diff --git a/server/src/adapters/codex-models.test.ts b/server/src/adapters/codex-models.test.ts
new file mode 100644
index 0000000000..a20d884f2c
--- /dev/null
+++ b/server/src/adapters/codex-models.test.ts
@@ -0,0 +1,131 @@
+import fs from "node:fs";
+import os from "node:os";
+import path from "node:path";
+
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
+
+import { models as codexFallbackModels } from "@paperclipai/adapter-codex-local";
+
+vi.mock("../config-file.js", () => ({
+  readConfigFile: () => null,
+}));
+
+import { listCodexModels, refreshCodexModels, resetCodexModelsCacheForTests } from "./codex-models.js";
+
+function writeCache(dir: string, payload: unknown): void {
+  fs.writeFileSync(
+    path.join(dir, "models_cache.json"),
+    typeof payload === "string" ? payload : JSON.stringify(payload),
+  );
+}
+
+function cacheEntry(overrides: Record<string, unknown>): Record<string, unknown> {
+  return {
+    slug: "gpt-test",
+    display_name: "GPT Test",
+    visibility: "list",
+    ...overrides,
+  };
+}
+
+describe("codex model discovery via the CLI models cache", () => {
+  let tempHome: string;
+  const originalCodexHome = process.env.CODEX_HOME;
+  const originalOpenAiKey = process.env.OPENAI_API_KEY;
+
+  beforeEach(() => {
+    tempHome = fs.mkdtempSync(path.join(os.tmpdir(), "paperclip-codex-models-"));
+    process.env.CODEX_HOME = tempHome;
+    delete process.env.OPENAI_API_KEY;
+    resetCodexModelsCacheForTests();
+  });
+
+  afterEach(() => {
+    fs.rmSync(tempHome, { recursive: true, force: true });
+    if (originalCodexHome === undefined) delete process.env.CODEX_HOME;
+    else process.env.CODEX_HOME = originalCodexHome;
+    if (originalOpenAiKey === undefined) delete process.env.OPENAI_API_KEY;
+    else process.env.OPENAI_API_KEY = originalOpenAiKey;
+  });
+
+  it("returns the static fallback in its original order when the cache file is missing", async () => {
+    const models = await listCodexModels();
+    expect(models).toEqual(codexFallbackModels);
+  });
+
+  it("lists cache entries with visibility list and keeps internal slugs out", async () => {
+    writeCache(tempHome, {
+      models: [
+        cacheEntry({ slug: "gpt-5.5", display_name: "GPT-5.5" }),
+        cacheEntry({ slug: "gpt-reserve", display_name: "GPT Reserve", visibility: "hide" }),
+        cacheEntry({ slug: "codex-auto-review", visibility: null }),
+      ],
+    });
+
+    const models = await listCodexModels();
+    const ids = models.map((m) => m.id);
+
+    expect(ids).toContain("gpt-5.5");
+    expect(ids).not.toContain("gpt-reserve");
+    expect(ids).not.toContain("codex-auto-review");
+    // Static fallback ids still merge in.
+    for (const fallback of codexFallbackModels) {
+      expect(ids).toContain(fallback.id);
+    }
+    expect(models.find((m) => m.id === "gpt-5.5")?.label).toBe("GPT-5.5");
+  });
+
+  it("falls back to the slug when a cache entry has no display name", async () => {
+    writeCache(tempHome, {
+      models: [cacheEntry({ slug: "o3-mini", display_name: "" })],
+    });
+
+    const models = await listCodexModels();
+    expect(models.find((m) => m.id === "o3-mini")?.label).toBe("o3-mini");
+  });
+
+  it("returns the static fallback in its original order when the cache file is malformed", async () => {
+    writeCache(tempHome, "{not json");
+
+    const models = await refreshCodexModels();
+    expect(models).toEqual(codexFallbackModels);
+  });
+
+  it("returns the static fallback when the cache payload has no models array", async () => {
+    writeCache(tempHome, { fetched_at: "2026-09-09T00:00:00Z" });
+
+    const models = await listCodexModels();
+    expect(models.length).toBe(codexFallbackModels.length);
+  });
+
+  it("ignores cache entries whose slug is missing or empty", async () => {
+    writeCache(tempHome, {
+      models: [
+        cacheEntry({ slug: "", display_name: "Empty" }),
+        { display_name: "No slug", visibility: "list" },
+        cacheEntry({ slug: "gpt-5" }),
+      ],
+    });
+
+    const models = await listCodexModels();
+    const ids = models.map((m) => m.id);
+    expect(ids).toContain("gpt-5");
+    expect(ids).not.toContain("");
+  });
+
+  it("serves repeated listings from the TTL cache and refreshes on demand", async () => {
+    writeCache(tempHome, { models: [cacheEntry({ slug: "gpt-5.5" })] });
+    const first = await listCodexModels();
+    expect(first.map((m) => m.id)).toContain("gpt-5.5");
+
+    // Rewrite the file: an ordinary listing still serves the cached catalog,
+    // while an explicit refresh re-reads the file.
+    writeCache(tempHome, { models: [cacheEntry({ slug: "gpt-6-new" })] });
+    const second = await listCodexModels();
+    expect(second.map((m) => m.id)).toContain("gpt-5.5");
+    expect(second.map((m) => m.id)).not.toContain("gpt-6-new");
+
+    const refreshed = await refreshCodexModels();
+    expect(refreshed.map((m) => m.id)).toContain("gpt-6-new");
+  });
+});
diff --git a/server/src/adapters/codex-models.ts b/server/src/adapters/codex-models.ts
index 872779144f..a896f9328e 100644
--- a/server/src/adapters/codex-models.ts
+++ b/server/src/adapters/codex-models.ts
@@ -1,3 +1,7 @@
+import fs from "node:fs";
+import os from "node:os";
+import path from "node:path";
+
 import type { AdapterModel } from "./types.js";
 import { models as codexFallbackModels } from "@paperclipai/adapter-codex-local";
 import { readConfigFile } from "../config-file.js";
@@ -7,6 +11,7 @@ const OPENAI_MODELS_TIMEOUT_MS = 5000;
 const OPENAI_MODELS_CACHE_TTL_MS = 60_000;
 
 let cached: { keyFingerprint: string; expiresAt: number; models: AdapterModel[] } | null = null;
+let codexCache: { cachePath: string; expiresAt: number; models: AdapterModel[] } | null = null;
 
 function fingerprint(apiKey: string): string {
   return `${apiKey.length}:${apiKey.slice(-6)}`;
@@ -31,6 +36,64 @@ function mergedWithFallback(models: AdapterModel[]): AdapterModel[] {
   ]).sort((a, b) => a.id.localeCompare(b.id, "en", { numeric: true, sensitivity: "base" }));
 }
 
+/**
+ * Read the Codex CLI's own model catalog, `$CODEX_HOME/models_cache.json`
+ * (default `~/.codex/models_cache.json`). The CLI refreshes this file from
+ * the ChatGPT backend during normal use, so it is authoritative for exactly
+ * the ChatGPT-authenticated installs that the OpenAI API-key path cannot
+ * serve. Entries the CLI does not list in its picker (`visibility` other
+ * than "list", e.g. internal slugs) stay out of the catalog.
+ *
+ * Returns an empty list when the file is missing, unreadable, malformed, or
+ * lists nothing usable — the caller merges the static fallback either way.
+ */
+function readCodexModelsCache(options?: { forceRefresh?: boolean }): AdapterModel[] {
+  try {
+    const codexHome = process.env.CODEX_HOME?.trim() || path.join(os.homedir(), ".codex");
+    const cachePath = path.join(codexHome, "models_cache.json");
+    const now = Date.now();
+    if (
+      !options?.forceRefresh
+      && codexCache
+      && codexCache.cachePath === cachePath
+      && codexCache.expiresAt > now
+    ) {
+      return codexCache.models;
+    }
+    const raw = fs.readFileSync(cachePath, "utf8");
+    const parsed: unknown = JSON.parse(raw);
+    if (typeof parsed !== "object" || parsed === null) return [];
+    const entries = (parsed as { models?: unknown }).models;
+    if (!Array.isArray(entries)) return [];
+
+    const models: AdapterModel[] = [];
+    for (const entry of entries) {
+      if (typeof entry !== "object" || entry === null) continue;
+      const { slug, display_name: displayName, visibility } = entry as {
+        slug?: unknown;
+        display_name?: unknown;
+        visibility?: unknown;
+      };
+      if (visibility !== "list") continue;
+      if (typeof slug !== "string" || slug.trim().length === 0) continue;
+      const label =
+        typeof displayName === "string" && displayName.trim().length > 0
+          ? displayName.trim()
+          : slug.trim();
+      models.push({ id: slug.trim(), label });
+    }
+    const catalog = dedupeModels(models);
+    codexCache = {
+      cachePath,
+      expiresAt: now + OPENAI_MODELS_CACHE_TTL_MS,
+      models: catalog,
+    };
+    return catalog;
+  } catch {
+    return [];
+  }
+}
+
 function resolveOpenAiApiKey(): string | null {
   const envKey = process.env.OPENAI_API_KEY?.trim();
   if (envKey) return envKey;
@@ -73,8 +136,18 @@ async function fetchOpenAiModels(apiKey: string): Promise<AdapterModel[]> {
 async function loadCodexModels(options?: { forceRefresh?: boolean }): Promise<AdapterModel[]> {
   const forceRefresh = options?.forceRefresh === true;
   const apiKey = resolveOpenAiApiKey();
+  if (!apiKey) {
+    // ChatGPT-authenticated installs have no API key. Their catalog lives in
+    // the Codex CLI's own models_cache.json, merged over the static fallback
+    // so no id that shipped in a release disappears. Keep the fallback's
+    // original order (no sort) so a missing cache degrades to exactly the
+    // pre-change result.
+    return dedupeModels([
+      ...readCodexModelsCache({ forceRefresh }),
+      ...codexFallbackModels,
+    ]);
+  }
   const fallback = dedupeModels(codexFallbackModels);
-  if (!apiKey) return fallback;
 
   const now = Date.now();
   const keyFingerprint = fingerprint(apiKey);
@@ -110,4 +183,5 @@ export async function refreshCodexModels(): Promise<AdapterModel[]> {
 
 export function resetCodexModelsCacheForTests() {
   cached = null;
+  codexCache = null;
 }