fix(open-sse): treat CommandCode in-stream error events as request failures
Upstream emits AI SDK v5 {"type":"error"} events inside an HTTP 200 stream.
The translator turned them into fake success content ([CommandCode error: ...]
+ finish_reason stop), so account/model fallback never fired and logs showed
Status: success.
- translator: error events now emit an OpenAI-shaped error chunk (chunk.error)
instead of content; parseSSEToOpenAIResponse already detects chunk?.error
- executor: peek the first events before committing the response; an early
error event returns 502 so fallback runs before any byte reaches the client
This commit is contained in:
@@ -1,6 +1,7 @@
|
||||
import { randomUUID } from "crypto";
|
||||
import { BaseExecutor } from "./base.js";
|
||||
import { PROVIDERS } from "../config/providers.js";
|
||||
import { HTTP_STATUS } from "../config/runtimeConfig.js";
|
||||
import { commandCodeToOpenAIResponse } from "../translator/response/commandcode-to-openai.js";
|
||||
import { SSE_DONE } from "../utils/sseConstants.js";
|
||||
|
||||
@@ -14,81 +15,268 @@ import { SSE_DONE } from "../utils/sseConstants.js";
|
||||
* We translate each event to an OpenAI chat.completion.chunk and emit it as SSE so
|
||||
* both the streaming and non-streaming (forced SSE → JSON) downstream handlers in
|
||||
* 9router can consume it without further format translation.
|
||||
*
|
||||
* Terminal upstream failures arrive as `{"type":"error"}` events inside the HTTP
|
||||
* 200 stream, so a plain `response.ok` check cannot see them. We peek the first
|
||||
* events before committing the response (see peekForUpstreamError) so a stream
|
||||
* that starts with an error fails fast — the normal `!response.ok` path then
|
||||
* triggers account/model fallback instead of streaming fake success content.
|
||||
*/
|
||||
export class CommandCodeExecutor extends BaseExecutor {
|
||||
constructor() {
|
||||
super("commandcode", PROVIDERS.commandcode);
|
||||
}
|
||||
constructor() {
|
||||
super("commandcode", PROVIDERS.commandcode);
|
||||
}
|
||||
|
||||
transformRequest(model, body, stream, credentials) {
|
||||
body.stream = true;
|
||||
return body;
|
||||
}
|
||||
transformRequest(model, body, stream, credentials) {
|
||||
body.stream = true;
|
||||
return body;
|
||||
}
|
||||
|
||||
buildHeaders(credentials, stream = true) {
|
||||
const headers = {
|
||||
"Content-Type": "application/json",
|
||||
...(this.config.headers || {}),
|
||||
"x-session-id": randomUUID(),
|
||||
};
|
||||
buildHeaders(credentials, stream = true) {
|
||||
const headers = {
|
||||
"Content-Type": "application/json",
|
||||
...(this.config.headers || {}),
|
||||
"x-session-id": randomUUID(),
|
||||
};
|
||||
|
||||
const token = credentials?.apiKey || credentials?.accessToken;
|
||||
if (token) headers["Authorization"] = `Bearer ${token}`;
|
||||
const token = credentials?.apiKey || credentials?.accessToken;
|
||||
if (token) headers["Authorization"] = `Bearer ${token}`;
|
||||
|
||||
if (stream) headers["Accept"] = "text/event-stream";
|
||||
return headers;
|
||||
}
|
||||
if (stream) headers["Accept"] = "text/event-stream";
|
||||
return headers;
|
||||
}
|
||||
|
||||
async execute(opts) {
|
||||
const result = await super.execute(opts);
|
||||
if (!result?.response?.ok || !result.response.body) return result;
|
||||
result.response = wrapNdjsonAsOpenAISse(result.response, opts.model);
|
||||
return result;
|
||||
}
|
||||
async execute(opts) {
|
||||
const result = await super.execute(opts);
|
||||
if (!result?.response?.ok || !result.response.body) return result;
|
||||
result.response = await peekForUpstreamError(result.response, opts.model, {
|
||||
signal: opts.signal,
|
||||
});
|
||||
return result;
|
||||
}
|
||||
}
|
||||
|
||||
// How long to hold the response open while peeking the first upstream events.
|
||||
// An upstream error event ("Network connection lost") is emitted at stream
|
||||
// start, so the peek is fast; the bound just prevents a slow-started stream
|
||||
// from being held hostage. Env: COMMANDCODE_PEEK_TIMEOUT_MS.
|
||||
const PEEK_TIMEOUT_MS = (() => {
|
||||
const raw = process.env.COMMANDCODE_PEEK_TIMEOUT_MS;
|
||||
const n = raw ? parseInt(raw, 10) : NaN;
|
||||
return Number.isFinite(n) && n > 0 ? n : 10 * 1000;
|
||||
})();
|
||||
|
||||
// Event types that count as "the stream has started producing". Everything
|
||||
// else (start, start-step, reasoning-start, text-start, ...) is metadata and
|
||||
// does not end the peek.
|
||||
const MEANINGFUL_EVENT_TYPES = new Set([
|
||||
"text-delta",
|
||||
"reasoning-delta",
|
||||
"tool-input-start",
|
||||
"tool-input-delta",
|
||||
"tool-input-end",
|
||||
"tool-call",
|
||||
"finish-step",
|
||||
"finish",
|
||||
]);
|
||||
|
||||
function makeAbortError(reason) {
|
||||
const error = new Error(reason?.message || reason || "Request aborted");
|
||||
error.name = "AbortError";
|
||||
return error;
|
||||
}
|
||||
|
||||
function tryParseEvent(line) {
|
||||
const trimmed = line.trim();
|
||||
if (!trimmed) return null;
|
||||
const json = trimmed.startsWith("data:") ? trimmed.slice(5).trim() : trimmed;
|
||||
if (!json || json === "[DONE]") return null;
|
||||
try {
|
||||
return JSON.parse(json);
|
||||
} catch {
|
||||
return null;
|
||||
}
|
||||
}
|
||||
|
||||
function formatErrorValue(errVal) {
|
||||
const errStr =
|
||||
typeof errVal === "string"
|
||||
? errVal
|
||||
: typeof errVal?.message === "string"
|
||||
? errVal.message
|
||||
: JSON.stringify(errVal);
|
||||
const errType =
|
||||
typeof errVal === "string"
|
||||
? "upstream_error"
|
||||
: errVal?.type || "upstream_error";
|
||||
return { message: errStr, type: errType };
|
||||
}
|
||||
|
||||
/**
|
||||
* Read the first upstream events before committing the response.
|
||||
*
|
||||
* - `{"type":"error"}` as the first meaningful event → return a 502 Response so
|
||||
* chatCore's `!response.ok` path parses the error and triggers fallback.
|
||||
* - Otherwise → re-emit the buffered bytes + the rest of the stream through the
|
||||
* normal NDJSON → OpenAI SSE wrapper and return it untouched in spirit.
|
||||
*
|
||||
* Bounded by `timeoutMs` (default PEEK_TIMEOUT_MS): if no meaningful event
|
||||
* arrives in time, or the request signal aborts, we commit whatever we have and
|
||||
* let the regular stream pipeline (stall detection, abort handling) take over.
|
||||
*/
|
||||
export async function peekForUpstreamError(
|
||||
originalResponse,
|
||||
model,
|
||||
{ signal = null, timeoutMs = PEEK_TIMEOUT_MS } = {},
|
||||
) {
|
||||
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 });
|
||||
|
||||
let peeked = "";
|
||||
let errorEvent = null;
|
||||
let committed = false;
|
||||
|
||||
const readWithTimeout = (ms) => {
|
||||
if (abortController.signal.aborted) {
|
||||
return Promise.reject(makeAbortError(abortController.signal.reason));
|
||||
}
|
||||
const timeoutPromise = new Promise((_, reject) => {
|
||||
const t = setTimeout(() => reject(new Error("peek timeout")), ms);
|
||||
t.unref?.();
|
||||
});
|
||||
const abortPromise = new Promise((_, reject) => {
|
||||
abortController.signal.addEventListener(
|
||||
"abort",
|
||||
() => reject(makeAbortError(abortController.signal.reason)),
|
||||
{ once: true },
|
||||
);
|
||||
});
|
||||
return Promise.race([reader.read(), timeoutPromise, abortPromise]);
|
||||
};
|
||||
|
||||
try {
|
||||
const deadline = Date.now() + timeoutMs;
|
||||
while (!errorEvent && !committed && Date.now() < deadline) {
|
||||
const { done, value } = await readWithTimeout(
|
||||
Math.max(deadline - Date.now(), 1),
|
||||
);
|
||||
if (done) break;
|
||||
peeked += decoder.decode(value, { stream: true });
|
||||
const lines = peeked.split("\n");
|
||||
// The last segment may be a partial line — only parse complete ones.
|
||||
for (const line of lines.slice(0, -1)) {
|
||||
const event = tryParseEvent(line);
|
||||
if (!event?.type) continue;
|
||||
if (event.type === "error") {
|
||||
errorEvent = event;
|
||||
break;
|
||||
}
|
||||
if (MEANINGFUL_EVENT_TYPES.has(event.type)) {
|
||||
committed = true;
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
} catch {
|
||||
// timeout / abort / read failure during the peek → commit whatever we have;
|
||||
// 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) {
|
||||
await reader.cancel("commandcode early error detected").catch(() => {});
|
||||
const { message, type } = formatErrorValue(
|
||||
errorEvent.error ?? errorEvent.message ?? "unknown",
|
||||
);
|
||||
return new Response(JSON.stringify({ error: { message, type } }), {
|
||||
status: HTTP_STATUS.BAD_GATEWAY,
|
||||
statusText: message.slice(0, 200),
|
||||
headers: { "Content-Type": "application/json" },
|
||||
});
|
||||
}
|
||||
|
||||
const remaining = new ReadableStream({
|
||||
start(controller) {
|
||||
(async () => {
|
||||
try {
|
||||
if (peeked) controller.enqueue(encoder.encode(peeked));
|
||||
while (true) {
|
||||
const { done, value } = await reader.read();
|
||||
if (done) break;
|
||||
controller.enqueue(value);
|
||||
}
|
||||
controller.close();
|
||||
} catch (err) {
|
||||
controller.error(err);
|
||||
}
|
||||
})();
|
||||
},
|
||||
cancel() {
|
||||
reader.cancel("commandcode stream cancelled").catch(() => {});
|
||||
},
|
||||
});
|
||||
|
||||
const combined = new Response(remaining, {
|
||||
status: originalResponse.status,
|
||||
statusText: originalResponse.statusText,
|
||||
headers: originalResponse.headers,
|
||||
});
|
||||
return wrapNdjsonAsOpenAISse(combined, model);
|
||||
}
|
||||
|
||||
function wrapNdjsonAsOpenAISse(originalResponse, model) {
|
||||
const decoder = new TextDecoder();
|
||||
const encoder = new TextEncoder();
|
||||
let buffer = "";
|
||||
const state = { model };
|
||||
const decoder = new TextDecoder();
|
||||
const encoder = new TextEncoder();
|
||||
let buffer = "";
|
||||
const state = { model };
|
||||
|
||||
const emitChunks = (chunks, controller) => {
|
||||
if (!chunks) return;
|
||||
const list = Array.isArray(chunks) ? chunks : [chunks];
|
||||
for (const c of list) {
|
||||
if (c == null) continue;
|
||||
controller.enqueue(encoder.encode(`data: ${JSON.stringify(c)}\n\n`));
|
||||
}
|
||||
};
|
||||
const emitChunks = (chunks, controller) => {
|
||||
if (!chunks) return;
|
||||
const list = Array.isArray(chunks) ? chunks : [chunks];
|
||||
for (const c of list) {
|
||||
if (c == null) continue;
|
||||
controller.enqueue(encoder.encode(`data: ${JSON.stringify(c)}\n\n`));
|
||||
}
|
||||
};
|
||||
|
||||
const transform = new TransformStream({
|
||||
transform(chunk, controller) {
|
||||
buffer += decoder.decode(chunk, { stream: true });
|
||||
const lines = buffer.split("\n");
|
||||
buffer = lines.pop() || "";
|
||||
for (const line of lines) {
|
||||
const trimmed = line.trim();
|
||||
if (!trimmed) continue;
|
||||
// Translate AI SDK v5 NDJSON line to one or more OpenAI chunks
|
||||
emitChunks(commandCodeToOpenAIResponse(trimmed, state), controller);
|
||||
}
|
||||
},
|
||||
flush(controller) {
|
||||
const trimmed = buffer.trim();
|
||||
if (trimmed) {
|
||||
emitChunks(commandCodeToOpenAIResponse(trimmed, state), controller);
|
||||
}
|
||||
controller.enqueue(encoder.encode(SSE_DONE));
|
||||
},
|
||||
});
|
||||
const transform = new TransformStream({
|
||||
transform(chunk, controller) {
|
||||
buffer += decoder.decode(chunk, { stream: true });
|
||||
const lines = buffer.split("\n");
|
||||
buffer = lines.pop() || "";
|
||||
for (const line of lines) {
|
||||
const trimmed = line.trim();
|
||||
if (!trimmed) continue;
|
||||
// Translate AI SDK v5 NDJSON line to one or more OpenAI chunks
|
||||
emitChunks(commandCodeToOpenAIResponse(trimmed, state), controller);
|
||||
}
|
||||
},
|
||||
flush(controller) {
|
||||
const trimmed = buffer.trim();
|
||||
if (trimmed) {
|
||||
emitChunks(commandCodeToOpenAIResponse(trimmed, state), controller);
|
||||
}
|
||||
controller.enqueue(encoder.encode(SSE_DONE));
|
||||
},
|
||||
});
|
||||
|
||||
const newBody = originalResponse.body.pipeThrough(transform);
|
||||
return new Response(newBody, {
|
||||
status: originalResponse.status,
|
||||
statusText: originalResponse.statusText,
|
||||
headers: originalResponse.headers,
|
||||
});
|
||||
const newBody = originalResponse.body.pipeThrough(transform);
|
||||
return new Response(newBody, {
|
||||
status: originalResponse.status,
|
||||
statusText: originalResponse.statusText,
|
||||
headers: originalResponse.headers,
|
||||
});
|
||||
}
|
||||
|
||||
export default CommandCodeExecutor;
|
||||
|
||||
Reference in New Issue
Block a user