dengyie · Oct 4, 2026 · Oct 4, 2026
diff --git a/.env.example b/.env.example
@@ -18,8 +18,11 @@ CNB_REPO_SLUG=your-org/ai-proxy
 #QUOTA_ORG=your-org
 
 # ---- optional tuning ----
-# Comma-separated model ids advertised on /v1/models (routing is done by CNB).
-#PROXY_MODELS=deepseek-v4-flash,glm-5.3-flash,kimi-k3
+# Fallback model ids for /v1/models, used only until a response reveals the
+# real upstream model name (which is then cached). See src/models.mjs.
+#PROXY_MODELS=deepseek-v4.1-flash,glm-5.3-flash,kimi-k3
+# Where the sniffed-model registry is persisted across workspace restarts.
+#PROXY_MODELS_STATE=models-state.json
 #PROXY_PORT=9001
 # Upstream connect/first-byte timeout (ms)
 #PROXY_UPSTREAM_TIMEOUT_MS=15000

diff --git a/.gitignore b/.gitignore
@@ -3,3 +3,5 @@ node_modules/
 .DS_Store
 __pycache__/
 *.pyc
+models-state.json
+models-state.json.tmp
diff --git a/README.md b/README.md
@@ -100,22 +100,22 @@ Three simple steps, only one API key to configure:
 ```bash
 ANTHROPIC_BASE_URL=https://ai.example.com \
 ANTHROPIC_AUTH_TOKEN=your-proxy-key \
-ANTHROPIC_MODEL=deepseek-v4-flash \
+ANTHROPIC_MODEL=deepseek-v4.1-flash \
 claude -p "Hello! Introduce yourself."
 ```
 
 ### 2. OpenAI Clients & Gateways
 Point BaseURL to `https://ai.example.com/v1` and set API Key to `PROXY_KEY`:
 - **Chat Apps**: NextChat, LobeChat, Cherry Studio, Open WebUI
-- **Gateways**: New API, One API (Channel type: OpenAI, model: `deepseek-v4-flash`)
+- **Gateways**: New API, One API (Channel type: OpenAI, model: `deepseek-v4.1-flash`)
 - **Coding Agents**: Cursor, Continue, Codex CLI
 
 ### 3. curl Verification
 ```bash
 curl https://ai.example.com/v1/chat/completions \
   -H "Authorization: Bearer your-proxy-key" \
   -H "Content-Type: application/json" \
-  -d '{"model":"deepseek-v4-flash","messages":[{"role":"user","content":"hi"}],"stream":false}'
+  -d '{"model":"deepseek-v4.1-flash","messages":[{"role":"user","content":"hi"}],"stream":false}'
 ```
 
 ---
@@ -131,7 +131,7 @@ Real production billing data measured over months of live operation (not marketi
 | **Compute Core-Hours** | **1,600 core-hours** (shared dev + CI pool) | A `runner.cpus: 2` workspace running 24/7 for 30 days costs **1,440 core-hours** (safely within the 1,600 pool). No need to stop the proxy. |
 
 > 💡 **The Math Behind 200M Free Tokens/Month**:
-> 1. **Base Upstream Rate**: Live measurements on `deepseek-v4-flash` show an uncached 9,000-token request costs ~0.39 credit (approx **23,000 tokens / credit**).
+> 1. **Base Upstream Rate**: Live measurements on `deepseek-v4.1-flash` show an uncached 9,000-token request costs ~0.39 credit (approx **23,000 tokens / credit**).
 > 2. **Official Prompt Cache 30x Discount**: Once prompt prefixes hit CNB's prompt cache, cost drops to **~0.01 credit**, slashing cost to **1/30** of base price (~**700,000 tokens / credit**).
 > 3. **Real-World Scenarios**: In coding assistants, Claude Code, and multi-turn Agent conversations with long system prompts and history reuse, cache hit rates typically sit at 80%~95% (assuming a realistic **90% cache hit rate**):
 >    $$\text{Blended Cost} = 10\% \times 1 + 90\% \times \frac{1}{30} \approx 13\% \text{ base cost} \implies \approx 177,000\text{ tokens / credit}$$
@@ -171,7 +171,8 @@ npm run quota                 # or: npx cnb2api-quota
 | `PROXY_KEY` | — (required) | Bearer key / x-api-key sent by clients. Refuses to start if missing. |
 | `CNB_TOKEN` | — (required) | Upstream token injected by CNB pipeline environment. |
 | `CNB_REPO_SLUG` | (built-in) | `org/repo` used for upstream routing. Auto-filled by CNB. |
-| `PROXY_MODELS` | `deepseek-v4-flash,glm-5.3-flash,kimi-k3` | Model IDs advertised on `/v1/models`. |
+| `PROXY_MODELS` | `deepseek-v4.1-flash,glm-5.3-flash,kimi-k3` | **Fallback** model ids for `/v1/models` — used only until a response reveals the real upstream model name. |
+| `PROXY_MODELS_STATE` | `models-state.json` | Where the sniffed-model registry is persisted across workspace restarts. Relative paths resolve against the run dir. |
 | `PROXY_PORT` | `9001` | Listen port. |
 | `PROXY_UPSTREAM_TIMEOUT_MS` | `15000` | Upstream connect / first-byte timeout (ms). |
 | `PROXY_IDLE_TIMEOUT_MS` | `300000` | Per-stream idle watchdog (ms). |
@@ -183,7 +184,7 @@ npm run quota                 # or: npx cnb2api-quota
 ## FAQ
 
 **Q: Why does the model list include glm and kimi when the backend is DeepSeek?**  
-A: The official CNB AI gateway currently routes all models to `deepseek-v4-flash`. Exposing multiple aliases ensures compatibility with various client presets out of the box.
+A: `/v1/models` is **self-discovering**: the CNB gateway has no list endpoint, so the proxy learns the real model name from every completion response (the gateway echoes the resolved id, e.g. `deepseek-v4.1-flash`) and advertises that. `PROXY_MODELS` is only a fallback for the very first requests before anything has been sniffed — it exists so client presets that name glm/kimi keep working out of the box, and you can freeze it to whatever `/v1/models` has been returning.
 
 **Q: Are native Tool Calls / Function Calling supported?**  
 A: Fully supported. In streaming mode, `input_json_delta` chunks are cleanly combined; in non-streaming mode, standard `tool_calls` structures are reassembled.

diff --git a/README.zh-CN.md b/README.zh-CN.md
@@ -100,22 +100,22 @@ https://api.cnb.cool/<org>/<repo>/-/ai/chat/completions (官方 AI 核心端点)
 ```bash
 ANTHROPIC_BASE_URL=https://ai.example.com \
 ANTHROPIC_AUTH_TOKEN=your-proxy-key \
-ANTHROPIC_MODEL=deepseek-v4-flash \
+ANTHROPIC_MODEL=deepseek-v4.1-flash \
 claude -p "你好，请介绍一下你自己"
 ```
 
 ### 2. OpenAI 客户端 / 聚合网关
 将 BaseURL 指向 `https://ai.example.com/v1`，API Key 填入 `PROXY_KEY`：
 - **聊天应用**：NextChat、LobeChat、Cherry Studio、Open WebUI
-- **网关渠道**：New API、One API（渠道类型选择 OpenAI，模型填 `deepseek-v4-flash`）
+- **网关渠道**：New API、One API（渠道类型选择 OpenAI，模型填 `deepseek-v4.1-flash`）
 - **代码插件**：Cursor、Continue、Codex CLI
 
 ### 3. curl 调用验证
 ```bash
 curl https://ai.example.com/v1/chat/completions \
   -H "Authorization: Bearer your-proxy-key" \
   -H "Content-Type: application/json" \
-  -d '{"model":"deepseek-v4-flash","messages":[{"role":"user","content":"hi"}],"stream":false}'
+  -d '{"model":"deepseek-v4.1-flash","messages":[{"role":"user","content":"hi"}],"stream":false}'
 ```
 
 ---
@@ -131,7 +131,7 @@ curl https://ai.example.com/v1/chat/completions \
 | **算力核时** | **1,600 核时**（dev + CI 共享池） | `runner.cpus: 2` 反代工作区整月 30 天 7×24h 不间断运行仅消耗 **1,440 核时**（完全在免费池内），无需关机省核时。 |
 
 > 💡 **2 亿 Tokens/月的实测换算逻辑**：
-> 1. **官方基础定价**：对 `deepseek-v4-flash` 实测，全新未命中缓存的请求约 9,000 tokens 花费 0.39 credit（约 **2.3 万 tokens / credit**）。
+> 1. **官方基础定价**：对 `deepseek-v4.1-flash` 实测，全新未命中缓存的请求约 9,000 tokens 花费 0.39 credit（约 **2.3 万 tokens / credit**）。
 > 2. **官方 Prompt Cache 30倍优惠**：命中 CNB 的 prompt 缓存后降至 **~0.01 credit**，相当于基准价格的 **1/30**（约 **70 万 tokens / credit**）。
 > 3. **综合场景折算**：在 Cursor、Claude Code、Agent 编码等多轮对话和长 System Prompt 场景下，上下文缓存命中率通常在 80%~95%（取典型 **90% 命中率**）：
 >    $$\text{综合成本} = 10\% \times 1 + 90\% \times \frac{1}{30} \approx 13\% \text{ 原价} \implies \approx 17.7\text{ 万 tokens / credit}$$
@@ -171,7 +171,8 @@ npm run quota                 # 或: npx cnb2api-quota
 | `PROXY_KEY` | — (必填) | 客户端访问反代网关的 Bearer Token / x-api-key。 |
 | `CNB_TOKEN` | — (必填) | CNB 内网端点鉴权令牌，流水线环境自动注入。 |
 | `CNB_REPO_SLUG` | (内置) | 当前仓库 `org/repo`，流水线内置自动填充。 |
-| `PROXY_MODELS` | `deepseek-v4-flash,glm-5.3-flash,kimi-k3` | `/v1/models` 暴露的模型列表。 |
+| `PROXY_MODELS` | `deepseek-v4.1-flash,glm-5.3-flash,kimi-k3` | **兜底**模型列表：仅在尚未嗅探到真实上游模型名时用于 `/v1/models`。 |
+| `PROXY_MODELS_STATE` | `models-state.json` | 嗅探到的模型名注册表的持久化路径，跨工作区重启保留。相对路径相对于运行目录解析。 |
 | `PROXY_PORT` | `9001` | 服务监听端口。 |
 | `PROXY_UPSTREAM_TIMEOUT_MS` | `15000` | 上游连接首字节超时 (ms)。 |
 | `PROXY_IDLE_TIMEOUT_MS` | `300000` | 流式事件空闲超时看门狗 (ms)。 |
@@ -183,7 +184,7 @@ npm run quota                 # 或: npx cnb2api-quota
 ## 常见问题 (FAQ)
 
 **Q: 既然上游是 DeepSeek，为什么模型列表有 glm 和 kimi？**  
-A: CNB 官方网关目前在底层统一将请求路由至 `deepseek-v4-flash`。暴露多个模型 ID 仅为了兼容各类客户端的默认预设，后续上游开放多模型时反代无需修改即可支持。
+A: `/v1/models` 是**自发现**的：CNB 网关没有列表接口，所以反代从每一次补全响应里嗅探网关回带的真实模型 id（例如 `deepseek-v4.1-flash`）并对外暴露，且持久化到磁盘跨重启保留。`PROXY_MODELS` 只是首个请求发生前的**兜底**列表——它保证各类客户端默认预设开箱即用，你也可以随时把它固化成 `/v1/models` 实际返回过的值。
 
 **Q: 原生函数/工具调用（Tool Calls）支持怎么样？**  
 A: 完整原生支持。流式模式下支持 `input_json_delta` 增量合并，非流式模式下忠实聚合出标准的 `tool_calls` 结构。

diff --git a/docs/DESIGN.md b/docs/DESIGN.md
@@ -23,6 +23,7 @@ src/
 ├── config.mjs   env loading + fail-fast validation (no default key, ever)
 ├── auth.mjs     Bearer key check (timing-safe) + sliding-window fail limiter
 ├── sse.mjs      SSE event parsing + non-stream aggregation
+├── models.mjs   passive model-name registry (sniffed from responses, disk-persisted)
 ├── proxy.mjs    upstream forwarding: AbortSignal timeout + two-way cancel
 ├── server.mjs   http server + routing (/health, /v1/models, /v1/chat/completions)
 └── log.mjs      structured single-line JSON logs, correlated by reqId
@@ -64,6 +65,19 @@ The non-stream aggregator is a small state machine, not a naive
 - `id`/`model`/`created` → upstream values win
 - comment/heartbeat lines (`: keep-alive`) → ignored
 
+### Model discovery (no list endpoint upstream)
+
+The CNB gateway exposes **no model-list endpoint** (every candidate path 404s),
+so the only trustworthy source of the real model name is the `model` field the
+gateway echoes on every completion. `models.mjs` therefore **sniffs passively**:
+`proxy.mjs` calls `observe(j.model)` on streamed chunks and on the non-stream
+aggregate, and the registry keeps first-seen order, dedups, and atomically
+persists to `PROXY_MODELS_STATE` (`writeFileSync` to `.tmp` + `renameSync`) so a
+workspace hot-restart keeps history. `/v1/models` returns the sniffed names, and
+only falls back to the frozen `PROXY_MODELS` list while nothing has been sniffed
+yet. There is deliberately **no startup probe** — we never spend credits on a
+liveness request. Cost: zero.
+
 ## 4. Self-healing address (keepalive)
 
 Three pieces, no long-lived CNB token in the loop:

diff --git a/docs/SETUP.md b/docs/SETUP.md
@@ -52,10 +52,18 @@ the secrets-repo flow below — delete the inline `env:` line, uncomment
 - A [cnb.cool](https://cnb.cool) account whose org has **AI credits** enabled
   (the "天才程序员" / built-in AI offering). Without it the upstream endpoint
   `https://api.cnb.cool/<org>/<repo>/-/ai/chat/completions` returns errors.
-- Know which models your account exposes. Set `PROXY_MODELS` to that list —
-  the repo default is `deepseek-v4-flash,glm-5.3-flash,kimi-k3` (what our
-  account advertises; the gateway currently routes all three to the same
-  upstream model) and requests naming an unavailable model will be rejected.
+- Know which models your account exposes. `/v1/models` **self-discovers** the
+  real upstream model name: the CNB gateway has no list endpoint, so the proxy
+  sniffs the resolved `model` id echoed on every completion response (currently
+  `deepseek-v4.1-flash`) and advertises that, persisting it across workspace
+  restarts. `PROXY_MODELS` is only the **fallback** used until the first
+  response has been sniffed — the repo default is
+  `deepseek-v4.1-flash,glm-5.3-flash,kimi-k3` (what our account advertises) so
+  client presets keep working before any request has run. As of 2026-10 CNB
+  upgraded the DeepSeek backend from `deepseek-v4-flash` to
+  `deepseek-v4.1-flash`; because the gateway echoes the resolved model id on
+  every response, the proxy picks up future renames automatically — no code or
+  config change needed.
 
 **Quota & compute economics** (measured on our deployment):
 - **Compute**: a workspace pinned at `runner.cpus: 2` burns 48 core-hours/day;
@@ -99,7 +107,7 @@ inlined in the code repo — is the cleaner layout for teams or public forks.
 |---|---|
 | both `imports:` stanzas | `https://cnb.cool/your-org/cnb2api-secrets/-/blob/main/proxy.yml` |
 | vscode `env:` | delete the inline `PROXY_KEY` (it now comes from the secrets repo) |
-| (secrets repo) | `PROXY_MODELS: "model-x,model-y"` if your models differ |
+| (secrets repo) | `PROXY_MODELS: "model-x,model-y"` — fallback list only (optional) |
 
 `REPO` for the keepalive needs no edit either way — it reads the built-in
 `CNB_REPO_SLUG`. Note the keepalive cron lives under the `main:` branch key —

diff --git a/src/config.mjs b/src/config.mjs
@@ -7,9 +7,15 @@ export const config = {
   repo: process.env.CNB_REPO_SLUG || process.env.CNB_BUILD_REPO || '',
   proxyKey: process.env.PROXY_KEY || '',
   upstreamToken: process.env.CNB_TOKEN || '',
-  // Model ids advertised on /v1/models. Actual routing is done by the CNB gateway;
-  // set PROXY_MODELS to whatever your account exposes.
-  models: (process.env.PROXY_MODELS || 'deepseek-v4-flash,glm-5.3-flash,kimi-k3').split(',').map((s) => s.trim()).filter(Boolean),
+  // Model ids advertised on /v1/models when nothing has been sniffed yet.
+  // The gateway has no list endpoint, so once a response reveals the real
+  // upstream model name (e.g. deepseek-v4.1-flash) it wins; this default is the
+  // frozen fallback (see src/models.mjs). Set PROXY_MODELS to whatever your
+  // account exposes.
+  models: (process.env.PROXY_MODELS || 'deepseek-v4.1-flash,glm-5.3-flash,kimi-k3').split(',').map((s) => s.trim()).filter(Boolean),
+  // Where the sniffed-model registry is persisted (survives a workspace hot
+  // restart). Relative paths resolve against the process CWD (the run dir).
+  modelsStatePath: process.env.PROXY_MODELS_STATE || 'models-state.json',
   maxBodyBytes: 4 * 1024 * 1024,
   upstreamTimeoutMs: Number(process.env.PROXY_UPSTREAM_TIMEOUT_MS || 15_000), // connect + first byte
   idleTimeoutMs: Number(process.env.PROXY_IDLE_TIMEOUT_MS || 300_000),        // per-stream idle cap (reset each chunk)

diff --git a/src/models.mjs b/src/models.mjs
@@ -0,0 +1,70 @@
+// 上游模型名嗅探注册表。
+//
+// 背景：CNB 的 AI 网关没有「列模型」接口（/-/ai/models 等一律 404），
+// 但每个 chat/completions 响应都会回显它真正路由到的模型名（如
+// deepseek-v4.1-flash）。因此唯一的真实来源是「读响应里的 model 字段」。
+//
+// 策略（纯被动，零额外请求，不主动测活）：
+//   - 每次转发成功的响应里观察到的 model 名登记进内存注册表（首见顺序去重）；
+//   - 新模型出现时落盘持久化，重启后先加载，避免冷启动丢历史；
+//   - GET /v1/models 优先返回嗅探到的真实名；尚未嗅探到任何名字时回退
+//     config.models（即 PROXY_MODELS，人工固化自 /v1/models 的历史返回值）。
+import fs from 'node:fs';
+import { config } from './config.mjs';
+import { log } from './log.mjs';
+
+/** @type {string[]} 首见顺序的去重上游模型名 */
+let observed = [];
+const seen = new Set();
+let loaded = false;
+
+function load() {
+  loaded = true;
+  try {
+    const raw = fs.readFileSync(config.modelsStatePath, 'utf8');
+    const parsed = JSON.parse(raw);
+    const list = Array.isArray(parsed?.models) ? parsed.models : [];
+    for (const m of list) {
+      if (typeof m === 'string' && m && !seen.has(m)) { seen.add(m); observed.push(m); }
+    }
+    if (observed.length) log.info('-', 'models: loaded state', { count: observed.length, path: config.modelsStatePath });
+  } catch (e) {
+    if (e.code !== 'ENOENT') log.warn('-', 'models: state load failed', { err: String(e).slice(0, 160), path: config.modelsStatePath });
+  }
+}
+
+function persist() {
+  try {
+    const tmp = `${config.modelsStatePath}.tmp`;
+    fs.writeFileSync(tmp, JSON.stringify({ models: observed, updated_at: new Date().toISOString() }, null, 2));
+    fs.renameSync(tmp, config.modelsStatePath); // 原子替换，避免半写文件
+  } catch (e) {
+    // 持久化失败不致命：内存注册表仍可用，只是重启后可能冷启动
+    log.warn('-', 'models: state persist failed', { err: String(e).slice(0, 160), path: config.modelsStatePath });
+  }
+}
+
+// 观察一个上游回显的模型名。首次见到才登记 + 落盘（后续命中直接返回）。
+export function observe(model) {
+  if (typeof model !== 'string') return;
+  const m = model.trim();
+  if (!m || m === 'unknown') return;
+  if (!loaded) load();
+  if (seen.has(m)) return;
+  seen.add(m);
+  observed.push(m);
+  log.info('-', 'models: discovered upstream model', { model: m });
+  persist();
+}
+
+// /v1/models 的模型列表：优先真实嗅探名，尚无样本时回退 PROXY_MODELS。
+export function list() {
+  if (!loaded) load();
+  return observed.length ? [...observed] : [...config.models];
+}
+
+// 快照：供调试/测试观察注册表来源。
+export function snapshot() {
+  if (!loaded) load();
+  return { source: observed.length ? 'upstream' : 'configured', observed: [...observed], fallback: [...config.models] };
+}
diff --git a/src/proxy.mjs b/src/proxy.mjs
@@ -3,6 +3,7 @@ import { config } from './config.mjs';
 import { log } from './log.mjs';
 import { aggregate } from './sse.mjs';
 import { record as recordUsage } from './usage.mjs';
+import { observe as observeModel } from './models.mjs';
 
 export async function forward({ reqId, payload, wantStream, res, transform }) {
   // transform（可选）：Anthropic /v1/messages 适配挂载点。
@@ -77,6 +78,7 @@ export async function forward({ reqId, payload, wantStream, res, transform }) {
       let j;
       try { j = JSON.parse(data); } catch { return; /* 心跳/非 JSON 行忽略 */ }
       if (j.usage) usage = j.usage;
+      if (j.model) observeModel(j.model); // 嗅探上游真实模型名（无 list 接口，只能从响应读）
       if (transform) writeEvents(transform.stream.feed(j));
     };
     const parseLines = (lines) => {
@@ -156,6 +158,7 @@ export async function forward({ reqId, payload, wantStream, res, transform }) {
   }
   clearTimeout(idleTimer);
   const body = aggregate(text);
+  if (body.model) observeModel(body.model); // 嗅探上游真实模型名（非流式聚合结果里的 model）
   recordUsage({ prompt: body.usage.prompt_tokens || 0, completion: body.usage.completion_tokens || 0 });
   log.info(reqId, 'chat done', { ms: Date.now() - started, out: body.usage.completion_tokens });
   if (transform) return { handled: false, status: 200, body: transform.response(body) };