fix(zen): strip images for the Zen models that reject them (#1043)

src/providers/registry.ts

@@ -348,6 +348,34 @@ const THINKING_BUDGET_MODELS = [
 const OPENCODE_GO_THINKING_BUDGET_MODELS = ["qwen3.5-plus", "qwen3.6-plus", "qwen3.7-max", "qwen3.7-plus"];
 const DEEPSEEK_THINKING_MODELS = ["deepseek-v4-pro", "deepseek-v4-flash"];
 const OPENCODE_FREE_DEEPSEEK_MODELS = ["deepseek-v4-flash-free"];
+/*
+ * Zen free models that reject `image_url` upstream (#1043, and the reproducible
+ * half of #1024).
+ *
+ * Zen publishes NO modality metadata — its `/v1/models` returns only id, object,
+ * created, owned_by — so this list is measured, not derived. Each id was probed
+ * once against https://opencode.ai/zen/v1 on 2026-08-05 with a text control first
+ * and then a 1x1 PNG; the six below failed the image request, four of them with
+ * `[404] No endpoints found that support image input` and `big-pickle` with the
+ * exact deserialize error quoted in #1043.
+ *
+ * `mimo-v2.5-free` and `longcat-2.0-free` ACCEPT images and are deliberately
+ * absent. Adding them would silently replace a working image with a caption,
+ * which is worse than the loud 400 this list exists to prevent — see the negative
+ * assertion in tests/provider-registry-parity.test.ts.
+ *
+ * Zen's roster is discovered live while this list is static, so it is a dated
+ * exception list, not a capability model. Re-probe before extending it.
+ * Evidence: devlog/_plan/260805_bug_fix_stack/002_zen_modality_probe.md
+ */
+const OPENCODE_ZEN_TEXT_ONLY_MODELS = [
+  "big-pickle",
+  "nemotron-3-ultra-free",
+  "ling-3.0-flash-free",
+  "north-mini-code-free",
+  "laguna-s-2.1-free",
+  "deepseek-v4-flash-free",
+];
 /*
  * DeepSeek's Codex ladder is low/high/max, and the two V4 models resolve it
  * DIFFERENTLY. From the official thinking-mode table (api-docs.deepseek.com,
@@ -1694,7 +1722,17 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     autoToolChoiceOnlyModels: KIMI_AUTO_TOOL_CHOICE_ONLY_MODELS,
     preserveReasoningContentModels: KIMI_THINKING_MODELS,
   },
-  { id: "opencode-zen", label: "opencode zen", baseUrl: "https://opencode.ai/zen/v1", adapter: "openai-chat", authKind: "key", dashboardUrl: "https://opencode.ai/auth" },
+  {
+    id: "opencode-zen",
+    label: "opencode zen",
+    baseUrl: "https://opencode.ai/zen/v1",
+    adapter: "openai-chat",
+    authKind: "key",
+    dashboardUrl: "https://opencode.ai/auth",
+    // #1043: without this the proxy forwards image parts to text-only Zen models and
+    // the upstream rejects the whole request with a 400.
+    noVisionModels: OPENCODE_ZEN_TEXT_ONLY_MODELS,
+  },
   { id: "vercel-ai-gateway", label: "Vercel AI Gateway", baseUrl: "https://ai-gateway.vercel.sh/v1", adapter: "openai-chat", authKind: "key", dashboardUrl: "https://vercel.com/dashboard" },
   {
     id: "opencode-free",
@@ -1713,7 +1751,9 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     modelReasoningEfforts: Object.fromEntries(OPENCODE_FREE_DEEPSEEK_MODELS.map(id => [id, deepseekThinkingEffortsFor(id)])),
     modelReasoningEffortMap: Object.fromEntries(OPENCODE_FREE_DEEPSEEK_MODELS.map(id => [id, deepseekReasoningMapFor(id)])),
     preserveReasoningContentModels: OPENCODE_FREE_DEEPSEEK_MODELS,
-    noVisionModels: OPENCODE_FREE_DEEPSEEK_MODELS,
+    // Same Zen roster behind the same base URL, so it carries the same measured
+    // text-only list rather than only its DeepSeek member (#1043).
+    noVisionModels: OPENCODE_ZEN_TEXT_ONLY_MODELS,
   },
   { id: "xiaomi", label: "Xiaomi MiMo", baseUrl: "https://api.xiaomimimo.com/anthropic", adapter: "anthropic", authKind: "key", dashboardUrl: "https://xiaomimimo.com", defaultModel: "mimo-v2.5-pro" },
   { id: "kilo", label: "Kilo", baseUrl: "https://api.kilo.ai/api/gateway", adapter: "openai-chat", authKind: "key", dashboardUrl: "https://kilo.ai" },

tests/provider-registry-parity.test.ts

@@ -864,6 +864,47 @@ describe("provider registry parity", () => {
     // so the request still reaches Command Code as `deepseek/deepseek-v4-flash`.
     expect(entries.find(e => e.slug === "commandcode/deepseek-deepseek-v4-flash")).toBeTruthy();
   });
+  /*
+   * #1043. Zen publishes no modality metadata, so the classification below is an
+   * empirical list measured against the live endpoint on 2026-08-05, not something
+   * derived from provider data.
+   *
+   * The negative half is the part that matters. `mimo-v2.5-free` and
+   * `longcat-2.0-free` accept images; listing them would silently swap a working
+   * image for a caption, which is a worse failure than the loud 400 this fixes
+   * because nothing surfaces it. This test exists so a future "classify all the
+   * free models" patch fails here instead of shipping.
+   */
+  test("Zen text-only classification covers the measured models and excludes the vision ones", () => {
+    const measuredTextOnly = [
+      "big-pickle",
+      "nemotron-3-ultra-free",
+      "ling-3.0-flash-free",
+      "north-mini-code-free",
+      "laguna-s-2.1-free",
+      "deepseek-v4-flash-free",
+    ];
+    // Measured as ACCEPTING images. Never add these to a noVisionModels list.
+    const measuredVisionCapable = ["mimo-v2.5-free", "longcat-2.0-free"];
+
+    for (const providerId of ["opencode-zen", "opencode-free"]) {
+      const entry = PROVIDER_REGISTRY.find(p => p.id === providerId);
+      expect(entry, `registry entry ${providerId} is missing`).toBeTruthy();
+      expect(entry?.baseUrl, `${providerId} should serve the Zen roster`)
+        .toBe("https://opencode.ai/zen/v1");
+
+      const listed = entry?.noVisionModels ?? [];
+      for (const model of measuredTextOnly) {
+        expect(listed, `${providerId} must strip images for text-only ${model}`)
+          .toContain(model);
+      }
+      for (const model of measuredVisionCapable) {
+        expect(listed, `${providerId} must NOT strip images for vision-capable ${model}`)
+          .not.toContain(model);
+      }
+    }
+  });
+
 });
 
 describe("free-provider directory isolation", () => {

tests/vision-sidecar-e2e.test.ts

@@ -4,6 +4,7 @@ import { tmpdir } from "node:os";
 import { join } from "node:path";
 import { saveConfig } from "../src/config";
 import { startServer } from "../src/server";
+import { PROVIDER_REGISTRY } from "../src/providers/registry";
 import type { OcxConfig } from "../src/types";
 import { installIsolatedCodexHome, type IsolatedCodexHome } from "./helpers/isolated-codex-home";
 import { fakeChatGptJwt } from "./helpers/fake-chatgpt-jwt";
@@ -195,4 +196,92 @@ describe("vision sidecar fallback (issue #88, end-to-end)", () => {
       await server.stop(true);
     }
   });
+
+  /*
+   * #1043 activation evidence. The registry classification is only useful if the
+   * strip actually fires for a Zen model, so this drives the real path with the
+   * built-in `opencode-zen` list rather than a fixture list, and asserts the
+   * observable effect: the image bytes are gone from the upstream body and the
+   * omission marker is there instead.
+   *
+   * `big-pickle` is the id that reproduced the reported 400 verbatim against the
+   * live endpoint (devlog/_plan/260805_bug_fix_stack/002_zen_modality_probe.md).
+   */
+  test("a text-only Zen model has its image stripped before the upstream request (#1043)", async () => {
+    let upstreamBody = "";
+    upstream = serveUpstream(b => { upstreamBody = b; });
+
+    const zen = PROVIDER_REGISTRY.find(p => p.id === "opencode-zen");
+    expect(zen?.noVisionModels).toContain("big-pickle");
+
+    const config: OcxConfig = {
+      port: 0, hostname: "127.0.0.1", defaultProvider: "zenlike", openaiProviderTierVersion: 2,
+      providers: {
+        // A custom provider carrying the REGISTRY's list verbatim. The built-in
+        // opencode-zen entry pins its own baseUrl, so it cannot be aimed at a local
+        // upstream; what is under test is the classification, which is read from the
+        // registry above rather than written out here.
+        zenlike: {
+          adapter: "openai-chat",
+          baseUrl: `http://127.0.0.1:${upstream.port}/v1`,
+          allowPrivateNetwork: true,
+          apiKey: "key-alpha-000111222333",
+          noVisionModels: zen?.noVisionModels,
+        },
+        openai: { adapter: "openai-responses", authMode: "forward", baseUrl: "https://chatgpt.com/backend-api/codex" },
+      },
+    } as OcxConfig;
+    saveConfig(config);
+    const server = startServer(0);
+    try {
+      const res = await fetch(new URL("/v1/responses", server.url), {
+        method: "POST",
+        headers: { "content-type": "application/json", authorization: "Bearer forward-oauth-token" },
+        body: JSON.stringify(baseRequest("zenlike/big-pickle")),
+      });
+      expect(res.status).toBe(200);
+      // The effect, not merely a 200: no image bytes on the wire, marker present.
+      expect(upstreamBody).not.toContain("aGVsbG8taW1hZ2UtYnl0ZXM=");
+      expect(upstreamBody).toContain("[image omitted");
+    } finally {
+      await server.stop(true);
+    }
+  });
+
+  test("a vision-capable Zen model keeps its image (#1043 negative case)", async () => {
+    let upstreamBody = "";
+    upstream = serveUpstream(b => { upstreamBody = b; });
+
+    const zen = PROVIDER_REGISTRY.find(p => p.id === "opencode-zen");
+    // Measured as accepting images; classifying it would silently degrade it.
+    expect(zen?.noVisionModels).not.toContain("mimo-v2.5-free");
+
+    const config: OcxConfig = {
+      port: 0, hostname: "127.0.0.1", defaultProvider: "zenlike", openaiProviderTierVersion: 2,
+      providers: {
+        zenlike: {
+          adapter: "openai-chat",
+          baseUrl: `http://127.0.0.1:${upstream.port}/v1`,
+          allowPrivateNetwork: true,
+          apiKey: "key-alpha-000111222333",
+          noVisionModels: zen?.noVisionModels,
+        },
+        openai: { adapter: "openai-responses", authMode: "forward", baseUrl: "https://chatgpt.com/backend-api/codex" },
+      },
+    } as OcxConfig;
+    saveConfig(config);
+    const server = startServer(0);
+    try {
+      const res = await fetch(new URL("/v1/responses", server.url), {
+        method: "POST",
+        headers: { "content-type": "application/json", authorization: "Bearer forward-oauth-token" },
+        body: JSON.stringify(baseRequest("zenlike/mimo-v2.5-free")),
+      });
+      expect(res.status).toBe(200);
+      expect(upstreamBody).toContain("aGVsbG8taW1hZ2UtYnl0ZXM=");
+      expect(upstreamBody).not.toContain("[image omitted");
+    } finally {
+      await server.stop(true);
+    }
+  });
 });