feat: auto-append [1m] to models with a 1M context window, v1.19.0

Claude Code clamps third-party models to ~200k behind a custom base URL and
only opens the full window when the model name ends in [1m]. Users had to
retype `deepseek-flash[1m]` on every launch and after every /model switch.

The suffix is a client-side marker, never an upstream id: native Claude Code
strips it before calling the API, and Kimi answers `k3[1m]` with 401 "model
id does not exist". So it is generated at the two Claude-Code-facing exits
and stripped at the entrance:

  - the four model slots injected by `ccmr claude` (ANTHROPIC_MODEL plus the
    SONNET/OPUS/HAIKU defaults) -- fixes "200k on launch"
  - the ids served by GET /v1/models -- fixes "200k again after /model"
  - ConfigManager.resolveModelName strips [1m] before alias resolution, so
    both `deepseek-flash` and `deepseek-flash[1m]` route, existing configs
    and scripts are unaffected, and upstream always receives the bare id

Threshold is context_window >= 1000000 (vendors report 1M as either 1000000
or 1048576), which covers 27 of 42 variants. Models at 256k and below stay
bare on purpose: a false [1m] would make Claude Code overrun the real window,
which is worse than the original bug.

Verified on a scratch gateway: 27 ids suffixed, 15 bare, and a POST with a
suffixed name reaches upstream as the bare model id.

Also adds the diagnosis note for the related report where WebSearch/WebFetch
kept hitting an exhausted Kimi plan: those tools run on the haiku slot, which
`/model` cannot retarget, so they stayed on the launch default.

Author: 鲁工

README.md

@@ -172,8 +172,8 @@ ccmr claude --print --output-format json "你的问题"
 - `CLAUDE_CONFIG_DIR=~/.claude-gateway`
 - `ANTHROPIC_BASE_URL=http://127.0.0.1:<gateway-port>`
 - `ANTHROPIC_AUTH_TOKEN=ccmr-local-gateway`
-- `ANTHROPIC_MODEL` 和 Claude 默认模型变量会指向当前 `default_model`
-- `CLAUDE_CODE_AUTO_COMPACT_WINDOW` 会按启动模型的 `context_window` 自动注入（如 Seed/Kimi 256K、DeepSeek 1M），避免 Claude Code 把第三方模型默认裁剪到 ~200k。若你已自行设置该变量，则尊重你的值不覆盖。注意它是会话级全局：运行时用 `/model` 切到不同上下文窗口的模型时，请重启或用 `ccmr claude --model <模型>` 重新启动以匹配
+- `ANTHROPIC_MODEL` 和 Claude 默认模型变量会指向当前 `default_model`；若该模型的 `context_window` ≥ 1M，会自动带上 `[1m]` 后缀（如 `deepseek-flash[1m]`）。Claude Code 只有在模型名带这个后缀时才会真正开启 1M 上下文，并且会在请求发出前把后缀剥掉，因此不影响上游模型 ID
+- `CLAUDE_CODE_AUTO_COMPACT_WINDOW` 会按启动模型的 `context_window` 自动注入（如 Seed/Kimi 256K、DeepSeek 1M），避免 Claude Code 把第三方模型默认裁剪到 ~200k。若你已自行设置该变量，则尊重你的值不覆盖。注意它是会话级全局：运行时用 `/model` 切到**不同上下文窗口**的模型时，压缩阈值不会跟着变，需要重启或用 `ccmr claude --model <模型>` 重新启动以匹配（1M 模型之间互相切换不受影响）
 
 这些变量只作用于 `ccmr claude` 启动的 Claude Code 子进程，不会修改你的官方 Claude Code 配置，也不会影响直接运行 `claude` 的官方订阅模式。
 
@@ -543,6 +543,14 @@ DeepSeek Anthropic 兼容接口会忽略 `metadata` 字段，但某些 Claude Co
 
 ## 更新日志
 
+### v1.19.0
+
+- **1M 模型自动带 `[1m]` 后缀**：Claude Code 在自定义 base_url 后面默认按 ~200k 处理第三方模型，只有模型名以 `[1m]` 结尾时才会开启百万上下文——此前需要每次手动 `/model deepseek-flash[1m]`。现在两个出口都会自动补后缀：
+  - `ccmr claude` 注入的 `ANTHROPIC_MODEL` / `ANTHROPIC_DEFAULT_SONNET_MODEL` / `..._OPUS_MODEL` / `..._HAIKU_MODEL`（解决"刚进界面就是 200K"）
+  - `GET /v1/models` 返回的模型 id（解决"用 `/model` 切换后又变回 200K"）
+- **判定规则**：`context_window ≥ 1000000` 才加后缀，当前 42 个模型中 27 个符合（DeepSeek、Kimi K3、Qwen3.8、GLM-5.x、MiniMax M3、MiMo 等）；256K 及以下的模型（Seed、Kimi K2.x、Step）保持裸名——给小窗口模型挂 `[1m]` 会让 Claude Code 超发上下文
+- **后缀只对客户端有意义**：网关在解析模型名时会先剥掉 `[1m]` 再查别名，上游收到的始终是裸模型 ID（Kimi 曾因 `k3[1m]` 报 401）。因此 `deepseek-flash` 和 `deepseek-flash[1m]` 都能正常路由，老配置、老脚本不受影响
+
 ### v1.18.0
 
 - **DeepSeek V4.1 Flash 正式版接入**（官方 2026-09-10 发布公告 / [定价页](https://api-docs.deepseek.com/quick_start/pricing)）：正式模型名 `deepseek-flash`，取代 09-08 的内测 id（`deepseek-v4.1-flash-expires-on-0910`，已过期）。552B MoE、Causal-Encoder-Decoder 非对称结构（输入激活 8B / 输出激活 16B）、原生多模态视觉理解，1M 上下文、最大输出 384K。实测：文本 200（0.62s）；base64 image block 真实进入上下文（8×8 纯蓝图正确答 "Blue"，input_tokens 227）

dist/cli.js

@@ -0,0 +1,18 @@
+/**
+ * The "[1m]" model-name suffix.
+ *
+ * Claude Code opens its 1M context window only when the model name it was
+ * given carries this suffix, and it strips the suffix before calling the API.
+ * The suffix is therefore a client-side marker: it never belongs in an
+ * upstream model id (Kimi answers `k3[1m]` with 401 "model id does not
+ * exist"), and the gateway must be able to take it back off again.
+ */
+/** Remove a trailing "[1m]" so the name can be resolved as a model or alias. */
+export declare function stripContextSuffix(name: string): string;
+/**
+ * Add "[1m]" when the model's context window is at least 1M, so Claude Code
+ * uses the full window instead of assuming ~200k. Already-suffixed and
+ * smaller-window names are returned unchanged.
+ */
+export declare function withContextSuffix(name: string, contextWindow?: number): string;
+//# sourceMappingURL=model-suffix.d.ts.map