Distillation

Date: Sep 16, 2026

• 🔴 (14:39) User directive: embedded plugins must never request lifecycle control; only CLI startup owns StartOptions.processBoundary. Quiet in-process callers use { quiet:true, local:true }, route notices through file logging, and must not write lifecycle text into full-screen TUIs.
• 🔴 (14:39) User directive: OpenCode v2 replay must never resume an older assistant projection. Exactly one durable settlement event should determine replayable state; a newer turn supersedes stale incomplete rows, while reasoning and compaction deltas remain live-only progress.
• 🔴 (14:39) User directive for parseOpenAIResponsesRequestChunks: use a hybrid bounded parser rather than always buffering or always using a JS parser. Below STREAMING_PARSE_SPOOL_BYTES (256 KiB), use JSON.parse(Buffer.concat(...).toString("utf8")); after crossing the limit, replay the bounded prefix once and stream $.input.* using Tokenizer/TokenParser with keepStack:false. Share createInputItemsBuilder() across both paths; do not retain a decoded-string accumulator or full raw input tree.
• 🔴 (14:39) User directive: recall continuation regressions must exercise retries with indexed hidden/sparse events and resolve an empty-name Codex candidate into a visible tool; a terminal-only sparse fixture never reaches promotion and is therefore insufficient.
• 🔴 (14:39) User directive: streaming request failures must never reflect codec or parser diagnostics. Decode, decompression, unsupported-encoding, and parser failures map to the fixed HTTP 400 invalid_request_error envelope.
• 🔴 (14:39) User directive: streaming JSON decoding must require exactly one root value; successful tokenizer/parser end calls do not make whitespace-only input valid. Gateway-owned Content-Encoding must be normalized, and a client encoding must never be forwarded when it does not match transmitted bytes.
• 🟡 (14:39) Documented recall continuation resource policy: maxDeferredBytes is the 1 MiB cap for untrusted unresolved tool input, while validated chunks reserve against maxTransactionalBytes. Sparse non-recall resolution transfers deferred bytes to transactional bytes exactly once before flush; retries restore transactional events/bytes, retained and hidden bytes, and output/reference identity sets to baseline before reopening.
• 🟡 (14:42) Jujutsu working copy lrqlyykx / commit d890b184d786a4348e30bcdac7412baa7ae01ae9 was empty and based on tklkqxyt / ce88c614c0ae55d229cb605aa7c7369d918d3650; bookmark fix/responses-private-references had conflicts. The parent diff involved packages/gateway/src/pipeline.ts, packages/gateway/src/stream/openai-responses.ts, and packages/gateway/test/openai-responses-recall-aware-stream.test.ts.
• 🟡 (14:44) Inspected boundary tests in packages/gateway/test/openai-responses-recall-aware-stream.test.ts: sparse continuation coordinates use first invalid values 99_998 for output and 100_000 for content/summary; accepted boundary coordinates are one lower. Accepted oversized state should fail as resource_limit, rejected coordinates as follow_up_protocol, and private sparse values must not leak.
• 🟡 (14:44) Inspected test retains a sparse continuation Codex function call resolved by the limit-crossing done event, which starts a function call with empty call_id, name, and arguments, streams private arguments, then resolves it via response.output_item.done to call_id: "call_sparse_continuation_limit", name: "read", and status: "completed".
• 🟡 (14:47) Inspected Codex terminal compatibility tests in packages/gateway/test/openai-responses-recall-aware-stream.test.ts: an empty terminal output: [] and a partial terminal output are accepted after completed streamed items; principal response.done statuses failed and cancelled suppress recall and emit exactly one response.failed.
• 🟡 (14:52) packages/gateway/src/recall-budget.ts defines recall limits: MAX_RECALL_EXECUTIONS = 24, MAX_RECALL_CHAIN_TOKENS = 128_000, MAX_RECALL_CHAIN_RESULT_BYTES = 512 * 1024, MAX_RECALL_CHAIN_ITEMS = 64, MAX_RECALL_SEARCH_ITEMS = 30, RECALL_FINALIZATION_RESERVE_MS = 20_000, RECALL_FINALIZATION_RESERVE_TOKENS = 4_096, and MAX_CONSECUTIVE_RECALL_NO_PROGRESS = 2.
• 🟡 (14:52) RecallChainBudget stop reasons are "time", "tokens", "result_bytes", "items", "stalled", and "execution". Its default foreground deadline is startedAt + 300_000 - RECALL_FINALIZATION_RESERVE_MS; admit() reserves item capacity before execution, recordUsage() counts input/output/cache tokens, and record() replaces reservations with distinct delivered coverage and stops after consecutive no-progress exceeds the configured maximum.
• 🟡 (14:59) An ad hoc tsx reproduction failed before execution because top-level await is unsupported with the "cjs" output format.
• 🟡 (15:00) Wrapping the ad hoc reproduction to avoid top-level await reached module loading but failed with MODULE_NOT_FOUND for /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/node_modules/@loreai/core/dist/node/index.js under Node.js v24.16.0.
• 🟡 (15:01) Attempting pnpm execution with vite-node failed with ERR_PNPM_RECURSIVE_EXEC_FIRST_FAIL because command vite-node was not found; pnpm suggested pnpm exec vitest.