟡 (15:21) packages/gateway/src/stream/openai-responses.ts defines ResponsesAccState at line 48 and finalizeResponsesAcc() at line 815.
• 🟡 (15:21) ResponsesAccState retains response identity/model/stop reason/usage, optional terminal event and response snapshot, optional codexRateLimits, original output in rawItems, strict identity maps (itemIndexById, callIndexById, effectiveToolIndexById), active/unbound/done item sets, and normalized text/tool-use items indexed by output_index.
• 🟡 (15:22) finalizeResponsesAcc() in packages/gateway/src/stream/openai-responses.ts:815-901 sorts state.items by output index; emits output_text parts as gateway text blocks; preserves non-text parts as opaque Responses items; converts refusals into opaque completed message items; parses tool arguments as JSON with raw-string fallback; changes end_turn to tool_use when tool output exists; and returns sorted rawOutputItems excluding item_reference, plus optional codexRateLimits.
• 🟡 (15:22) validatePublicResponsesEvent() begins at packages/gateway/src/stream/openai-responses.ts:908; an output_index is valid only when it is a safe integer, nonnegative, and less than maxSparseIndex.
• 🟡 (15:23) buildTerminal() in packages/gateway/src/pipeline.ts:9943-9984 reconstructs the terminal Responses event using buildOutputItems(), maps status from the stop reason, computes inclusive input tokens as normal input + cache-read + cache-creation tokens, emits input_tokens_details when cache usage exists, preserves the terminal snapshot, and defaults created_at to Math.floor(Date.now() / 1000).
• 🟡 (15:23) buildOutputItems() in packages/gateway/src/pipeline.ts:9986-10048 merges sorted indices from state.rawItems and state.items, skips hiddenIndices and item_reference, preserves raw-only items, reconstructs message text/refusal content, and reconstructs function calls with id, call_id, name, arguments, and a default "completed" status.
• 🟡 (15:23) finalizedMessageContent() in packages/gateway/src/pipeline.ts:7971-7984 creates a sparse array keyed by contentIndex, including only parts where valueDone or partDone is true and materializing either { type: "refusal", refusal } or { type: "output_text", text }.
• 🟡 (15:23) materializeCompletedCodexMessages() in packages/gateway/src/pipeline.ts:7985-7997 runs only when opts.validation === "codex"; for unfinished message output lifecycles with finalized content, it updates both acc.rawItems and the normalized text item’s content.
• 🟡 (15:23) Transaction rollback in packages/gateway/src/pipeline.ts:8071-8098 restores id, model, stopReason, terminalEvent, terminalResponse, cloned usage, items, and raw items from transactionBaseline, then executes registered rollback callbacks in reverse order and logs callback failures.
• 🟡 (15:23) Recall-stream defaults in packages/gateway/src/pipeline.ts:8101-8121 are: maxDeferredBytes = 1024 * 1024, maxHiddenRecallBytes = maxDeferredBytes, maxRetainedStateBytes = 16 * 1024 * 1024, maxTransactionalBytes = maxRetainedStateBytes, maxStreamBytes = 64 * 1024 * 1024, maxSparseIndex = Math.min(maxSSEFrames, DEFAULT_MAX_SSE_FRAMES), maxPrincipalTransportRetries = 1, and maxRecallContinuationTransportRetries = 1.
• 🟡 (15:24) When recall is detected in packages/gateway/src/pipeline.ts:10694-10733, more than 1 pending recall throws RecallContinuationFailure("parallel_recall"); the transaction baseline shallow-copies state and clones usage/items/rawItems; principal usage is recorded once; transactional chunks count against maxTransactionalBytes and overflow throws RecallContinuationFailure("resource_limit"); failure to locate a recall tool block throws RecallContinuationFailure("missing_recall_block").
• 🔴 (15:24) User stated the distilled prefix always occupies transformed-array positions [0,1], durable knowledge deltas remain near the tail, and orphan removal never touches the prefix.
• 🔴 (15:24) User stated resolveTrustedRemote must ALWAYS use the on-disk git remote when one exists; a client-provided x-lore-git-remote must never override on-disk truth. In local mode with no disk remote, the client remote must be dropped entirely to prevent stale-header remote selection.
• 🔴 (15:24) User stated shared memoized arrays such as the syncedColumns PRAGMA cache must always be frozen; mutation should throw TypeError in strict mode rather than silently corrupt later results.
• 🔴 [enforced-workflow] (15:24) User requires ALWAYS running git add . before committing because partial staging can omit related tests, fixtures, or configuration and produce broken CI or non-reproducible commits.
• 🔴 [enforced-workflow] (15:24) User’s pre-commit workflow is: format → lint → typecheck → tests on the current base; also check for a stray root-level opencode.json, because BYK/loreai has none and such a file contaminates cache/auth tests.
• 🔴 (15:24) User stated .lore.md is auto-managed by the background lore process; stash-pop conflicts must be resolved to “ours” (origin/main) rather than hand-edited, and regenerated changes are committed with git add ..