feat: OpenResponses

docs/plan/worker_api_plan.md

@@ -0,0 +1,83 @@
+# Worker API/DSL 実装計画
+
+## 目的
+
+- [Open Responses](https://www.openresponses.org)(以後"OR")に準拠した正規化を前提に、
+  Item/Part の2段スコープを扱える Worker API を設計する。
+- APIの煩雑化を防ぐため、worker.on_xxx として公開するのを避けつつ、
+  Text/Thinking/Tool など型の違いを静的に扱える DSL を提供する。
+
+## 方針
+
+- 内部は Timeline が Event を正規化し、Item/Part/Meta
+  を単一ストリームとして扱う。
+- API では Item/Part 型ごとに ctx を持てるようにし、DSL
+  で記述の冗長さを削減する。
+- まず macro_rules! 版を作り、必要なら proc-macro に拡張する。
+- Item/Part の型パラメータはクレートが公開する Kind 型を使う。
+
+## 仕様の前提
+
+- Item は OR の item (message, function_call, reasoning など) に対応する。
+- Part は OR の content part (output_text, reasoning_text など) に対応する。
+- Item は必ず start/stop を持つ。Part は Item 内で複数発生し得る。
+- Item/Part の型指定は `Item<Message>` / `Part<ReasoningText>` のように書く。
+
+## 設計ステップ
+
+### 1. 内部イベントモデルの整理
+
+- Event を Item/Part/Meta の3層に整理する。
+- ItemEvent / PartEvent は型パラメータで区別する。
+- 例: ItemEvent<Message>, PartEvent<Message, OutputText>
+
+### 2. スコープの二段化
+
+- Item ctx: Item 型ごとに1つ
+- Part ctx: Part 型ごとに1つ
+- Part のイベントでは常に item ctx と part ctx の両方を渡す。
+
+### 3. Handler trait の再定義
+
+- Item/Part を型で指定できる trait を導入する。
+- 例:
+  - trait ItemHandler<I>
+  - trait PartHandler<I, P>
+- PartHandler には ItemHandler の ItemCtx を必須で渡す。
+- Part の ctx 型は `PartKind::Ctx` 方式 or enum 方式で切り替える。
+
+### 4. Timeline との結合
+
+- Timeline は ItemStart で ItemCtx を生成
+- PartStart で PartCtx を生成
+- Delta/Stop は対応 ctx に流す
+- ItemStop で ItemCtx を破棄
+
+### 5. DSL (macro_rules!) の導入
+
+- まず宣言的 DSL を提供する。
+- 例:
+  - handler! { Item<Message> { type ItemCtx = ...; Part<OutputText> { type
+    PartCtx = ...; } } }
+- DSL は ItemHandler / PartHandler 実装を生成する。
+- Item/Part の Kind 型はクレートが公開する型を参照する。
+
+### 6. 拡張ポイント
+
+- 追加 Part (output_image など) を DSL に追加しやすい形にする。
+- 必要なら proc-macro に移行して構文自由度を上げる。
+
+## 実装順序
+
+1. Event/Item/Part の型定義の整理
+2. Item/Part ctx を持つ Timeline 実装
+3. Handler trait の定義・既存コードの移行
+4. macro_rules! DSL の実装
+5. 既存ユースケースの移植
+
+## TODO
+
+- Item と Part の型対応表を整理する
+- OR と既存 llm_client の差分を再確認する
+- Tool args の delta を OR 拡張として扱うか検討する
+- macro_rules! で表現可能な DSL の最小文法を確定する

docs/research/openresponses_mapping.md

@@ -0,0 +1,80 @@
+# Open Responses mapping (llm_client -> Open Responses)
+
+This document maps the current `llm_client` event model to Open Responses items
+and streaming events. It focuses on output streaming; input items are noted
+where they are the closest semantic match.
+
+## Legend
+
+- **OR item**: Open Responses item types used in `response.output`.
+- **OR event**: Open Responses streaming events (`response.*`).
+- **Note**: Gaps or required adaptation decisions.
+
+## Response lifecycle / meta events
+
+| llm_client               | Open Responses                                                | Note                                                                                           |
+| ------------------------ | ------------------------------------------------------------- | ---------------------------------------------------------------------------------------------- |
+| `StatusEvent::Started`   | `response.created`, `response.queued`, `response.in_progress` | OR has finer-grained lifecycle states; pick a subset or map Started -> `response.in_progress`. |
+| `StatusEvent::Completed` | `response.completed`                                          |                                                                                                |
+| `StatusEvent::Failed`    | `response.failed`                                             |                                                                                                |
+| `StatusEvent::Cancelled` | (no direct event)                                             | Could map to `response.incomplete` or `response.failed` depending on semantics.                |
+| `UsageEvent`             | `response.completed` payload usage                            | OR reports usage on the response object, not as a dedicated streaming event.                   |
+| `ErrorEvent`             | `error` event                                                 | OR has a dedicated error streaming event.                                                      |
+| `PingEvent`              | (no direct event)                                             | OR does not define a heartbeat event.                                                          |
+
+## Output block lifecycle
+
+### Text block
+
+| llm_client                                        | Open Responses                                                                           | Note                                                                                |
+| ------------------------------------------------- | ---------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------- |
+| `BlockStart { block_type: Text, metadata: Text }` | `response.output_item.added` with item type `message` (assistant)                        | OR output items are message/function_call/reasoning. This creates the message item. |
+| `BlockDelta { delta: Text(..) }`                  | `response.output_text.delta`                                                             | Text deltas map 1:1 to output text deltas.                                          |
+| `BlockStop { block_type: Text }`                  | `response.output_text.done` + `response.content_part.done` + `response.output_item.done` | OR emits separate done events for content parts and items.                          |
+
+### Tool use (function call)
+
+| llm_client                                                           | Open Responses                                                        | Note                                                                                                  |
+| -------------------------------------------------------------------- | --------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------- |
+| `BlockStart { block_type: ToolUse, metadata: ToolUse { id, name } }` | `response.output_item.added` with item type `function_call`           | OR uses `call_id` + `name` + `arguments` string. Map `id` -> `call_id`.                               |
+| `BlockDelta { delta: InputJson(..) }`                                | `response.function_call_arguments.delta`                              | OR spec does not explicitly require argument deltas; treat as OpenAI-compatible extension if adopted. |
+| `BlockStop { block_type: ToolUse }`                                  | `response.function_call_arguments.done` + `response.output_item.done` | Item status can be set to `completed` or `incomplete`.                                                |
+
+### Tool result (function call output)
+
+| llm_client                                                                    | Open Responses                        | Note                                                                                     |
+| ----------------------------------------------------------------------------- | ------------------------------------- | ---------------------------------------------------------------------------------------- |
+| `BlockStart { block_type: ToolResult, metadata: ToolResult { tool_use_id } }` | **Input item** `function_call_output` | OR treats tool results as input items, not output items. This is a request-side mapping. |
+| `BlockDelta`                                                                  | (no direct output event)              | OR does not stream tool output deltas as response events.                                |
+| `BlockStop`                                                                   | (no direct output event)              | Tool output lives on the next request as an input item.                                  |
+
+### Thinking / reasoning
+
+| llm_client                                                | Open Responses                                          | Note                                                                                                                                                          |
+| --------------------------------------------------------- | ------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `BlockStart { block_type: Thinking, metadata: Thinking }` | `response.output_item.added` with item type `reasoning` | OR models reasoning as a separate item type.                                                                                                                  |
+| `BlockDelta { delta: Thinking(..) }`                      | `response.reasoning.delta`                              | OR has dedicated reasoning delta events.                                                                                                                      |
+| `BlockStop { block_type: Thinking }`                      | `response.reasoning.done`                               | OR separates reasoning summary events (`response.reasoning_summary_*`) from reasoning deltas. Decide whether Thinking maps to full reasoning or summary only. |
+
+## Stop reasons
+
+| llm_client `StopReason` | Open Responses                                                                 | Note                                           |
+| ----------------------- | ------------------------------------------------------------------------------ | ---------------------------------------------- |
+| `EndTurn`               | `response.completed` + item status `completed`                                 |                                                |
+| `MaxTokens`             | `response.incomplete` + item status `incomplete`                               |                                                |
+| `StopSequence`          | `response.completed`                                                           |                                                |
+| `ToolUse`               | `response.completed` for message item, followed by `function_call` output item | OR models tool call as a separate output item. |
+
+## Gaps / open decisions
+
+- `PingEvent` has no OR equivalent. If needed, keep as internal only.
+- `Cancelled` status needs a policy: map to `response.incomplete` or
+  `response.failed`.
+- OR has `response.refusal.delta` / `response.refusal.done`. `llm_client` has no
+  refusal delta type; consider adding a new block or delta variant if needed.
+- OR splits _item_ and _content part_ lifecycles. `llm_client` currently has a
+  single block lifecycle, so mapping should decide whether to synthesize
+  `content_part.*` events or ignore them.
+- The OR specification does not state how `function_call.arguments` stream
+  deltas; `response.function_call_arguments.*` should be treated as a compatible
+  extension if required.

flake.lock

@@ -35,11 +35,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1767116409,
+        "lastModified": 1771369470,
-        "narHash": "sha256-5vKw92l1GyTnjoLzEagJy5V5mDFck72LiQWZSOnSicw=",
+        "narHash": "sha256-0NBlEBKkN3lufyvFegY4TYv5mCNHbi5OmBDrzihbBMQ=",
         "owner": "nixos",
         "repo": "nixpkgs",
-        "rev": "cad22e7d996aea55ecab064e84834289143e44a0",
+        "rev": "0182a361324364ae3f436a63005877674cf45efb",
         "type": "github"
       },
       "original": {

llm-worker/src/hook.rs

@@ -52,7 +52,7 @@ pub enum PostToolCallResult {
 #[derive(Debug, Clone)]
 pub enum OnTurnEndResult {
     Finish,
-    ContinueWithMessages(Vec<crate::Message>),
+    ContinueWithMessages(Vec<crate::Item>),
     Paused,
 }
 
@@ -83,12 +83,12 @@ pub struct PostToolCallContext {
 }
 
 impl HookEventKind for OnPromptSubmit {
-    type Input = crate::Message;
+    type Input = crate::Item;
     type Output = OnPromptSubmitResult;
 }
 
 impl HookEventKind for PreLlmRequest {
-    type Input = Vec<crate::Message>;
+    type Input = Vec<crate::Item>;
     type Output = PreLlmRequestResult;
 }
 
@@ -103,7 +103,7 @@ impl HookEventKind for PostToolCall {
 }
 
 impl HookEventKind for OnTurnEnd {
-    type Input = Vec<crate::Message>;
+    type Input = Vec<crate::Item>;
     type Output = OnTurnEndResult;
 }
 

llm-worker/src/lib.rs

@@ -12,7 +12,7 @@
 //! # Quick Start
 //!
 //! ```ignore
-//! use llm_worker::{Worker, Message};
+//! use llm_worker::{Worker, Item};
 //!
 //! // Create a Worker
 //! let mut worker = Worker::new(client)
@@ -47,5 +47,5 @@ pub mod subscriber;
 pub mod timeline;
 pub mod tool;
 
-pub use message::{ContentPart, Message, MessageContent, Role};
+pub use message::{ContentPart, Item, Message, Role};
 pub use worker::{ToolRegistryError, Worker, WorkerConfig, WorkerError, WorkerResult};

llm-worker/src/llm_client/scheme/anthropic/request.rs

@@ -1,15 +1,17 @@
-//! Anthropic リクエスト生成
+//! Anthropic Request Builder
+//!
+//! Converts Open Responses native Item model to Anthropic Messages API format.
 
 use serde::Serialize;
 
 use crate::llm_client::{
+    types::{ContentPart, Item, Role, ToolDefinition},
     Request,
-    types::{ContentPart, Message, MessageContent, Role, ToolDefinition},
 };
 
 use super::AnthropicScheme;
 
-/// Anthropic APIへのリクエストボディ
+/// Anthropic API request body
 #[derive(Debug, Serialize)]
 pub(crate) struct AnthropicRequest {
     pub model: String,
@@ -30,14 +32,14 @@ pub(crate) struct AnthropicRequest {
     pub stream: bool,
 }
 
-/// Anthropic メッセージ
+/// Anthropic message
 #[derive(Debug, Serialize)]
 pub(crate) struct AnthropicMessage {
     pub role: String,
     pub content: AnthropicContent,
 }
 
-/// Anthropic コンテンツ
+/// Anthropic content
 #[derive(Debug, Serialize)]
 #[serde(untagged)]
 pub(crate) enum AnthropicContent {
@@ -45,7 +47,7 @@ pub(crate) enum AnthropicContent {
     Parts(Vec<AnthropicContentPart>),
 }
 
-/// Anthropic コンテンツパーツ
+/// Anthropic content part
 #[derive(Debug, Serialize)]
 #[serde(tag = "type")]
 pub(crate) enum AnthropicContentPart {
@@ -58,13 +60,10 @@ pub(crate) enum AnthropicContentPart {
         input: serde_json::Value,
     },
     #[serde(rename = "tool_result")]
-    ToolResult {
+    ToolResult { tool_use_id: String, content: String },
-        tool_use_id: String,
-        content: String,
-    },
 }
 
-/// Anthropic ツール定義
+/// Anthropic tool definition
 #[derive(Debug, Serialize)]
 pub(crate) struct AnthropicTool {
     pub name: String,
@@ -74,14 +73,9 @@ pub(crate) struct AnthropicTool {
 }
 
 impl AnthropicScheme {
-    /// RequestからAnthropicのリクエストボディを構築
+    /// Build Anthropic request from Request
     pub(crate) fn build_request(&self, model: &str, request: &Request) -> AnthropicRequest {
-        let messages = request
+        let messages = self.convert_items_to_messages(&request.items);
-            .messages
-            .iter()
-            .map(|m| self.convert_message(m))
-            .collect();
-
         let tools = request.tools.iter().map(|t| self.convert_tool(t)).collect();
 
         AnthropicRequest {
@@ -98,49 +92,160 @@ impl AnthropicScheme {
         }
     }
 
-    fn convert_message(&self, message: &Message) -> AnthropicMessage {
+    /// Convert Open Responses Items to Anthropic Messages
-        let role = match message.role {
+    ///
+    /// Anthropic uses a message-based model where:
+    /// - User messages have role "user"
+    /// - Assistant messages have role "assistant"
+    /// - Tool calls are content parts within assistant messages
+    /// - Tool results are content parts within user messages
+    fn convert_items_to_messages(&self, items: &[Item]) -> Vec<AnthropicMessage> {
+        let mut messages = Vec::new();
+        let mut pending_assistant_parts: Vec<AnthropicContentPart> = Vec::new();
+        let mut pending_user_parts: Vec<AnthropicContentPart> = Vec::new();
+
+        for item in items {
+            match item {
+                Item::Message { role, content, .. } => {
+                    // Flush pending parts before a new message
+                    self.flush_pending_parts(
+                        &mut messages,
+                        &mut pending_assistant_parts,
+                        &mut pending_user_parts,
+                    );
+
+                    let anthropic_role = match role {
                         Role::User => "user",
                         Role::Assistant => "assistant",
+                        Role::System => continue, // Skip system role items
                     };
 
-        let content = match &message.content {
+                    let parts: Vec<AnthropicContentPart> = content
-            MessageContent::Text(text) => AnthropicContent::Text(text.clone()),
-            MessageContent::ToolResult {
-                tool_use_id,
-                content,
-            } => AnthropicContent::Parts(vec![AnthropicContentPart::ToolResult {
-                tool_use_id: tool_use_id.clone(),
-                content: content.clone(),
-            }]),
-            MessageContent::Parts(parts) => {
-                let converted: Vec<_> = parts
                         .iter()
                         .map(|p| match p {
-                        ContentPart::Text { text } => {
+                            ContentPart::InputText { text } => {
                                 AnthropicContentPart::Text { text: text.clone() }
                             }
-                        ContentPart::ToolUse { id, name, input } => AnthropicContentPart::ToolUse {
+                            ContentPart::OutputText { text } => {
-                            id: id.clone(),
+                                AnthropicContentPart::Text { text: text.clone() }
-                            name: name.clone(),
+                            }
-                            input: input.clone(),
+                            ContentPart::Refusal { refusal } => {
-                        },
+                                AnthropicContentPart::Text {
-                        ContentPart::ToolResult {
+                                    text: refusal.clone(),
-                            tool_use_id,
+                                }
-                            content,
+                            }
-                        } => AnthropicContentPart::ToolResult {
-                            tool_use_id: tool_use_id.clone(),
-                            content: content.clone(),
-                        },
                         })
                         .collect();
-                AnthropicContent::Parts(converted)
-            }
-        };
 
-        AnthropicMessage {
+                    if parts.len() == 1 {
-            role: role.to_string(),
+                        if let AnthropicContentPart::Text { text } = &parts[0] {
-            content,
+                            messages.push(AnthropicMessage {
+                                role: anthropic_role.to_string(),
+                                content: AnthropicContent::Text(text.clone()),
+                            });
+                        } else {
+                            messages.push(AnthropicMessage {
+                                role: anthropic_role.to_string(),
+                                content: AnthropicContent::Parts(parts),
+                            });
+                        }
+                    } else {
+                        messages.push(AnthropicMessage {
+                            role: anthropic_role.to_string(),
+                            content: AnthropicContent::Parts(parts),
+                        });
+                    }
+                }
+
+                Item::FunctionCall {
+                    call_id,
+                    name,
+                    arguments,
+                    ..
+                } => {
+                    // Flush pending user parts first
+                    if !pending_user_parts.is_empty() {
+                        messages.push(AnthropicMessage {
+                            role: "user".to_string(),
+                            content: AnthropicContent::Parts(std::mem::take(
+                                &mut pending_user_parts,
+                            )),
+                        });
+                    }
+
+                    // Parse arguments JSON string to Value
+                    let input = serde_json::from_str(arguments)
+                        .unwrap_or_else(|_| serde_json::Value::Object(serde_json::Map::new()));
+
+                    pending_assistant_parts.push(AnthropicContentPart::ToolUse {
+                        id: call_id.clone(),
+                        name: name.clone(),
+                        input,
+                    });
+                }
+
+                Item::FunctionCallOutput { call_id, output, .. } => {
+                    // Flush pending assistant parts first
+                    if !pending_assistant_parts.is_empty() {
+                        messages.push(AnthropicMessage {
+                            role: "assistant".to_string(),
+                            content: AnthropicContent::Parts(std::mem::take(
+                                &mut pending_assistant_parts,
+                            )),
+                        });
+                    }
+
+                    pending_user_parts.push(AnthropicContentPart::ToolResult {
+                        tool_use_id: call_id.clone(),
+                        content: output.clone(),
+                    });
+                }
+
+                Item::Reasoning { text, .. } => {
+                    // Flush pending user parts first
+                    if !pending_user_parts.is_empty() {
+                        messages.push(AnthropicMessage {
+                            role: "user".to_string(),
+                            content: AnthropicContent::Parts(std::mem::take(
+                                &mut pending_user_parts,
+                            )),
+                        });
+                    }
+
+                    // Reasoning is treated as assistant text in Anthropic
+                    // (actual thinking blocks are handled differently in streaming)
+                    pending_assistant_parts.push(AnthropicContentPart::Text { text: text.clone() });
+                }
+            }
+        }
+
+        // Flush remaining pending parts
+        self.flush_pending_parts(
+            &mut messages,
+            &mut pending_assistant_parts,
+            &mut pending_user_parts,
+        );
+
+        messages
+    }
+
+    fn flush_pending_parts(
+        &self,
+        messages: &mut Vec<AnthropicMessage>,
+        pending_assistant_parts: &mut Vec<AnthropicContentPart>,
+        pending_user_parts: &mut Vec<AnthropicContentPart>,
+    ) {
+        if !pending_assistant_parts.is_empty() {
+            messages.push(AnthropicMessage {
+                role: "assistant".to_string(),
+                content: AnthropicContent::Parts(std::mem::take(pending_assistant_parts)),
+            });
+        }
+        if !pending_user_parts.is_empty() {
+            messages.push(AnthropicMessage {
+                role: "user".to_string(),
+                content: AnthropicContent::Parts(std::mem::take(pending_user_parts)),
+            });
         }
     }
 
@@ -195,4 +300,24 @@ mod tests {
         assert_eq!(anthropic_req.tools.len(), 1);
         assert_eq!(anthropic_req.tools[0].name, "get_weather");
     }
+
+    #[test]
+    fn test_function_call_and_output() {
+        let scheme = AnthropicScheme::new();
+        let request = Request::new()
+            .user("What's the weather?")
+            .item(Item::function_call(
+                "call_123",
+                "get_weather",
+                r#"{"city":"Tokyo"}"#,
+            ))
+            .item(Item::function_call_output("call_123", "Sunny, 25°C"));
+
+        let anthropic_req = scheme.build_request("claude-sonnet-4-20250514", &request);
+
+        assert_eq!(anthropic_req.messages.len(), 3);
+        assert_eq!(anthropic_req.messages[0].role, "user");
+        assert_eq!(anthropic_req.messages[1].role, "assistant");
+        assert_eq!(anthropic_req.messages[2].role, "user");
+    }
 }

llm-worker/src/llm_client/scheme/gemini/request.rs

@@ -1,130 +1,130 @@
-//! Gemini リクエスト生成
+//! Gemini Request Builder
 //!
-//! Google Gemini APIへのリクエストボディを構築
+//! Converts Open Responses native Item model to Google Gemini API format.
 
 use serde::Serialize;
 use serde_json::Value;
 
 use crate::llm_client::{
+    types::{Item, Role, ToolDefinition},
     Request,
-    types::{ContentPart, Message, MessageContent, Role, ToolDefinition},
 };
 
 use super::GeminiScheme;
 
-/// Gemini APIへのリクエストボディ
+/// Gemini API request body
 #[derive(Debug, Serialize)]
 #[serde(rename_all = "camelCase")]
 pub(crate) struct GeminiRequest {
-    /// コンテンツ（会話履歴）
+    /// Contents (conversation history)
     pub contents: Vec<GeminiContent>,
-    /// システム指示
+    /// System instruction
     #[serde(skip_serializing_if = "Option::is_none")]
     pub system_instruction: Option<GeminiContent>,
-    /// ツール定義
+    /// Tool definitions
     #[serde(skip_serializing_if = "Vec::is_empty")]
     pub tools: Vec<GeminiTool>,
-    /// ツール設定
+    /// Tool config
     #[serde(skip_serializing_if = "Option::is_none")]
     pub tool_config: Option<GeminiToolConfig>,
-    /// 生成設定
+    /// Generation config
     #[serde(skip_serializing_if = "Option::is_none")]
     pub generation_config: Option<GeminiGenerationConfig>,
 }
 
-/// Gemini コンテンツ
+/// Gemini content
 #[derive(Debug, Serialize)]
 pub(crate) struct GeminiContent {
-    /// ロール
+    /// Role
     pub role: String,
-    /// パーツ
+    /// Parts
     pub parts: Vec<GeminiPart>,
 }
 
-/// Gemini パーツ
+/// Gemini part
 #[derive(Debug, Serialize)]
 #[serde(untagged)]
 pub(crate) enum GeminiPart {
-    /// テキストパーツ
+    /// Text part
     Text { text: String },
-    /// 関数呼び出しパーツ
+    /// Function call part
     FunctionCall {
         #[serde(rename = "functionCall")]
         function_call: GeminiFunctionCall,
     },
-    /// 関数レスポンスパーツ
+    /// Function response part
     FunctionResponse {
         #[serde(rename = "functionResponse")]
         function_response: GeminiFunctionResponse,
     },
 }
 
-/// Gemini 関数呼び出し
+/// Gemini function call
 #[derive(Debug, Serialize)]
 pub(crate) struct GeminiFunctionCall {
     pub name: String,
     pub args: Value,
 }
 
-/// Gemini 関数レスポンス
+/// Gemini function response
 #[derive(Debug, Serialize)]
 pub(crate) struct GeminiFunctionResponse {
     pub name: String,
     pub response: GeminiFunctionResponseContent,
 }
 
-/// Gemini 関数レスポンス内容
+/// Gemini function response content
 #[derive(Debug, Serialize)]
 pub(crate) struct GeminiFunctionResponseContent {
     pub name: String,
     pub content: Value,
 }
 
-/// Gemini ツール定義
+/// Gemini tool definition
 #[derive(Debug, Serialize)]
 #[serde(rename_all = "camelCase")]
 pub(crate) struct GeminiTool {
-    /// 関数宣言
+    /// Function declarations
     pub function_declarations: Vec<GeminiFunctionDeclaration>,
 }
 
-/// Gemini 関数宣言
+/// Gemini function declaration
 #[derive(Debug, Serialize)]
 pub(crate) struct GeminiFunctionDeclaration {
-    /// 関数名
+    /// Function name
     pub name: String,
-    /// 説明
+    /// Description
     #[serde(skip_serializing_if = "Option::is_none")]
     pub description: Option<String>,
-    /// パラメータスキーマ
+    /// Parameter schema
     pub parameters: Value,
 }
 
-/// Gemini ツール設定
+/// Gemini tool config
 #[derive(Debug, Serialize)]
 #[serde(rename_all = "camelCase")]
 pub(crate) struct GeminiToolConfig {
-    /// 関数呼び出し設定
+    /// Function calling config
     pub function_calling_config: GeminiFunctionCallingConfig,
 }
 
-/// Gemini 関数呼び出し設定
+/// Gemini function calling config
 #[derive(Debug, Serialize)]
 #[serde(rename_all = "camelCase")]
 pub(crate) struct GeminiFunctionCallingConfig {
-    /// モード: AUTO, ANY, NONE
+    /// Mode: AUTO, ANY, NONE
     #[serde(skip_serializing_if = "Option::is_none")]
     pub mode: Option<String>,
-    /// ストリーミング関数呼び出し引数を有効にするか
+    /// Enable streaming function call arguments
     #[serde(skip_serializing_if = "Option::is_none")]
     pub stream_function_call_arguments: Option<bool>,
 }
 
-/// Gemini 生成設定
+/// Gemini generation config
 #[derive(Debug, Serialize)]
 #[serde(rename_all = "camelCase")]
 pub(crate) struct GeminiGenerationConfig {
-    /// 最大出力トークン数
+    /// Max output tokens
     #[serde(skip_serializing_if = "Option::is_none")]
     pub max_output_tokens: Option<u32>,
     /// Temperature
@@ -136,27 +136,23 @@ pub(crate) struct GeminiGenerationConfig {
     /// Top K
     #[serde(skip_serializing_if = "Option::is_none")]
     pub top_k: Option<u32>,
-    /// ストップシーケンス
+    /// Stop sequences
     #[serde(skip_serializing_if = "Vec::is_empty")]
     pub stop_sequences: Vec<String>,
 }
 
 impl GeminiScheme {
-    /// RequestからGeminiのリクエストボディを構築
+    /// Build Gemini request from Request
     pub(crate) fn build_request(&self, request: &Request) -> GeminiRequest {
-        let mut contents = Vec::new();
+        let contents = self.convert_items_to_contents(&request.items);
 
-        for message in &request.messages {
+        // System prompt
-            contents.push(self.convert_message(message));
-        }
-
-        // システムプロンプト
         let system_instruction = request.system_prompt.as_ref().map(|s| GeminiContent {
-            role: "user".to_string(), // system_instructionではroleは"user"か省略
+            role: "user".to_string(),
             parts: vec![GeminiPart::Text { text: s.clone() }],
         });
 
-        // ツール
+        // Tools
         let tools = if request.tools.is_empty() {
             vec![]
         } else {
@@ -165,7 +161,7 @@ impl GeminiScheme {
             }]
         };
 
-        // ツール設定
+        // Tool config
         let tool_config = if !request.tools.is_empty() {
             Some(GeminiToolConfig {
                 function_calling_config: GeminiFunctionCallingConfig {
@@ -181,7 +177,7 @@ impl GeminiScheme {
             None
         };
 
-        // 生成設定
+        // Generation config
         let generation_config = Some(GeminiGenerationConfig {
             max_output_tokens: request.config.max_tokens,
             temperature: request.config.temperature,
@@ -199,58 +195,126 @@ impl GeminiScheme {
         }
     }
 
-    fn convert_message(&self, message: &Message) -> GeminiContent {
+    /// Convert Open Responses Items to Gemini Contents
-        let role = match message.role {
+    ///
+    /// Gemini uses:
+    /// - role "user" for user messages and function responses
+    /// - role "model" for assistant messages and function calls
+    fn convert_items_to_contents(&self, items: &[Item]) -> Vec<GeminiContent> {
+        let mut contents = Vec::new();
+        let mut pending_model_parts: Vec<GeminiPart> = Vec::new();
+        let mut pending_user_parts: Vec<GeminiPart> = Vec::new();
+
+        for item in items {
+            match item {
+                Item::Message { role, content, .. } => {
+                    // Flush pending parts
+                    self.flush_pending_parts(
+                        &mut contents,
+                        &mut pending_model_parts,
+                        &mut pending_user_parts,
+                    );
+
+                    let gemini_role = match role {
                         Role::User => "user",
                         Role::Assistant => "model",
+                        Role::System => continue, // Skip system role items
                     };
 
-        let parts = match &message.content {
+                    let parts: Vec<GeminiPart> = content
-            MessageContent::Text(text) => vec![GeminiPart::Text { text: text.clone() }],
-            MessageContent::ToolResult {
-                tool_use_id,
-                content,
-            } => {
-                // Geminiでは関数レスポンスとしてマップ
-                vec![GeminiPart::FunctionResponse {
-                    function_response: GeminiFunctionResponse {
-                        name: tool_use_id.clone(),
-                        response: GeminiFunctionResponseContent {
-                            name: tool_use_id.clone(),
-                            content: serde_json::Value::String(content.clone()),
-                        },
-                    },
-                }]
-            }
-            MessageContent::Parts(parts) => parts
                         .iter()
-                .map(|p| match p {
+                        .map(|p| GeminiPart::Text {
-                    ContentPart::Text { text } => GeminiPart::Text { text: text.clone() },
+                            text: p.as_text().to_string(),
-                    ContentPart::ToolUse { id: _, name, input } => GeminiPart::FunctionCall {
+                        })
+                        .collect();
+
+                    contents.push(GeminiContent {
+                        role: gemini_role.to_string(),
+                        parts,
+                    });
+                }
+
+                Item::FunctionCall {
+                    name, arguments, ..
+                } => {
+                    // Flush pending user parts first
+                    if !pending_user_parts.is_empty() {
+                        contents.push(GeminiContent {
+                            role: "user".to_string(),
+                            parts: std::mem::take(&mut pending_user_parts),
+                        });
+                    }
+
+                    // Parse arguments
+                    let args = serde_json::from_str(arguments)
+                        .unwrap_or_else(|_| Value::Object(serde_json::Map::new()));
+
+                    pending_model_parts.push(GeminiPart::FunctionCall {
                         function_call: GeminiFunctionCall {
                             name: name.clone(),
-                            args: input.clone(),
+                            args,
                         },
-                    },
+                    });
-                    ContentPart::ToolResult {
+                }
-                        tool_use_id,
-                        content,
-                    } => GeminiPart::FunctionResponse {
-                        function_response: GeminiFunctionResponse {
-                            name: tool_use_id.clone(),
-                            response: GeminiFunctionResponseContent {
-                                name: tool_use_id.clone(),
-                                content: serde_json::Value::String(content.clone()),
-                            },
-                        },
-                    },
-                })
-                .collect(),
-        };
 
-        GeminiContent {
+                Item::FunctionCallOutput { call_id, output, .. } => {
-            role: role.to_string(),
+                    // Flush pending model parts first
-            parts,
+                    if !pending_model_parts.is_empty() {
+                        contents.push(GeminiContent {
+                            role: "model".to_string(),
+                            parts: std::mem::take(&mut pending_model_parts),
+                        });
+                    }
+
+                    pending_user_parts.push(GeminiPart::FunctionResponse {
+                        function_response: GeminiFunctionResponse {
+                            name: call_id.clone(),
+                            response: GeminiFunctionResponseContent {
+                                name: call_id.clone(),
+                                content: Value::String(output.clone()),
+                            },
+                        },
+                    });
+                }
+
+                Item::Reasoning { text, .. } => {
+                    // Flush pending user parts first
+                    if !pending_user_parts.is_empty() {
+                        contents.push(GeminiContent {
+                            role: "user".to_string(),
+                            parts: std::mem::take(&mut pending_user_parts),
+                        });
+                    }
+
+                    // Reasoning is treated as model text in Gemini
+                    pending_model_parts.push(GeminiPart::Text { text: text.clone() });
+                }
+            }
+        }
+
+        // Flush remaining pending parts
+        self.flush_pending_parts(&mut contents, &mut pending_model_parts, &mut pending_user_parts);
+
+        contents
+    }
+
+    fn flush_pending_parts(
+        &self,
+        contents: &mut Vec<GeminiContent>,
+        pending_model_parts: &mut Vec<GeminiPart>,
+        pending_user_parts: &mut Vec<GeminiPart>,
+    ) {
+        if !pending_model_parts.is_empty() {
+            contents.push(GeminiContent {
+                role: "model".to_string(),
+                parts: std::mem::take(pending_model_parts),
+            });
+        }
+        if !pending_user_parts.is_empty() {
+            contents.push(GeminiContent {
+                role: "user".to_string(),
+                parts: std::mem::take(pending_user_parts),
+            });
         }
     }
 
@@ -318,4 +382,24 @@ mod tests {
         assert_eq!(gemini_req.contents[0].role, "user");
         assert_eq!(gemini_req.contents[1].role, "model");
     }
+
+    #[test]
+    fn test_function_call_and_output() {
+        let scheme = GeminiScheme::new();
+        let request = Request::new()
+            .user("What's the weather?")
+            .item(Item::function_call(
+                "call_123",
+                "get_weather",
+                r#"{"city":"Tokyo"}"#,
+            ))
+            .item(Item::function_call_output("call_123", "Sunny, 25°C"));
+
+        let gemini_req = scheme.build_request(&request);
+
+        assert_eq!(gemini_req.contents.len(), 3);
+        assert_eq!(gemini_req.contents[0].role, "user");
+        assert_eq!(gemini_req.contents[1].role, "model");
+        assert_eq!(gemini_req.contents[2].role, "user");
+    }
 }

llm-worker/src/llm_client/scheme/openai/request.rs

@@ -1,21 +1,23 @@
-//! OpenAI リクエスト生成
+//! OpenAI Request Builder
+//!
+//! Converts Open Responses native Item model to OpenAI Chat Completions API format.
 
 use serde::Serialize;
 use serde_json::Value;
 
 use crate::llm_client::{
+    types::{Item, Role, ToolDefinition},
     Request,
-    types::{ContentPart, Message, MessageContent, Role, ToolDefinition},
 };
 
 use super::OpenAIScheme;
 
-/// OpenAI APIへのリクエストボディ
+/// OpenAI API request body
 #[derive(Debug, Serialize)]
 pub(crate) struct OpenAIRequest {
     pub model: String,
     #[serde(skip_serializing_if = "Option::is_none")]
-    pub max_completion_tokens: Option<u32>, // max_tokens is deprecated for newer models, generally max_completion_tokens is preferred
+    pub max_completion_tokens: Option<u32>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub max_tokens: Option<u32>, // Legacy field for compatibility (e.g. Ollama)
     #[serde(skip_serializing_if = "Option::is_none")]
@@ -31,7 +33,7 @@ pub(crate) struct OpenAIRequest {
     #[serde(skip_serializing_if = "Vec::is_empty")]
     pub tools: Vec<OpenAITool>,
     #[serde(skip_serializing_if = "Option::is_none")]
-    pub tool_choice: Option<String>, // "auto", "none", or specific
+    pub tool_choice: Option<String>,
 }
 
 #[derive(Debug, Serialize)]
@@ -39,20 +41,21 @@ pub(crate) struct StreamOptions {
     pub include_usage: bool,
 }
 
-/// OpenAI メッセージ
+/// OpenAI message
 #[derive(Debug, Serialize)]
 pub(crate) struct OpenAIMessage {
     pub role: String,
-    pub content: Option<OpenAIContent>, // Optional for assistant tool calls
+    pub content: Option<OpenAIContent>,
     #[serde(skip_serializing_if = "Vec::is_empty")]
     pub tool_calls: Vec<OpenAIToolCall>,
     #[serde(skip_serializing_if = "Option::is_none")]
-    pub tool_call_id: Option<String>, // For tool_result (role: tool)
+    pub tool_call_id: Option<String>,
     #[serde(skip_serializing_if = "Option::is_none")]
-    pub name: Option<String>, // Optional name
+    pub name: Option<String>,
 }
 
-/// OpenAI コンテンツ
+/// OpenAI content
+#[allow(dead_code)]
 #[derive(Debug, Serialize)]
 #[serde(untagged)]
 pub(crate) enum OpenAIContent {
@@ -60,7 +63,7 @@ pub(crate) enum OpenAIContent {
     Parts(Vec<OpenAIContentPart>),
 }
 
-/// OpenAI コンテンツパーツ
+/// OpenAI content part
 #[allow(dead_code)]
 #[derive(Debug, Serialize)]
 #[serde(tag = "type")]
@@ -76,7 +79,7 @@ pub(crate) struct ImageUrl {
     pub url: String,
 }
 
-/// OpenAI ツール定義
+/// OpenAI tool definition
 #[derive(Debug, Serialize)]
 pub(crate) struct OpenAITool {
     pub r#type: String,
@@ -91,7 +94,7 @@ pub(crate) struct OpenAIToolFunction {
     pub parameters: Value,
 }
 
-/// OpenAI ツール呼び出し（メッセージ内）
+/// OpenAI tool call in message
 #[derive(Debug, Serialize)]
 pub(crate) struct OpenAIToolCall {
     pub id: String,
@@ -106,10 +109,11 @@ pub(crate) struct OpenAIToolCallFunction {
 }
 
 impl OpenAIScheme {
-    /// RequestからOpenAIのリクエストボディを構築
+    /// Build OpenAI request from Request
     pub(crate) fn build_request(&self, model: &str, request: &Request) -> OpenAIRequest {
         let mut messages = Vec::new();
 
+        // Add system message if present
         if let Some(system) = &request.system_prompt {
             messages.push(OpenAIMessage {
                 role: "system".to_string(),
@@ -120,7 +124,8 @@ impl OpenAIScheme {
             });
         }
 
-        messages.extend(request.messages.iter().map(|m| self.convert_message(m)));
+        // Convert items to messages
+        messages.extend(self.convert_items_to_messages(&request.items));
 
         let tools = request.tools.iter().map(|t| self.convert_tool(t)).collect();
 
@@ -143,106 +148,122 @@ impl OpenAIScheme {
             }),
             messages,
             tools,
-            tool_choice: None, // Default to auto if tools are present? Or let API decide (which is auto)
+            tool_choice: None,
         }
     }
 
-    fn convert_message(&self, message: &Message) -> OpenAIMessage {
+    /// Convert Open Responses Items to OpenAI Messages
-        match &message.content {
+    ///
-            MessageContent::ToolResult {
+    /// OpenAI uses a message-based model where:
-                tool_use_id,
+    /// - User messages have role "user"
-                content,
+    /// - Assistant messages have role "assistant"
-            } => OpenAIMessage {
+    /// - Tool calls are within assistant messages as tool_calls array
-                role: "tool".to_string(),
+    /// - Tool results have role "tool" with tool_call_id
-                content: Some(OpenAIContent::Text(content.clone())),
+    fn convert_items_to_messages(&self, items: &[Item]) -> Vec<OpenAIMessage> {
-                tool_calls: vec![],
+        let mut messages = Vec::new();
-                tool_call_id: Some(tool_use_id.clone()),
+        let mut pending_tool_calls: Vec<OpenAIToolCall> = Vec::new();
-                name: None,
+        let mut pending_assistant_text: Option<String> = None;
-            },
+
-            MessageContent::Text(text) => {
+        for item in items {
-                let role = match message.role {
+            match item {
+                Item::Message { role, content, .. } => {
+                    // Flush pending tool calls
+                    self.flush_pending_assistant(
+                        &mut messages,
+                        &mut pending_tool_calls,
+                        &mut pending_assistant_text,
+                    );
+
+                    let openai_role = match role {
                         Role::User => "user",
                         Role::Assistant => "assistant",
+                        Role::System => "system",
                     };
-                OpenAIMessage {
+
-                    role: role.to_string(),
+                    let text_content: String = content
-                    content: Some(OpenAIContent::Text(text.clone())),
+                        .iter()
+                        .map(|p| p.as_text())
+                        .collect::<Vec<_>>()
+                        .join("");
+
+                    messages.push(OpenAIMessage {
+                        role: openai_role.to_string(),
+                        content: Some(OpenAIContent::Text(text_content)),
                         tool_calls: vec![],
                         tool_call_id: None,
                         name: None,
+                    });
                 }
-            }
-            MessageContent::Parts(parts) => {
-                let role = match message.role {
-                    Role::User => "user",
-                    Role::Assistant => "assistant",
-                };
 
-                let mut content_parts = Vec::new();
+                Item::FunctionCall {
-                let mut tool_calls = Vec::new();
+                    call_id,
-                let mut is_tool_result = false;
+                    name,
-                let mut tool_result_id = None;
+                    arguments,
-                let mut tool_result_content = String::new();
+                    ..
-
+                } => {
-                for part in parts {
+                    pending_tool_calls.push(OpenAIToolCall {
-                    match part {
+                        id: call_id.clone(),
-                        ContentPart::Text { text } => {
-                            content_parts.push(OpenAIContentPart::Text { text: text.clone() });
-                        }
-                        ContentPart::ToolUse { id, name, input } => {
-                            tool_calls.push(OpenAIToolCall {
-                                id: id.clone(),
                         r#type: "function".to_string(),
                         function: OpenAIToolCallFunction {
                             name: name.clone(),
-                                    arguments: input.to_string(),
+                            arguments: arguments.clone(),
                         },
                     });
                 }
-                        ContentPart::ToolResult {
-                            tool_use_id,
-                            content,
-                        } => {
-                            // OpenAI doesn't support mixed content with ToolResult in the same message easily if not careful
-                            // But strictly speaking, a Message with ToolResult should be its own message with role "tool"
-                            is_tool_result = true;
-                            tool_result_id = Some(tool_use_id.clone());
-                            tool_result_content = content.clone();
-                        }
-                    }
-                }
 
-                if is_tool_result {
+                Item::FunctionCallOutput { call_id, output, .. } => {
-                    OpenAIMessage {
+                    // Flush pending tool calls before tool result
+                    self.flush_pending_assistant(
+                        &mut messages,
+                        &mut pending_tool_calls,
+                        &mut pending_assistant_text,
+                    );
+
+                    messages.push(OpenAIMessage {
                         role: "tool".to_string(),
-                        content: Some(OpenAIContent::Text(tool_result_content)),
+                        content: Some(OpenAIContent::Text(output.clone())),
                         tool_calls: vec![],
-                        tool_call_id: tool_result_id,
+                        tool_call_id: Some(call_id.clone()),
                         name: None,
+                    });
                 }
-                } else {
-                    let content = if content_parts.is_empty() {
-                        None
-                    } else if content_parts.len() == 1 {
-                        // Simplify single text part to just Text content if preferred, or keep as Parts
-                        if let OpenAIContentPart::Text { text } = &content_parts[0] {
-                            Some(OpenAIContent::Text(text.clone()))
-                        } else {
-                            Some(OpenAIContent::Parts(content_parts))
-                        }
-                    } else {
-                        Some(OpenAIContent::Parts(content_parts))
-                    };
 
-                    OpenAIMessage {
+                Item::Reasoning { text, .. } => {
-                        role: role.to_string(),
+                    // Reasoning is treated as assistant text in OpenAI
-                        content,
+                    // (OpenAI doesn't have native reasoning support like Claude)
-                        tool_calls,
+                    if let Some(ref mut existing) = pending_assistant_text {
+                        existing.push_str(text);
+                    } else {
+                        pending_assistant_text = Some(text.clone());
+                    }
+                }
+            }
+        }
+
+        // Flush remaining pending items
+        self.flush_pending_assistant(
+            &mut messages,
+            &mut pending_tool_calls,
+            &mut pending_assistant_text,
+        );
+
+        messages
+    }
+
+    fn flush_pending_assistant(
+        &self,
+        messages: &mut Vec<OpenAIMessage>,
+        pending_tool_calls: &mut Vec<OpenAIToolCall>,
+        pending_assistant_text: &mut Option<String>,
+    ) {
+        if !pending_tool_calls.is_empty() || pending_assistant_text.is_some() {
+            messages.push(OpenAIMessage {
+                role: "assistant".to_string(),
+                content: pending_assistant_text.take().map(OpenAIContent::Text),
+                tool_calls: std::mem::take(pending_tool_calls),
                 tool_call_id: None,
                 name: None,
-                    }
+            });
-                }
-            }
         }
     }
 
@@ -274,7 +295,6 @@ mod tests {
         assert_eq!(body.messages[0].role, "system");
         assert_eq!(body.messages[1].role, "user");
 
-        // Check system content
         if let Some(OpenAIContent::Text(text)) = &body.messages[0].content {
             assert_eq!(text, "System prompt");
         } else {
@@ -301,20 +321,39 @@ mod tests {
 
         let body = scheme.build_request("llama3", &request);
 
-        // max_tokens should be set, max_completion_tokens should be None
         assert_eq!(body.max_tokens, Some(100));
         assert!(body.max_completion_tokens.is_none());
     }
 
     #[test]
     fn test_build_request_modern_max_tokens() {
-        let scheme = OpenAIScheme::new(); // Default matches modern (legacy=false)
+        let scheme = OpenAIScheme::new();
         let request = Request::new().user("Hello").max_tokens(100);
 
         let body = scheme.build_request("gpt-4o", &request);
 
-        // max_completion_tokens should be set, max_tokens should be None
         assert_eq!(body.max_completion_tokens, Some(100));
         assert!(body.max_tokens.is_none());
     }
+
+    #[test]
+    fn test_function_call_and_output() {
+        let scheme = OpenAIScheme::new();
+        let request = Request::new()
+            .user("Check weather")
+            .item(Item::function_call(
+                "call_123",
+                "get_weather",
+                r#"{"city":"Tokyo"}"#,
+            ))
+            .item(Item::function_call_output("call_123", "Sunny, 25°C"));
+
+        let body = scheme.build_request("gpt-4o", &request);
+
+        assert_eq!(body.messages.len(), 3);
+        assert_eq!(body.messages[0].role, "user");
+        assert_eq!(body.messages[1].role, "assistant");
+        assert_eq!(body.messages[1].tool_calls.len(), 1);
+        assert_eq!(body.messages[2].role, "tool");
+    }
 }

llm-worker/src/llm_client/scheme/openresponses/events.rs

@@ -0,0 +1,494 @@
+//! Open Responses Event Parser
+//!
+//! Parses SSE events from the Open Responses API into internal Event types.
+
+use serde::Deserialize;
+
+use crate::llm_client::{
+    event::{
+        BlockMetadata, BlockStart, BlockStop, DeltaContent, ErrorEvent, Event, ResponseStatus,
+        StatusEvent, StopReason, UsageEvent,
+    },
+    ClientError,
+};
+
+// =============================================================================
+// Open Responses SSE Event Types
+// =============================================================================
+
+/// Response created event
+#[derive(Debug, Deserialize)]
+pub struct ResponseCreatedEvent {
+    pub response: ResponseObject,
+}
+
+/// Response object
+#[derive(Debug, Deserialize)]
+pub struct ResponseObject {
+    pub id: String,
+    pub status: String,
+    #[serde(default)]
+    pub output: Vec<OutputItem>,
+    pub usage: Option<UsageObject>,
+}
+
+/// Output item in response
+#[derive(Debug, Deserialize)]
+#[serde(tag = "type", rename_all = "snake_case")]
+pub enum OutputItem {
+    Message {
+        id: String,
+        role: String,
+        #[serde(default)]
+        content: Vec<ContentPartObject>,
+    },
+    FunctionCall {
+        id: String,
+        call_id: String,
+        name: String,
+        arguments: String,
+    },
+    Reasoning {
+        id: String,
+        #[serde(default)]
+        text: String,
+    },
+}
+
+/// Content part object
+#[derive(Debug, Deserialize)]
+#[serde(tag = "type", rename_all = "snake_case")]
+pub enum ContentPartObject {
+    OutputText { text: String },
+    InputText { text: String },
+    Refusal { refusal: String },
+}
+
+/// Usage object
+#[derive(Debug, Deserialize)]
+pub struct UsageObject {
+    pub input_tokens: Option<u64>,
+    pub output_tokens: Option<u64>,
+    pub total_tokens: Option<u64>,
+}
+
+/// Output item added event
+#[derive(Debug, Deserialize)]
+pub struct OutputItemAddedEvent {
+    pub output_index: usize,
+    pub item: OutputItem,
+}
+
+/// Text delta event
+#[derive(Debug, Deserialize)]
+pub struct TextDeltaEvent {
+    pub output_index: usize,
+    pub content_index: usize,
+    pub delta: String,
+}
+
+/// Text done event
+#[derive(Debug, Deserialize)]
+pub struct TextDoneEvent {
+    pub output_index: usize,
+    pub content_index: usize,
+    pub text: String,
+}
+
+/// Function call arguments delta event
+#[derive(Debug, Deserialize)]
+pub struct FunctionCallArgumentsDeltaEvent {
+    pub output_index: usize,
+    pub call_id: String,
+    pub delta: String,
+}
+
+/// Function call arguments done event
+#[derive(Debug, Deserialize)]
+pub struct FunctionCallArgumentsDoneEvent {
+    pub output_index: usize,
+    pub call_id: String,
+    pub arguments: String,
+}
+
+/// Reasoning delta event
+#[derive(Debug, Deserialize)]
+pub struct ReasoningDeltaEvent {
+    pub output_index: usize,
+    pub delta: String,
+}
+
+/// Reasoning done event
+#[derive(Debug, Deserialize)]
+pub struct ReasoningDoneEvent {
+    pub output_index: usize,
+    pub text: String,
+}
+
+/// Content part done event
+#[derive(Debug, Deserialize)]
+pub struct ContentPartDoneEvent {
+    pub output_index: usize,
+    pub content_index: usize,
+    pub part: ContentPartObject,
+}
+
+/// Output item done event
+#[derive(Debug, Deserialize)]
+pub struct OutputItemDoneEvent {
+    pub output_index: usize,
+    pub item: OutputItem,
+}
+
+/// Response done event
+#[derive(Debug, Deserialize)]
+pub struct ResponseDoneEvent {
+    pub response: ResponseObject,
+}
+
+/// Error event from API
+#[derive(Debug, Deserialize)]
+pub struct ApiErrorEvent {
+    pub error: ApiError,
+}
+
+/// API error details
+#[derive(Debug, Deserialize)]
+pub struct ApiError {
+    pub code: Option<String>,
+    pub message: String,
+}
+
+// =============================================================================
+// Event Parsing
+// =============================================================================
+
+/// Parse SSE event into internal Event(s)
+///
+/// Returns `Ok(None)` for events that should be ignored (e.g., heartbeats)
+/// Returns `Ok(Some(vec))` for events that produce one or more internal Events
+pub fn parse_event(event_type: &str, data: &str) -> Result<Option<Vec<Event>>, ClientError> {
+    // Skip empty data
+    if data.is_empty() || data == "[DONE]" {
+        return Ok(None);
+    }
+
+    let events = match event_type {
+        // Response lifecycle
+        "response.created" => {
+            let _event: ResponseCreatedEvent = parse_json(data)?;
+            Some(vec![Event::Status(StatusEvent {
+                status: ResponseStatus::Started,
+            })])
+        }
+
+        "response.in_progress" => {
+            // Just a status update, no action needed
+            None
+        }
+
+        "response.completed" | "response.done" => {
+            let event: ResponseDoneEvent = parse_json(data)?;
+            let mut events = Vec::new();
+
+            // Emit usage if present
+            if let Some(usage) = event.response.usage {
+                events.push(Event::Usage(UsageEvent {
+                    input_tokens: usage.input_tokens,
+                    output_tokens: usage.output_tokens,
+                    total_tokens: usage.total_tokens,
+                    cache_read_input_tokens: None,
+                    cache_creation_input_tokens: None,
+                }));
+            }
+
+            events.push(Event::Status(StatusEvent {
+                status: ResponseStatus::Completed,
+            }));
+            Some(events)
+        }
+
+        "response.failed" => {
+            // Try to parse error
+            if let Ok(error_event) = parse_json::<ApiErrorEvent>(data) {
+                Some(vec![
+                    Event::Error(ErrorEvent {
+                        code: error_event.error.code,
+                        message: error_event.error.message,
+                    }),
+                    Event::Status(StatusEvent {
+                        status: ResponseStatus::Failed,
+                    }),
+                ])
+            } else {
+                Some(vec![Event::Status(StatusEvent {
+                    status: ResponseStatus::Failed,
+                })])
+            }
+        }
+
+        // Output item events
+        "response.output_item.added" => {
+            let event: OutputItemAddedEvent = parse_json(data)?;
+            Some(vec![convert_item_added(&event)])
+        }
+
+        "response.output_item.done" => {
+            let event: OutputItemDoneEvent = parse_json(data)?;
+            Some(vec![convert_item_done(&event)])
+        }
+
+        // Text content events
+        "response.output_text.delta" => {
+            let event: TextDeltaEvent = parse_json(data)?;
+            Some(vec![Event::text_delta(event.output_index, &event.delta)])
+        }
+
+        "response.output_text.done" => {
+            // Text done - we'll handle stop in output_item.done
+            let _event: TextDoneEvent = parse_json(data)?;
+            None
+        }
+
+        // Content part events
+        "response.content_part.added" => {
+            // Content part added - we handle this via output_item.added
+            None
+        }
+
+        "response.content_part.done" => {
+            // Content part done - we handle stop in output_item.done
+            None
+        }
+
+        // Function call events
+        "response.function_call_arguments.delta" => {
+            let event: FunctionCallArgumentsDeltaEvent = parse_json(data)?;
+            Some(vec![Event::BlockDelta(crate::llm_client::event::BlockDelta {
+                index: event.output_index,
+                delta: DeltaContent::InputJson(event.delta),
+            })])
+        }
+
+        "response.function_call_arguments.done" => {
+            // Arguments done - we handle stop in output_item.done
+            let _event: FunctionCallArgumentsDoneEvent = parse_json(data)?;
+            None
+        }
+
+        // Reasoning events
+        "response.reasoning.delta" | "response.reasoning_summary_text.delta" => {
+            let event: ReasoningDeltaEvent = parse_json(data)?;
+            Some(vec![Event::BlockDelta(crate::llm_client::event::BlockDelta {
+                index: event.output_index,
+                delta: DeltaContent::Thinking(event.delta),
+            })])
+        }
+
+        "response.reasoning.done" | "response.reasoning_summary_text.done" => {
+            // Reasoning done - we handle stop in output_item.done
+            let _event: ReasoningDoneEvent = parse_json(data)?;
+            None
+        }
+
+        // Error event
+        "error" => {
+            let event: ApiErrorEvent = parse_json(data)?;
+            Some(vec![Event::Error(ErrorEvent {
+                code: event.error.code,
+                message: event.error.message,
+            })])
+        }
+
+        // Unknown event type - ignore
+        _ => {
+            tracing::debug!(event_type = event_type, "Unknown Open Responses event type");
+            None
+        }
+    };
+
+    Ok(events)
+}
+
+fn parse_json<T: serde::de::DeserializeOwned>(data: &str) -> Result<T, ClientError> {
+    serde_json::from_str(data).map_err(|e| ClientError::Parse(e.to_string()))
+}
+
+fn convert_item_added(event: &OutputItemAddedEvent) -> Event {
+    match &event.item {
+        OutputItem::Message { id, role: _, content: _ } => Event::BlockStart(BlockStart {
+            index: event.output_index,
+            block_type: crate::llm_client::event::BlockType::Text,
+            metadata: BlockMetadata::Text,
+        }),
+
+        OutputItem::FunctionCall {
+            id,
+            call_id,
+            name,
+            arguments: _,
+        } => Event::BlockStart(BlockStart {
+            index: event.output_index,
+            block_type: crate::llm_client::event::BlockType::ToolUse,
+            metadata: BlockMetadata::ToolUse {
+                id: call_id.clone(),
+                name: name.clone(),
+            },
+        }),
+
+        OutputItem::Reasoning { id, text: _ } => Event::BlockStart(BlockStart {
+            index: event.output_index,
+            block_type: crate::llm_client::event::BlockType::Thinking,
+            metadata: BlockMetadata::Thinking,
+        }),
+    }
+}
+
+fn convert_item_done(event: &OutputItemDoneEvent) -> Event {
+    let stop_reason = match &event.item {
+        OutputItem::FunctionCall { .. } => Some(StopReason::ToolUse),
+        _ => Some(StopReason::EndTurn),
+    };
+
+    Event::BlockStop(BlockStop {
+        index: event.output_index,
+        stop_reason,
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_parse_response_created() {
+        let data = r#"{"response":{"id":"resp_123","status":"in_progress","output":[]}}"#;
+        let events = parse_event("response.created", data).unwrap().unwrap();
+        assert_eq!(events.len(), 1);
+        assert!(matches!(
+            events[0],
+            Event::Status(StatusEvent {
+                status: ResponseStatus::Started
+            })
+        ));
+    }
+
+    #[test]
+    fn test_parse_text_delta() {
+        let data = r#"{"output_index":0,"content_index":0,"delta":"Hello"}"#;
+        let events = parse_event("response.output_text.delta", data)
+            .unwrap()
+            .unwrap();
+        assert_eq!(events.len(), 1);
+        if let Event::BlockDelta(delta) = &events[0] {
+            assert_eq!(delta.index, 0);
+            assert!(matches!(&delta.delta, DeltaContent::Text(t) if t == "Hello"));
+        } else {
+            panic!("Expected BlockDelta");
+        }
+    }
+
+    #[test]
+    fn test_parse_output_item_added_message() {
+        let data = r#"{"output_index":0,"item":{"type":"message","id":"msg_123","role":"assistant","content":[]}}"#;
+        let events = parse_event("response.output_item.added", data)
+            .unwrap()
+            .unwrap();
+        assert_eq!(events.len(), 1);
+        if let Event::BlockStart(start) = &events[0] {
+            assert_eq!(start.index, 0);
+            assert!(matches!(
+                start.block_type,
+                crate::llm_client::event::BlockType::Text
+            ));
+        } else {
+            panic!("Expected BlockStart");
+        }
+    }
+
+    #[test]
+    fn test_parse_output_item_added_function_call() {
+        let data = r#"{"output_index":1,"item":{"type":"function_call","id":"fc_123","call_id":"call_456","name":"get_weather","arguments":""}}"#;
+        let events = parse_event("response.output_item.added", data)
+            .unwrap()
+            .unwrap();
+        assert_eq!(events.len(), 1);
+        if let Event::BlockStart(start) = &events[0] {
+            assert_eq!(start.index, 1);
+            assert!(matches!(
+                start.block_type,
+                crate::llm_client::event::BlockType::ToolUse
+            ));
+            if let BlockMetadata::ToolUse { id, name } = &start.metadata {
+                assert_eq!(id, "call_456");
+                assert_eq!(name, "get_weather");
+            } else {
+                panic!("Expected ToolUse metadata");
+            }
+        } else {
+            panic!("Expected BlockStart");
+        }
+    }
+
+    #[test]
+    fn test_parse_function_call_arguments_delta() {
+        let data = r#"{"output_index":1,"call_id":"call_456","delta":"{\"city\":"}"#;
+        let events = parse_event("response.function_call_arguments.delta", data)
+            .unwrap()
+            .unwrap();
+        assert_eq!(events.len(), 1);
+        if let Event::BlockDelta(delta) = &events[0] {
+            assert_eq!(delta.index, 1);
+            assert!(matches!(
+                &delta.delta,
+                DeltaContent::InputJson(s) if s == "{\"city\":"
+            ));
+        } else {
+            panic!("Expected BlockDelta");
+        }
+    }
+
+    #[test]
+    fn test_parse_response_completed() {
+        let data = r#"{"response":{"id":"resp_123","status":"completed","output":[],"usage":{"input_tokens":10,"output_tokens":20,"total_tokens":30}}}"#;
+        let events = parse_event("response.completed", data).unwrap().unwrap();
+        assert_eq!(events.len(), 2);
+
+        // First event should be usage
+        if let Event::Usage(usage) = &events[0] {
+            assert_eq!(usage.input_tokens, Some(10));
+            assert_eq!(usage.output_tokens, Some(20));
+            assert_eq!(usage.total_tokens, Some(30));
+        } else {
+            panic!("Expected Usage event");
+        }
+
+        // Second event should be status
+        assert!(matches!(
+            events[1],
+            Event::Status(StatusEvent {
+                status: ResponseStatus::Completed
+            })
+        ));
+    }
+
+    #[test]
+    fn test_parse_error() {
+        let data = r#"{"error":{"code":"rate_limit","message":"Too many requests"}}"#;
+        let events = parse_event("error", data).unwrap().unwrap();
+        assert_eq!(events.len(), 1);
+        if let Event::Error(err) = &events[0] {
+            assert_eq!(err.code, Some("rate_limit".to_string()));
+            assert_eq!(err.message, "Too many requests");
+        } else {
+            panic!("Expected Error event");
+        }
+    }
+
+    #[test]
+    fn test_parse_unknown_event() {
+        let data = r#"{}"#;
+        let events = parse_event("some.unknown.event", data).unwrap();
+        assert!(events.is_none());
+    }
+}

llm-worker/src/llm_client/scheme/openresponses/mod.rs

@@ -0,0 +1,49 @@
+//! Open Responses Scheme
+//!
+//! Handles request/response conversion for the Open Responses API.
+//! Since our internal types are already Open Responses native, this scheme
+//! primarily passes through data with minimal transformation.
+
+mod events;
+mod request;
+
+use crate::llm_client::{ClientError, Request};
+
+pub use events::*;
+pub use request::*;
+
+/// Open Responses Scheme
+///
+/// Handles conversion between internal types and the Open Responses wire format.
+#[derive(Debug, Clone, Default)]
+pub struct OpenResponsesScheme {
+    /// Optional model override
+    pub model: Option<String>,
+}
+
+impl OpenResponsesScheme {
+    /// Create a new OpenResponsesScheme
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// Set the model
+    pub fn with_model(mut self, model: impl Into<String>) -> Self {
+        self.model = Some(model.into());
+        self
+    }
+
+    /// Build Open Responses request from internal Request
+    pub fn build_request(&self, model: &str, request: &Request) -> OpenResponsesRequest {
+        build_request(model, request)
+    }
+
+    /// Parse SSE event data into internal Event(s)
+    pub fn parse_event(
+        &self,
+        event_type: &str,
+        data: &str,
+    ) -> Result<Option<Vec<crate::llm_client::Event>>, ClientError> {
+        parse_event(event_type, data)
+    }
+}

llm-worker/src/llm_client/scheme/openresponses/request.rs

@@ -0,0 +1,285 @@
+//! Open Responses Request Builder
+//!
+//! Converts internal Request/Item types to Open Responses API format.
+//! Since our internal types are already Open Responses native, this is
+//! mostly a direct serialization with some field renaming.
+
+use serde::Serialize;
+use serde_json::Value;
+
+use crate::llm_client::{types::Item, Request, ToolDefinition};
+
+/// Open Responses API request body
+#[derive(Debug, Serialize)]
+pub struct OpenResponsesRequest {
+    /// Model identifier
+    pub model: String,
+
+    /// Input items (conversation history)
+    pub input: Vec<OpenResponsesItem>,
+
+    /// System instructions
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub instructions: Option<String>,
+
+    /// Tool definitions
+    #[serde(skip_serializing_if = "Vec::is_empty")]
+    pub tools: Vec<OpenResponsesTool>,
+
+    /// Enable streaming
+    pub stream: bool,
+
+    /// Maximum output tokens
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub max_output_tokens: Option<u32>,
+
+    /// Temperature
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub temperature: Option<f32>,
+
+    /// Top P (nucleus sampling)
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub top_p: Option<f32>,
+}
+
+/// Open Responses input item
+#[derive(Debug, Serialize)]
+#[serde(tag = "type", rename_all = "snake_case")]
+pub enum OpenResponsesItem {
+    /// Message item
+    Message {
+        #[serde(skip_serializing_if = "Option::is_none")]
+        id: Option<String>,
+        role: String,
+        content: Vec<OpenResponsesContentPart>,
+    },
+
+    /// Function call item
+    FunctionCall {
+        #[serde(skip_serializing_if = "Option::is_none")]
+        id: Option<String>,
+        call_id: String,
+        name: String,
+        arguments: String,
+    },
+
+    /// Function call output item
+    FunctionCallOutput {
+        #[serde(skip_serializing_if = "Option::is_none")]
+        id: Option<String>,
+        call_id: String,
+        output: String,
+    },
+
+    /// Reasoning item
+    Reasoning {
+        #[serde(skip_serializing_if = "Option::is_none")]
+        id: Option<String>,
+        text: String,
+    },
+}
+
+/// Open Responses content part
+#[derive(Debug, Serialize)]
+#[serde(tag = "type", rename_all = "snake_case")]
+pub enum OpenResponsesContentPart {
+    /// Input text (for user messages)
+    InputText { text: String },
+
+    /// Output text (for assistant messages)
+    OutputText { text: String },
+
+    /// Refusal
+    Refusal { refusal: String },
+}
+
+/// Open Responses tool definition
+#[derive(Debug, Serialize)]
+pub struct OpenResponsesTool {
+    /// Tool type (always "function")
+    pub r#type: String,
+
+    /// Function definition
+    pub name: String,
+
+    /// Description
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub description: Option<String>,
+
+    /// Parameters schema
+    pub parameters: Value,
+}
+
+/// Build Open Responses request from internal Request
+pub fn build_request(model: &str, request: &Request) -> OpenResponsesRequest {
+    let input = request.items.iter().map(convert_item).collect();
+    let tools = request.tools.iter().map(convert_tool).collect();
+
+    OpenResponsesRequest {
+        model: model.to_string(),
+        input,
+        instructions: request.system_prompt.clone(),
+        tools,
+        stream: true,
+        max_output_tokens: request.config.max_tokens,
+        temperature: request.config.temperature,
+        top_p: request.config.top_p,
+    }
+}
+
+fn convert_item(item: &Item) -> OpenResponsesItem {
+    match item {
+        Item::Message {
+            id,
+            role,
+            content,
+            status: _,
+        } => {
+            let role_str = match role {
+                crate::llm_client::types::Role::User => "user",
+                crate::llm_client::types::Role::Assistant => "assistant",
+                crate::llm_client::types::Role::System => "system",
+            };
+
+            let parts = content
+                .iter()
+                .map(|p| match p {
+                    crate::llm_client::types::ContentPart::InputText { text } => {
+                        OpenResponsesContentPart::InputText { text: text.clone() }
+                    }
+                    crate::llm_client::types::ContentPart::OutputText { text } => {
+                        OpenResponsesContentPart::OutputText { text: text.clone() }
+                    }
+                    crate::llm_client::types::ContentPart::Refusal { refusal } => {
+                        OpenResponsesContentPart::Refusal {
+                            refusal: refusal.clone(),
+                        }
+                    }
+                })
+                .collect();
+
+            OpenResponsesItem::Message {
+                id: id.clone(),
+                role: role_str.to_string(),
+                content: parts,
+            }
+        }
+
+        Item::FunctionCall {
+            id,
+            call_id,
+            name,
+            arguments,
+            status: _,
+        } => OpenResponsesItem::FunctionCall {
+            id: id.clone(),
+            call_id: call_id.clone(),
+            name: name.clone(),
+            arguments: arguments.clone(),
+        },
+
+        Item::FunctionCallOutput {
+            id,
+            call_id,
+            output,
+        } => OpenResponsesItem::FunctionCallOutput {
+            id: id.clone(),
+            call_id: call_id.clone(),
+            output: output.clone(),
+        },
+
+        Item::Reasoning {
+            id,
+            text,
+            status: _,
+        } => OpenResponsesItem::Reasoning {
+            id: id.clone(),
+            text: text.clone(),
+        },
+    }
+}
+
+fn convert_tool(tool: &ToolDefinition) -> OpenResponsesTool {
+    OpenResponsesTool {
+        r#type: "function".to_string(),
+        name: tool.name.clone(),
+        description: tool.description.clone(),
+        parameters: tool.input_schema.clone(),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::llm_client::types::Item;
+
+    #[test]
+    fn test_build_simple_request() {
+        let request = Request::new()
+            .system("You are a helpful assistant.")
+            .user("Hello!");
+
+        let or_req = build_request("gpt-4o", &request);
+
+        assert_eq!(or_req.model, "gpt-4o");
+        assert_eq!(
+            or_req.instructions,
+            Some("You are a helpful assistant.".to_string())
+        );
+        assert_eq!(or_req.input.len(), 1);
+        assert!(or_req.stream);
+    }
+
+    #[test]
+    fn test_build_request_with_tool() {
+        let request = Request::new().user("What's the weather?").tool(
+            ToolDefinition::new("get_weather")
+                .description("Get current weather")
+                .input_schema(serde_json::json!({
+                    "type": "object",
+                    "properties": {
+                        "location": { "type": "string" }
+                    },
+                    "required": ["location"]
+                })),
+        );
+
+        let or_req = build_request("gpt-4o", &request);
+
+        assert_eq!(or_req.tools.len(), 1);
+        assert_eq!(or_req.tools[0].name, "get_weather");
+        assert_eq!(or_req.tools[0].r#type, "function");
+    }
+
+    #[test]
+    fn test_function_call_and_output() {
+        let request = Request::new()
+            .user("What's the weather?")
+            .item(Item::function_call(
+                "call_123",
+                "get_weather",
+                r#"{"city":"Tokyo"}"#,
+            ))
+            .item(Item::function_call_output("call_123", "Sunny, 25°C"));
+
+        let or_req = build_request("gpt-4o", &request);
+
+        assert_eq!(or_req.input.len(), 3);
+
+        // Check function call
+        if let OpenResponsesItem::FunctionCall { call_id, name, .. } = &or_req.input[1] {
+            assert_eq!(call_id, "call_123");
+            assert_eq!(name, "get_weather");
+        } else {
+            panic!("Expected FunctionCall");
+        }
+
+        // Check function call output
+        if let OpenResponsesItem::FunctionCallOutput { call_id, output, .. } = &or_req.input[2] {
+            assert_eq!(call_id, "call_123");
+            assert_eq!(output, "Sunny, 25°C");
+        } else {
+            panic!("Expected FunctionCallOutput");
+        }
+    }
+}

llm-worker/src/llm_client/types.rs

@@ -1,189 +1,491 @@
-//! LLMクライアント共通型定義
+//! LLM Client Common Types - Open Responses Native
+//!
+//! This module defines types that are natively aligned with the Open Responses specification.
+//! The core abstraction is `Item` which represents different types of conversation elements:
+//! - Message items (user/assistant messages with content parts)
+//! - FunctionCall items (tool invocations)
+//! - FunctionCallOutput items (tool results)
+//! - Reasoning items (extended thinking)
 
 use serde::{Deserialize, Serialize};
 
-/// リクエスト構造体
+// ============================================================================
+// Item - The core unit of conversation
+// ============================================================================
+
+/// Item ID type for tracking items in a conversation
+pub type ItemId = String;
+
+/// Call ID type for linking function calls to their outputs
+pub type CallId = String;
+
+/// Conversation item - the primary unit in Open Responses
+///
+/// Items represent discrete elements in a conversation. Unlike traditional
+/// message-based APIs, Open Responses treats tool calls and reasoning as
+/// first-class items rather than parts of messages.
+///
+/// # Examples
+///
+/// ```ignore
+/// use llm_worker::Item;
+///
+/// // User message
+/// let user_item = Item::user_message("Hello!");
+///
+/// // Assistant message
+/// let assistant_item = Item::assistant_message("Hi there!");
+///
+/// // Function call
+/// let call = Item::function_call("call_123", "get_weather", json!({"city": "Tokyo"}));
+///
+/// // Function call output
+/// let result = Item::function_call_output("call_123", "Sunny, 25°C");
+/// ```
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+#[serde(tag = "type", rename_all = "snake_case")]
+pub enum Item {
+    /// User or assistant message with content parts
+    Message {
+        /// Optional item ID
+        #[serde(skip_serializing_if = "Option::is_none")]
+        id: Option<ItemId>,
+        /// Message role
+        role: Role,
+        /// Content parts
+        content: Vec<ContentPart>,
+        /// Item status
+        #[serde(skip_serializing_if = "Option::is_none")]
+        status: Option<ItemStatus>,
+    },
+
+    /// Function (tool) call from the assistant
+    FunctionCall {
+        /// Optional item ID
+        #[serde(skip_serializing_if = "Option::is_none")]
+        id: Option<ItemId>,
+        /// Call ID for linking to output
+        call_id: CallId,
+        /// Function name
+        name: String,
+        /// Function arguments as JSON string
+        arguments: String,
+        /// Item status
+        #[serde(skip_serializing_if = "Option::is_none")]
+        status: Option<ItemStatus>,
+    },
+
+    /// Function (tool) call output/result
+    FunctionCallOutput {
+        /// Optional item ID
+        #[serde(skip_serializing_if = "Option::is_none")]
+        id: Option<ItemId>,
+        /// Call ID linking to the function call
+        call_id: CallId,
+        /// Output content
+        output: String,
+    },
+
+    /// Reasoning/thinking item
+    Reasoning {
+        /// Optional item ID
+        #[serde(skip_serializing_if = "Option::is_none")]
+        id: Option<ItemId>,
+        /// Reasoning text
+        text: String,
+        /// Item status
+        #[serde(skip_serializing_if = "Option::is_none")]
+        status: Option<ItemStatus>,
+    },
+}
+
+impl Item {
+    // ========================================================================
+    // Message constructors
+    // ========================================================================
+
+    /// Create a user message item with text content
+    pub fn user_message(text: impl Into<String>) -> Self {
+        Self::Message {
+            id: None,
+            role: Role::User,
+            content: vec![ContentPart::InputText {
+                text: text.into(),
+            }],
+            status: None,
+        }
+    }
+
+    /// Create a user message item with multiple content parts
+    pub fn user_message_parts(parts: Vec<ContentPart>) -> Self {
+        Self::Message {
+            id: None,
+            role: Role::User,
+            content: parts,
+            status: None,
+        }
+    }
+
+    /// Create an assistant message item with text content
+    pub fn assistant_message(text: impl Into<String>) -> Self {
+        Self::Message {
+            id: None,
+            role: Role::Assistant,
+            content: vec![ContentPart::OutputText {
+                text: text.into(),
+            }],
+            status: None,
+        }
+    }
+
+    /// Create an assistant message item with multiple content parts
+    pub fn assistant_message_parts(parts: Vec<ContentPart>) -> Self {
+        Self::Message {
+            id: None,
+            role: Role::Assistant,
+            content: parts,
+            status: None,
+        }
+    }
+
+    // ========================================================================
+    // Function call constructors
+    // ========================================================================
+
+    /// Create a function call item
+    pub fn function_call(
+        call_id: impl Into<String>,
+        name: impl Into<String>,
+        arguments: impl Into<String>,
+    ) -> Self {
+        Self::FunctionCall {
+            id: None,
+            call_id: call_id.into(),
+            name: name.into(),
+            arguments: arguments.into(),
+            status: None,
+        }
+    }
+
+    /// Create a function call item from a JSON value
+    pub fn function_call_json(
+        call_id: impl Into<String>,
+        name: impl Into<String>,
+        arguments: serde_json::Value,
+    ) -> Self {
+        Self::function_call(call_id, name, arguments.to_string())
+    }
+
+    /// Create a function call output item
+    pub fn function_call_output(call_id: impl Into<String>, output: impl Into<String>) -> Self {
+        Self::FunctionCallOutput {
+            id: None,
+            call_id: call_id.into(),
+            output: output.into(),
+        }
+    }
+
+    // ========================================================================
+    // Reasoning constructors
+    // ========================================================================
+
+    /// Create a reasoning item
+    pub fn reasoning(text: impl Into<String>) -> Self {
+        Self::Reasoning {
+            id: None,
+            text: text.into(),
+            status: None,
+        }
+    }
+
+    // ========================================================================
+    // Builder methods
+    // ========================================================================
+
+    /// Set the item ID
+    pub fn with_id(mut self, id: impl Into<String>) -> Self {
+        match &mut self {
+            Self::Message { id: item_id, .. } => *item_id = Some(id.into()),
+            Self::FunctionCall { id: item_id, .. } => *item_id = Some(id.into()),
+            Self::FunctionCallOutput { id: item_id, .. } => *item_id = Some(id.into()),
+            Self::Reasoning { id: item_id, .. } => *item_id = Some(id.into()),
+        }
+        self
+    }
+
+    /// Set the item status
+    pub fn with_status(mut self, new_status: ItemStatus) -> Self {
+        match &mut self {
+            Self::Message { status, .. } => *status = Some(new_status),
+            Self::FunctionCall { status, .. } => *status = Some(new_status),
+            Self::FunctionCallOutput { .. } => {} // Output items don't have status
+            Self::Reasoning { status, .. } => *status = Some(new_status),
+        }
+        self
+    }
+
+    // ========================================================================
+    // Accessors
+    // ========================================================================
+
+    /// Get the item ID if set
+    pub fn id(&self) -> Option<&str> {
+        match self {
+            Self::Message { id, .. } => id.as_deref(),
+            Self::FunctionCall { id, .. } => id.as_deref(),
+            Self::FunctionCallOutput { id, .. } => id.as_deref(),
+            Self::Reasoning { id, .. } => id.as_deref(),
+        }
+    }
+
+    /// Get the item type as a string
+    pub fn item_type(&self) -> &'static str {
+        match self {
+            Self::Message { .. } => "message",
+            Self::FunctionCall { .. } => "function_call",
+            Self::FunctionCallOutput { .. } => "function_call_output",
+            Self::Reasoning { .. } => "reasoning",
+        }
+    }
+
+    /// Check if this is a user message
+    pub fn is_user_message(&self) -> bool {
+        matches!(self, Self::Message { role: Role::User, .. })
+    }
+
+    /// Check if this is an assistant message
+    pub fn is_assistant_message(&self) -> bool {
+        matches!(self, Self::Message { role: Role::Assistant, .. })
+    }
+
+    /// Check if this is a function call
+    pub fn is_function_call(&self) -> bool {
+        matches!(self, Self::FunctionCall { .. })
+    }
+
+    /// Check if this is a function call output
+    pub fn is_function_call_output(&self) -> bool {
+        matches!(self, Self::FunctionCallOutput { .. })
+    }
+
+    /// Check if this is a reasoning item
+    pub fn is_reasoning(&self) -> bool {
+        matches!(self, Self::Reasoning { .. })
+    }
+
+    /// Get text content if this is a simple text message
+    pub fn as_text(&self) -> Option<&str> {
+        match self {
+            Self::Message { content, .. } if content.len() == 1 => match &content[0] {
+                ContentPart::InputText { text } => Some(text),
+                ContentPart::OutputText { text } => Some(text),
+                _ => None,
+            },
+            _ => None,
+        }
+    }
+}
+
+// ============================================================================
+// Content Parts - Components within message items
+// ============================================================================
+
+/// Content part within a message item
+///
+/// Open Responses distinguishes between input and output content types.
+/// Input types are used in user messages, output types in assistant messages.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+#[serde(tag = "type", rename_all = "snake_case")]
+pub enum ContentPart {
+    /// Input text (for user messages)
+    InputText {
+        /// The text content
+        text: String,
+    },
+
+    /// Output text (for assistant messages)
+    OutputText {
+        /// The text content
+        text: String,
+    },
+
+    /// Refusal content (for assistant messages)
+    Refusal {
+        /// The refusal message
+        refusal: String,
+    },
+    // Future: InputAudio, OutputAudio, etc.
+}
+
+impl ContentPart {
+    /// Create an input text part
+    pub fn input_text(text: impl Into<String>) -> Self {
+        Self::InputText { text: text.into() }
+    }
+
+    /// Create an output text part
+    pub fn output_text(text: impl Into<String>) -> Self {
+        Self::OutputText { text: text.into() }
+    }
+
+    /// Create a refusal part
+    pub fn refusal(refusal: impl Into<String>) -> Self {
+        Self::Refusal {
+            refusal: refusal.into(),
+        }
+    }
+
+    /// Get the text content regardless of type
+    pub fn as_text(&self) -> &str {
+        match self {
+            Self::InputText { text } => text,
+            Self::OutputText { text } => text,
+            Self::Refusal { refusal } => refusal,
+        }
+    }
+}
+
+// ============================================================================
+// Role and Status
+// ============================================================================
+
+/// Message role
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "lowercase")]
+pub enum Role {
+    /// User
+    User,
+    /// Assistant
+    Assistant,
+    /// System (for system prompts, not typically used in items)
+    System,
+}
+
+/// Item status
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "lowercase")]
+pub enum ItemStatus {
+    /// Item is being generated
+    InProgress,
+    /// Item completed successfully
+    Completed,
+    /// Item was truncated (e.g., max tokens)
+    Incomplete,
+}
+
+// ============================================================================
+// Request Types
+// ============================================================================
+
+/// LLM Request
 #[derive(Debug, Clone, Default)]
 pub struct Request {
-    /// システムプロンプト
+    /// System prompt (instructions)
     pub system_prompt: Option<String>,
-    /// メッセージ履歴
+    /// Input items (conversation history)
-    pub messages: Vec<Message>,
+    pub items: Vec<Item>,
-    /// ツール定義
+    /// Tool definitions
     pub tools: Vec<ToolDefinition>,
-    /// リクエスト設定
+    /// Request configuration
     pub config: RequestConfig,
 }
 
 impl Request {
-    /// 新しいリクエストを作成
+    /// Create a new empty request
     pub fn new() -> Self {
         Self::default()
     }
 
-    /// システムプロンプトを設定
+    /// Set the system prompt
     pub fn system(mut self, prompt: impl Into<String>) -> Self {
         self.system_prompt = Some(prompt.into());
         self
     }
 
-    /// ユーザーメッセージを追加
+    /// Add a user message
     pub fn user(mut self, content: impl Into<String>) -> Self {
-        self.messages.push(Message::user(content));
+        self.items.push(Item::user_message(content));
         self
     }
 
-    /// アシスタントメッセージを追加
+    /// Add an assistant message
     pub fn assistant(mut self, content: impl Into<String>) -> Self {
-        self.messages.push(Message::assistant(content));
+        self.items.push(Item::assistant_message(content));
         self
     }
 
-    /// メッセージを追加
+    /// Add an item
-    pub fn message(mut self, message: Message) -> Self {
+    pub fn item(mut self, item: Item) -> Self {
-        self.messages.push(message);
+        self.items.push(item);
         self
     }
 
-    /// ツールを追加
+    /// Add multiple items
+    pub fn items(mut self, items: impl IntoIterator<Item = Item>) -> Self {
+        self.items.extend(items);
+        self
+    }
+
+    /// Add a tool definition
     pub fn tool(mut self, tool: ToolDefinition) -> Self {
         self.tools.push(tool);
         self
     }
 
-    /// 設定を適用
+    /// Set the request config
     pub fn config(mut self, config: RequestConfig) -> Self {
         self.config = config;
         self
     }
 
-    /// max_tokensを設定
+    /// Set max tokens
     pub fn max_tokens(mut self, max_tokens: u32) -> Self {
         self.config.max_tokens = Some(max_tokens);
         self
     }
 
-    /// temperatureを設定
+    /// Set temperature
     pub fn temperature(mut self, temperature: f32) -> Self {
         self.config.temperature = Some(temperature);
         self
     }
 
-    /// top_pを設定
+    /// Set top_p
     pub fn top_p(mut self, top_p: f32) -> Self {
         self.config.top_p = Some(top_p);
         self
     }
 
-    /// top_kを設定
+    /// Set top_k
     pub fn top_k(mut self, top_k: u32) -> Self {
         self.config.top_k = Some(top_k);
         self
     }
 
-    /// ストップシーケンスを追加
+    /// Add a stop sequence
     pub fn stop_sequence(mut self, sequence: impl Into<String>) -> Self {
         self.config.stop_sequences.push(sequence.into());
         self
     }
 }
 
-/// メッセージ
+// ============================================================================
-#[derive(Debug, Clone, Serialize, Deserialize)]
+// Tool Definition
-pub struct Message {
+// ============================================================================
-    /// ロール
-    pub role: Role,
-    /// コンテンツ
-    pub content: MessageContent,
-}
 
-impl Message {
+/// Tool (function) definition
-    /// ユーザーメッセージを作成
-    pub fn user(content: impl Into<String>) -> Self {
-        Self {
-            role: Role::User,
-            content: MessageContent::Text(content.into()),
-        }
-    }
-
-    /// アシスタントメッセージを作成
-    pub fn assistant(content: impl Into<String>) -> Self {
-        Self {
-            role: Role::Assistant,
-            content: MessageContent::Text(content.into()),
-        }
-    }
-
-    /// ツール結果メッセージを作成
-    pub fn tool_result(tool_use_id: impl Into<String>, content: impl Into<String>) -> Self {
-        Self {
-            role: Role::User,
-            content: MessageContent::ToolResult {
-                tool_use_id: tool_use_id.into(),
-                content: content.into(),
-            },
-        }
-    }
-}
-
-/// ロール
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
-#[serde(rename_all = "lowercase")]
-pub enum Role {
-    User,
-    Assistant,
-}
-
-/// メッセージコンテンツ
-#[derive(Debug, Clone, Serialize, Deserialize)]
-#[serde(untagged)]
-pub enum MessageContent {
-    /// テキストコンテンツ
-    Text(String),
-    /// ツール結果
-    ToolResult {
-        tool_use_id: String,
-        content: String,
-    },
-    /// 複合コンテンツ (テキスト + ツール使用等)
-    Parts(Vec<ContentPart>),
-}
-
-/// コンテンツパーツ
-#[derive(Debug, Clone, Serialize, Deserialize)]
-#[serde(tag = "type")]
-pub enum ContentPart {
-    /// テキスト
-    #[serde(rename = "text")]
-    Text { text: String },
-    /// ツール使用
-    #[serde(rename = "tool_use")]
-    ToolUse {
-        id: String,
-        name: String,
-        input: serde_json::Value,
-    },
-    /// ツール結果
-    #[serde(rename = "tool_result")]
-    ToolResult {
-        tool_use_id: String,
-        content: String,
-    },
-}
-
-/// ツール定義
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct ToolDefinition {
-    /// ツール名
+    /// Tool name
     pub name: String,
-    /// 説明
+    /// Tool description
     pub description: Option<String>,
-    /// 入力スキーマ (JSON Schema)
+    /// Input schema (JSON Schema)
     pub input_schema: serde_json::Value,
 }
 
 impl ToolDefinition {
-    /// 新しいツール定義を作成
+    /// Create a new tool definition
     pub fn new(name: impl Into<String>) -> Self {
         Self {
             name: name.into(),
@@ -195,65 +497,69 @@ impl ToolDefinition {
         }
     }
 
-    /// 説明を設定
+    /// Set the description
     pub fn description(mut self, desc: impl Into<String>) -> Self {
         self.description = Some(desc.into());
         self
     }
 
-    /// 入力スキーマを設定
+    /// Set the input schema
     pub fn input_schema(mut self, schema: serde_json::Value) -> Self {
         self.input_schema = schema;
         self
     }
 }
 
-/// リクエスト設定
+// ============================================================================
+// Request Config
+// ============================================================================
+
+/// Request configuration
 #[derive(Debug, Clone, Default)]
 pub struct RequestConfig {
-    /// 最大トークン数
+    /// Maximum tokens to generate
     pub max_tokens: Option<u32>,
-    /// Temperature
+    /// Temperature (randomness)
     pub temperature: Option<f32>,
     /// Top P (nucleus sampling)
     pub top_p: Option<f32>,
     /// Top K
     pub top_k: Option<u32>,
-    /// ストップシーケンス
+    /// Stop sequences
     pub stop_sequences: Vec<String>,
 }
 
 impl RequestConfig {
-    /// 新しいデフォルト設定を作成
+    /// Create a new default config
     pub fn new() -> Self {
         Self::default()
     }
 
-    /// 最大トークン数を設定
+    /// Set max tokens
     pub fn with_max_tokens(mut self, max_tokens: u32) -> Self {
         self.max_tokens = Some(max_tokens);
         self
     }
 
-    /// temperatureを設定
+    /// Set temperature
     pub fn with_temperature(mut self, temperature: f32) -> Self {
         self.temperature = Some(temperature);
         self
     }
 
-    /// top_pを設定
+    /// Set top_p
     pub fn with_top_p(mut self, top_p: f32) -> Self {
         self.top_p = Some(top_p);
         self
     }
 
-    /// top_kを設定
+    /// Set top_k
     pub fn with_top_k(mut self, top_k: u32) -> Self {
         self.top_k = Some(top_k);
         self
     }
 
-    /// ストップシーケンスを追加
+    /// Add a stop sequence
     pub fn with_stop_sequence(mut self, sequence: impl Into<String>) -> Self {
         self.stop_sequences.push(sequence.into());
         self

llm-worker/src/message.rs

@@ -1,116 +1,16 @@
-//! Message Types
+//! Message and Item Types
 //!
-//! Message structure used in conversations with LLM.
+//! This module provides the core types for representing conversation items
-//! Can be easily created using [`Message::user`] or [`Message::assistant`].
+//! in the Open Responses format.
+//!
+//! The primary type is [`Item`], which represents different kinds of conversation
+//! elements: messages, function calls, function call outputs, and reasoning.
 
-use serde::{Deserialize, Serialize};
+// Re-export all types from llm_client::types
+pub use crate::llm_client::types::{ContentPart, Item, Role};
 
-/// Message role
+/// Convenience alias for backward compatibility
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
-#[serde(rename_all = "lowercase")]
-pub enum Role {
-    /// User
-    User,
-    /// Assistant
-    Assistant,
-}
-
-/// Conversation message
 ///
-/// # Examples
+/// In the Open Responses model, messages are just one type of Item.
-///
+/// This alias allows code that expects a "Message" type to continue working.
-/// ```ignore
+pub type Message = Item;
-/// use llm_worker::Message;
-///
-/// // User message
-/// let user_msg = Message::user("Hello!");
-///
-/// // Assistant message
-/// let assistant_msg = Message::assistant("Hi there!");
-/// ```
-#[derive(Debug, Clone, Serialize, Deserialize)]
-pub struct Message {
-    /// Role
-    pub role: Role,
-    /// Content
-    pub content: MessageContent,
-}
-
-/// Message content
-#[derive(Debug, Clone, Serialize, Deserialize)]
-#[serde(untagged)]
-pub enum MessageContent {
-    /// Text content
-    Text(String),
-    /// Tool result
-    ToolResult {
-        tool_use_id: String,
-        content: String,
-    },
-    /// Composite content (text + tool use, etc.)
-    Parts(Vec<ContentPart>),
-}
-
-/// Content part
-#[derive(Debug, Clone, Serialize, Deserialize)]
-#[serde(tag = "type")]
-pub enum ContentPart {
-    /// Text
-    #[serde(rename = "text")]
-    Text { text: String },
-    /// Tool use
-    #[serde(rename = "tool_use")]
-    ToolUse {
-        id: String,
-        name: String,
-        input: serde_json::Value,
-    },
-    /// Tool result
-    #[serde(rename = "tool_result")]
-    ToolResult {
-        tool_use_id: String,
-        content: String,
-    },
-}
-
-impl Message {
-    /// Create a user message
-    ///
-    /// # Examples
-    ///
-    /// ```ignore
-    /// use llm_worker::Message;
-    /// let msg = Message::user("Hello");
-    /// ```
-    pub fn user(content: impl Into<String>) -> Self {
-        Self {
-            role: Role::User,
-            content: MessageContent::Text(content.into()),
-        }
-    }
-
-    /// Create an assistant message
-    ///
-    /// Usually auto-generated inside Worker,
-    /// but can be manually created for history initialization, etc.
-    pub fn assistant(content: impl Into<String>) -> Self {
-        Self {
-            role: Role::Assistant,
-            content: MessageContent::Text(content.into()),
-        }
-    }
-
-    /// Create a tool result message
-    ///
-    /// Auto-generated inside Worker after tool execution.
-    /// Usually no need to create manually.
-    pub fn tool_result(tool_use_id: impl Into<String>, content: impl Into<String>) -> Self {
-        Self {
-            role: Role::User,
-            content: MessageContent::ToolResult {
-                tool_use_id: tool_use_id.into(),
-                content: content.into(),
-            },
-        }
-    }
-}

llm-worker/src/worker.rs

@@ -7,23 +7,20 @@ use tokio::sync::mpsc;
 use tracing::{debug, info, trace, warn};
 
 use crate::{
-    ContentPart, Message, MessageContent, Role,
+    Item,
     hook::{
         Hook, HookError, HookRegistry, OnAbort, OnPromptSubmit, OnPromptSubmitResult, OnTurnEnd,
         OnTurnEndResult, PostToolCall, PostToolCallContext, PostToolCallResult, PreLlmRequest,
         PreLlmRequestResult, PreToolCall, PreToolCallResult, ToolCall, ToolCallContext, ToolResult,
     },
-    llm_client::{
+    llm_client::{ClientError, ConfigWarning, LlmClient, Request, RequestConfig, ToolDefinition},
-        ClientError, ConfigWarning, LlmClient, Request, RequestConfig,
-        ToolDefinition as LlmToolDefinition,
-    },
     state::{CacheLocked, Mutable, WorkerState},
     subscriber::{
         ErrorSubscriberAdapter, StatusSubscriberAdapter, TextBlockSubscriberAdapter,
         ToolUseBlockSubscriberAdapter, UsageSubscriberAdapter, WorkerSubscriber,
     },
     timeline::{TextBlockCollector, Timeline, ToolCallCollector},
-    tool::{Tool, ToolDefinition, ToolError, ToolMeta},
+    tool::{Tool, ToolDefinition as WorkerToolDefinition, ToolError, ToolMeta},
 };
 
 // =============================================================================
@@ -136,7 +133,7 @@ impl<S: WorkerSubscriber + 'static> TurnNotifier for SubscriberTurnNotifier<S> {
 /// # Examples
 ///
 /// ```ignore
-/// use llm_worker::{Worker, Message};
+/// use llm_worker::{Worker, Item};
 ///
 /// // Create a Worker and register tools
 /// let mut worker = Worker::new(client)
@@ -172,8 +169,8 @@ pub struct Worker<C: LlmClient, S: WorkerState = Mutable> {
     hooks: HookRegistry,
     /// System prompt
     system_prompt: Option<String>,
-    /// Message history (owned by Worker)
+    /// Item history (owned by Worker)
-    history: Vec<Message>,
+    history: Vec<Item>,
     /// History length at lock time (only meaningful in CacheLocked state)
     locked_prefix_len: usize,
     /// Turn count
@@ -210,8 +207,8 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
     ) -> Result<WorkerResult, WorkerError> {
         self.reset_interruption_state();
         // Hook: on_prompt_submit
-        let mut user_message = Message::user(user_input);
+        let mut user_item = Item::user_message(user_input);
-        let result = self.run_on_prompt_submit_hooks(&mut user_message).await;
+        let result = self.run_on_prompt_submit_hooks(&mut user_item).await;
         let result = match result {
             Ok(value) => value,
             Err(err) => return self.finalize_interruption(Err(err)).await,
@@ -223,7 +220,7 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
             }
             OnPromptSubmitResult::Continue => {}
         }
-        self.history.push(user_message);
+        self.history.push(user_item);
         let result = self.run_turn_loop().await;
         self.finalize_interruption(result).await
     }
@@ -318,7 +315,7 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
     /// });
     /// worker.register_tool(def)?;
     /// ```
-    pub fn register_tool(&mut self, factory: ToolDefinition) -> Result<(), ToolRegistryError> {
+    pub fn register_tool(&mut self, factory: WorkerToolDefinition) -> Result<(), ToolRegistryError> {
         let (meta, instance) = factory();
         if self.tools.contains_key(&meta.name) {
             return Err(ToolRegistryError::DuplicateName(meta.name.clone()));
@@ -330,7 +327,7 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
     /// Register multiple tools
     pub fn register_tools(
         &mut self,
-        factories: impl IntoIterator<Item = ToolDefinition>,
+        factories: impl IntoIterator<Item = WorkerToolDefinition>,
     ) -> Result<(), ToolRegistryError> {
         for factory in factories {
             self.register_tool(factory)?;
@@ -378,7 +375,7 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
     }
 
     /// Get a reference to the history
-    pub fn history(&self) -> &[Message] {
+    pub fn history(&self) -> &[Item] {
         &self.history
     }
 
@@ -510,64 +507,48 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
     }
 
     /// Generate list of ToolDefinitions for LLM from registered tools
-    fn build_tool_definitions(&self) -> Vec<LlmToolDefinition> {
+    fn build_tool_definitions(&self) -> Vec<ToolDefinition> {
         self.tools
             .values()
             .map(|(meta, _)| {
-                LlmToolDefinition::new(&meta.name)
+                ToolDefinition::new(&meta.name)
                     .description(&meta.description)
                     .input_schema(meta.input_schema.clone())
             })
             .collect()
     }
 
-    /// Build assistant message from text blocks and tool calls
+    /// Build assistant response items from text blocks and tool calls
-    fn build_assistant_message(
+    fn build_assistant_items(
         &self,
         text_blocks: &[String],
         tool_calls: &[ToolCall],
-    ) -> Option<Message> {
+    ) -> Vec<Item> {
-        // Return None if no text or tool calls
+        let mut items = Vec::new();
-        if text_blocks.is_empty() && tool_calls.is_empty() {
-            return None;
-        }
 
-        // Simple text message if text only
+        // Add text as assistant message if present
-        if tool_calls.is_empty() {
         let text = text_blocks.join("");
-            return Some(Message::assistant(text));
-        }
-
-        // Build as Parts if tool calls are present
-        let mut parts = Vec::new();
-
-        // Add text parts
-        for text in text_blocks {
         if !text.is_empty() {
-                parts.push(ContentPart::Text { text: text.clone() });
+            items.push(Item::assistant_message(text));
-            }
         }
 
-        // Add tool call parts
+        // Add tool calls as FunctionCall items
         for call in tool_calls {
-            parts.push(ContentPart::ToolUse {
+            items.push(Item::function_call_json(
-                id: call.id.clone(),
+                &call.id,
-                name: call.name.clone(),
+                &call.name,
-                input: call.input.clone(),
+                call.input.clone(),
-            });
+            ));
         }
 
-        Some(Message {
+        items
-            role: Role::Assistant,
-            content: MessageContent::Parts(parts),
-        })
     }
 
     /// Build a request
     fn build_request(
         &self,
-        tool_definitions: &[LlmToolDefinition],
+        tool_definitions: &[ToolDefinition],
-        context: &[Message],
+        context: &[Item],
     ) -> Request {
         let mut request = Request::new();
 
@@ -576,50 +557,8 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
             request = request.system(system);
         }
 
-        // Add messages
+        // Add items directly (Request now uses Items natively)
-        for msg in context {
+        request = request.items(context.iter().cloned());
-            // Convert Message to llm_client::Message
-            request = request.message(crate::llm_client::Message {
-                role: match msg.role {
-                    Role::User => crate::llm_client::Role::User,
-                    Role::Assistant => crate::llm_client::Role::Assistant,
-                },
-                content: match &msg.content {
-                    MessageContent::Text(t) => crate::llm_client::MessageContent::Text(t.clone()),
-                    MessageContent::ToolResult {
-                        tool_use_id,
-                        content,
-                    } => crate::llm_client::MessageContent::ToolResult {
-                        tool_use_id: tool_use_id.clone(),
-                        content: content.clone(),
-                    },
-                    MessageContent::Parts(parts) => crate::llm_client::MessageContent::Parts(
-                        parts
-                            .iter()
-                            .map(|p| match p {
-                                ContentPart::Text { text } => {
-                                    crate::llm_client::ContentPart::Text { text: text.clone() }
-                                }
-                                ContentPart::ToolUse { id, name, input } => {
-                                    crate::llm_client::ContentPart::ToolUse {
-                                        id: id.clone(),
-                                        name: name.clone(),
-                                        input: input.clone(),
-                                    }
-                                }
-                                ContentPart::ToolResult {
-                                    tool_use_id,
-                                    content,
-                                } => crate::llm_client::ContentPart::ToolResult {
-                                    tool_use_id: tool_use_id.clone(),
-                                    content: content.clone(),
-                                },
-                            })
-                            .collect(),
-                    ),
-                },
-            });
-        }
 
         // Add tool definitions
         for tool_def in tool_definitions {
@@ -637,10 +576,10 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
     /// Called immediately after receiving a user message in `run()` (first time only).
     async fn run_on_prompt_submit_hooks(
         &self,
-        message: &mut Message,
+        item: &mut Item,
     ) -> Result<OnPromptSubmitResult, WorkerError> {
         for hook in &self.hooks.on_prompt_submit {
-            let result = hook.call(message).await?;
+            let result = hook.call(item).await?;
             match result {
                 OnPromptSubmitResult::Continue => continue,
                 OnPromptSubmitResult::Cancel(reason) => {
@@ -656,7 +595,7 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
     /// Called before sending an LLM request for each turn.
     async fn run_pre_llm_request_hooks(
         &self,
-    ) -> Result<(PreLlmRequestResult, Vec<Message>), WorkerError> {
+    ) -> Result<(PreLlmRequestResult, Vec<Item>), WorkerError> {
         let mut temp_context = self.history.clone();
         for hook in &self.hooks.pre_llm_request {
             let result = hook.call(&mut temp_context).await?;
@@ -672,13 +611,13 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
 
     /// Hooks: on_turn_end
     async fn run_on_turn_end_hooks(&self) -> Result<OnTurnEndResult, WorkerError> {
-        let mut temp_messages = self.history.clone();
+        let mut temp_items = self.history.clone();
         for hook in &self.hooks.on_turn_end {
-            let result = hook.call(&mut temp_messages).await?;
+            let result = hook.call(&mut temp_items).await?;
             match result {
                 OnTurnEndResult::Finish => continue,
-                OnTurnEndResult::ContinueWithMessages(msgs) => {
+                OnTurnEndResult::ContinueWithMessages(items) => {
-                    return Ok(OnTurnEndResult::ContinueWithMessages(msgs));
+                    return Ok(OnTurnEndResult::ContinueWithMessages(items));
                 }
                 OnTurnEndResult::Paused => return Ok(OnTurnEndResult::Paused),
             }
@@ -719,25 +658,43 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
 
     /// Check for pending tool calls (for resuming from Pause)
     fn get_pending_tool_calls(&self) -> Option<Vec<ToolCall>> {
-        let last_msg = self.history.last()?;
+        // Find the last FunctionCall items that don't have corresponding FunctionCallOutput
-        if last_msg.role != Role::Assistant {
+        let mut pending_calls = Vec::new();
-            return None;
+        let mut answered_call_ids = std::collections::HashSet::new();
+
+        // First pass: collect all answered call IDs
+        for item in &self.history {
+            if let Item::FunctionCallOutput { call_id, .. } = item {
+                answered_call_ids.insert(call_id.clone());
+            }
         }
 
-        let mut calls = Vec::new();
+        // Second pass: find unanswered function calls
-        if let MessageContent::Parts(parts) = &last_msg.content {
+        for item in &self.history {
-            for part in parts {
+            if let Item::FunctionCall {
-                if let ContentPart::ToolUse { id, name, input } = part {
+                call_id,
-                    calls.push(ToolCall {
+                name,
-                        id: id.clone(),
+                arguments,
+                ..
+            } = item
+            {
+                if !answered_call_ids.contains(call_id) {
+                    let input = serde_json::from_str(arguments)
+                        .unwrap_or_else(|_| serde_json::Value::Object(serde_json::Map::new()));
+                    pending_calls.push(ToolCall {
+                        id: call_id.clone(),
                         name: name.clone(),
-                        input: input.clone(),
+                        input,
                     });
                 }
             }
         }
 
-        if calls.is_empty() { None } else { Some(calls) }
+        if pending_calls.is_empty() {
+            None
+        } else {
+            Some(pending_calls)
+        }
     }
 
     /// Execute tools in parallel
@@ -882,7 +839,7 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
         let tool_definitions = self.build_tool_definitions();
 
         info!(
-            message_count = self.history.len(),
+            item_count = self.history.len(),
             tool_count = tool_definitions.len(),
             "Starting worker run"
         );
@@ -898,7 +855,7 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
                 Ok(ToolExecutionResult::Completed(results)) => {
                     for result in results {
                         self.history
-                            .push(Message::tool_result(&result.tool_use_id, &result.content));
+                            .push(Item::function_call_output(&result.tool_use_id, &result.content));
                     }
                     // Continue to loop
                 }
@@ -945,7 +902,7 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
             // Build request
             let request = self.build_request(&tool_definitions, &request_context);
             debug!(
-                message_count = request.messages.len(),
+                item_count = request.items.len(),
                 tool_count = request.tools.len(),
                 has_system = request.system_prompt.is_some(),
                 "Sending request to LLM"
@@ -1015,11 +972,9 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
             let text_blocks = self.text_block_collector.take_collected();
             let tool_calls = self.tool_call_collector.take_collected();
 
-            // Add assistant message to history
+            // Add assistant items to history
-            let assistant_message = self.build_assistant_message(&text_blocks, &tool_calls);
+            let assistant_items = self.build_assistant_items(&text_blocks, &tool_calls);
-            if let Some(msg) = assistant_message {
+            self.history.extend(assistant_items);
-                self.history.push(msg);
-            }
 
             if tool_calls.is_empty() {
                 // No tool calls → determine turn end
@@ -1052,7 +1007,7 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
                 Ok(ToolExecutionResult::Completed(results)) => {
                     for result in results {
                         self.history
-                            .push(Message::tool_result(&result.tool_use_id, &result.content));
+                            .push(Item::function_call_output(&result.tool_use_id, &result.content));
                     }
                 }
                 Err(err) => {
@@ -1212,35 +1167,35 @@ impl<C: LlmClient> Worker<C, Mutable> {
     /// Get a mutable reference to history
     ///
     /// Available only in Mutable state. History can be freely edited.
-    pub fn history_mut(&mut self) -> &mut Vec<Message> {
+    pub fn history_mut(&mut self) -> &mut Vec<Item> {
         &mut self.history
     }
 
     /// Set history
-    pub fn set_history(&mut self, messages: Vec<Message>) {
+    pub fn set_history(&mut self, items: Vec<Item>) {
-        self.history = messages;
+        self.history = items;
     }
 
-    /// Add a message to history (builder pattern)
+    /// Add an item to history (builder pattern)
-    pub fn with_message(mut self, message: Message) -> Self {
+    pub fn with_item(mut self, item: Item) -> Self {
-        self.history.push(message);
+        self.history.push(item);
         self
     }
 
-    /// Add a message to history
+    /// Add an item to history
-    pub fn push_message(&mut self, message: Message) {
+    pub fn push_item(&mut self, item: Item) {
-        self.history.push(message);
+        self.history.push(item);
     }
 
-    /// Add multiple messages to history (builder pattern)
+    /// Add multiple items to history (builder pattern)
-    pub fn with_messages(mut self, messages: impl IntoIterator<Item = Message>) -> Self {
+    pub fn with_items(mut self, items: impl IntoIterator<Item = Item>) -> Self {
-        self.history.extend(messages);
+        self.history.extend(items);
         self
     }
 
-    /// Add multiple messages to history
+    /// Add multiple items to history
-    pub fn extend_history(&mut self, messages: impl IntoIterator<Item = Message>) {
+    pub fn extend_history(&mut self, items: impl IntoIterator<Item = Item>) {
-        self.history.extend(messages);
+        self.history.extend(items);
     }
 
     /// Clear history
@@ -1279,7 +1234,6 @@ impl<C: LlmClient> Worker<C, Mutable> {
             _state: PhantomData,
         }
     }
-
 }
 
 // =============================================================================

llm-worker/tests/worker_state_test.rs

@@ -8,7 +8,7 @@ mod common;
 use common::MockLlmClient;
 use llm_worker::Worker;
 use llm_worker::llm_client::event::{Event, ResponseStatus, StatusEvent};
-use llm_worker::{Message, MessageContent};
+use llm_worker::Item;
 
 // =============================================================================
 // Mutable State Tests
@@ -39,12 +39,12 @@ fn test_mutable_history_manipulation() {
     assert!(worker.history().is_empty());
 
     // Add to history
-    worker.push_message(Message::user("Hello"));
+    worker.push_item(Item::user_message("Hello"));
-    worker.push_message(Message::assistant("Hi there!"));
+    worker.push_item(Item::assistant_message("Hi there!"));
     assert_eq!(worker.history().len(), 2);
 
     // Mutable access to history
-    worker.history_mut().push(Message::user("How are you?"));
+    worker.history_mut().push(Item::user_message("How are you?"));
     assert_eq!(worker.history().len(), 3);
 
     // Clear history
@@ -52,8 +52,8 @@ fn test_mutable_history_manipulation() {
     assert!(worker.history().is_empty());
 
     // Set history
-    let messages = vec![Message::user("Test"), Message::assistant("Response")];
+    let items = vec![Item::user_message("Test"), Item::assistant_message("Response")];
-    worker.set_history(messages);
+    worker.set_history(items);
     assert_eq!(worker.history().len(), 2);
 }
 
@@ -63,29 +63,29 @@ fn test_mutable_builder_pattern() {
     let client = MockLlmClient::new(vec![]);
     let worker = Worker::new(client)
         .system_prompt("System prompt")
-        .with_message(Message::user("Hello"))
+        .with_item(Item::user_message("Hello"))
-        .with_message(Message::assistant("Hi!"))
+        .with_item(Item::assistant_message("Hi!"))
-        .with_messages(vec![
+        .with_items(vec![
-            Message::user("How are you?"),
+            Item::user_message("How are you?"),
-            Message::assistant("I'm fine!"),
+            Item::assistant_message("I'm fine!"),
         ]);
 
     assert_eq!(worker.get_system_prompt(), Some("System prompt"));
     assert_eq!(worker.history().len(), 4);
 }
 
-/// Verify that multiple messages can be added with extend_history
+/// Verify that multiple items can be added with extend_history
 #[test]
 fn test_mutable_extend_history() {
     let client = MockLlmClient::new(vec![]);
     let mut worker = Worker::new(client);
 
-    worker.push_message(Message::user("First"));
+    worker.push_item(Item::user_message("First"));
 
     worker.extend_history(vec![
-        Message::assistant("Response 1"),
+        Item::assistant_message("Response 1"),
-        Message::user("Second"),
+        Item::user_message("Second"),
-        Message::assistant("Response 2"),
+        Item::assistant_message("Response 2"),
     ]);
 
     assert_eq!(worker.history().len(), 4);
@@ -102,8 +102,8 @@ fn test_lock_transition() {
     let mut worker = Worker::new(client);
 
     worker.set_system_prompt("System");
-    worker.push_message(Message::user("Hello"));
+    worker.push_item(Item::user_message("Hello"));
-    worker.push_message(Message::assistant("Hi"));
+    worker.push_item(Item::assistant_message("Hi"));
 
     // Lock
     let locked_worker = worker.lock();
@@ -120,14 +120,14 @@ fn test_unlock_transition() {
     let client = MockLlmClient::new(vec![]);
     let mut worker = Worker::new(client);
 
-    worker.push_message(Message::user("Hello"));
+    worker.push_item(Item::user_message("Hello"));
     let locked_worker = worker.lock();
 
     // Unlock
     let mut worker = locked_worker.unlock();
 
     // History operations are available again in Mutable state
-    worker.push_message(Message::assistant("Hi"));
+    worker.push_item(Item::assistant_message("Hi"));
     worker.clear_history();
     assert!(worker.history().is_empty());
 }
@@ -160,16 +160,10 @@ async fn test_mutable_run_updates_history() {
     assert_eq!(history.len(), 2); // user + assistant
 
     // User message
-    assert!(matches!(
+    assert_eq!(history[0].as_text(), Some("Hi there"));
-        &history[0].content,
-        MessageContent::Text(t) if t == "Hi there"
-    ));
 
     // Assistant message
-    assert!(matches!(
+    assert_eq!(history[1].as_text(), Some("Hello, I'm an assistant!"));
-        &history[1].content,
-        MessageContent::Text(t) if t == "Hello, I'm an assistant!"
-    ));
 }
 
 /// Verify that history accumulates correctly over multiple turns in CacheLocked state
@@ -201,7 +195,7 @@ async fn test_locked_multi_turn_history_accumulation() {
 
     // Lock (after setting system prompt)
     let mut locked_worker = worker.lock();
-    assert_eq!(locked_worker.locked_prefix_len(), 0); // No messages yet
+    assert_eq!(locked_worker.locked_prefix_len(), 0); // No items yet
 
     // Turn 1
     let result1 = locked_worker.run("Hello!").await;
@@ -217,16 +211,16 @@ async fn test_locked_multi_turn_history_accumulation() {
     let history = locked_worker.history();
 
     // Turn 1 user message
-    assert!(matches!(&history[0].content, MessageContent::Text(t) if t == "Hello!"));
+    assert_eq!(history[0].as_text(), Some("Hello!"));
 
     // Turn 1 assistant message
-    assert!(matches!(&history[1].content, MessageContent::Text(t) if t == "Nice to meet you!"));
+    assert_eq!(history[1].as_text(), Some("Nice to meet you!"));
 
     // Turn 2 user message
-    assert!(matches!(&history[2].content, MessageContent::Text(t) if t == "Can you help me?"));
+    assert_eq!(history[2].as_text(), Some("Can you help me?"));
 
     // Turn 2 assistant message
-    assert!(matches!(&history[3].content, MessageContent::Text(t) if t == "I can help with that."));
+    assert_eq!(history[3].as_text(), Some("I can help with that."));
 }
 
 /// Verify that locked_prefix_len correctly records history length at lock time
@@ -253,15 +247,15 @@ async fn test_locked_prefix_len_tracking() {
 
     let mut worker = Worker::new(client);
 
-    // Add messages beforehand
+    // Add items beforehand
-    worker.push_message(Message::user("Pre-existing message 1"));
+    worker.push_item(Item::user_message("Pre-existing message 1"));
-    worker.push_message(Message::assistant("Pre-existing response 1"));
+    worker.push_item(Item::assistant_message("Pre-existing response 1"));
 
     assert_eq!(worker.history().len(), 2);
 
     // Lock
     let mut locked_worker = worker.lock();
-    assert_eq!(locked_worker.locked_prefix_len(), 2); // 2 messages at lock time
+    assert_eq!(locked_worker.locked_prefix_len(), 2); // 2 items at lock time
 
     // Execute turn
     locked_worker.run("New message").await.unwrap();
@@ -317,8 +311,8 @@ async fn test_unlock_edit_relock() {
     ]]);
 
     let worker = Worker::new(client)
-        .with_message(Message::user("Hello"))
+        .with_item(Item::user_message("Hello"))
-        .with_message(Message::assistant("Hi"));
+        .with_item(Item::assistant_message("Hi"));
 
     // Lock -> Unlock
     let locked = worker.lock();
@@ -328,7 +322,7 @@ async fn test_unlock_edit_relock() {
 
     // Edit history
     unlocked.clear_history();
-    unlocked.push_message(Message::user("Fresh start"));
+    unlocked.push_item(Item::user_message("Fresh start"));
 
     // Re-lock
     let relocked = unlocked.lock();