Distillation

Date: Sep 16, 2026

• 🔴 (14:12) Streaming frame-limit references total 14 matches: DEFAULT_MAX_SSE_FRAMES = 100_000 is exported from /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/src/stream/anthropic.ts; it is also used by anthropic.ts, openai.ts, gemini.ts, and openai-responses.ts, including as the default maxSparseIndex in openai-responses.ts.
• 🔴 (14:15) /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/test/openai-responses-recall-aware-stream.test.ts test "computes the continuation offset without spreading 126000 retained item keys" creates visibleItemCount = 63_000, validates and accumulates all 63,000 items, then provides two complete 63,001-entry maps to the production max-index path; its total frame count must remain <= 100_000.
• 🔴 (14:15) The 126,000-key continuation-offset regression test temporarily spies on Map.prototype[Symbol.iterator] to isolate the offset calculation from an existing quadratic duplicate-identity scan, restores the iterator before releasing the recall tail, uses maxRetainedStateBytes: 64 * 1024 * 1024, and has a 60_000 ms timeout.
• 🔴 (14:15) The 126,000-key continuation-offset regression expects the follow-up text "answer" in completedResponse.content, successful === true, output containing "response.completed", and no "response.failed".
• 🔴 (14:15) The test "accounts for terminal Codex continuation usage once when candidate flush exceeds the transaction limit" uses principal usage { input_tokens: 2, output_tokens: 3, total_tokens: 5 } and continuation usage { input_tokens: 5, output_tokens: 7, total_tokens: 12 }; its private terminal argument contains private-terminal-flush- followed by 4,096 x characters.
• 🔴 (14:17) Targeted Vitest run passed: 1 test file, 255 tests passed, duration 15.69s; transform 3.83s, setup 570ms, import 6.65s, tests 8.13s, environment 0ms.
• 🔴 (14:18) Search found recall-aware tests "rejects unfinished reasoning omitted from output_item.done" at line 4195 and "rejects unfinished reasoning when output_item.done omits summary" at line 4227 in /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/test/openai-responses-recall-aware-stream.test.ts.
• 🔴 (14:18) Corresponding non-recall Codex coverage in /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/test/openai-responses-stream.test.ts includes "Codex binds an omitted added function-call ID from output_item.done" at line 2990 and "Codex validation infers omitted output/content indices and permits an added item without done" at line 3134.
• 🔴 (14:21) Tool output reported four exact SHA-256-like digests: 4b247d3850983b60a249cd0be3a6c005509d5fb26b41da259c8ed1c6e9ede154, 39593a5f72dec344a670f89125245f3cc4ad75337a1b9a3611aa1a99256bea34, 504b5448846f43333bccba80f9a262171a3b6e0fde6146252ca6334297d3ecaa, and a83e976242db98545e26f5c53618f8b5713ab6b5d60cba449ac9ae2996d31333.
• 🔴 (14:23) /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/test/openai-responses-recall-aware-stream.test.ts has 6 maxStreamBytes references, including "bounds request-wide no-index stream bytes" at line 7163 and "caps raw stream bytes across a continuation" at line 7183.
• 🔴 (14:24) Test "keeps sparse continuation candidates under the deferred input limit" uses a 1 * 1024 * 1024-character argument delta, maxDeferredBytes: 1024 * 1024, and maxRetainedStateBytes: 8 * 1024 * 1024; it expects PUBLIC_RECALL_ERROR and failure category "resource_limit".
• 🔴 (14:24) Test "charges a resolved sparse continuation candidate to the transactional spool" uses a 4 KiB anchorText, a sparse visible function-call path containing 10 KiB of x, maxDeferredBytes: 16 * 1024, and maxRetainedStateBytes: 32 * 1024; it expects PUBLIC_RECALL_ERROR and failure category "resource_limit".
• 🔴 (14:24) Test "delivers a resolved sparse continuation candidate exactly once" uses the same 10 KiB sparse arguments and 4 KiB anchor with maxDeferredBytes: 16 * 1024 and maxRetainedStateBytes: 48 * 1024; it expects exactly one response.function_call_arguments.delta frame with output_index: 1, item_id: "fc_sparse_0", and no PUBLIC_RECALL_ERROR.
• 🔴 (14:24) Test "stops reading upstream while the client applies backpressure" confirms the source remains at exactly 3 pulls until the client reader is acquired and canceled.
• 🔴 (14:24) Test "bounds request-wide no-index stream bytes" sets maxStreamBytes to the encoded lifecycle frame byte length plus 1 and expects "response.failed".
• 🔴 (14:24) Test "caps raw stream bytes across a continuation" sets maxStreamBytes to the encoded principal plus continuation-created frame length plus 1; it expects PUBLIC_RECALL_ERROR and failure category "resource_limit".
• 🔴 (14:24) Test "caps retained indexed state across a continuation" sets maxRetainedStateBytes to retained principal indexed bytes plus 1; it expects PUBLIC_RECALL_ERROR and failure category "resource_limit".
• 🔴 (14:26) Targeted Vitest run passed: 1 test file, 169 tests passed, duration 7.19s; transform 3.93s, setup 607ms, import 5.80s, tests 569ms, environment 0ms.
• 🔴 (14:27) Search found parseSSEStream() in /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/src/stream/anthropic.ts at line 150; its shared frame counter is typed as frameCounter?: { count: number }, initialized at line 176, incremented at line 198, and checked against the limit at line 199.
• 🔴 (14:28) parseSSEStream() accepts maxEventBytes, maxFrames, frameCounter, inactivityMs, signal, requireEventTerminator, maxTotalBytes, and fatalUtf8; it defaults maxEventBytes to 4 * 1024 * 1024 and maxFrames to DEFAULT_MAX_SSE_FRAMES.
• 🔴 (14:28) parseSSEStream() validates that maxFrames is a non-negative safe integer, otherwise throws "SSE frame limit must be a non-negative safe integer"; exceeding it throws SSEStreamLimitError with message SSE stream exceeded ${maxFrames} frame limit.
• 🔴 (14:28) parseSSEStream() uses TextDecoder("utf-8", { fatal: opts.fatalUtf8, ignoreBOM: true }), performs its own initial UTF-8 BOM removal and byte accounting, joins multiple data: lines with \n, recognizes blank-line delimiters across CRLF/LF/CR forms, and defaults an event without an event: line to "message".
• 🔴 (14:29) User asserted the recall-aware stream never commits when an abort-ignoring recall callback resolves late; the corresponding test is at line 8626 of /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/test/openai-responses-recall-aware-stream.test.ts.
• 🔴 (14:29) User asserted the recall-aware stream never waits for a non-settling recall callback; the corresponding cancellation test "cancellation never waits for a non-settling recall callback" is at line 8715 of /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/test/openai-responses-recall-aware-stream.test.ts.
• 🔴 (14:29) Additional recall-aware transactional/cancellation coverage includes "rolls back every recall mutation when a commit callback fails" at line 8435, "aborts an in-flight recall callback when the client disconnects" at line 9150, and "classifies recall execution failure without exposing callback details" at line 9476; the latter uses private error text "private callback failure" and asserts it does not appear in output.