Distillation

Date: Sep 16, 2026

• 🟡 (14:13) In /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/src/pipeline.ts, successful streaming post-response persistence runs postResponseForTenant() inside withTenant() and savepoint "responses_recall_post_response"; it requires a truthy persistence result before recallPersistenceTransaction?.commit(), otherwise rolls back through rollbackRecallPersistence().
• 🟡 (14:13) In /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/src/pipeline.ts, unsuccessful streaming finalization waits for downstreamSettled and one setImmediate, rolls back recall persistence, drops stale/aborted finalizers, and calls accountUnsuccessfulResponse() only when the request generation remains current and sessionSignal is not aborted.
• 🟡 (14:13) captureUnsuccessfulResponses() catches only ResponsesTerminalError: it schedules unsuccessful finalization, returns { response, successful: false } for status "incomplete", returns undefined for other terminal statuses, and rethrows unrelated errors.
• 🟡 (14:13) The OpenAI Responses streaming branch in /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/src/pipeline.ts uses streamResponsesRecallAware() when the client protocol is "openai-responses", no warning must be injected, and the modified request contains the recall tool. Its options include Codex/public validation, completion handling, transaction replacement with prior rollback, session ID, configured chain execution limit, noStore, foreground abort signal/deadline, principal retry, and an onRecall callback using executeRecall().
• 🟡 (14:13) The streaming onRecall path constructs IDs as recall_stream_${query}_${scope ?? ""}_${id ?? ""}_${ids?.join(",") ?? ""}, defers transfer recordings, verifies the accumulated block is the expected recall tool_use, creates a UUID anchor plus responsesAnchorContext(), records companion tool uses with "before"/"after" position, stores under anchor:${anchorId}, and builds visible provenance with buildRecallAnchor() and responsesProvenanceContent().
• 🟡 (14:14) /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/src/recall-continuation-failure.ts defines 16 ordered RECALL_CONTINUATION_FAILURE_CATEGORIES: 1. "recall_execution", 2. "follow_up_setup", 3. "follow_up_transport", 4. "follow_up_protocol", 5. "follow_up_failed", 6. "follow_up_missing_output", 7. "follow_up_incomplete_arguments", 8. "parallel_recall", 9. "nested_recall_incomplete", 10. "nested_recall_execution", 11. "depth_exhausted", 12. "missing_recall_block", 13. "resource_limit", 14. "delivery", 15. "unexpected".
• 🟡 (14:14) RecallContinuationFailure always exposes the generic message "recall continuation failed" while retaining a typed category; reportRecallContinuationFailure() ignores unknown categories and catches exceptions thrown by the registered failure hook.
• 🔴 (14:14) User asserted that diagnostics “never affect the response path.”
• 🟡 (14:15) /home/byk/Code/opencode-lore-responses-accumulation/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.
• 🟡 (14:15) RecallStopReason values are "time", "tokens", "result_bytes", "items", "stalled", and "execution". RecallChainBudget retains only counts and short-lived hashed/opaque coverage keys and does not log or export source IDs, queries, rendered bodies, or fingerprints.
• 🟡 (14:15) RecallChainBudget defaults its deadline to startedAt + 300_000 - RECALL_FINALIZATION_RESERVE_MS; an explicit deadline is reduced by the same reserve but clamped to no earlier than startedAt.
• 🟡 (14:15) RecallChainBudget.admit() validates maxDeliveredItems as a positive safe integer, then checks time, token admission limit, result bytes, item capacity including reservations, and execution count in that order. On admission it increments executions and reserves the operation’s maximum source count.
• 🟡 (14:15) Recall token accounting includes inputTokens, outputTokens, cacheReadInputTokens, and cacheCreationInputTokens; only positive safe integers count. Admission reserves min(4_096, floor(maxTokens / 4)) tokens, while mustFinalizeNext() becomes true on any existing stop, expiration, or total usage reaching maxTokens - 2 * reserve.
• 🟡 (14:15) RecallChainBudget.record() validates resultBytes as a non-negative safe integer, replaces the operation’s conservative item reservation with distinct delivered coverage, keys coverage by identity/revision/kind/offset/length, and keys item counts by identity/revision. Missing coverage metadata consumes the reservation and is treated as progress for compatibility; explicit [] is no progress.
• 🔴 (14:15) User asserted that search previews “always carry a non-zero marker”; zero-length coverage is accepted only for a complete detail page with item.kind === "detail", item.complete, and item.length === 0.
• 🟡 (14:15) Recall budget stop precedence after recording is time, execution, result bytes, items, tokens, then stalled. Stalling occurs when consecutiveNoProgress > maxConsecutiveNoProgress, meaning the default threshold of 2 permits two consecutive no-progress results and stops after the next.
• 🟡 (14:16) /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/test/openai-responses-recall-aware-stream.test.ts verifies malformed terminal reasoning ciphertext such as encrypted_content: 42 is rejected before recall execution: output contains PUBLIC_RECALL_ERROR, with recalls === 0 and followUps === 0.
• 🟡 (14:16) Recall-aware stream validation tests reject unfinished reasoning when a response.reasoning_summary_text.delta containing "partial" is omitted from the corresponding response.output_item.done, both when summary: [] is present and when the summary field is absent; the client stream must contain "response.failed".
• 🟡 (14:16) Recall-aware stream validation includes a test rejecting output-text lifecycle inconsistency where response.output_text.delta emits "different" but the done/final item reports "final".
• 🟡 (14:16) The current change set spans 3 files: packages/gateway/src/pipeline.ts with 519 changed-line indicators, packages/gateway/src/stream/openai-responses.ts with 1, and packages/gateway/test/openai-responses-recall-aware-stream.test.ts with 1,684; total diff is 2,068 insertions and 136 deletions.
• 🟡 (14:16) Relevant commit history, newest first: ce88c614c0ae55d229cb605aa7c7369d918d3650 (fix(responses): complete lifecycle reconciliation), 683dff02101f15198b18e3088a0b57dade290056 (fix(responses): bound sparse lifecycle coordinates), df49b0cec904a3f432a0f7f29ea398840d62fe9e (fix(responses): preserve limit-crossing state), and 2cbf40060c67df229731cf1b9e34d5bf22ec2145 (feat(cch): add seeds for Claude Code up to 2.1.272 (#1780)).
• 🟡 (14:17) accountUnsuccessfulResponse() in /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/src/pipeline.ts records provider usage only when at least one token value is positive, optionally marks session state dirty, updates conversation/Sentry usage, then always marks the generation span status as code 2 with message "upstream response did not complete" and ends the span.
• 🔴 (14:18) User asserted that streamResponsesRecallAware() “never commits when an abort-ignoring recall callback resolves late”; the corresponding test expects committed to remain 0.
• 🔴 (14:18) User asserted that cancellation “never waits for a non-settling recall callback.”
• 🔴 (14:18) User asserted that cancellation “never waits for non-settling follow-up setup.”
• 🟡 (14:18) Additional recall-aware stream cancellation/timeout tests cover rollback observation after cancellation, cancellation of a dropped continuation after a "socket reset", foreground timeout settling non-settling follow-up setup, and foreground abort cancelling and unlocking a hostile continuation reader whose pull never settles.
• 🟡 (14:18) handleRequestForTenant() in /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/src/pipeline.ts returns HTTP 400 for missing headers and HTTP 503 while pipelineResetInProgress; otherwise it creates a per-request downstreamSettled promise and downstreamCancelled flag, passes both into handleRequestInner(), resolves settlement from runActivePipelineRequest() cleanup, and marks cancellation through a separate callback.
• 🟡 (14:18) handleRequestInner() rejects conflicting x-api-key and Authorization headers with HTTP 400, validates explicit provider/upstream selection before alternate branches, intercepts /lore:* commands, passes Claude Code side-channel requests upstream untouched before compaction detection, excludes Claude Code subagents from structural compaction detection, passes meta requests through, and routes normal turns to handleConversationTurn() with downstream settlement/cancellation tracking.
• 🟡 (14:18) Client disconnects represented by DOMException with name "AbortError" are logged at info level as "pipeline aborted (client disconnect)"; pressure-aware abort capture is invoked, while arbitrary parser/fetch error details are restricted because upstream response content must never reach logs.
• 🟡 (14:19) runActivePipelineRequest() is defined in /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/src/pipeline.ts at line 15372.