@kiritoko1029/opencode-add-oai-instruction 1.0.0 → 1.0.1

package/README.md

@@ -0,0 +1,77 @@
+# opencode-add-oai-instruction
+
+[简体中文](README.zh-CN.md)
+
+[OpenCode](https://github.com/anomalyco/opencode) plugin that injects an `instructions` string into OpenAI-compatible API requests.
+
+The instructions are loaded from a Markdown file based on the request's `model`.
+
+It also removes `max_output_tokens` / `max_tokens` for `gpt*` models (useful for some OpenAI-compatible providers that reject these fields).
+
+## Installation
+
+```bash
+cd ~/.config/opencode
+bun add github:kiritoko1029/opencode-add-oai-instruction
+```
+
+## Usage
+
+Add to your `opencode.json`:
+
+```jsonc
+{
+  "plugins": [
+    "@kiritoko1029/opencode-add-oai-instruction" // Add this line
+  ],
+  "provider": {
+    "my-provider": {
+      "npm": "@ai-sdk/openai",
+      "api": "https://your-api.com/v1",
+      "env": ["MY_API_KEY"],
+      "options": {
+        "addInstruction": true // Enable the plugin
+      },
+      "models": {
+        "gpt-5.2": {}
+      }
+    }
+  }
+}
+```
+
+## Prompt Files
+
+This package ships with built-in prompt Markdown files (so you typically don't need to create anything).
+
+Prompt source: https://github.com/openai/codex/tree/main/codex-rs/core
+
+### Custom prompts (optional)
+
+If you want to use your own prompt (override the built-in one, or add a prompt for another model), put a Markdown file in the same directory as this plugin's `index.mjs`.
+
+File name format:
+
+- `<model>_prompt.md`
+
+Examples:
+
+- `model="gpt-5.2-codex"` → `gpt-5.2-codex_prompt.md`
+- `model="gpt-5.2"` → `gpt-5.2_prompt.md` (or `gpt_5_2_prompt.md`)
+
+Notes:
+
+- Matching is case-insensitive.
+- If no matching file is found, the plugin falls back to a short built-in `instructions` (unless the request already includes a non-empty `instructions`).
+
+## How It Works
+
+1. Hooks into `chat.params` to read `provider.options.addInstruction`
+2. Intercepts `globalThis.fetch` and (when enabled):
+   - Removes `max_output_tokens` / `max_tokens` when `model` contains `gpt`
+   - Loads `instructions` from the model-specific `_prompt.md` file
+3. Caches the loaded instructions per model (in-memory)
+
+## License
+
+MIT

package/README.zh-CN.md

@@ -0,0 +1,78 @@
+# opencode-add-oai-instruction
+
+[English](README.md) | [简体中文](README.zh-CN.md)
+
+[OpenCode](https://github.com/anomalyco/opencode) 插件：会向 OpenAI 兼容接口的请求体中注入 `instructions` 字段。
+
+`instructions` 的内容会根据请求里的 `model` 从对应的 Markdown 文件加载。
+
+另外，对于 `gpt*` 模型（`model` 包含 `gpt`），插件也会移除 `max_output_tokens` / `max_tokens`，用于兼容一些不支持这些字段的 OpenAI-compatible provider。
+
+## 安装
+
+```bash
+cd ~/.config/opencode
+bun add github:kiritoko1029/opencode-add-oai-instruction
+```
+
+## 使用
+
+在你的 `opencode.json` 中加入：
+
+```jsonc
+{
+  "plugins": [
+    "@kiritoko1029/opencode-add-oai-instruction" // 添加这一行
+  ],
+  "provider": {
+    "my-provider": {
+      "npm": "@ai-sdk/openai",
+      "api": "https://your-api.com/v1",
+      "env": ["MY_API_KEY"],
+      "options": {
+        "addInstruction": true // 启用插件
+      },
+      "models": {
+        "gpt-5.2": {}
+      }
+    }
+  }
+}
+```
+
+## 提示词文件（Prompt Files）
+
+本包已内置提示词 Markdown 文件（通常不需要用户额外创建）。
+
+提示词来源：https://github.com/openai/codex/tree/main/codex-rs/core
+
+### 自定义提示词（可选）
+
+如果你希望使用自己的提示词（覆盖内置提示词，或为其它 model 增加提示词），可以在插件目录（与 `index.mjs` 同级）放置/替换对应的 Markdown 文件。
+
+文件命名规则：
+
+- `<model>_prompt.md`
+
+示例：
+
+- `model="gpt-5.2-codex"` → `gpt-5.2-codex_prompt.md`
+- `model="gpt-5.2"` → `gpt-5.2_prompt.md`（或 `gpt_5_2_prompt.md`）
+
+说明：
+
+- 匹配不区分大小写。
+- 如果没有找到匹配的文件，插件会使用一段内置的简短 `instructions`（除非请求里已经包含非空的 `instructions`）。
+- 读取到的提示词会按 `model` 做内存缓存（in-memory cache）。
+
+## 工作原理
+
+1. 在 `chat.params` hook 中读取 `provider.options.addInstruction`。
+2. 拦截 `globalThis.fetch`，当插件启用时：
+   - 当 `model` 包含 `gpt` 时，移除 `max_output_tokens` / `max_tokens`。
+   - 根据 `model` 加载对应的 `_prompt.md` 文件内容，并写入请求体 `instructions`。
+3. 对每个 `model` 的提示词内容进行缓存，减少文件读取次数。
+
+## License
+
+MIT

package/index.mjs

@@ -1,5 +1,7 @@
 /**
- * OpenCode plugin to omit max_output_tokens/max_tokens from API requests.
+ * OpenCode plugin to inject model-specific instructions from Markdown files.
+ *
+ * It also removes max_output_tokens/max_tokens for gpt* models.
  *
  * Configuration in opencode.json:
  *
@@ -29,6 +31,14 @@ let addInstruction = false;
 /** @type {Map<string, string | null>} */
 const instructionCache = new Map();
 
+const DEFAULT_INSTRUCTIONS = `You are a helpful coding assistant.
+- Be concise and direct.
+- Follow the user's requirements; ask one clarifying question if needed.
+- Don't guess or fabricate details.
+- Prefer minimal, safe changes.
+- Don't commit or push unless explicitly asked.
+- Reply in plain text.`;
+
 function toSafeFilenameKey(value) {
   return value
     .toLowerCase()
@@ -111,9 +121,16 @@ if (typeof originalFetch === "function") {
     }
 
     // 根据模型不同 使用不同的提示词
-    const instructions = await loadInstructionForModel(model);
-    if (instructions !== null) {
-      parsed.instructions = instructions;
+    const fileInstructions = await loadInstructionForModel(model);
+    if (typeof fileInstructions === "string" && fileInstructions.trim() !== "") {
+      parsed.instructions = fileInstructions;
+    } else {
+      const existingInstructions = parsed.instructions;
+      const hasExistingInstructions =
+        typeof existingInstructions === "string" && existingInstructions.trim() !== "";
+      if (!hasExistingInstructions) {
+        parsed.instructions = DEFAULT_INSTRUCTIONS;
+      }
     }
 
     return originalFetch.call(this, input, { ...init, body: JSON.stringify(parsed) });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@kiritoko1029/opencode-add-oai-instruction",
-  "version": "1.0.0",
+  "version": "1.0.1",
   "description": "OpenCode plugin for adding instructions from OpenAI-compatible API requests",
   "type": "module",
   "main": "index.mjs",