A stream that stalled or lost its upstream was closed with no terminal frame
at all, so clients saw "200 OK, a few chunks, then nothing" and could not tell
a truncated reply from a finished one. The Responses passthrough path already
synthesized response.failed; every other client format got nothing.
The watchdog now hands its reason ("stream stall timeout" or "upstream
connection lost") to onAbortTerminal, and buildStreamErrorBytes frames it per
client format: OpenAI-compatible clients get data: {"error":{...}} followed by
data: [DONE], Anthropic clients get `event: error`. The error frame always
precedes [DONE] (openai-python raises APIError on any data payload carrying an
error key), and no synthetic finish_reason is ever emitted — a truncated
stream must not look like a clean stop.
Co-Authored-By: Claude Code <noreply@anthropic.com>
148 lines
4.8 KiB
JavaScript
148 lines
4.8 KiB
JavaScript
import { FORMATS } from "../translator/formats.js";
|
|
import { buildErrorBody } from "./error.js";
|
|
import { SSE_DONE } from "./sseConstants.js";
|
|
|
|
const sharedEncoder = new TextEncoder();
|
|
|
|
// Parse SSE data line
|
|
export function parseSSELine(line, format = null) {
|
|
if (!line) return null;
|
|
|
|
// NDJSON format (Ollama): raw JSON lines without "data:" prefix
|
|
if (format === FORMATS.OLLAMA) {
|
|
const trimmed = line.trim();
|
|
if (trimmed.startsWith("{")) {
|
|
try {
|
|
return JSON.parse(trimmed);
|
|
} catch (error) {
|
|
return null;
|
|
}
|
|
}
|
|
return null;
|
|
}
|
|
|
|
// Standard SSE format: "data: {...}"
|
|
if (line.charCodeAt(0) !== 100) return null; // 'd' = 100
|
|
|
|
const data = line.slice(5).trim();
|
|
if (data === "[DONE]") return { done: true };
|
|
|
|
try {
|
|
return JSON.parse(data);
|
|
} catch (error) {
|
|
if (data.length > 0 && data.length < 1000) {
|
|
console.log(`[WARN] Failed to parse SSE line (${data.length} chars): ${data.substring(0, 100)}...`);
|
|
}
|
|
return null;
|
|
}
|
|
}
|
|
|
|
// Check if chunk has valuable content (not empty)
|
|
export function hasValuableContent(chunk, format) {
|
|
// OpenAI format
|
|
if (format === FORMATS.OPENAI && chunk.choices?.[0]?.delta) {
|
|
const delta = chunk.choices[0].delta;
|
|
return delta.content && delta.content !== "" ||
|
|
delta.reasoning_content && delta.reasoning_content !== "" ||
|
|
delta.tool_calls && delta.tool_calls.length > 0 ||
|
|
chunk.choices[0].finish_reason ||
|
|
delta.role;
|
|
}
|
|
|
|
// Claude format
|
|
if (format === FORMATS.CLAUDE) {
|
|
const isContentBlockDelta = chunk.type === "content_block_delta";
|
|
const hasText = chunk.delta?.text && chunk.delta.text !== "";
|
|
const hasThinking = chunk.delta?.thinking && chunk.delta.thinking !== "";
|
|
const hasInputJson = chunk.delta?.partial_json && chunk.delta.partial_json !== "";
|
|
|
|
if (isContentBlockDelta && !hasText && !hasThinking && !hasInputJson) {
|
|
return false;
|
|
}
|
|
return true;
|
|
}
|
|
|
|
return true; // Other formats: keep all chunks
|
|
}
|
|
|
|
// Fix invalid id (generic or too short)
|
|
export function fixInvalidId(parsed) {
|
|
if (parsed.id && (parsed.id === "chat" || parsed.id === "completion" || parsed.id.length < 8)) {
|
|
const fallbackId = parsed.extend_fields?.requestId ||
|
|
parsed.extend_fields?.traceId ||
|
|
Date.now().toString(36);
|
|
parsed.id = `chatcmpl-${fallbackId}`;
|
|
return true;
|
|
}
|
|
return false;
|
|
}
|
|
|
|
function cleanUsagePayload(payload) {
|
|
if (!payload || typeof payload !== "object" || Array.isArray(payload)) {
|
|
return payload;
|
|
}
|
|
|
|
let cleaned = payload;
|
|
|
|
if ("usage" in cleaned) {
|
|
if (cleaned.usage === null) {
|
|
const { usage, ...payloadWithoutUsage } = cleaned;
|
|
cleaned = payloadWithoutUsage;
|
|
} else if (typeof cleaned.usage === "object" && cleaned.usage.perf_metrics === null) {
|
|
const { perf_metrics, ...usageWithoutPerf } = cleaned.usage;
|
|
cleaned = { ...cleaned, usage: usageWithoutPerf };
|
|
}
|
|
}
|
|
|
|
if (cleaned.response && typeof cleaned.response === "object" && !Array.isArray(cleaned.response)) {
|
|
const cleanedResponse = cleanUsagePayload(cleaned.response);
|
|
if (cleanedResponse !== cleaned.response) {
|
|
cleaned = { ...cleaned, response: cleanedResponse };
|
|
}
|
|
}
|
|
|
|
return cleaned;
|
|
}
|
|
|
|
// Format output as SSE
|
|
export function formatSSE(data, sourceFormat) {
|
|
if (data === null || data === undefined) return "data: null\n\n";
|
|
if (data && data.done) return "data: [DONE]\n\n";
|
|
|
|
// OpenAI Responses API format
|
|
if (data && data.event && data.data) {
|
|
const cleanedEventData = cleanUsagePayload(data.data);
|
|
return `event: ${data.event}\ndata: ${JSON.stringify(cleanedEventData)}\n\n`;
|
|
}
|
|
|
|
data = cleanUsagePayload(data);
|
|
|
|
// Claude format
|
|
if (sourceFormat === FORMATS.CLAUDE && data && data.type) {
|
|
return `event: ${data.type}\ndata: ${JSON.stringify(data)}\n\n`;
|
|
}
|
|
|
|
return `data: ${JSON.stringify(data)}\n\n`;
|
|
}
|
|
|
|
// Terminal frames for a stream that aborted after HTTP 200 was already sent, so
|
|
// the status code can no longer change. OpenAI-compatible clients (openai-python
|
|
// raises APIError on any `data:` payload carrying an `error` key, checked before
|
|
// [DONE]) need the error frame first, then [DONE]; Anthropic clients need
|
|
// `event: error`. Never fabricate a successful finish_reason instead.
|
|
//
|
|
// Returns encoded bytes: onAbortTerminal callbacks are enqueued verbatim, same
|
|
// as buildAbortedResponsesTerminalBytes.
|
|
//
|
|
// NOTE: non-SSE client formats (Ollama NDJSON) get an SSE frame here — dead in
|
|
// practice because detectFormatByEndpoint never resolves to OLLAMA.
|
|
export function buildStreamErrorBytes(statusCode, message, clientFormat) {
|
|
const { error } = buildErrorBody(statusCode, message);
|
|
|
|
const sse = clientFormat === FORMATS.CLAUDE
|
|
? formatSSE({ type: "error", error }, FORMATS.CLAUDE)
|
|
: formatSSE({ error }, clientFormat) + SSE_DONE;
|
|
|
|
return sharedEncoder.encode(sse);
|
|
}
|