fix(provider): match ChatGPT reasoning efforts to the served catalog

The gpt-6 and gpt-6.1 models offered fewer levels than the Codex backend
serves (xhigh, or max for gpt-6-astra), while gpt-5.6-luna was offered
ultra, which the backend does not serve.

Author: bwachman

src/provider/reasoning-effort.test.ts

@@ -1,4 +1,5 @@
 import { afterEach, describe, test, expect } from "bun:test";
+import { CODEX_DEFAULT_MODELS } from "../auth/codex/constants.js";
 import {
   isReasoningEffort,
   normalizeProviderEfforts,
@@ -54,32 +55,53 @@ describe("supportedEfforts", () => {
     ]);
   });
 
-  test("gpt-6-astra takes low through max on both API and Codex paths", () => {
+  test("gpt-6-astra takes low through max on the API path", () => {
     expect(supportedEfforts("gpt-6-astra")).toEqual([
       "low",
       "medium",
       "high",
       "xhigh",
       "max",
     ]);
-    expect(supportedEfforts("gpt-6-astra", undefined, true)).toEqual([
-      "low",
-      "medium",
-      "high",
-      "xhigh",
-      "max",
-    ]);
   });
 
-  test("gpt-5.6 family additionally takes max and ultra on the codex backend", () => {
-    expect(supportedEfforts("gpt-5.6-sol", undefined, true)).toEqual([
-      "low",
-      "medium",
-      "high",
-      "xhigh",
-      "max",
-      "ultra",
-    ]);
+  test("ChatGPT models the backend serves up to ultra offer max and ultra", () => {
+    for (const model of [
+      "gpt-6.1-sol",
+      "gpt-6-astra",
+      "gpt-6.1-terra",
+      "gpt-6-sol",
+      "gpt-5.6-sol",
+      "gpt-5.6-terra",
+    ]) {
+      expect(supportedEfforts(model, undefined, true)).toEqual([
+        "low",
+        "medium",
+        "high",
+        "xhigh",
+        "max",
+        "ultra",
+      ]);
+    }
+  });
+
+  test("ChatGPT Luna models stop at max", () => {
+    for (const model of ["gpt-6-luna", "gpt-5.6-luna"]) {
+      const efforts = supportedEfforts(model, undefined, true);
+      expect(efforts).toEqual(["low", "medium", "high", "xhigh", "max"]);
+      expect(efforts).not.toContain("ultra");
+    }
+  });
+
+  test("every built-in ChatGPT model has an assigned effort ladder", () => {
+    // A model added to the fallback list without a ladder would silently
+    // stop at xhigh. A model the backend really serves only up to xhigh
+    // belongs here, so adding one is a decision rather than an accident.
+    const servedOnlyUpToXhigh: readonly string[] = [];
+    for (const model of CODEX_DEFAULT_MODELS) {
+      if (servedOnlyUpToXhigh.includes(model)) continue;
+      expect(supportedEfforts(model, undefined, true)).toContain("max");
+    }
   });
 
   test("the model name alone does not imply codex levels", () => {

src/provider/reasoning-effort.ts

@@ -56,19 +56,30 @@ const FULL_EFFORT_MODELS: readonly string[] = [
   "gpt-5.1-codex-max",
 ];
 
-// Codex backend models take low/medium/high/xhigh — no `minimal`, no `none`.
+// ChatGPT (Codex backend) models not listed below: low/medium/high/xhigh,
+// no `minimal`, no `none`.
 const CODEX_EFFORTS: readonly ReasoningEffort[] = [
   "low",
   "medium",
   "high",
   "xhigh",
 ];
 
-// gpt-5.6 family adds `max` and `ultra`; older Codex models do not support
-// them.
-const MAX_EFFORT_CODEX_MODELS: readonly string[] = [
+// Per-model levels the Codex backend lists (`supported_reasoning_levels` in
+// GET /codex/models, client_version 0.165.0, 2026-10-28). Exact ids rather
+// than a name rule: an id not seen yet gets the conservative fallback above
+// instead of a guessed tier, because the adapter does not clamp and a level
+// the backend does not serve breaks the turn.
+const ULTRA_EFFORT_CODEX_MODELS: readonly string[] = [
+  "gpt-6.1-sol",
+  "gpt-6-astra",
+  "gpt-6.1-terra",
+  "gpt-6-sol",
   "gpt-5.6-sol",
   "gpt-5.6-terra",
+];
+const MAX_EFFORT_CODEX_MODELS: readonly string[] = [
+  "gpt-6-luna",
   "gpt-5.6-luna",
 ];
 
@@ -96,8 +107,8 @@ function isMuseSparkModel(model: string): boolean {
 const GROK_46_AND_47_EFFORTS: readonly ReasoningEffort[] = CODEX_EFFORTS;
 const GROK_46_AND_47_MODELS: readonly string[] = ["grok-4.6", "grok-4.7"];
 
-// GPT-6 Astra accepts low–max on both surfaces; not ultra (gpt-5.6
-// Codex-only), minimal, or none.
+// GPT-6 Astra on the API path accepts low–max; not ultra, minimal, or none.
+// On ChatGPT sign-in it follows the served-levels lists above.
 const GPT6_ASTRA_EFFORTS: readonly ReasoningEffort[] = [
   "low",
   "medium",
@@ -152,14 +163,17 @@ export function supportedEfforts(
   if (providerEfforts.length > 0) {
     return [...providerEfforts];
   }
-  if (model === "gpt-6-astra") {
-    return [...GPT6_ASTRA_EFFORTS];
-  }
   if (isCodex) {
+    if (ULTRA_EFFORT_CODEX_MODELS.includes(model)) {
+      return [...CODEX_EFFORTS, "max", "ultra"];
+    }
     return MAX_EFFORT_CODEX_MODELS.includes(model)
-      ? [...CODEX_EFFORTS, "max", "ultra"]
+      ? [...CODEX_EFFORTS, "max"]
       : [...CODEX_EFFORTS];
   }
+  if (model === "gpt-6-astra") {
+    return [...GPT6_ASTRA_EFFORTS];
+  }
   if (FULL_EFFORT_MODELS.includes(model)) {
     return ["none", ...DEFAULT_EFFORTS, "xhigh"];
   }