Skip to content

Commit daed83d

Browse files
jpr5jpr5
committed
fix: normalize OpenAI-compatible paths with arbitrary prefixes (#107)
Providers like BigModel (/v4/chat/completions), DeepSeek, and others use non-standard base URLs. Normalize any path ending in a known endpoint suffix to /v1/<endpoint> before routing. Closes #107
1 parent 0495773 commit daed83d

File tree

2 files changed

+299
-9
lines changed

2 files changed

+299
-9
lines changed

‎src/__tests__/provider-compat.test.ts‎

Lines changed: 244 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,4 +1,6 @@
11
import { describe, it, expect, afterEach } from "vitest";
2+
import http from "node:http";
3+
import crypto from "node:crypto";
24
import { createServer, type ServerInstance } from "../server.js";
35
import type { Fixture } from "../types.js";
46

@@ -209,6 +211,248 @@ describe("Together AI compatibility", () => {
209211
});
210212
});
211213

214+
describe("OpenAI-compatible path prefix normalization", () => {
215+
it("normalizes /v4/chat/completions to /v1/chat/completions", async () => {
216+
instance = await createServer(CATCH_ALL_FIXTURES);
217+
218+
const { status, body } = await httpPost(`${instance.url}/v4/chat/completions`, {
219+
model: "bigmodel-4",
220+
stream: false,
221+
messages: [{ role: "user", content: "hello" }],
222+
});
223+
224+
expect(status).toBe(200);
225+
const parsed = JSON.parse(body);
226+
expect(parsed.choices).toBeDefined();
227+
expect(parsed.choices[0].message.content).toBe("Hello from aimock!");
228+
expect(parsed.object).toBe("chat.completion");
229+
});
230+
231+
it("normalizes /api/coding/paas/v4/chat/completions to /v1/chat/completions", async () => {
232+
instance = await createServer(CATCH_ALL_FIXTURES);
233+
234+
const { status, body } = await httpPost(`${instance.url}/api/coding/paas/v4/chat/completions`, {
235+
model: "bigmodel-4",
236+
stream: false,
237+
messages: [{ role: "user", content: "hello" }],
238+
});
239+
240+
expect(status).toBe(200);
241+
const parsed = JSON.parse(body);
242+
expect(parsed.choices).toBeDefined();
243+
expect(parsed.choices[0].message.content).toBe("Hello from aimock!");
244+
expect(parsed.object).toBe("chat.completion");
245+
});
246+
247+
it("still handles standard /v1/chat/completions (regression)", async () => {
248+
instance = await createServer(CATCH_ALL_FIXTURES);
249+
250+
const { status, body } = await httpPost(`${instance.url}/v1/chat/completions`, {
251+
model: "gpt-4o",
252+
stream: false,
253+
messages: [{ role: "user", content: "hello" }],
254+
});
255+
256+
expect(status).toBe(200);
257+
const parsed = JSON.parse(body);
258+
expect(parsed.choices).toBeDefined();
259+
expect(parsed.choices[0].message.content).toBe("Hello from aimock!");
260+
expect(parsed.object).toBe("chat.completion");
261+
});
262+
263+
it("normalizes /custom/embeddings to /v1/embeddings", async () => {
264+
instance = await createServer(CATCH_ALL_FIXTURES);
265+
266+
const { status, body } = await httpPost(`${instance.url}/custom/embeddings`, {
267+
model: "text-embedding-3-small",
268+
input: "test embedding via custom prefix",
269+
});
270+
271+
expect(status).toBe(200);
272+
const parsed = JSON.parse(body);
273+
expect(parsed.object).toBe("list");
274+
expect(parsed.data[0].embedding).toBeInstanceOf(Array);
275+
});
276+
277+
it("combines /openai/ prefix strip with normalization for non-v1 paths", async () => {
278+
instance = await createServer(CATCH_ALL_FIXTURES);
279+
280+
// /openai/v4/chat/completions — strip /openai/ then normalize /v4/ to /v1/
281+
const { status, body } = await httpPost(
282+
`${instance.url}/openai/v4/chat/completions`,
283+
{
284+
model: "llama-3.3-70b-versatile",
285+
stream: false,
286+
messages: [{ role: "user", content: "hello" }],
287+
},
288+
{ Authorization: "Bearer mock-groq-key" },
289+
);
290+
291+
expect(status).toBe(200);
292+
const parsed = JSON.parse(body);
293+
expect(parsed.choices).toBeDefined();
294+
expect(parsed.choices[0].message.content).toBe("Hello from aimock!");
295+
});
296+
297+
it("normalizes /custom/responses to /v1/responses", async () => {
298+
instance = await createServer(CATCH_ALL_FIXTURES);
299+
300+
const { body } = await httpPost(`${instance.url}/custom/responses`, {
301+
model: "gpt-4o",
302+
input: "hello",
303+
stream: false,
304+
});
305+
306+
// Normalization works: we get "No fixture matched" from the Responses handler
307+
// (not "Not found" which would mean the path wasn't routed at all)
308+
const parsed = JSON.parse(body);
309+
expect(parsed.error.type).toBe("invalid_request_error");
310+
expect(parsed.error.code).toBe("no_fixture_match");
311+
});
312+
313+
it("normalizes /custom/audio/speech to /v1/audio/speech", async () => {
314+
instance = await createServer(CATCH_ALL_FIXTURES);
315+
316+
const { body } = await httpPost(`${instance.url}/custom/audio/speech`, {
317+
model: "tts-1",
318+
input: "test speech",
319+
voice: "alloy",
320+
});
321+
322+
// Normalization works: handler reached (not "Not found")
323+
const parsed = JSON.parse(body);
324+
expect(parsed.error.type).toBe("invalid_request_error");
325+
});
326+
327+
it("normalizes /custom/audio/transcriptions to /v1/audio/transcriptions", async () => {
328+
instance = await createServer(CATCH_ALL_FIXTURES);
329+
330+
const { body } = await httpPost(`${instance.url}/custom/audio/transcriptions`, {
331+
model: "whisper-1",
332+
file: "test",
333+
});
334+
335+
// Normalization works: handler reached (not "Not found")
336+
const parsed = JSON.parse(body);
337+
expect(parsed.error.type).toBe("invalid_request_error");
338+
});
339+
340+
it("normalizes /custom/images/generations to /v1/images/generations", async () => {
341+
instance = await createServer(CATCH_ALL_FIXTURES);
342+
343+
const { body } = await httpPost(`${instance.url}/custom/images/generations`, {
344+
model: "dall-e-3",
345+
prompt: "test",
346+
});
347+
348+
// Normalization works: handler reached (not "Not found")
349+
const parsed = JSON.parse(body);
350+
expect(parsed.error.type).toBe("invalid_request_error");
351+
});
352+
353+
it("does NOT normalize /v2/chat/completions (/v2/ guard for Cohere convention)", async () => {
354+
instance = await createServer(CATCH_ALL_FIXTURES);
355+
356+
const { status } = await httpPost(`${instance.url}/v2/chat/completions`, {
357+
model: "command-r-plus",
358+
stream: false,
359+
messages: [{ role: "user", content: "hello" }],
360+
});
361+
362+
// /v2/chat/completions should NOT be rewritten to /v1/chat/completions
363+
// — the /v2/ guard prevents normalization, so this falls through to 404
364+
expect(status).toBe(404);
365+
});
366+
367+
it("routes /v2/chat to Cohere handler (not normalization concern)", async () => {
368+
instance = await createServer(CATCH_ALL_FIXTURES);
369+
370+
// /v2/chat is Cohere's endpoint — reaches the Cohere handler directly
371+
const { status } = await httpPost(`${instance.url}/v2/chat`, {
372+
model: "command-r-plus",
373+
stream: false,
374+
messages: [{ role: "user", content: "hello" }],
375+
});
376+
377+
expect(status).toBe(200);
378+
});
379+
380+
it("returns 404 for unrecognized paths that don't match any suffix", async () => {
381+
instance = await createServer(CATCH_ALL_FIXTURES);
382+
383+
const { status } = await httpPost(`${instance.url}/custom/foo/bar`, {
384+
model: "test",
385+
messages: [{ role: "user", content: "hello" }],
386+
});
387+
388+
expect(status).toBe(404);
389+
});
390+
});
391+
392+
describe("WebSocket path normalization", () => {
393+
/**
394+
* Send an HTTP upgrade request and return the resulting status code.
395+
* 101 = upgrade succeeded (WebSocket), anything else = rejected.
396+
*/
397+
function wsUpgrade(url: string, path: string): Promise<{ statusCode: number }> {
398+
return new Promise((resolve, reject) => {
399+
const parsed = new URL(url);
400+
const req = http.request({
401+
hostname: parsed.hostname,
402+
port: parsed.port,
403+
path,
404+
headers: {
405+
Connection: "Upgrade",
406+
Upgrade: "websocket",
407+
"Sec-WebSocket-Key": Buffer.from(crypto.randomBytes(16)).toString("base64"),
408+
"Sec-WebSocket-Version": "13",
409+
},
410+
});
411+
req.on("upgrade", (_res, socket) => {
412+
socket.destroy();
413+
resolve({ statusCode: 101 });
414+
});
415+
req.on("response", (res) => {
416+
resolve({ statusCode: res.statusCode ?? 0 });
417+
});
418+
req.on("error", reject);
419+
req.end();
420+
});
421+
}
422+
423+
it("WS upgrade to /custom/responses normalizes to /v1/responses", async () => {
424+
instance = await createServer(CATCH_ALL_FIXTURES);
425+
const { statusCode } = await wsUpgrade(instance.url, "/custom/responses");
426+
expect(statusCode).toBe(101);
427+
});
428+
429+
it("WS upgrade to /openai/v1/responses works (/openai/ strip)", async () => {
430+
instance = await createServer(CATCH_ALL_FIXTURES);
431+
const { statusCode } = await wsUpgrade(instance.url, "/openai/v1/responses");
432+
expect(statusCode).toBe(101);
433+
});
434+
435+
it("WS upgrade to /v2/responses is NOT normalized (returns 404)", async () => {
436+
instance = await createServer(CATCH_ALL_FIXTURES);
437+
const { statusCode } = await wsUpgrade(instance.url, "/v2/responses");
438+
expect(statusCode).toBe(404);
439+
});
440+
441+
it("WS upgrade to Azure deployment path is NOT normalized", async () => {
442+
instance = await createServer(CATCH_ALL_FIXTURES);
443+
444+
// Azure deployment WebSocket path should NOT have /openai/ stripped
445+
// or be normalized — it should 404 cleanly (Azure WS not supported)
446+
const { statusCode } = await wsUpgrade(
447+
instance.url,
448+
"/openai/deployments/gpt-4o/chat/completions",
449+
);
450+
451+
// Not upgraded (Azure deployment paths don't support WS)
452+
expect(statusCode).toBe(404);
453+
});
454+
});
455+
212456
describe("vLLM compatibility", () => {
213457
// vLLM uses standard /v1/chat/completions with custom model names
214458
it("handles vLLM-style request via /v1/chat/completions", async () => {

‎src/server.ts‎

Lines changed: 55 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -78,6 +78,46 @@ const VIDEOS_STATUS_RE = /^\/v1\/videos\/([^/]+)$/;
7878
const GEMINI_PREDICT_RE = /^\/v1beta\/models\/([^:]+):predict$/;
7979
const DEFAULT_CHUNK_SIZE = 20;
8080

81+
// OpenAI-compatible endpoint suffixes for path prefix normalization.
82+
// Providers like BigModel (/v4/) use non-standard base URL prefixes.
83+
// Only includes endpoints that third-party OpenAI-compatible providers are
84+
// likely to serve — excludes provider-specific paths (/messages, /realtime)
85+
// and endpoints unlikely to appear behind non-standard prefixes
86+
// (/moderations, /videos, /models).
87+
const COMPAT_SUFFIXES = [
88+
"/chat/completions",
89+
"/embeddings",
90+
"/responses",
91+
"/audio/speech",
92+
"/audio/transcriptions",
93+
"/images/generations",
94+
];
95+
96+
/**
97+
* Normalize OpenAI-compatible paths with arbitrary prefixes.
98+
* Strips /openai/ prefix and rewrites paths ending in known suffixes to /v1/<suffix>.
99+
* Skips /v1/ (already standard) and /v2/ (Cohere convention).
100+
*/
101+
function normalizeCompatPath(pathname: string, logger?: Logger): string {
102+
// Strip /openai/ prefix (Groq/OpenAI-compat alias)
103+
if (pathname.startsWith("/openai/")) {
104+
pathname = pathname.slice(7);
105+
}
106+
107+
// Normalize arbitrary prefixes to /v1/
108+
if (!pathname.startsWith("/v1/") && !pathname.startsWith("/v2/")) {
109+
for (const suffix of COMPAT_SUFFIXES) {
110+
if (pathname.endsWith(suffix)) {
111+
if (logger) logger.debug(`Path normalized: ${pathname} → /v1${suffix}`);
112+
pathname = "/v1" + suffix;
113+
break;
114+
}
115+
}
116+
}
117+
118+
return pathname;
119+
}
120+
81121
const GEMINI_PATH_RE = /^\/v1beta\/models\/([^:]+):(generateContent|streamGenerateContent)$/;
82122
const AZURE_DEPLOYMENT_RE = /^\/openai\/deployments\/([^/]+)\/(chat\/completions|embeddings)$/;
83123
const BEDROCK_INVOKE_RE = /^\/model\/([^/]+)\/invoke$/;
@@ -691,12 +731,13 @@ export async function createServer(
691731
const parsedUrl = new URL(req.url ?? "/", `http://${req.headers.host ?? "localhost"}`);
692732
let pathname = parsedUrl.pathname;
693733

694-
// Instrument response completion for metrics
734+
// Instrument response completion for metrics. The finish callback reads
735+
// pathname via closure after normalizeCompatPath has rewritten it, so
736+
// metrics record the canonical /v1/... path.
695737
if (registry) {
696-
const rawPathname = pathname;
697738
res.on("finish", () => {
698739
try {
699-
const normalizedPath = normalizePathLabel(rawPathname);
740+
const normalizedPath = normalizePathLabel(pathname);
700741
const method = req.method ?? "UNKNOWN";
701742
const status = String(res.statusCode);
702743
registry.incrementCounter("aimock_requests_total", {
@@ -743,10 +784,9 @@ export async function createServer(
743784
pathname = `/v1/${operation}`;
744785
}
745786

746-
// Groq/OpenAI-compatible alias: strip /openai prefix so that
747-
// /openai/v1/chat/completions → /v1/chat/completions, etc.
748-
if (!azureDeploymentId && pathname.startsWith("/openai/")) {
749-
pathname = pathname.slice(7); // remove "/openai" prefix, keep the rest
787+
// Normalize OpenAI-compatible paths (strip /openai/ prefix + rewrite arbitrary prefixes)
788+
if (!azureDeploymentId) {
789+
pathname = normalizeCompatPath(pathname, logger);
750790
}
751791

752792
// Health / readiness probes
@@ -1508,9 +1548,9 @@ export async function createServer(
15081548
head: Buffer,
15091549
): Promise<void> {
15101550
const parsedUrl = new URL(req.url ?? "/", `http://${req.headers.host ?? "localhost"}`);
1511-
const pathname = parsedUrl.pathname;
1551+
let pathname = parsedUrl.pathname;
15121552

1513-
// Dispatch to mounted services
1553+
// Dispatch to mounted services before any path rewrites
15141554
if (mounts) {
15151555
for (const { path: mountPath, handler } of mounts) {
15161556
if (
@@ -1523,6 +1563,12 @@ export async function createServer(
15231563
}
15241564
}
15251565

1566+
// Normalize OpenAI-compatible paths (strip /openai/ prefix + rewrite arbitrary prefixes)
1567+
// Skip Azure deployment paths — they have their own rewrite in the HTTP handler
1568+
if (!pathname.match(AZURE_DEPLOYMENT_RE)) {
1569+
pathname = normalizeCompatPath(pathname, logger);
1570+
}
1571+
15261572
if (
15271573
pathname !== RESPONSES_PATH &&
15281574
pathname !== REALTIME_PATH &&

0 commit comments

Comments
 (0)