Provider: 支持官方 OpenAI Chat Completions

README.md

@@ -4,26 +4,27 @@
 [`learn-claude-code`](https://github.com/shareAI-lab/learn-claude-code) 的 Harness
 Engineering 思路：模型负责判断下一步做什么，Harness 负责提供工具、上下文、权限边界、持久化和用户界面。
 
-当前版本为 `0.1.0`，使用 DeepSeek 模型，提供一次性命令行执行和交互式 TUI 两种使用方式。
+当前版本为 `0.1.0`，默认使用 DeepSeek，也支持 OpenAI Chat Completions，
+提供一次性命令行执行和交互式 TUI 两种使用方式。
 
 ## 工作区结构
 
 ```text
-kuncode-cli ──▶ kuncode-agent ──▶ kuncode-core ──▶ DeepSeek API
+kuncode-cli ──▶ kuncode-agent ──▶ kuncode-core ──▶ LLM API
     │                 │                 │
     │                 │                 └─ 消息、Completion、流式协议、Provider
     │                 └─ Agent Loop、工具、权限、会话、压缩与编排
     └─ 参数、配置、审批、一次性输出与 TUI
 ```
 
-- `kuncode-core`：Provider-neutral 的消息与 Completion 抽象，以及 DeepSeek Provider。
+- `kuncode-core`：Provider-neutral 的消息与 Completion 抽象，以及 DeepSeek、OpenAI Provider。
 - `kuncode-agent`：Agent 运行时、工具调度、权限、Hook、Todo、会话持久化和上下文压缩。
 - `kuncode-cli`：命令行参数、项目配置、终端审批、普通输出和交互式 TUI。
 
 ## 环境要求
 
 - Rust stable，项目使用 Rust 2024 edition。
-- DeepSeek API Key。
+- DeepSeek 或 OpenAI API Key。
 - 支持 ANSI 终端；交互模式需要 stdin 和 stdout 都连接到真实终端。
 
 ## 快速开始
@@ -40,6 +41,24 @@ export DEEPSEEK_API_KEY="your-api-key"
 DEEPSEEK_API_KEY=your-api-key
 ```
 
+使用 OpenAI 官方接口时，在 `.kuncode/settings.json` 配置：
+
+```json
+{
+  "model": {
+    "provider": "openai",
+    "name": "gpt-5.1",
+    "maxTokens": 16384
+  }
+}
+```
+
+并设置对应环境变量：
+
+```bash
+export OPENAI_API_KEY="your-api-key"
+```
+
 一次性执行任务：
 
 ```bash
@@ -76,6 +95,7 @@ cargo build --release -p kuncode-cli
     "defaultMode": "default"
   },
   "model": {
+    "provider": "deepseek",
     "name": "deepseek-v4-pro",
     "maxTokens": 65536
   },
@@ -94,8 +114,14 @@ cargo build --release -p kuncode-cli
 
 补充说明：
 
-- `DEEPSEEK_MODEL` 可以覆盖配置文件中的模型名称。
+- `KUNCODE_MODEL` 可以覆盖配置文件中的模型名称；`DEEPSEEK_MODEL` 作为兼容别名保留。
+- `model.provider` 支持 `deepseek` 和 `openai`；两者分别使用固定官方 endpoint，
+  并读取 `DEEPSEEK_API_KEY` 或 `OPENAI_API_KEY`。
 - 内置模型配置包括 `deepseek-v4-pro` 和 `deepseek-v4-flash`。
+- 没有内置能力档案的模型，`model.maxTokens` 默认值为 `16384`；从旧版
+  `32768` 默认值升级时，如配置了 `compaction.reservedOutput`，需同步调整或显式设置
+  `model.maxTokens`。
+- 非内置模型启用上下文压缩时，需要显式设置 `compaction.contextLimit`。
 - `compaction.mode` 支持 `disabled`、`shadow` 和 `enabled`，默认是 `disabled`。
 - `shadow` 只计算和报告压缩候选，不替换当前上下文。
 - `enabled` 会在达到预算阈值时执行压缩，并要求会话持久化状态保持健康。

crates/kuncode-agent/src/compaction/summary/summarizer.rs

@@ -219,7 +219,7 @@ mod tests {
     use kuncode_core::{
         completion::{
             AssistantContent, CompletionError, CompletionModel, CompletionRequest,
-            CompletionResponse, CompletionStream, Message, ReasoningEffort, ToolChoice, Usage,
+            CompletionResponse, CompletionStream, Message, ReasoningEffort, Usage,
         },
         non_empty_vec::NonEmptyVec,
     };
@@ -308,7 +308,7 @@ mod tests {
         assert!(sent.model.is_none());
         assert_eq!(sent.chat_history.len(), 2);
         assert!(sent.tools.is_empty());
-        assert_eq!(sent.tool_choice, Some(ToolChoice::None));
+        assert!(sent.tool_choice.is_none());
         assert_eq!(sent.temperature, Some(0.0));
         assert_eq!(sent.max_tokens, Some(2_048));
         assert_eq!(sent.reasoning, Some(ReasoningEffort::Off));

crates/kuncode-agent/src/compaction/summary/summarizer/attempt.rs

@@ -3,7 +3,7 @@
 use kuncode_core::{
     completion::{
         AssistantContent, CompletionError, CompletionModel, CompletionRequestBuilder, Message,
-        ReasoningEffort, ToolChoice, Usage,
+        ReasoningEffort, Usage,
     },
     non_empty_vec::NonEmptyVec,
 };
@@ -42,11 +42,11 @@ pub(super) async fn run_attempt<M>(
 where
     M: CompletionModel,
 {
+    // An empty tool set already makes `none` the provider default.
     let completion = CompletionRequestBuilder::from_messages(prompt)
         .temperature(Some(0.0))
         .max_tokens(Some(max_output_tokens))
         .reasoning(Some(ReasoningEffort::Off))
-        .tool_choice(Some(ToolChoice::None))
         .output_schema(Some(schema))
         .build();
     let response = model

crates/kuncode-cli/src/runtime.rs

@@ -27,10 +27,14 @@ use kuncode_agent::system_prompt::{
 };
 use kuncode_agent::workspace::Workspace;
 use kuncode_core::completion::{CompletionModel, RetryModel, RetryPolicy};
-use kuncode_core::providers::deepseek::{DeepSeekClient, DeepSeekCompletionModel};
+use kuncode_core::providers::{
+    any_chat::{AnyChatClient, AnyChatCompletionModel},
+    deepseek::DeepSeekClient,
+    openai::OpenAiClient,
+};
 
 use crate::config::{PermissionFlags, resolve_permissions};
-use crate::settings::{ProjectSettings, ProjectTrust, load_project_settings};
+use crate::settings::{ProjectSettings, ProjectTrust, ProviderKind, load_project_settings};
 use crate::{Cli, logging::LoggingObserver};
 
 /// Identity and behavioral instructions rendered as the first system-prompt
@@ -46,7 +50,7 @@ Keep working until the task is done, then give a short, direct final answer.";
 /// observer + approver, plus the bits a frontend renders directly
 /// ([`model_name`](Self::model_name), [`mode`](Self::mode)). Generic over the
 /// model so a test or a future provider can supply its own `M`; [`assemble`]
-/// pins it to the CLI's [`DeepSeekCompletionModel`] wrapped in a
+/// pins it to the configured [`AnyChatCompletionModel`] wrapped in a
 /// [`RetryModel`] so transient provider failures are retried transparently.
 ///
 /// [`assemble`]: Self::assemble
@@ -64,21 +68,21 @@ pub struct CliRuntime<M> {
     persistence_error: Option<String>,
 }
 
-impl CliRuntime<RetryModel<DeepSeekCompletionModel>> {
+impl CliRuntime<RetryModel<AnyChatCompletionModel>> {
     /// Builds the runtime from parsed CLI args and the project settings file.
     ///
     /// Resolves permissions from built-in ∪ project file ∪ CLI flags (mode
     /// precedence CLI > project > Default), assembles the system prompt from its
-    /// identity/environment/tools sections, and wires the DeepSeek model + the
-    /// default workspace tool registry.
+    /// identity/environment/tools sections, and wires the configured model +
+    /// the default workspace tool registry.
     ///
     /// # Errors
     ///
     /// Fails if the current directory is not a usable workspace, the project
     /// settings or resolved permissions are invalid, active compaction cannot
-    /// be bound to the selected model, or the DeepSeek client cannot be built
-    /// from the environment. Failure to open the optional session store is
-    /// retained as degraded persistence state rather than failing assembly.
+    /// be bound to the selected model, or the provider client cannot be built
+    /// from its fixed credential environment. Failure to open the optional session
+    /// store is retained as degraded persistence state rather than failing assembly.
     pub async fn assemble(cli: &Cli) -> Result<Self, Box<dyn std::error::Error>> {
         let workspace = Workspace::from_current_dir().await?;
         tracing::debug!(
@@ -97,6 +101,7 @@ impl CliRuntime<RetryModel<DeepSeekCompletionModel>> {
         let project = load_project_settings(workspace.root(), project_trust)?;
         let model_name = project.model_name.clone();
         let config = agent_config(&project)?;
+        let client = provider_client(&project)?;
         let flags = PermissionFlags {
             allow: &cli.allow,
             ask: &cli.ask,
@@ -160,11 +165,10 @@ impl CliRuntime<RetryModel<DeepSeekCompletionModel>> {
                     (None, Some("home directory unavailable".to_string()))
                 }
             };
-        let client = DeepSeekClient::from_env()?;
         // Normal turns inherit the default retry budget. Semantic summaries use
         // a separate one-retry wrapper so their fallback latency is bounded
         // independently of ordinary model calls.
-        let provider = DeepSeekCompletionModel::make(&client, model_name.clone());
+        let provider = AnyChatCompletionModel::make(&client, model_name.clone());
         let model = RetryModel::with_policy(provider.clone(), RetryPolicy::default());
         let summary_model = RetryModel::with_policy(provider, summary_retry_policy());
         let registry = ToolRegistry::with_default_workspace_tools(workspace)?;
@@ -185,6 +189,13 @@ impl CliRuntime<RetryModel<DeepSeekCompletionModel>> {
     }
 }
 
+fn provider_client(project: &ProjectSettings) -> Result<AnyChatClient, Box<dyn std::error::Error>> {
+    match project.provider {
+        ProviderKind::DeepSeek => Ok(AnyChatClient::DeepSeek(DeepSeekClient::from_env()?)),
+        ProviderKind::OpenAi => Ok(AnyChatClient::OpenAi(OpenAiClient::from_env()?)),
+    }
+}
+
 fn agent_config(project: &ProjectSettings) -> Result<AgentConfig, AgentCompactionConfigError> {
     let compaction = project
         .compaction
@@ -282,7 +293,7 @@ impl<M: CompletionModel> CliRuntime<M> {
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::settings::{ProjectSettings, load_project_settings_from};
+    use crate::settings::{ModelOverrides, ProjectSettings, load_project_settings_from};
     use std::fs;
 
     fn compaction_settings(tag: &str) -> ProjectSettings {
@@ -301,7 +312,8 @@ mod tests {
         )
         .expect("write settings");
         let settings =
-            load_project_settings_from(&dir, None, ProjectTrust::Untrusted).expect("load settings");
+            load_project_settings_from(&dir, ModelOverrides::default(), ProjectTrust::Untrusted)
+                .expect("load settings");
         let _ = fs::remove_dir_all(&dir);
         settings
     }