Read max_completion_tokens list through the OpenAI preset entry

openAISourceQuirks now reads the api path entry's maxCompletionTokensModels field instead of the static const, so the entry is the single source of truth. A new union test pins flagged plus explicit-exempt against the preset model list so an undecided model fails loudly.

Author: TheGreatAxios

packages/first-class-providers/src/providers.ts

@@ -24,7 +24,9 @@ export const OPENAI_API_BASE_URL = "https://api.openai.com/v1";
  * Preset models whose first-party endpoint rejects `max_tokens` and requires
  * `max_completion_tokens`. Explicit per-model list: adding a model here
  * declares its own requirement, never inferred from name prefixes. gpt-4.1
- * is non-reasoning and stays on `max_tokens`.
+ * is non-reasoning and stays on `max_tokens`. This const is only the api
+ * path entry's initial value — runtime reads the entry's
+ * `maxCompletionTokensModels` field, so that field is the source of truth.
  */
 export const OPENAI_API_MAX_COMPLETION_TOKENS_MODELS: readonly string[] = [
   "gpt-6-astra",

src/config/inference-sources.test.ts

@@ -129,6 +129,21 @@ describe("OpenAI reasoning max_completion_tokens quirk (CL-7785)", () => {
     }
   });
 
+  test("every preset model has an explicit quirk decision", () => {
+    const flagged = new Set(openaiApi?.maxCompletionTokensModels ?? []);
+    // Explicit max_tokens decision: non-reasoning preset models stay on
+    // max_tokens. Adding a preset model requires a decision here AND in the
+    // preset's maxCompletionTokensModels — the union below fails loudly
+    // otherwise instead of silently sending max_tokens.
+    const explicitMaxTokensModels = new Set(["gpt-4.1"]);
+    expect([...flagged, ...explicitMaxTokensModels].sort()).toEqual(
+      [...new Set(presetModels)].sort(),
+    );
+    expect([...flagged].filter((m) => explicitMaxTokensModels.has(m))).toEqual(
+      [],
+    );
+  });
+
   test("reasoning preset models emit max_completion_tokens, never max_tokens", () => {
     for (const model of openaiApi?.maxCompletionTokensModels ?? []) {
       const body = wireBody(model, presetBaseURL);

src/config/inference-sources.ts

@@ -11,7 +11,7 @@ import {
 } from "./index.js";
 import {
   OPENAI_API_BASE_URL,
-  OPENAI_API_MAX_COMPLETION_TOKENS_MODELS,
+  firstClassProviderById,
 } from "../../packages/first-class-providers/src/index.js";
 import { normalizeOpenAICompatibleBaseURL, type Settings } from "./settings.js";
 import {
@@ -42,10 +42,18 @@ function catalogEntry(
 
 // First-party OpenAI reasoning models reject `max_tokens` and require
 // `max_completion_tokens`. The requirement is declared per model on the
-// first-class OpenAI API-key preset — never inferred from name prefixes —
-// and the quirk attaches to the source actually in use: it follows the
-// first-party endpoint, so relays serving the same model names through the
-// same adapter keep `max_tokens`.
+// first-class OpenAI API-key path's `maxCompletionTokensModels` field — never
+// inferred from name prefixes — and read here through that entry, so the
+// entry stays the single source of truth. The quirk attaches to the source
+// actually in use: it follows the first-party endpoint, so relays serving
+// the same model names through the same adapter keep `max_tokens`.
+function openAIAPIPathMaxCompletionTokensModels(): readonly string[] {
+  return (
+    firstClassProviderById("openai")?.paths?.find((p) => p.id === "api")
+      ?.maxCompletionTokensModels ?? []
+  );
+}
+
 function openAISourceQuirks(
   baseURL: string,
   model: string,
@@ -54,7 +62,7 @@ function openAISourceQuirks(
   if (normalized !== normalizeOpenAICompatibleBaseURL(OPENAI_API_BASE_URL)) {
     return undefined;
   }
-  if (!OPENAI_API_MAX_COMPLETION_TOKENS_MODELS.includes(model)) {
+  if (!openAIAPIPathMaxCompletionTokensModels().includes(model)) {
     return undefined;
   }
   return { maxTokensField: "max_completion_tokens" };