feat(open-sse): early-peek stream error detection + fix UTF-8 loss in CommandCode peek

- new open-sse/utils/streamErrorPeek.js: bounded peek of the first bytes of a
  200 stream; configured pattern match → 502 so account/combo fallback can run
  before any byte reaches the client (streaming included). Re-emits RAW bytes
  so split multi-byte UTF-8 sequences survive the peek (never re-encode
  decoded text — TextDecoder flush corrupts a lone leading byte to U+FFFD).
- chatCore: run the peek after executor.execute when the provider has
  streamErrorPatterns configured; chat.js passes the settings through.
- commandcode executor: same raw-bytes fix in peekForUpstreamError + regression
  test that fails against the old flush-based re-encode.
This commit is contained in:
2026-08-04 23:33:12 +07:00
parent 008e0ef311
commit e438a03f96
6 changed files with 1301 additions and 583 deletions

View File

@@ -27,7 +27,7 @@ export class CommandCodeExecutor extends BaseExecutor {
super("commandcode", PROVIDERS.commandcode);
}
transformRequest(model, body, stream, credentials) {
transformRequest(_model, body, _stream, _credentials) {
body.stream = true;
return body;
}
@@ -131,13 +131,17 @@ export async function peekForUpstreamError(
) {
const reader = originalResponse.body.getReader();
const decoder = new TextDecoder();
const encoder = new TextEncoder();
const abortController = new AbortController();
const forwardAbort = () => abortController.abort(signal?.reason);
if (signal?.aborted) abortController.abort(signal?.reason);
else if (signal)
signal.addEventListener("abort", forwardAbort, { once: true });
// Raw bytes for lossless re-emission; decoded text is only used for line
// parsing / error detection. Never re-encode decoded text: TextDecoder
// holds a split multi-byte char internally and flush() would replace it
// with U+FFFD, corrupting the stream.
const rawChunks = [];
let peeked = "";
let errorEvent = null;
let committed = false;
@@ -167,6 +171,7 @@ export async function peekForUpstreamError(
Math.max(deadline - Date.now(), 1),
);
if (done) break;
rawChunks.push(value);
peeked += decoder.decode(value, { stream: true });
const lines = peeked.split("\n");
// The last segment may be a partial line — only parse complete ones.
@@ -188,10 +193,6 @@ export async function peekForUpstreamError(
// the downstream stream pipeline (stall detection, abort handling) takes over.
}
// Flush any partial multi-byte UTF-8 sequence held by the decoder so the
// re-encoded peeked bytes round-trip losslessly.
peeked += decoder.decode();
if (signal) signal.removeEventListener("abort", forwardAbort);
if (errorEvent) {
@@ -210,7 +211,9 @@ export async function peekForUpstreamError(
start(controller) {
(async () => {
try {
if (peeked) controller.enqueue(encoder.encode(peeked));
// Re-emit RAW bytes (never re-encoded decoded text) so split
// multi-byte UTF-8 sequences survive the peek untouched.
for (const c of rawChunks) controller.enqueue(c);
while (true) {
const { done, value } = await reader.read();
if (done) break;