feat: PLANNING体系首版（00路线图/02任务总表/03执行协议/I-03/I-04/registry42口径）+M1止血交付

.mimo2codex-audit/REPORT.md

@@ -0,0 +1,62 @@
+# mimo2codex 模型能力实测与配置报告
+
+日期：2026-09-06 ｜ 网关：OpenCode Zen (Console Go) `https://opencode.ai/zen/go/v1` ｜ 代理：mimo2codex v0.5.29 @ `http://127.0.0.1:8788`
+
+## 一、做了什么
+
+对 mimo2codex 中**所有分组**（deepseek / mimo / generic）可达的全部 **35 个上游模型**逐一实测：
+
+- **思考（reasoning）检测**：分别用 `thinking:{type:"enabled"}`、`reasoning_effort:"high"`、两者组合探测，检查响应中的 `reasoning_content` / `reasoning` 字段或 `<think>` 标签。
+- **多模态（vision）检测**：发送 64×64 红色 PNG（data URL），验证模型能否真正说出"红色"。仅返回 200 但回答"没看到图片"或出现 `[Unsupported Image]` 占位符的，判为**不支持**（排除了第一轮的误报：deepseek-v4-pro / hy4-preview / longcat-2.0）。
+
+## 二、实测结果（25 个可用模型）
+
+| 模型 | 思考 | 多模态 | 备注 |
+|---|---|---|---|
+| mimo-v2.5 | ✅ | ✅ | 代理默认模型 |
+| deepseek-v4-pro | ✅ | ❌ | 图片被上游替换为占位符 |
+| deepseek-v4-flash | ✅ | ❌ | 400 text-only |
+| deepseek-v4-flash-vision-exp | ✅ | ✅ | 实验版视觉模型 |
+| glm-5.3 | ✅ | ❌ | 已同步修正 Codex catalog 为 text-only |
+| glm-5.3-flash | ✅ | ✅ | |
+| glm-5.2 / 5.1 / 5 | ✅ | ❌ | |
+| kimi-k3 / k2.7-code / k2.6 | ✅ | ✅ | reasoning 字段风格 |
+| qwen3.8-max / 3.8-flash / 3.7-plus / 3.6-plus / 3.5-plus | ✅ | ✅ | 1×1 图会因尺寸限制被拒，需 ≥64px |
+| qwen3.7-max | ✅ | ❌ | 拒绝 image_url 内容格式 |
+| minimax-m3 | ✅ | ✅ | `<think>` 标签风格；不认 thinking.type=enabled |
+| minimax-m2.5 | ✅ | ❌ | |
+| mimo-v2.5-pro | ✅ | ❌ | 404 no image endpoints |
+| hy3 / hy4-preview | ✅ | ❌ | hy4 图片不透传 |
+| longcat-2.0 | ✅ | ❌ | 声称看不到图片 |
+| omen-alpha | ✅ | ✅ | 仅 thinking 参数可触发 |
+
+**当前不可用（无法实测，保留路由，出错时上游给出明确报错）**：
+gpt-5.6-luna（上游 500）、grok-4.5（503）、grok-4.6（不支持 oa-compat）、kimi-k2.5（无可用供应商）、minimax-m2.7（500）、mimo-v2-pro / mimo-v2-omni（套餐不支持）、hy3-preview（不可用）、muse-spark-1.2/1.3-contributor（地区限制 403）。
+
+## 三、配置变更
+
+1. **新建 `C:\Users\12914\.mimo2codex\providers.json`**：generic 分组（别名 zen）指向 Zen 网关，登记全部 35 个模型并写入实测的 `supportsReasoning` / `supportsImages` / `contextWindow` 标志。代理启动时自动种子进 `data.db` 的 models 表（已验证）。
+   - **修复了核心路由 bug**：此前 `glm-5.3`、`deepseek-v4-pro` 等不在任何分组目录中的模型会被静默改写成 `mimo-v2.5`（chat_logs 有 `gpt-5.6-luna -> mimo-v2.5` 记录为证）。现在全部**按原样透传**（已验证 `glm-5.3 -> glm-5.3`、`deepseek-v4-pro -> deepseek-v4-pro`，均 200）。
+2. **`.env` 追加 `GENERIC_API_KEY`**（与 MIMO_API_KEY 相同的 Zen key）。
+3. **启用多模态兜底**（settings 表 `codex.visionFallbackEnabled=1`，`codex.visionFallbackModel=mimo-v2.5`）。注意：该兜底按代理实现只对内置 mimo 分组生效；generic 分组下非视觉模型的图片请求走"图片落盘 + OCR 提示"的降级路径（已实测）。
+4. **修正 Codex catalog** `C:\Users\12914\.codex\cc-switch-model-catalog.json`：`glm-5.3` 的 `input_modalities` 由 `[text,image]` 改为 `[text]`（与实测一致）。
+5. 内置 deepseek 分组无官方 API key（只有 Zen key），其 4 个模型经 generic 分组照常可用；若日后配置 `DS_API_KEY`，需从 providers.json 移除 deepseek 条目才会切回官方直连。
+
+## 四、端到端验证（全部通过）
+
+| 测试 | 结果 |
+|---|---|
+| `/v1/chat/completions` glm-5.3 | 200，upstream_model=glm-5.3（无改写） |
+| `/v1/chat/completions` deepseek-v4-pro | 200，upstream_model=deepseek-v4-pro |
+| `/v1/responses` mimo-v2.5 + 图片 | 200，回答"红色" |
+| `/v1/responses` mimo-v2.5 文本 | 200，回答"12" |
+| `/v1/responses` kimi-k3 / glm-5.3 + 图片 | 200，图片落盘 `~/.mimo2codex/cache/images/` 并附 OCR 指引 |
+| `/v1/models` | 列出全部 35 个模型 |
+
+代理以独立进程常驻（pid 见端口 8788 监听），不依赖本会话。管理面板：http://127.0.0.1:8788/admin/
+
+## 五、产物文件
+
+- `providers.json` → `C:\Users\12914\.mimo2codex\providers.json`
+- 实测原始数据 → `.mimo2codex-audit/probe-results.json`、`vision-retest.json`
+- 测试脚本 → `.mimo2codex-audit/probe-zen.mjs`、`vision-retest.mjs`、`responses-test.mjs`、`verify-db.mjs`

.mimo2codex-audit/dump-db.mjs

@@ -0,0 +1,30 @@
+import { createRequire } from "node:module";
+const require = createRequire("C:/Users/12914/AppData/Roaming/npm/node_modules/mimo2codex/package.json");
+const Database = require("better-sqlite3");
+
+const db = new Database("C:/Users/12914/.mimo2codex/data.db", { readonly: true, fileMustExist: true });
+
+const providers = db.prepare("SELECT * FROM providers ORDER BY id").all();
+console.log("=== PROVIDERS (分组) ===");
+for (const p of providers) {
+  console.log(JSON.stringify(p));
+}
+
+console.log("\n=== MODELS ===");
+const models = db.prepare("SELECT * FROM models ORDER BY provider_id, sort_order, upstream_id").all();
+for (const m of models) {
+  console.log(JSON.stringify({
+    id: m.id, provider: m.provider_id, upstream_id: m.upstream_id,
+    display_name: m.display_name,
+    images: m.supports_images, reasoning: m.supports_reasoning, web: m.supports_web_search,
+    ctx: m.context_window, builtin: m.is_builtin
+  }));
+}
+
+console.log("\n=== ALIASES ===");
+for (const a of db.prepare("SELECT * FROM model_aliases").all()) console.log(JSON.stringify(a));
+
+console.log("\n=== SETTINGS (keys) ===");
+for (const s of db.prepare("SELECT key FROM settings").all()) console.log(s.key);
+
+db.close();

.mimo2codex-audit/lt-fl-probe.json

@@ -0,0 +1,74 @@
+{
+  "glm-5.3-flash": {
+    "baseline": {
+      "status": 200,
+      "content": "5",
+      "err": null
+    },
+    "enable_thinking": {
+      "status": 200,
+      "reasoning": "reasoning_content",
+      "err": null
+    },
+    "thinking_type": {
+      "status": 200,
+      "reasoning": "reasoning_content",
+      "err": null
+    },
+    "reasoning_effort_high": {
+      "status": 200,
+      "reasoning": "reasoning_content",
+      "err": null
+    },
+    "vision": {
+      "status": 200,
+      "answer": "红色",
+      "err": null
+    },
+    "developer_role": {
+      "status": 400,
+      "content": null,
+      "err": "角色信息不正确"
+    },
+    "max_tokens_131072": {
+      "status": 200,
+      "err": null
+    }
+  },
+  "qwen3.8-flash": {
+    "baseline": {
+      "status": 200,
+      "content": "\n\n5",
+      "err": null
+    },
+    "enable_thinking": {
+      "status": 200,
+      "reasoning": "reasoning",
+      "err": null
+    },
+    "thinking_type": {
+      "status": 400,
+      "reasoning": null,
+      "err": "\"thinking\" is not supported on /v1/chat/completions and was not applied. Use \"reasoning_effort\" (or \"***.effort\") to control thinking."
+    },
+    "reasoning_effort_high": {
+      "status": 400,
+      "reasoning": null,
+      "err": "Unexpected reasoning effort high. Supported types are xhigh (default), medium, and low."
+    },
+    "vision": {
+      "status": 200,
+      "answer": "\n\n红色",
+      "err": null
+    },
+    "developer_role": {
+      "status": 400,
+      "content": null,
+      "err": "Unexpected message role."
+    },
+    "max_tokens_131072": {
+      "status": 200,
+      "err": null
+    }
+  }
+}

.mimo2codex-audit/probe-ltfl.mjs

@@ -0,0 +1,82 @@
+// Probe lt-fl (platform.lantian.pro) models: thinking / vision / developer-role / max_tokens cap.
+import { readFileSync, writeFileSync } from "node:fs";
+
+const KEY = "sk-yXtdmUJNnJQKHPKhZGsDrTgbQtg46choVzfqueBnIyuXMybB";
+const BASE = "https://platform.lantian.pro/v1";
+const B64 = readFileSync("C:/Users/12914/Desktop/vscode/.mimo2codex-audit/red64.png").toString("base64");
+const MODELS = ["glm-5.3-flash", "qwen3.8-flash"];
+
+async function call(model, extra = {}, messages = null, maxTokens = 900) {
+  const body = {
+    model,
+    messages: messages ?? [{ role: "user", content: "只回答一个数字：2+3=?" }],
+    max_tokens: maxTokens,
+    stream: false,
+    ...extra,
+  };
+  const started = Date.now();
+  const res = await fetch(`${BASE}/chat/completions`, {
+    method: "POST",
+    headers: { Authorization: `Bearer ${KEY}`, "Content-Type": "application/json" },
+    body: JSON.stringify(body),
+  });
+  const ms = Date.now() - started;
+  const raw = await res.text();
+  let json = null;
+  try { json = JSON.parse(raw); } catch {}
+  return { status: res.status, ms, json, raw: raw.slice(0, 260) };
+}
+
+function reasoningKind(json) {
+  const m = json?.choices?.[0]?.message;
+  if (!m) return null;
+  if (typeof m.reasoning_content === "string" && m.reasoning_content.trim()) return "reasoning_content";
+  if (typeof m.reasoning === "string" && m.reasoning.trim()) return "reasoning";
+  if (typeof m.content === "string" && /<think>/i.test(m.content)) return "think_tags";
+  return null;
+}
+const err = (r) => r.json?.error?.message?.slice(0, 180) ?? (r.status !== 200 ? r.raw : null);
+
+const results = {};
+for (const model of MODELS) {
+  const out = {};
+  // baseline
+  const p0 = await call(model);
+  out.baseline = { status: p0.status, content: p0.json?.choices?.[0]?.message?.content?.slice(0, 60) ?? null, err: err(p0) };
+  // thinking variants
+  for (const [name, extra] of [
+    ["enable_thinking", { enable_thinking: true }],
+    ["thinking_type", { thinking: { type: "enabled" } }],
+    ["reasoning_effort_high", { reasoning_effort: "high" }],
+  ]) {
+    const r = await call(model, extra);
+    out[name] = { status: r.status, reasoning: r.json ? reasoningKind(r.json) : null, err: err(r) };
+  }
+  // vision
+  const v = await call(model, {}, [{
+    role: "user",
+    content: [
+      { type: "text", text: "这张图片是什么颜色？只回答颜色词。" },
+      { type: "image_url", image_url: { url: `data:image/png;base64,${B64}` } },
+    ],
+  }], 400);
+  out.vision = {
+    status: v.status,
+    answer: v.json?.choices?.[0]?.message?.content?.slice(0, 60) ?? null,
+    err: err(v),
+  };
+  // developer role
+  const d = await call(model, {}, [
+    { role: "developer", content: "You are a helpful assistant." },
+    { role: "user", content: "只回答一个数字：1+1=?" },
+  ], 200);
+  out.developer_role = { status: d.status, content: d.json?.choices?.[0]?.message?.content?.slice(0, 40) ?? null, err: err(d) };
+  // max_tokens ceiling probe
+  const big = await call(model, {}, [{ role: "user", content: "只回答一个数字：9+9=?" }], 131072);
+  out.max_tokens_131072 = { status: big.status, err: err(big) };
+  results[model] = out;
+  writeFileSync("C:/Users/12914/Desktop/vscode/.mimo2codex-audit/lt-fl-probe.json", JSON.stringify(results, null, 2));
+  console.log(`--- ${model} ---`);
+  console.log(JSON.stringify(out, null, 2));
+}
+console.log("DONE");

.mimo2codex-audit/probe-zen.mjs

@@ -0,0 +1,143 @@
+// Probe every OpenCode Zen model for: (a) thinking/reasoning support, (b) vision (multimodal) support.
+// Run with Node 20:  "C:\Program Files\nodejs\node.exe" probe-zen.mjs
+import { readFileSync, writeFileSync } from "node:fs";
+
+const envText = readFileSync("C:/Users/12914/.mimo2codex/.env", "utf8");
+const KEY = envText.match(/^MIMO_API_KEY=(.+)$/m)?.[1]?.trim();
+const BASE = "https://opencode.ai/zen/go/v1";
+
+if (!KEY) { console.error("no key"); process.exit(1); }
+
+const TINY_PNG = "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mP8z8BQDwAEhQGAhKmMIQAAAABJRU5ErkJggg==";
+
+async function call(body, timeoutMs = 90000) {
+  const ctrl = new AbortController();
+  const t = setTimeout(() => ctrl.abort(), timeoutMs);
+  const started = Date.now();
+  try {
+    const res = await fetch(`${BASE}/chat/completions`, {
+      method: "POST",
+      headers: { Authorization: `Bearer ${KEY}`, "Content-Type": "application/json" },
+      body: JSON.stringify(body),
+      signal: ctrl.signal,
+    });
+    const ms = Date.now() - started;
+    let json = null, text = null;
+    const raw = await res.text();
+    try { json = JSON.parse(raw); } catch { text = raw.slice(0, 300); }
+    return { status: res.status, ms, json, text };
+  } catch (e) {
+    return { status: 0, ms: Date.now() - started, error: String(e?.message || e) };
+  } finally { clearTimeout(t); }
+}
+
+function extractReasoning(json) {
+  try {
+    const msg = json?.choices?.[0]?.message;
+    if (!msg) return null;
+    if (typeof msg.reasoning_content === "string" && msg.reasoning_content.trim()) return "reasoning_content";
+    if (typeof msg.reasoning === "string" && msg.reasoning.trim()) return "reasoning";
+    if (Array.isArray(msg.reasoning_content) && msg.reasoning_content.length) return "reasoning_content(array)";
+    if (typeof msg.content === "string" && /<think>/i.test(msg.content)) return "think_tags";
+    return null;
+  } catch { return null; }
+}
+
+async function probeBaseline(model) {
+  const r = await call({
+    model,
+    messages: [{ role: "user", content: "只回答一个数字：1+1=?" }],
+    max_tokens: 512,
+    stream: false,
+  });
+  return r;
+}
+
+// Try several common "enable thinking" conventions; any success that yields
+// reasoning output means the model supports thinking.
+async function probeThinking(model) {
+  const variants = [
+    { name: "thinking+effort", body: { thinking: { type: "enabled" }, reasoning_effort: "high" } },
+    { name: "effort-only", body: { reasoning_effort: "high" } },
+    { name: "thinking-only", body: { thinking: { type: "enabled" } } },
+  ];
+  const detail = [];
+  for (const v of variants) {
+    const r = await call({
+      model,
+      messages: [{ role: "user", content: "只回答一个数字：2+2=?" }],
+      max_tokens: 900,
+      stream: false,
+      ...v.body,
+    });
+    const reason = r.json ? extractReasoning(r.json) : null;
+    detail.push({
+      variant: v.name,
+      status: r.status,
+      ms: r.ms,
+      reasoning: reason,
+      err: r.json?.error?.message?.slice(0, 200) ?? r.error ?? r.text ?? null,
+      contentPreview: r.json?.choices?.[0]?.message?.content?.slice(0, 80) ?? null,
+    });
+    if (reason) return { supported: true, via: v.name, reason, detail };
+    // If this variant 400'd on unknown field, try the next variant.
+  }
+  return { supported: false, via: null, reason: null, detail };
+}
+
+async function probeVision(model) {
+  const r = await call({
+    model,
+    messages: [{
+      role: "user",
+      content: [
+        { type: "text", text: "这是什么颜色的图片？只回答颜色词。" },
+        { type: "image_url", image_url: { url: `data:image/png;base64,${TINY_PNG}` } },
+      ],
+    }],
+    max_tokens: 300,
+    stream: false,
+  });
+  const ok = r.status === 200 && r.json?.choices?.[0]?.message;
+  return {
+    supported: !!ok,
+    status: r.status,
+    ms: r.ms,
+    answer: ok ? String(r.json.choices[0].message.content).slice(0, 80) : null,
+    err: r.json?.error?.message?.slice(0, 200) ?? r.error ?? r.text ?? null,
+  };
+}
+
+async function main() {
+  // 1. list models
+  const lr = await fetch(`${BASE}/models`, { headers: { Authorization: `Bearer ${KEY}` } });
+  const lj = await lr.json();
+  const models = (lj?.data ?? []).map((m) => m.id).sort();
+  console.log(`models: ${models.length}`);
+
+  const results = {};
+  const CONC = 4;
+  const queue = [...models];
+  async function worker(id) {
+    while (queue.length) {
+      const m = queue.shift();
+      if (!m) break;
+      process.stdout.write(`[${m}] probing...\n`);
+      const base = await probeBaseline(m);
+      const think = await probeThinking(m);
+      const vision = await probeVision(m);
+      results[m] = {
+        baseline: { status: base.status, ms: base.ms, content: base.json?.choices?.[0]?.message?.content?.slice(0, 60) ?? null, err: base.json?.error?.message?.slice(0, 150) ?? base.error ?? null },
+        thinking: think,
+        vision,
+      };
+      process.stdout.write(`[${m}] think=${think.supported ? think.reason : false} vision=${vision.supported}\n`);
+      writeFileSync("C:/Users/12914/Desktop/vscode/.mimo2codex-audit/probe-results.json", JSON.stringify(results, null, 2));
+    }
+  }
+  await Promise.all(Array.from({ length: CONC }, (_, i) => worker(i)));
+  writeFileSync("C:/Users/12914/Desktop/vscode/.mimo2codex-audit/probe-results.json", JSON.stringify(results, null, 2));
+  console.log("DONE");
+}
+
+main();

.mimo2codex-audit/responses-test.mjs

@@ -0,0 +1,43 @@
+// Responses API end-to-end test through the proxy (UTF-8 safe).
+import { readFileSync } from "node:fs";
+const B64 = readFileSync("C:/Users/12914/Desktop/vscode/.mimo2codex-audit/red64.png").toString("base64");
+
+async function responsesTest(model, content) {
+  const item = typeof content === "string"
+    ? { type: "message", role: "user", content }
+    : { type: "message", role: "user", content };
+  const res = await fetch("http://127.0.0.1:8788/v1/responses", {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ model, stream: false, input: [item] }),
+  });
+  const j = await res.json().catch(() => null);
+  const msg = j?.output?.find((o) => o.type === "message");
+  const text = msg?.content?.find((c) => c.type === "output_text")?.text;
+  return { status: res.status, text: text?.slice(0, 160), err: j?.error?.message?.slice(0, 200) };
+}
+
+// 1. multimodal to mimo-v2.5 (vision-capable, Responses wire)
+const r1 = await responsesTest("mimo-v2.5", [
+  { type: "input_text", text: "这张图片是什么颜色？只回答颜色词。" },
+  { type: "input_image", image_url: `data:image/png;base64,${B64}` },
+]);
+console.log("mimo-v2.5 + image  :", JSON.stringify(r1));
+
+// 2. plain text to mimo-v2.5
+const r2 = await responsesTest("mimo-v2.5", "只回答一个数字：6+6=?");
+console.log("mimo-v2.5 text     :", JSON.stringify(r2));
+
+// 3. image to kimi-k3 (vision-capable upstream but translator strips for non-mimo) — expect OCR note
+const r3 = await responsesTest("kimi-k3", [
+  { type: "input_text", text: "这张图片是什么颜色？" },
+  { type: "input_image", image_url: `data:image/png;base64,${B64}` },
+]);
+console.log("kimi-k3 + image    :", JSON.stringify(r3));
+
+// 4. image to glm-5.3 (non-vision) — expect OCR strip note
+const r4 = await responsesTest("glm-5.3", [
+  { type: "input_text", text: "这张图片是什么颜色？" },
+  { type: "input_image", image_url: `data:image/png;base64,${B64}` },
+]);
+console.log("glm-5.3 + image    :", JSON.stringify(r4));

.mimo2codex-audit/set-vision-fallback.mjs

@@ -0,0 +1,15 @@
+// Enable multimodal vision fallback: image requests to non-vision models are
+// rerouted to mimo-v2.5 (verified vision-capable). Same settings table the admin UI writes.
+import { createRequire } from "node:module";
+const require = createRequire("C:/Users/12914/AppData/Roaming/npm/node_modules/mimo2codex/package.json");
+const Database = require("better-sqlite3");
+
+const db = new Database("C:/Users/12914/.mimo2codex/data.db");
+const up = db.prepare("INSERT INTO settings (key, value, updated_at) VALUES (?, ?, ?) ON CONFLICT(key) DO UPDATE SET value = excluded.value, updated_at = excluded.updated_at");
+const now = Date.now();
+up.run("codex.visionFallbackEnabled", "1", now);
+up.run("codex.visionFallbackModel", "mimo-v2.5", now);
+for (const row of db.prepare("SELECT key, value FROM settings WHERE key LIKE 'codex.%'").all()) {
+  console.log(row.key, "=", row.value);
+}
+db.close();

.mimo2codex-audit/verify-db.mjs

@@ -0,0 +1,15 @@
+import { createRequire } from "node:module";
+const require = createRequire("C:/Users/12914/AppData/Roaming/npm/node_modules/mimo2codex/package.json");
+const Database = require("better-sqlite3");
+const db = new Database("C:/Users/12914/.mimo2codex/data.db", { readonly: true });
+
+console.log("=== last 6 chat_logs ===");
+for (const l of db.prepare("SELECT ts, provider_id, client_model, upstream_model, status_code, error_code FROM chat_logs ORDER BY id DESC LIMIT 6").all()) {
+  console.log(new Date(l.ts).toISOString().slice(11, 19), l.provider_id, `${l.client_model} -> ${l.upstream_model}`, l.status_code, l.error_code ?? "");
+}
+
+console.log("\n=== models table by provider ===");
+for (const m of db.prepare("SELECT provider_id, upstream_id, supports_images AS img, supports_reasoning AS r, context_window AS ctx FROM models ORDER BY provider_id, sort_order").all()) {
+  console.log(`${m.provider_id.padEnd(8)} ${m.upstream_id.padEnd(30)} img=${m.img} reasoning=${m.r} ctx=${m.ctx ?? "-"}`);
+}
+db.close();

.mimo2codex-audit/vision-inspect.mjs

@@ -0,0 +1,27 @@
+import { readFileSync } from "node:fs";
+const envText = readFileSync("C:/Users/12914/.mimo2codex/.env", "utf8");
+const KEY = envText.match(/^MIMO_API_KEY=(.+)$/m)?.[1]?.trim();
+const BASE = "https://opencode.ai/zen/go/v1";
+const B64 = readFileSync("C:/Users/12914/Desktop/vscode/.mimo2codex-audit/red64.png").toString("base64");
+
+for (const m of ["deepseek-v4-pro", "longcat-2.0", "hy4-preview"]) {
+  const res = await fetch(`${BASE}/chat/completions`, {
+    method: "POST",
+    headers: { Authorization: `Bearer ${KEY}`, "Content-Type": "application/json" },
+    body: JSON.stringify({
+      model: m,
+      messages: [{
+        role: "user",
+        content: [
+          { type: "text", text: "这张图片是什么颜色？只回答颜色词。" },
+          { type: "image_url", image_url: { url: `data:image/png;base64,${B64}` } },
+        ],
+      }],
+      max_tokens: 500,
+      stream: false,
+    }),
+  });
+  const j = await res.json().catch(() => null);
+  console.log(`=== ${m} (${res.status}) ===`);
+  console.log(JSON.stringify(j?.choices?.[0]?.message ?? j, null, 2).slice(0, 900));
+}

.mimo2codex-audit/vision-retest.json

@@ -0,0 +1,80 @@
+{
+  "gpt-5.6-luna": {
+    "status": 500,
+    "ok": false,
+    "answer": null,
+    "err": "Internal server error"
+  },
+  "deepseek-v4-pro": {
+    "status": 200,
+    "ok": true,
+    "answer": null,
+    "err": null
+  },
+  "deepseek-v4-flash": {
+    "status": 400,
+    "ok": false,
+    "answer": null,
+    "err": "Error from provider (Console Go): Upstream request failed: [400] Model only supports text input; received unsupported content type 'image_url'."
+  },
+  "longcat-2.0": {
+    "status": 200,
+    "ok": true,
+    "answer": null,
+    "err": null
+  },
+  "hy4-preview": {
+    "status": 200,
+    "ok": true,
+    "answer": null,
+    "err": null
+  },
+  "qwen3.5-plus": {
+    "status": 200,
+    "ok": true,
+    "answer": "红色",
+    "err": null
+  },
+  "qwen3.6-plus": {
+    "status": 200,
+    "ok": true,
+    "answer": "红色",
+    "err": null
+  },
+  "qwen3.7-max": {
+    "status": 400,
+    "ok": false,
+    "answer": null,
+    "err": "Error from provider (Console Go): Upstream request failed: [invalid_parameter_error] <400> InternalError.Algo.InvalidParameter: The provided messages input is invalid. The error info is [Unexpected item type in content.]"
+  },
+  "qwen3.7-plus": {
+    "status": 200,
+    "ok": true,
+    "answer": "红色",
+    "err": null
+  },
+  "qwen3.8-flash": {
+    "status": 200,
+    "ok": true,
+    "answer": "红色",
+    "err": null
+  },
+  "qwen3.8-max": {
+    "status": 200,
+    "ok": true,
+    "answer": "红色",
+    "err": null
+  },
+  "glm-5.3": {
+    "status": 400,
+    "ok": false,
+    "answer": null,
+    "err": "Error from provider (Console Go): Upstream request failed: [invalid_request_error] This model does not support image inputs"
+  },
+  "minimax-m2.5": {
+    "status": 400,
+    "ok": false,
+    "answer": null,
+    "err": "Error from provider (Console Go): Upstream request failed: [404] No endpoints found that support image input"
+  }
+}

.mimo2codex-audit/vision-retest.mjs

@@ -0,0 +1,70 @@
+// Retest vision for ambiguous models with a 64x64 red PNG.
+import { readFileSync, writeFileSync } from "node:fs";
+
+const envText = readFileSync("C:/Users/12914/.mimo2codex/.env", "utf8");
+const KEY = envText.match(/^MIMO_API_KEY=(.+)$/m)?.[1]?.trim();
+const BASE = "https://opencode.ai/zen/go/v1";
+const B64 = readFileSync("C:/Users/12914/Desktop/vscode/.mimo2codex-audit/red64.png").toString("base64");
+
+const MODELS = [
+  "gpt-5.6-luna",       // retry in case 500 was transient
+  "deepseek-v4-pro",
+  "deepseek-v4-flash",
+  "longcat-2.0",
+  "hy4-preview",
+  "qwen3.5-plus",
+  "qwen3.6-plus",
+  "qwen3.7-max",
+  "qwen3.7-plus",
+  "qwen3.8-flash",
+  "qwen3.8-max",
+  "glm-5.3",            // sanity: previously rejected image input
+  "minimax-m2.5",       // sanity: previously rejected image input
+];
+
+async function call(body, timeoutMs = 90000) {
+  const ctrl = new AbortController();
+  const t = setTimeout(() => ctrl.abort(), timeoutMs);
+  try {
+    const res = await fetch(`${BASE}/chat/completions`, {
+      method: "POST",
+      headers: { Authorization: `Bearer ${KEY}`, "Content-Type": "application/json" },
+      body: JSON.stringify(body),
+      signal: ctrl.signal,
+    });
+    const raw = await res.text();
+    let json = null;
+    try { json = JSON.parse(raw); } catch {}
+    return { status: res.status, json, raw: raw.slice(0, 200) };
+  } catch (e) {
+    return { status: 0, json: null, raw: String(e?.message || e) };
+  } finally { clearTimeout(t); }
+}
+
+const results = {};
+for (const m of MODELS) {
+  const r = await call({
+    model: m,
+    messages: [{
+      role: "user",
+      content: [
+        { type: "text", text: "这张图片是什么颜色？只回答颜色词。" },
+        { type: "image_url", image_url: { url: `data:image/png;base64,${B64}` } },
+      ],
+    }],
+    max_tokens: 400,
+    stream: false,
+  });
+  const msg = r.json?.choices?.[0]?.message;
+  const content = typeof msg?.content === "string" ? msg.content : null;
+  const thinkTag = content && /<think>/i.test(content);
+  results[m] = {
+    status: r.status,
+    ok: r.status === 200 && !!msg,
+    answer: content ? content.replace(/<think>[\s\S]*?<\/think>/g, "").trim().slice(0, 100) : null,
+    err: r.json?.error?.message?.slice(0, 220) ?? (r.status !== 200 ? r.raw : null),
+  };
+  console.log(`${m}: ${r.status} ${results[m].ok ? "answer=" + results[m].answer : "err=" + results[m].err}`);
+  writeFileSync("C:/Users/12914/Desktop/vscode/.mimo2codex-audit/vision-retest.json", JSON.stringify(results, null, 2));
+}
+console.log("DONE");