Fix MiniMax Coding Plan image input in Combos

docs-site/src/content/docs/guides/minimax.md

@@ -12,6 +12,11 @@ the protocol boundary it actually exposes:
 
 ## MiniMax Code
 
+In the OpenCodex Models page, both MiniMax Coding Plan regions advertise image input for
+`MiniMax-M3` and `MiniMax-M3.1-Flash-Preview`. A Combo's **Image / multimodal** switch is
+available only when every selected model supports image input. The switch controls image
+attachments; it does not enable video uploads.
+
 Install and sign in to MiniMax Code using MiniMax's instructions first. Then start
 OpenCodex and connect the reversible file integration:
 

docs-site/src/content/docs/zh-cn/guides/minimax.md

@@ -10,6 +10,10 @@ MiniMax 发布两种不同的命令行产品。OpenCodex 在它们实际提供
 
 ## MiniMax Code
 
+在 OpenCodex 的模型管理页面，MiniMax Coding Plan 两个区域的 `MiniMax-M3` 和
+`MiniMax-M3.1-Flash-Preview` 均声明支持图片输入。只有所有选定模型均支持图片输入时，
+Combo 的“图片 / 多模态”开关才可用。该开关控制图片附件，不提供视频上传能力。
+
 先按照 MiniMax 的说明安装并登录 MiniMax Code，然后启动 OpenCodex 并连接可逆的文件集成：
 
 ```bash

scripts/test-layout/layout.json

@@ -11,6 +11,7 @@
     "desktop-app-server-shim-launcher.test.ts": "clients",
     "desktop-chatgpt-config.test.ts": "clients",
     "owner-registry-acl.test.ts": "config",
+    "minimax-combo-image-input.test.ts": "providers",
     "claude-devin-output-order.test.ts": "claude-integration",
     "codex-config-preservation.test.ts": "codex-integration",
     "codex-credits.test.ts": "codex-integration",

src/providers/registry/entries-extended.ts

@@ -35,6 +35,7 @@ import {
   ZAI_GLM_5X_REASONING_EFFORTS,
   MINIMAX_MODELS,
   MINIMAX_MODEL_CONTEXT_WINDOWS,
+  MINIMAX_MODEL_INPUT_MODALITIES,
   MINIMAX_M3_REASONING_EFFORTS,
   MINIMAX_M3_REASONING_EFFORT_MAP,
   MINIMAX_M31_FLASH_PREVIEW,
@@ -1024,6 +1025,7 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     id: "minimax", label: "MiniMax — Coding Plan", baseUrl: "https://api.minimax.io/v1", adapter: "openai-chat", authKind: "key",
     dashboardUrl: "https://platform.minimax.io", defaultModel: "MiniMax-M3", models: MINIMAX_MODELS,
     modelContextWindows: MINIMAX_MODEL_CONTEXT_WINDOWS,
+    modelInputModalities: MINIMAX_MODEL_INPUT_MODALITIES,
     modelReasoningEfforts: { "MiniMax-M3": MINIMAX_M3_REASONING_EFFORTS, [MINIMAX_M31_FLASH_PREVIEW]: MINIMAX_M31_REASONING_EFFORTS },
     modelDefaultReasoningEfforts: { "MiniMax-M3": "medium", [MINIMAX_M31_FLASH_PREVIEW]: MINIMAX_M31_DEFAULT_REASONING_EFFORT },
     modelReasoningEffortMap: { "MiniMax-M3": MINIMAX_M3_REASONING_EFFORT_MAP },
@@ -1049,6 +1051,7 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     id: "minimax-cn", label: "MiniMax — Coding Plan (CN)", baseUrl: "https://api.minimaxi.com/v1", adapter: "openai-chat", authKind: "key",
     dashboardUrl: "https://platform.minimaxi.com", defaultModel: "MiniMax-M3", models: MINIMAX_MODELS,
     modelContextWindows: MINIMAX_MODEL_CONTEXT_WINDOWS,
+    modelInputModalities: MINIMAX_MODEL_INPUT_MODALITIES,
     modelReasoningEfforts: { "MiniMax-M3": MINIMAX_M3_REASONING_EFFORTS, [MINIMAX_M31_FLASH_PREVIEW]: MINIMAX_M31_REASONING_EFFORTS },
     modelDefaultReasoningEfforts: { "MiniMax-M3": "medium", [MINIMAX_M31_FLASH_PREVIEW]: MINIMAX_M31_DEFAULT_REASONING_EFFORT },
     modelReasoningEffortMap: { "MiniMax-M3": MINIMAX_M3_REASONING_EFFORT_MAP },

src/providers/registry/model-seeds.ts

@@ -172,6 +172,11 @@ export const MINIMAX_REASONING_SPLIT_MODELS = MINIMAX_MODELS_BEFORE_M31;
 export const MINIMAX_MODEL_CONTEXT_WINDOWS: Record<string, number> = Object.fromEntries(
   MINIMAX_MODELS.map(id => [id, id === "MiniMax-M3" || id === MINIMAX_M31_FLASH_PREVIEW ? 1_000_000 : 204_800]),
 );
+/** MiniMax's OpenAI-compatible M3 endpoints accept image_url; keep video off Codex's input enum. */
+export const MINIMAX_MODEL_INPUT_MODALITIES: Record<string, string[]> = {
+  "MiniMax-M3": ["text", "image"],
+  [MINIMAX_M31_FLASH_PREVIEW]: ["text", "image"],
+};
 export const MINIMAX_M3_REASONING_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
 /** Identity efforts on the wire; no map, so none omits the field instead of disabling thinking. */
 export const MINIMAX_M31_REASONING_EFFORTS = ["low", "medium", "high", "xhigh", "max"];

structure/providers-and-adapters.md

@@ -354,10 +354,9 @@ axis that outranks every source here, so it is where a deliberate text-only over
 (`ocx provider edit <provider> --model <id> --text-only` writes it) and the one declaration a
 restart cannot take back.
 
-Roster additions share the blind spot when the vendor's `/models` omits the new id (MiniMax-M3.1-Flash-Preview):
-`src/providers/stale-model-roster-migration.ts` replaces a saved roster only while it is byte-for-byte the previous
-seed, filling the added id's window and default effort only inside records the row already has, in the same startup
-pass; `CALLABLE_CONFIGURED_COMPATIBILITY_MODELS` (`src/codex/catalog/model-hints.ts`) keeps it in the live catalog.
+When `/models` omits MiniMax-M3.1-Flash-Preview, `src/providers/stale-model-roster-migration.ts` replaces only a saved roster identical to the old seed and fills existing window and default-effort records during startup; `CALLABLE_CONFIGURED_COMPATIBILITY_MODELS` (`src/codex/catalog/model-hints.ts`) retains it in the live catalog.
+
+Both MiniMax Coding Plan presets declare `text` and `image` for M3 and M3.1 in `src/providers/registry/model-seeds.ts`; registry enrichment fills missing saved declarations, so id-only live rows enable the Combo image switch when every target supports images. Explicit overrides still win; video is outside Codex's catalog input enum.
 
 The BigModel Coding Plan Responses preset uses the separately documented
 `https://open.bigmodel.cn/api/v1` transport and a static catalog. Its provider row

tests/fixtures/test-layout-expected.json

@@ -8,6 +8,7 @@
   "desktop-app-server-shim-launcher.test.ts": "clients",
   "desktop-chatgpt-config.test.ts": "clients",
   "owner-registry-acl.test.ts": "config",
+  "minimax-combo-image-input.test.ts": "providers",
   "claude-devin-output-order.test.ts": "claude-integration",
   "codex-config-preservation.test.ts": "codex-integration",
     "codex-credits.test.ts": "codex-integration",

tests/providers/minimax-combo-image-input.test.ts

@@ -0,0 +1,101 @@
+import { afterEach, describe, expect, test } from "bun:test";
+import { comboImagesSupported } from "../../gui/src/combo-capabilities";
+import { applyProviderConfigHints, deriveComboCatalogModel, gatherRoutedModels } from "../../src/codex/catalog";
+import { clearModelCache } from "../../src/codex/model-cache";
+import { enrichProviderFromRegistry, providerConfigSeed } from "../../src/providers/derive";
+import { getProviderRegistryEntry } from "../../src/providers/registry";
+import { listManagementModelRows } from "../../src/server/management/model-rows";
+import { isModelVisionSidecarConsumer } from "../../src/vision/eligibility";
+import type { CatalogModel, OcxConfig, OcxProviderConfig } from "../../src/types";
+
+const M3 = "MiniMax-M3";
+const PREVIEW = "MiniMax-M3.1-Flash-Preview";
+const REGIONS = ["minimax", "minimax-cn"] as const;
+const originalFetch = globalThis.fetch;
+
+function seeded(region: typeof REGIONS[number]): OcxProviderConfig {
+  const entry = getProviderRegistryEntry(region);
+  if (!entry) throw new Error(`missing ${region} registry entry`);
+  return { ...providerConfigSeed(entry), apiKey: "test-key" };
+}
+
+function config(region: typeof REGIONS[number], provider = seeded(region)): OcxConfig {
+  return { port: 10100, defaultProvider: region, providers: { [region]: provider } } as OcxConfig;
+}
+
+afterEach(() => {
+  globalThis.fetch = originalFetch;
+  clearModelCache();
+});
+
+describe("MiniMax Coding Plan Combo image input", () => {
+  test("both presets seed M3 image input and enrich missing saved declarations without replacing overrides", () => {
+    for (const region of REGIONS) {
+      const fresh = seeded(region);
+      expect(fresh.modelInputModalities?.[M3]).toEqual(["text", "image"]);
+      expect(fresh.modelInputModalities?.[PREVIEW]).toEqual(["text", "image"]);
+      expect(fresh.modelInputModalities?.["MiniMax-M2.7"]).toBeUndefined();
+
+      const saved: OcxProviderConfig = {
+        adapter: "openai-chat", baseUrl: fresh.baseUrl,
+        modelInputModalities: { "MiniMax-M2.7": ["text"] },
+        modelCapabilities: { [M3]: { inputModalities: ["text"] } },
+      };
+      enrichProviderFromRegistry(region, saved);
+      expect(saved.modelInputModalities).toMatchObject({
+        [M3]: ["text", "image"], [PREVIEW]: ["text", "image"], "MiniMax-M2.7": ["text"],
+      });
+      expect(saved.modelCapabilities?.[M3]?.inputModalities).toEqual(["text"]);
+      expect(isModelVisionSidecarConsumer(saved, M3)).toBe(true);
+      expect(applyProviderConfigHints(region, saved, { provider: region, id: M3 }).inputModalities)
+        .toEqual(["text", "image"]); // explicit text-only uses the existing vision sidecar
+      expect(isModelVisionSidecarConsumer(fresh, M3)).toBe(false);
+    }
+  });
+
+  test("id-only discovery and a warm cache yield management rows accepted by the Combo picker", async () => {
+    globalThis.fetch = (async (input: RequestInfo | URL) => {
+      if (!String(input).includes("/models")) throw new Error("unexpected request");
+      return Response.json({ data: [{ id: M3 }] });
+    }) as typeof fetch;
+
+    for (const region of REGIONS) {
+      const live = config(region);
+      const first = await gatherRoutedModels(live);
+      const cached = await gatherRoutedModels(live);
+      for (const models of [first, cached]) {
+        const m3 = models.find(model => model.provider === region && model.id === M3);
+        expect(m3?.inputModalities).toEqual(["text", "image"]);
+        const rows = await listManagementModelRows(live, { models });
+        const row = rows.find(model => model.provider === region && model.id === M3);
+        expect(row?.inputModalities).toEqual(["text", "image"]);
+        expect(comboImagesSupported([{ provider: region, model: M3 }], rows)).toBe(true);
+        expect(comboImagesSupported(
+          [{ provider: region, model: M3 }, { provider: "other", model: "vision" }],
+          [...rows, { provider: "other", id: "vision", inputModalities: ["text", "image"] }],
+        )).toBe(true);
+        expect(comboImagesSupported(
+          [{ provider: region, model: M3 }, { provider: "other", model: "blind" }],
+          [...rows, { provider: "other", id: "blind", inputModalities: ["text"] }],
+        )).toBe(false);
+      }
+    }
+  });
+
+  test("static fallback and Combo catalog retain image unless the operator disables it", async () => {
+    for (const region of REGIONS) {
+      const provider = seeded(region);
+      provider.liveModels = false;
+      const models = await gatherRoutedModels(config(region, provider));
+      const member = models.find(model => model.provider === region && model.id === M3);
+      expect(member?.inputModalities).toEqual(["text", "image"]);
+      const combo = (imageInput: "auto" | "disabled") => ({
+        targets: [{ provider: region, model: M3 }], imageInput, defaultEffort: "medium",
+      }) as never;
+      expect(deriveComboCatalogModel("minimax-image", combo("auto"), [member as CatalogModel])?.inputModalities)
+        .toContain("image");
+      expect(deriveComboCatalogModel("minimax-image", combo("disabled"), [member as CatalogModel])?.inputModalities)
+        .toEqual(["text"]);
+    }
+  });
+});