Skip to content

Open-source AI coworkers that each get a computer of their own: a browser, files and tools, with every action decided before it happens and recorded after. Bring any AG-UI agent.

Name conversations on the endpoint OPENAI_BASE_URL names, not on OpenAI (#556)

Vaibhav Zoped496dfa6fed4
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 52b0731b..bc9e660b 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -8,6 +8,15 @@ Newest first. `Unreleased` is what is on `main` and not yet tagged.
## Unreleased
+### Naming a conversation asks the endpoint OPENAI_BASE_URL names, not OpenAI
+
+The job that names a conversation sent its request to api.openai.com whatever `OPENAI_BASE_URL` said,
+while the Bots, the router and tool selection all used the configured endpoint. A deployment behind a
+gateway, a proxy or a local model therefore sent its model key, and the opening of every
+conversation, to OpenAI; OpenAI refused the key, so no conversation was ever named. The request now
+goes to the same endpoint as every other model call. A deployment that never set `OPENAI_BASE_URL`
+behaves as before.
+
### A Bot's question to a person survives a route that fails
Who "a person" is, is a seam a deployment fills in with its own on-call rota or duty desk. If that
diff --git a/server/src/channels/titler.ts b/server/src/channels/titler.ts
index 6032935a..5556d968 100644
--- a/server/src/channels/titler.ts
+++ b/server/src/channels/titler.ts
@@ -2,11 +2,9 @@
* One model call for a few words. Not a Bot and not a turn: a title is not said to anybody and is
* not part of the conversation it names, so there is no runtime work and no thread to hold.
*/
+import { chatCompletionsUrl } from "../routing/model";
import { oneLine } from "./text";
-/** Where an OpenAI-compatible provider answers, overridable the way `agent-bot` overrides it. */
-const DEFAULT_BASE_URL = "https://api.openai.com/v1";
-
/** Rules, not an example: a model copies an example's subject as readily as its shape. */
const INSTRUCTION = [
"You name conversations, like a title in a sidebar.",
@@ -20,7 +18,7 @@ export type TitlerOptions = {
model: string;
/** Resolved per call, so a rotated credential is picked up without a restart. */
resolveApiKey: () => Promise<string | null>;
- baseUrl?: string;
+ environment?: Record<string, string | undefined>;
/** Injectable so a test drives this without a network. */
fetchImpl?: typeof fetch;
/** How long one call may take before it is given up on. */
@@ -30,29 +28,31 @@ export type TitlerOptions = {
/** No key resolves to null and is not retried; a provider error throws, so the queue retries it. */
export function createChannelTitler(options: TitlerOptions) {
const call = options.fetchImpl ?? fetch;
- const baseUrl = (options.baseUrl ?? DEFAULT_BASE_URL).replace(/\/+$/, "");
return async (excerpt: string): Promise<string | null> => {
const apiKey = await options.resolveApiKey();
if (!apiKey) return null;
- const response = await call(`${baseUrl}/chat/completions`, {
- method: "POST",
- headers: {
- "content-type": "application/json",
- authorization: `Bearer ${apiKey}`,
+ const response = await call(
+ chatCompletionsUrl(options.environment ?? process.env),
+ {
+ method: "POST",
+ headers: {
+ "content-type": "application/json",
+ authorization: `Bearer ${apiKey}`,
+ },
+ body: JSON.stringify({
+ model: options.model,
+ messages: [
+ { role: "system", content: INSTRUCTION },
+ { role: "user", content: excerpt },
+ ],
+ // Reasoning tokens come out of this budget, so a cap sized for the answer alone 400s.
+ max_completion_tokens: 512,
+ }),
+ signal: AbortSignal.timeout(options.timeoutMs ?? 20_000),
},
- body: JSON.stringify({
- model: options.model,
- messages: [
- { role: "system", content: INSTRUCTION },
- { role: "user", content: excerpt },
- ],
- // Reasoning tokens come out of this budget, so a cap sized for the answer alone 400s.
- max_completion_tokens: 512,
- }),
- signal: AbortSignal.timeout(options.timeoutMs ?? 20_000),
- });
+ );
if (!response.ok) {
// Capped: this lands on the work item's row, and an HTML error page would land there whole.
diff --git a/server/tests/channel-titler.test.ts b/server/tests/channel-titler.test.ts
index 953be304..60af4675 100644
--- a/server/tests/channel-titler.test.ts
+++ b/server/tests/channel-titler.test.ts
@@ -25,6 +25,7 @@ describe("asking the model for a title", () => {
const answer = await createChannelTitler({
model: "gpt-4.1-mini",
resolveApiKey: async () => "key-123",
+ environment: {},
fetchImpl,
})("Asked: which receipts count as travel?");
@@ -33,6 +34,43 @@ describe("asking the model for a title", () => {
expect(calls[0]?.body.model).toBe("gpt-4.1-mini");
});
+ test("asks the endpoint OPENAI_BASE_URL names, not OpenAI", async () => {
+ const { calls, fetchImpl } = respondWith({
+ choices: [{ message: { content: "Travel receipt rules" } }],
+ });
+
+ await createChannelTitler({
+ model: "openai/gpt-5.6-terra",
+ resolveApiKey: async () => "gateway-key",
+ environment: { OPENAI_BASE_URL: "https://gateway.internal/v1" },
+ fetchImpl,
+ })("Asked: which receipts count as travel?");
+
+ expect(calls.map((call) => call.url)).toEqual([
+ "https://gateway.internal/v1/chat/completions",
+ ]);
+ });
+
+ test("reads OPENAI_BASE_URL from the process when no environment is passed", async () => {
+ const { calls, fetchImpl } = respondWith({
+ choices: [{ message: { content: "Travel receipt rules" } }],
+ });
+ const previous = process.env.OPENAI_BASE_URL;
+ process.env.OPENAI_BASE_URL = "http://localhost:4010";
+ try {
+ await createChannelTitler({
+ model: "gpt-4.1-mini",
+ resolveApiKey: async () => "gateway-key",
+ fetchImpl,
+ })("Asked: anything?");
+ } finally {
+ if (previous === undefined) delete process.env.OPENAI_BASE_URL;
+ else process.env.OPENAI_BASE_URL = previous;
+ }
+
+ expect(calls[0]?.url).toBe("http://localhost:4010/v1/chat/completions");
+ });
+
test("asks nothing at all when the deployment has no key", async () => {
const { calls, fetchImpl } = respondWith({});