feat: prompt-based tool calling for non-DeepSeek models via custom AP…

README.zh.md

@@ -202,9 +202,9 @@ whale migrate-config
 
 更多说明见 [docs/configuration.md](docs/configuration.md)。
 
-## 是否支持 Coding Plan？
+## 是否支持 Coding Plan / 自定义 API 中转站？
 
-支持。Whale 可以通过自定义 API endpoint 连接兼容 `/chat/completions` 的 Coding Plan / 第三方 DeepSeek 接口。
+支持。Whale 可以通过自定义 API endpoint 连接兼容 `/chat/completions` 的 Coding Plan、第三方 DeepSeek 接口或 API 中转站。
 
 以阿里云百炼为例，在 `~/.whale/config.toml` 中配置：
 
@@ -217,6 +217,21 @@ thinking_enabled = true
 base_url = "https://dashscope.aliyuncs.com/compatible-mode/v1"
 ```
 
+也可以配置第三方 API 中转站（如 tokenshengsheng.com）：
+
+```toml
+model = "gpt-5.5"
+reasoning_effort = "medium"
+thinking_enabled = false
+
+[api]
+base_url = "https://tokenshengsheng.com/v1"
+
+[permissions]
+auto_accept = true
+default = "allow"
+```
+
 然后配置对应的 API key：
 
 ```bash
@@ -235,6 +250,16 @@ whale setup
 - `DEEPSEEK_BASE_URL` 环境变量会覆盖 `~/.whale/config.toml` 里的 `[api].base_url`。
 - Whale 当前仍然优先适配 DeepSeek 风格的模型、流式输出、thinking、tool calls 和 prefix-cache 工作流。第三方 endpoint 是否完整可用，取决于服务端兼容程度。
 
+### 非 DeepSeek 模型的工具调用
+
+对于不支持原生 `tools` 参数的模型（如通过中转站接入的 GPT 系列模型），Whale 会自动启用 **Prompt-based 工具调用**：
+
+1. 在消息开头注入独立的工具指令系统提示，引导模型用 JSON 代码块输出工具调用
+2. 支持多种 JSON 格式的自动解析，包括标准 `{"tool_call": {...}}` 格式和 `{"command": "..."}`、`{"name": "...", "arguments": {...}}` 等替代格式
+3. 工具调用结果会以自然语言形式反馈给模型，确保多轮对话上下文完整
+
+此机制在检测到模型非 DeepSeek/Codex 系列时自动激活，无需手动配置。
+
 ---
 
 ## Non-goals

docs/configuration.md

@@ -169,6 +169,52 @@ fallback_filenames = ["AGENTS.md", ".claude/instructions.md", "CLAUDE.md"]
 and `deepseek-v4-pro` get 1,000,000 tokens (1M); other models default to 128K. No
 manual configuration is needed.
 
+## Custom API Endpoints and Non-DeepSeek Models
+
+Whale supports connecting to compatible API endpoints by setting `[api].base_url`
+in `config.toml`. This includes:
+
+- Third-party DeepSeek-compatible services (e.g., Alibaba Bailian)
+- API relay / proxy services (e.g., tokenshengsheng.com)
+- Any service that implements the `/chat/completions` endpoint
+
+Example configuration for an API relay:
+
+```toml
+model = "gpt-5.5"
+reasoning_effort = "medium"
+thinking_enabled = false
+
+[api]
+base_url = "https://tokenshengsheng.com/v1"
+
+[permissions]
+auto_accept = true
+default = "allow"
+```
+
+When using non-DeepSeek models through an API relay, `thinking_enabled` should
+typically be set to `false`, as most third-party models do not support the
+DeepSeek thinking protocol.
+
+### Prompt-based Tool Calling
+
+For models that do not support the native `tools` parameter (detected
+automatically when the model name does not contain "deepseek" or "codex"),
+Whale activates **prompt-based tool calling**:
+
+1. A standalone system message with tool instructions is prepended to every
+   request, telling the model to output tool calls as JSON code blocks
+2. Whale parses multiple JSON formats from the model's text output:
+   - Standard: `{"tool_call": {"name": "...", "arguments": {...}}}`
+   - Command-style: `{"command": "..."}`
+   - Named: `{"name": "...", "arguments": {...}}` or `{"tool": "...", "args": {...}}`
+   - JSON objects with `path`, `query`, or `pattern` fields
+3. Tool results are fed back to the model as natural-language user messages,
+   preserving full conversational context
+
+This mechanism activates automatically — no manual configuration needed.
+
 ## Migrating old config
 
 Whale v0.1.8 and earlier used `preferences.json` and `settings.json`. New

internal/defaults/defaults.go

@@ -37,7 +37,8 @@ func IsSupportedModel(model string) bool {
 			return true
 		}
 	}
-	return false
+	// Allow any model for custom API endpoints
+	return true
 }
 
 func DefaultMemoryFileOrder() []string {
@@ -57,5 +58,6 @@ func ContextWindowForModel(model string) int {
 	if IsDeepSeekV4Model(model) {
 		return DeepSeekV4ContextWindow
 	}
-	return DefaultContextWindow
+	// Use large context window for custom models
+	return DeepSeekV4ContextWindow
 }