@nick3/copilot-api 1.10.4 → 1.10.8

package/README.md

@@ -354,7 +354,6 @@ The `<target>` can be either the account ID (GitHub login) or a 1-based index.
       "gpt-5.4": "xhigh"
     },
     "allowOriginalModelNamesForAliases": false,
-    "useFunctionApplyPatch": true,
     "forceAgent": false,
     "compactUseSmallModel": true,
     "messageStartInputTokensFallback": false,
@@ -446,7 +445,6 @@ The `<target>` can be either the account ID (GitHub login) or a 1-based index.
 - **apiKey (deprecated):** Legacy single-key field kept for migration compatibility. Prefer `auth.apiKeys`. When `auth.apiKeys` is empty, the server falls back to `COPILOT_API_KEY` and then `apiKey`.- **modelReasoningEfforts:** Per-model `reasoning.effort` sent to the Copilot Responses API. Allowed values are `none`, `minimal`, `low`, `medium`, `high`, and `xhigh`. If a model isn’t listed, `high` is used by default.
 - **modelAliases:** Map of `alias -> { target, allowOriginal? }` (legacy string values are still accepted). Alias keys are normalized (trim + lowercase) and must be non-empty; aliases cannot map to themselves (case-insensitive), and conflicting normalized aliases are rejected. `allowOriginal` overrides the global default per alias. If multiple aliases map to the same target, original names are allowed when any alias sets `allowOriginal: true` (allow-wins). Admin UI/API rejects blocked keys (`__proto__`, `constructor`, `prototype`). Aliases can be used in downstream requests, and targets may be configured `provider/model` aliases for top-level `/v1/messages` and `/v1/messages/count_tokens` routing.
 - **allowOriginalModelNamesForAliases:** Global default for aliases that omit `allowOriginal`. When `false` (default), targets are blocked unless an alias explicitly allows them; when `true`, targets are allowed unless all aliases explicitly block them.
-- **useFunctionApplyPatch:** When `true` (default), `POST /v1/responses` converts a `tools` entry with `{ "type": "custom", "name": "apply_patch" }` into an OpenAI-style `function` tool (with a parameter schema) for upstream compatibility. Set to `false` to leave custom tools untouched.
 - **forceAgent:** When `true`, `POST /v1/responses` treats a request as agent-initiated if **any** input item has `role: "assistant"`. When `false` (default), only the **last** input item is checked.
 - **compactUseSmallModel:** When `true`, detected "compact" requests (e.g., from Claude Code or opencode compact mode) will automatically use the configured `smallModel` to avoid consuming premium usage for short/background tasks. Defaults to `true`.
 - **messageStartInputTokensFallback:** When `true`, the Anthropic streaming translation layer estimates `message_start.input_tokens` when upstream stream events do not provide it. Defaults to `false`.
@@ -730,6 +728,47 @@ You can find more options here: [Claude Code settings](https://docs.anthropic.co
 
 You can also read more about IDE integration here: [Add Claude Code to your IDE](https://docs.anthropic.com/en/docs/claude-code/ide-integrations)
 
+## GPT Tool Search
+
+For GPT Responses models such as `gpt-5.4+`, this proxy can expose Responses `tool_search` through a small MCP bridge. The same bridge can be used by Claude Code and opencode, as long as the client loads MCP servers and sends Anthropic Messages traffic through this proxy.
+
+Do not set Claude Code's native `ENABLE_TOOL_SEARCH` for GPT models. That flag enables Claude Code's own client-side tool search mode, and it may stop forwarding deferred tool definitions. This proxy needs the full tool definitions so it can keep the small always-loaded tool set eager and translate every other tool into Responses deferred namespaces.
+
+If you install `tool-search@copilot-api-marketplace`, Claude Code receives this MCP bridge automatically and you can skip the manual Claude Code MCP setup below.
+
+Add the tool search bridge to the MCP config used by Claude Code:
+
+```json
+{
+  "mcpServers": {
+    "tool_search": {
+      "type": "stdio",
+      "command": "npx",
+      "args": ["-y", "@nick3/copilot-api@latest", "mcp"]
+    }
+  }
+}
+```
+
+Add the tool search bridge to the MCP config used by opencode:
+
+```json
+{
+  "mcp": {
+    "tool_search": {
+      "type": "local",
+      "command": ["npx", "-y", "@nick3/copilot-api@latest", "mcp"]
+    }
+  }
+}
+```
+
+For local development, use `bun` as the command and `["run", "./src/main.ts", "mcp"]` as the args.
+
+Internally, the proxy now configures OpenAI Responses `tool_search` in client-executed mode. Deferred tools are still exposed as searchable namespaces, but the model is explicitly asked to return the exact deferred tool names it wants to load next.
+
+The bridge uses direct tool selection, not query search. Its tool input is `names`, a comma-separated list of exact deferred tool names, for example `TaskList,TaskGet,mcp__fetch__fetch`.
+
 ## Using with OpenCode
 
 OpenCode already has a direct GitHub Copilot provider. Use this section when you want OpenCode to point at this proxy through `@ai-sdk/anthropic` and reuse the agent behaviors described earlier in this README.
@@ -891,10 +930,13 @@ Plugin integrations are available for Claude Code and opencode.
 
 #### Claude Code plugin integration (marketplace-based)
 
-The Claude Code integration is packaged as a plugin named `claude-plugin`.
+The Claude Code integration is packaged as two plugins:
+
+- `agent-inject` injects `__SUBAGENT_MARKER__...` on `SubagentStart`, so this proxy can infer `x-initiator: agent`.
+- `tool-search` registers the `tool_search` MCP bridge used for GPT Responses deferred tool loading.
 
 - Marketplace catalog in this repository: `.claude-plugin/marketplace.json`
-- Plugin source in this repository: `claude-plugin`
+- Plugin sources in this repository: `claude-plugin/agent-inject`, `claude-plugin/tool-search`
 
 Add the marketplace remotely:
 
@@ -902,19 +944,22 @@ Add the marketplace remotely:
 /plugin marketplace add https://github.com/nick3/copilot-api.git#all
 ```
 
-Install the plugin from the marketplace:
+Install the plugins from the marketplace:
 
 ```sh
-/plugin install claude-plugin@copilot-api-marketplace
+/plugin install agent-inject@copilot-api-marketplace
+/plugin install tool-search@copilot-api-marketplace
 ```
 
-After installation, the plugin injects `__SUBAGENT_MARKER__...` on `SubagentStart`, and this proxy uses it to infer `x-initiator: agent`.
+After installation, `agent-inject` injects `__SUBAGENT_MARKER__...` on `SubagentStart`, and this proxy uses it to infer `x-initiator: agent`.
 
-The plugin also registers a `UserPromptSubmit` hook that returns `{"continue": true}`, and it can inject `SessionStart` reminder rules through environment variables:
+The `agent-inject` plugin also registers a `UserPromptSubmit` hook that returns `{"continue": true}`, and it can inject `SessionStart` reminder rules through environment variables:
 
 - `CLAUDE_PLUGIN_ENABLE_QUESTION_RULES=1` enables the two reminders about using the `question` tool automatically for Claude Code. Alternatively, you can add the same reminders manually in `CLAUDE.md`; see [CLAUDE.md or AGENTS.md Recommended Content](#claudemd-or-agentsmd-recommended-content).
 - `CLAUDE_PLUGIN_ENABLE_NO_BACKGROUND_AGENTS_RULE=1` enables the `run_in_background: true` avoidance reminder for agent hooks.
 
+The `tool-search` plugin bundles the same MCP bridge described in [GPT Tool Search](#gpt-tool-search), so Claude Code users do not need to add the `tool_search` server manually when they install that plugin.
+
 #### Opencode plugin
 
 The subagent marker producer is packaged as an opencode plugin located at `.opencode/plugins/subagent-marker.js`.

package/README.zh-CN.md

@@ -363,7 +363,6 @@ Copilot API 现在使用子命令结构，主要命令包括：
       "gpt-5.4": "xhigh"
     },
     "allowOriginalModelNamesForAliases": false,
-    "useFunctionApplyPatch": true,
     "forceAgent": false,
     "compactUseSmallModel": true,
     "messageStartInputTokensFallback": false,
@@ -455,7 +454,6 @@ Copilot API 现在使用子命令结构，主要命令包括：
 - **apiKey（已弃用）：** 兼容迁移的旧单 key 字段。优先使用 `auth.apiKeys`。当 `auth.apiKeys` 为空时，服务端会回退到 `COPILOT_API_KEY`，再回退到 `apiKey`。- **modelReasoningEfforts：** 按模型配置发送到 Copilot Responses API 的 `reasoning.effort`。可选值包括 `none`、`minimal`、`low`、`medium`、`high` 和 `xhigh`。若某模型未配置，则默认使用 `high`。
 - **modelAliases：** `alias -> { target, allowOriginal? }` 的映射（也仍然接受旧的字符串写法）。别名 key 会先做标准化（trim + lowercase），且不能为空；别名不能映射回自己（大小写不敏感），冲突的标准化别名会被拒绝。`allowOriginal` 可为单个别名覆盖全局默认值。如果多个别名映射到同一个 target，只要其中任意一个设置了 `allowOriginal: true`，原始模型名就会被允许（allow-wins）。Admin UI/API 会拒绝被屏蔽的键（`__proto__`、`constructor`、`prototype`）。下游请求可以直接使用这些别名，target 也可以是 `provider/model` 形式，用于顶层 `/v1/messages` 与 `/v1/messages/count_tokens` 路由。
 - **allowOriginalModelNamesForAliases：** 对未显式设置 `allowOriginal` 的别名所采用的全局默认值。当其为 `false`（默认）时，target 原名默认被屏蔽，除非某个别名显式允许；当其为 `true` 时，target 原名默认可用，除非所有别名都显式阻止。
-- **useFunctionApplyPatch：** 当为 `true`（默认）时，`POST /v1/responses` 会把 `tools` 中形如 `{ "type": "custom", "name": "apply_patch" }` 的条目转换为 OpenAI 风格的 `function` 工具（带参数 schema），以便与上游更好兼容。设为 `false` 则保留自定义工具原样。
 - **forceAgent：** 当为 `true` 时，只要 `POST /v1/responses` 的任一 input item 带有 `role: "assistant"`，就会把请求视为由 agent 发起；当为 `false`（默认）时，只检查最后一个 input item。
 - **compactUseSmallModel：** 当为 `true` 时，检测到的“compact”请求（例如 Claude Code 或 opencode 的 compact 模式）会自动改用配置中的 `smallModel`，以避免短后台任务消耗 premium 使用量。默认值为 `true`。
 - **messageStartInputTokensFallback：** 当为 `true` 时，如果上游流式事件没有提供 `message_start.input_tokens`，Anthropic 流式翻译层会自行估算该值。默认值为 `false`。
@@ -742,6 +740,47 @@ npx @nick3/copilot-api@latest start --claude-code
 
 也可以参考 IDE 集成说明：[Add Claude Code to your IDE](https://docs.anthropic.com/en/docs/claude-code/ide-integrations)
 
+## GPT Tool Search
+
+对于 `gpt-5.4+` 这类 GPT Responses 模型，本代理可以通过一个很小的 MCP bridge 暴露 Responses `tool_search`。Claude Code 和 opencode 都可以使用同一个 bridge，前提是客户端会加载 MCP server，并且 Anthropic Messages 流量会经过本代理。
+
+GPT 模型不要设置 Claude Code 原生的 `ENABLE_TOOL_SEARCH`。这个开关启用的是 Claude Code 自己的客户端 tool search 模式，可能导致 deferred 工具定义不再转发给代理。本代理需要完整的工具定义，这样才能只保留那一小组常驻加载工具，其余工具统一转换为 Responses deferred namespace。
+
+如果你安装了 `tool-search@copilot-api-marketplace`，Claude Code 会自动带上这个 MCP bridge，可以跳过下面这段 Claude Code MCP 手动配置。
+
+请把 tool search bridge 加到 Claude Code 使用的 MCP 配置中：
+
+```json
+{
+  "mcpServers": {
+    "tool_search": {
+      "type": "stdio",
+      "command": "npx",
+      "args": ["-y", "@nick3/copilot-api@latest", "mcp"]
+    }
+  }
+}
+```
+
+请把 tool search bridge 加到 opencode 使用的 MCP 配置中：
+
+```json
+{
+  "mcp": {
+    "tool_search": {
+      "type": "local",
+      "command": ["npx", "-y", "@nick3/copilot-api@latest", "mcp"]
+    }
+  }
+}
+```
+
+本地开发时可以将命令换成 `bun`，参数换成 `["run", "./src/main.ts", "mcp"]`。
+
+代理内部现在会把 OpenAI Responses `tool_search` 配置成 client-executed 模式。deferred tools 仍然会作为可搜索 namespace 暴露给模型，但会明确要求模型直接返回下一步要加载的精确工具名列表。
+
+该 bridge 使用直接工具选择，不做 query 搜索。工具入参是 `names`，值为逗号分隔的精确 deferred 工具名，例如 `TaskList,TaskGet,mcp__fetch__fetch`。
+
 ## 与 OpenCode 一起使用
 
 OpenCode 已经内置了 GitHub Copilot provider。本节适用于你希望让 OpenCode 通过 `@ai-sdk/anthropic` 指向这个代理，并复用本 README 前面提到的 agent 行为时。
@@ -905,10 +944,13 @@ UI 会把 token 保存在 `sessionStorage` 中，并通过 `x-admin-token` 请
 
 #### Claude Code 插件集成（基于 marketplace）
 
-Claude Code 集成被打包为名为 `claude-plugin` 的插件。
+Claude Code 集成现在拆分为两个插件：
+
+- `agent-inject` 会在 `SubagentStart` 时注入 `__SUBAGENT_MARKER__...`，以便本代理推导 `x-initiator: agent`。
+- `tool-search` 会注册用于 GPT Responses deferred tool loading 的 `tool_search` MCP bridge。
 
 - 本仓库中的 marketplace catalog：`.claude-plugin/marketplace.json`
-- 本仓库中的插件源码：`claude-plugin`
+- 本仓库中的插件源码：`claude-plugin/agent-inject`、`claude-plugin/tool-search`
 
 远程添加 marketplace：
 
@@ -919,16 +961,19 @@ Claude Code 集成被打包为名为 `claude-plugin` 的插件。
 从 marketplace 安装插件：
 
 ```sh
-/plugin install claude-plugin@copilot-api-marketplace
+/plugin install agent-inject@copilot-api-marketplace
+/plugin install tool-search@copilot-api-marketplace
 ```
 
-安装后，插件会在 `SubagentStart` 时注入 `__SUBAGENT_MARKER__...`，该代理会利用它推导 `x-initiator: agent`。
+安装后，`agent-inject` 会在 `SubagentStart` 时注入 `__SUBAGENT_MARKER__...`，该代理会利用它推导 `x-initiator: agent`。
 
-插件还会注册一个 `UserPromptSubmit` hook，并返回 `{"continue": true}`；同时它也可以通过环境变量注入 `SessionStart` reminder 规则：
+`agent-inject` 还会注册一个 `UserPromptSubmit` hook，并返回 `{"continue": true}`；同时它也可以通过环境变量注入 `SessionStart` reminder 规则：
 
 - `CLAUDE_PLUGIN_ENABLE_QUESTION_RULES=1` 会自动为 Claude Code 启用两条关于使用 `question` 工具的提醒。你也可以把同样的提醒手动写进 `CLAUDE.md`；见 [CLAUDE.md 或 AGENTS.md 推荐内容](#claudemd-or-agentsmd-recommended-content)。
 - `CLAUDE_PLUGIN_ENABLE_NO_BACKGROUND_AGENTS_RULE=1` 会启用关于避免在 agent hooks 中使用 `run_in_background: true` 的提醒。
 
+`tool-search` 插件内置了 [GPT Tool Search](#gpt-tool-search) 一节描述的同一个 MCP bridge，因此安装该插件后，Claude Code 用户无需再手动配置 `tool_search` server。
+
 #### Opencode 插件
 
 subagent marker 生成器被打包为一个 opencode 插件，位于 `.opencode/plugins/subagent-marker.js`。